Improving robustness against electrode shift of high density EMG for myoelectric control through common spatial patterns

Lizhi Pan, Dingguo Zhang, Ning Jiang, Xinjun Sheng and Xiangyang ZhuEmail author
Journal of NeuroEngineering and Rehabilitation201512:110
DOI: 10.1186/s12984-015-0102-9© Pan et al. 2015
Received: 18 August 2015Accepted: 19 November 2015Published: 2 December 2015
Abstract

Background
Most prosthetic myoelectric control studies have concentrated on low density (less than 16 electrodes, LD) electromyography (EMG) signals, due to its better clinical applicability and low computation complexity compared with high density (more than 16 electrodes, HD) EMG signals. Since HD EMG electrodes have been developed more conveniently to wear with respect to the previous versions recently, HD EMG signals become an alternative for myoelectric prostheses. The electrode shift, which may occur during repositioning or donning/doffing of the prosthetic socket, is one of the main reasons for degradation in classification accuracy (CA).

Methods
HD EMG signals acquired from the forearm of the subjects were used for pattern recognition-based myoelectric control in this study. Multiclass common spatial patterns (CSP) with two types of schemes, namely one versus one (CSP-OvO) and one versus rest (CSP-OvR), were used for feature extraction to improve the robustness against electrode shift for myoelectric control. Shift transversal (ST1 and ST2) and longitudinal (SL1 and SL2) to the direction of the muscle fibers were taken into consideration. We tested nine intact-limb subjects for eleven hand and wrist motions. The CSP features (CSP-OvO and CSP-OvR) were compared with three commonly used features, namely time-domain (TD) features, time-domain autoregressive (TDAR) features and variogram (Variog) features.

Results
Compared with the TD features, the CSP features significantly improved the CA over 10 % in all shift configurations (ST1, ST2, SL1 and SL2). Compared with the TDAR features, a. the CSP-OvO feature significantly improved the average CA over 5 % in all shift configurations; b. the CSP-OvR feature significantly improved the average CA in shift configurations ST1, SL1 and SL2. Compared with the Variog features, the CSP features significantly improved the average CA in longitudinal shift configurations (SL1 and SL2).

Conclusion
The results demonstrated that the CSP features significantly improved the robustness against electrode shift for myoelectric control with respect to the commonly used features.

Keywords

Electromyography (EMG) Common spatial patterns (CSP) Electrode shift Pattern recognition Myoelectric control
Introduction

Surface electromyography (EMG) signals, which contain neural information [1], have long been used as control inputs of myoelectric prostheses [2–4]. With most conventional, commercially available myoelectric prostheses, a control scheme based on using amplitude or power of the EMG signals to control one degree-of-freedom (DOF) has been employed for several decades. To improve the functionality and provide more intuitive control of myoelectric prostheses, pattern recognition methods have been employed to classify EMG signals towards multifunctional prosthesis control for more than 20 years [5–9]. The pattern recognition-based control scheme is based on the assumption that amputees can activate consistent (same motion) and distinctive (different motions) EMG patterns using residual stump muscles [10].

In general, there are two types of surface EMG, low density (less than 16 electrodes, LD) EMG and high density (more than 16 electrodes, HD) EMG, which are classified by the number of electrodes. Electrode shift is an identified problem existing in both LD and HD EMG applications. It may occur during repositioning or donning/doffing of the prosthetic socket. It is one of the main reasons for degradation in classification accuracy (CA) [11]. In LD EMG, some researchers proposed efficient methods to reduce the CA degradation of electrode shift. Hargrove et al. proposed a strategy training the classifier with EMG signals from all expected displacement locations [12]. However, this strategy needing long-time training can be often frustrating for the user and leading to frequent device abandonment [13]. Young et al. demonstrated that electrode with larger size reduced the sensitivity of shift while performing worse than electrodes with smaller size without shift [11]. They suggested that electrodes oriented in longitudinal direction with the muscle fibers performed better than that oriented in transversal direction. They also showed that time-domain autoregressive (TDAR) features achieved the best classification performance and was least affected by electrode displacements. They further demonstrated that a greater interelectrode distance improved classification performance, and a combination of longitudinal and transversal electrode configurations also improved the performance in the presence of electrode shift [7].

Recently, HD EMG signals become an alternative for myoelectric prostheses [14–19]. Huang et al. showed that double differential spatial filter on HD EMG signals could improve the myoelectric control performance on targeted muscle reinnervation (TMR) patients [14]. However, for HD EMG application, the electrode shift is also very common and serious. Stango et al. used variogram (Variog) of HD EMG signals to provide features robust to electrode number and shift for myoelectric control [18]. The Variog is a statistical measure of the spatial correlation and widely used as spatial-domain feature for classification in geostatistic [20, 21]. It can be also called semivariance since it is a graph of the semivariance against the distance.

To solve the electrode shift problem of HD EMG, common spatial patterns (CSP), a method widely used in electroencephalogram (EEG) study has drawn our attention [22, 23]. In general, EEG has many electrodes (64 ∼ 128), which are similar to the HD EMG condition. Therefore, we expect the excellent capacity of CSP in EEG can also be suitable for HD EMG. Actually, Hahne et al. demonstrated that CSP feature showed a higher robustness against noise than time domain (TD) feature for myoelectric control [17]. However, they did not investigate the performance of CSP feature in the presence of electrode shift. Huang et al. also used an improved CSP for EMG classification, but they targeted LD EMG and did not consider the problem of electrode shift [24].

In this study, we investigate whether the CSP of HD EMG signals can improve the myoelectric control performance under electrode shift for eleven classes of hand and wrist motions. We test nine able-bodied subjects. The performance of CSP feature is compared with the commonly used TD, TDAR and Variog features. Linear discriminant analysis (LDA) classifiers are used to process the EMG data.

Methods

Subjects
Nine able-bodied subjects (eight males and one female; aged 22–27; referenced as Sub1-Sub9) participated in the experiment. The subjects had no neurological disorders. This work was approved by the Ethics Committee of Shanghai Jiao Tong University. All subjects participating in the experiment signed informed consent and the procedures were in compliance with the Declaration of Helsinki.

Experiment setup
Eleven classes of hand and wrist motions were performed by the subjects in order, i.e., hand close (HC), hand open (HO), key grip (KG), tip prehension (TP), wrist flexion (WF), wrist extension (WE), radial deviation (RD), ulnar deviation (UD), forearm supination (FS), forearm pronation (FP) and “no movement” (NM). In each trial, the subjects were asked to perform each motion for 10 s. Ten trials were performed by each subject. To avoid fatigue, the subjects had a 1-min rest between each trial.

Data acquisition
Monopolar surface EMG signals were measured and collected using a grid of 192 electrodes (3 semi-disposable adhesive matrix, 64 electrodes, ELSCH064NM3) composed by 8 rows and 24 columns, with 10 mm interelectrode distance (IED) (Fig. 1). The skin surface of forearm was rubbed lightly with alcohol to reduce impedance. The grid was mounted around the circumference of the forearm (Fig. 1), starting from the ulnar bone. The grid was mounted on the skin by adhesive foam and a reference electrode was mounted at the wrist. The matrixes were connected to a multichannel surface EMG amplifier (EMG-USB2 +, OT Bioelettronica, Torino, Italy) and the signals were amplified with a gain of 500, band-pass filtered (pass band 10–500 Hz), sampled at 2048 Hz, and A ∖D converted with 12-bit resolution.
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig1_HTML.gif
Fig. 1
Position of the HD EMG grid and the HD EMG grid of 192 electrodes used in the experiments
Common spatial patterns
CSP is a supervised two-class method to design linear spatial filters simultaneously maximizing the variance of one class and minimizing the variance of another class [22]. In this way, the classes can be maximally separated by their variances. CSP is widely used in motor imaginary-based brain computer interface (BCI) for classification of EEG signals [23, 25].

The raw EMG signals of class j and class k were represented as X j and X k with dimensions c×l, where c was the number of channels, and l was the number of samples per each channel (here l was 408). The objective was to find the ω of the spatial filter y=ω T X, which maximized the variance of class j and minimized the variance of class k. Thus, the optimization process was formulated as following:
ω=argmaxωωT∑jωωT∑kω
(1)
where ∑j=1/(n−1)∗Xj∗XjT and ∑k=1/(n−1)∗Xk∗XkT were the covariance matrix of class j and class k respectively.

This was realized by finding the matrix W that simultaneously diagonalized both ∑j and ∑k:
W∑jWT=Dj,
(2)
W∑kWT=Dk,
(3)
Dj+Dk=I.
(4)
The row vectors of W were c spatial filters. Applying the full filter matrix W to the raw EMG signals would give c output signals Y=W∗X, which were called components. The variance of each component for class j was indicated by the corresponding eigenvalue of D j , for class k of D k . With the constraint (4), the eigenvector corresponding to the largest eigenvalue for D j would had the smallest eigenvalues for D k , and the eigenvector corresponding to the largest eigenvalue for D k would had the smallest eigenvalues for D j . These two eigenvectors were chosen as the spatial filters in this study.

Multiclass CSP
Since there were eleven motion classes in this study, we extended the two-class CSP into multiclass CSP by using one versus one (CSP-OvO) and one versus rest (CSP-OvR) scheme [17].

In the CSP-OvO scheme, the two-class CSP was designed for all possible class combinations. The filters were chosen in the same way as in the two-class CSP. Thus, there were M=N∗(N−1)/2 combinations for N classes. The features of all selected components were concatenated into one feature vector.

In the CSP-OvR scheme, each filter was designed to maximize the variance of one class and minimize the average of the variances of all other classes. The filters were chosen in the same way as in the two-class CSP. This process was repeated for all classes. Thus, there was N combinations for N classes. The features of all selected components were concatenated into one feature vector.

Feature extraction
The logarithm of the variances of the selected CSP components were calculated as features in the CSP-OvO and CSP-OvR scheme. Here, the length of analysis window was set to 200 ms and the increment of two adjacent windows was set to 50 ms. The length and the increment were chosen to ensure response time of the system was below 300 ms for reducing users’ perceived lag [5]. A feature set was computed on each of the CSP component, and then concatenated to form a feature vector.

To compare the proposed feature extraction method with the state of the art technology, TD features, TDAR features and Variog features, which were effective and robust with electrode shift [2, 5, 11, 18, 26, 27], were used in this study. These features were extracted using the same window length and the same increment as those specified in above paragraph.

Classification
As a simple and efficient classifier, the LDA classifier has been widely used for pattern recognition of EMG signals [7, 28]. Researchers have presented in previous studies that the LDA classifier can have the comparable performance to other more sophisticated classifiers [29] and generalizes better than the nonlinear multilayer perceptron classifier with electrode shift [11]. Hence, the LDA classifier was employed to identify the CSP features (CSP-OvO and CSP-OvR) and the two classic features (TD and TDAR) in this study. Since the Variog features performed better with support vector machine (SVM) classifier compared with LDA classifier [18], the SVM classifier was employed to identify the Variog features in this study [30]. A five-fold cross-validation procedure was used. Four fifths of the data were randomly selected and used as a training set to train the LDA classifier, while the remaining one fifth were used as a testing set.

Electrodes shift
Shift transversal and longitudinal to the direction of the muscle fibers were taken into consideration. We expected that shift in longitudinal or transversal direction would be the extreme situation. Meanwhile, the influence of electrode shift occurring along both axes would be between the influences of electrode shift in longitudinal and transversal directions. Since a shift of 10 mm or less was considered more likely in clinical applications [11], the shift distance was chosen as 10 mm to simulate the worst shift situation in the current study. To simulate the shift transversal to the direction of the muscle fibers, half of the columns were used for training and the remaining half for testing, which corresponded to a 10-mm shift for a configuration of 96 electrodes. Figure 2 shows the shift in transversal direction of the muscle fibers. Shift leftwards (ST1): the white color electrodes were used for training, while the red color electrodes were used for testing. Shift rightwards (ST2): the red color electrodes were used for training, while the white color electrodes were used for testing. To provide a control for transversal direction shift, the same color electrodes in Fig. 2 were used for both training and testing, referred as ST. It should be noted that the electrodes distance in the transversal and longitudinal direction was 20 mm and 10 mm respectively. Similar method was used to simulate the shift in longitudinal direction of the muscle fibers (Fig. 3). To provide a control for longitudinal direction shift, the same color electrodes in Fig. 3 were used for both training and testing, referred as SL. It should be noted that the electrodes distance in the transversal and longitudinal direction was 10 mm and 20 mm respectively.
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig2_HTML.gif
Fig. 2
Shift transversal to the direction of the muscle fibers. Shift leftwards (ST1): the white color electrodes were used for training, while the red color electrodes were used for testing. Shift rightwards (ST2): the red color electrodes were used for training, while the white color electrodes were used for testing
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig3_HTML.gif
Fig. 3
Shift longitudinal to the direction of the muscle fibers. Shift downwards (SL1): the white color electrodes were used for training, while the red color electrodes were used for testing. Shift upwards (SL2): the red color electrodes were used for training, while the white color electrodes were used for testing
Quantification of feature space
To investigate the variations in the EMG feature space before and after the electrode shift, relative center shift (RCS) was defined in the current study. RCS was defined as the ratio between the mean value of the Mahalanobis distance of the same motion before and after the electrode shift across N motions and the mean value of the Mahalanobis distance of the different motions across N motions after the electrode shift:
RCS=(N−1)∑i=1N(μi−μsi)T(Si+Ssi2)−1(μi−μsi)−−−−−−−−−−−−−−−−−−−−−−−−√∑i=1N∑j=1,j≠iN(μsj−μsi)T(Ssj+Ssi2)−1(μsj−μsi)−−−−−−−−−−−−−−−−−−−−−−−−−−√
(5)
where μ i and μ si were the centroid of the ellipsoid of motion i before and after the electrode shift, S i and S si were the covariance of the data for motion i before and after the electrode shift.

The value of RCS was positively correlated to the relative center shift in the EMG feature space.

As different feature sets would have different dimensionality of feature vector, prior to computation of RCS, the Fisher linear discriminant (FLD) [31] was adopted to reduce the dimension of feature vectors to the same level of N−1, where N is the number of motions, which was eleven here. Since the Variog features were identified by SVM classifier but not LDA classifier, the FLD was not suitable to process the Variog features. Therefore, the RCS was not computed on the Variog features.

Visualization of CSP patterns
To understand the improvements of the CSP features, the corresponding patterns of the motions before and after the electrode shift were visualized for a representative subject (Sub3). CSP patterns were columns of the inverse of filter matrix W. The ith pattern represented the source signal distribution to the sensors that produced activity in the ith CSP component. CSP patterns provided valuable information about the underlying electrophysiology processes and the related muscles. Contrary to the EMG amplitude patterns, which only showed muscle activation information, the CSP patterns emphasized the locations that provided most information to discriminate different motions. Figures 4 and 5 show the last CSP patterns of motion 1 and motion 4 for CSP-OvO extension scheme in the transversal direction shift (ST1) and longitudinal direction shift (SL1) respectively. Figures 6 and 7 show the first CSP pattern of each active motion and rest motions for CSP-OvR extension scheme in the transversal direction shift (ST1) and longitudinal direction shift (SL1) respectively.
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig4_HTML.gif
Fig. 4
Last CSP pattern of motion 1 and motion 4 for CSP-OvO extension scheme in transversal direction shift (ST1). Left and right columns were the CSP patterns before and after electrode shift
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig5_HTML.gif
Fig. 5
Last CSP pattern of motion 1 and motion 4 for CSP-OvO extension scheme in longitudinal direction shift (SL1). Left and right columns were the CSP patterns before and after electrode shift
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig6_HTML.gif
Fig. 6
First CSP pattern of each active motion and rest motions for CSP-OvR extension scheme in transversal direction shift (ST1). First and second columns were the CSP patterns of the first five active motions (HC, HO, KG, TP and WF) before and after electrode shift. Third and fourth columns were the CSP patterns of the last five active motions (WE, RD, UD, FS and FP) before and after electrode shift
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig7_HTML.gif
Fig. 7
First CSP pattern of each active motion and rest motions for CSP-OvR extension scheme in longitudinal direction shift (SL1). Left and right columns were the CSP patterns before and after electrode shift
Statistical analysis
A two-way repeated measures ANOVA was used to analyze CA. The ANOVA included the following two factors: Shift (ST1, ST2, SL1 and SL2) and Feature (CSP-OvO, CSP-OvR, TD, TDAR and Variog). Similarly, a two-way repeated measures ANOVA was used to analyze RCS. The ANOVA included the following two factors: Shift (ST1, ST2, SL1 and SL2) and Feature (CSP-OvO, CSP-OvR, TD and TDAR). In all ANOVA tests, the full model was conducted first. When a significant interaction was detected, a simple-effects analysis was conducted by fixing the levels of one of the interacting factors. When no interaction was detected, a reduced ANOVA model with only the main factors was performed. Whenever significance was detected for the main factors, a Tukey comparison was performed. Only a significant difference was reported for these comparison tests. The significance level for all tests was p<0.05.

Results

Classification accuracy
Figure 8 shows the average CA of all features (CSP-OvO, CSP-OvR, TD, TDAR and Variog) across all subjects for the half grid configuration of 96 electrodes (ST and SL) and the different shift configurations (ST1, ST2, SL1 and SL2). The average CA of CSP-OvO and CSP-OvR was slightly higher than that of TD and was comparable with that of TDAR for the half grid configuration without electrode shift (ST and SL). The average CA of CSP-OvO, CSP-OvR, TD and TDAR was 8 % higher than that of Variog for the half grid configuration without electrode shift (ST and SL). Since the average CA of all features without electrode shift was over 90 %, it demonstrated that the half grid configuration without electrode shift was sufficient to provide good myoelectric control performance for all features (CSP-OvO, CSP-OvR, TD, TDAR and Variog). However, the average CA for TD was decreased to 67.2 % in ST1, 65.0 % in ST2, 81.9 % in SL1, and 85.4 % in SL2; the average CA for TDAR was decreased to 72.1 % in ST1, 74.5 % in ST2, 87.9 % in SL1, and 89.9 % in SL2; the average CA for Variog was decreased to 78 % in ST1, 78 % in ST2, 82.8 % in SL1, and 84.4 % in SL2. The average CA for CSP features (CSP-OvO and CSP-OvR) was ∼80 % in the electrode shift in transversal direction (ST1 and ST2) and ∼95 % in the electrode shift in longitudinal direction (SL1 and SL2) respectively. Thus, the CSP features (CSP-OvO and CSP-OvR) were more robust than the commonly used features (TD, TDAR and Variog) in the presence of electrode shift.
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig8_HTML.gif
Fig. 8
Average CA across all subjects for five features: CSP-OvO, CSP-OvR, TD, TDAR and Variog. Error bars represented the standard deviation. The tests were marked by * in which significance were found between different features
The two-way ANOVA revealed a statistically significant interaction between Shift and Feature (p<0.001). The simple-effects analysis was conducted to break down the ANOVA into subsequent one-way ANOVA, looking separately at the ST1, ST2, SL1 and SL2 for main effect of Feature.

For the Shift ST1, the one-way ANOVA revealed a main effect of Feature (p<0.001). Tukey comparison showed that the CA of CSP-OvO was not significantly different with that of CSP-OvR (p=0.997) and Variog (p=0.869) but significantly higher than that of TD (p<0.001) and TDAR (p=0.002); the CA of CSP-OvR was not significantly different with that of Variog (p=0.973) but significantly higher than that of TD (p<0.001) and TDAR (p=0.006); the CA of TD was not significantly different with that of TDAR (p=0.138) but significantly lower than that of Variog (p<0.001); the CA of TDAR was significantly lower than that of Variog (p=0.04).

For the Shift ST2, the one-way ANOVA revealed a main effect of Feature (p<0.001). Tukey comparison showed that the CA of CSP-OvO was not significantly different with that of CSP-OvR (p=0.531) and Variog (p=0.777) but significantly higher than that of TD (p<0.001) and TDAR (p=0.019); the CA of CSP-OvR was not significantly different with that of TDAR (p=0.536) and Variog (p=0.995) but significantly higher than that of TD (p<0.001) and TDAR (p=0.006); the CA of TD was significantly lower than that of TDAR (p<0.001) and Variog (p<0.001); the CA of TDAR was not significantly different with that of Variog (p=0.3).

For the Shift SL1, the one-way ANOVA revealed a main effect of Feature (p<0.001). Tukey comparison showed that the CA of CSP-OvO was not significantly different with that of CSP-OvR (p=0.995) but significantly higher than that of TD (p<0.001), TDAR (p<0.001) and Variog (p<0.001); the CA of CSP-OvR was significantly higher than that of TD (p<0.001), TDAR (p<0.001) and Variog (p<0.001); the CA of TD was not significantly different with that of Variog (p=0.944) but significantly lower than that of TDAR (p<0.001); the CA of TDAR was significantly higher than that of Variog (p<0.001).

For the Shift SL2, the one-way ANOVA revealed a main effect of Feature (p<0.001). Tukey comparison showed that the CA of CSP-OvO was not significantly different with that of CSP-OvR (p=0.783) but significantly higher than that of TD (p<0.001), TDAR (p<0.001) and Variog (p<0.001); the CA of CSP-OvR was significantly higher than that of TD (p<0.001), TDAR (p=0.002) and Variog (p<0.001); the CA of TD was not significantly different with that of Variog (p=0.933) but significantly lower than that of TDAR (p=0.006); the CA of TDAR was significantly higher than that of Variog (p<0.001).

Figures 9 and 10 show the average confusion matrix of the five features (CSP-OvO, CSP-OvR, TD, TDAR and Variog) across all subjects in ST1 and ST2 respectively. We could find that the improvements of CSP features were mainly from NM, WF and UD in ST1 and were mainly from NM, WF and WE in ST2. Figures 11 and 12 show the average confusion matrix of the five features across all subjects in SL1 and SL2 respectively. We could find that the improvements of CSP features were mainly from NM, WF and UD in SL1 and were mainly from NM, TP, and WE in SL2. Comparing (a) and (b) of Figs 9, 10, 11 and 12, we found that the CA for each motion was similar between the two CSP features in all shift configurations (ST1, ST2, SL1 and SL2). Furthermore, we found the misclassifications of one motion vs. another (e.g. HO vs. UD) were also similar between the two CSP features in all shift configurations. These results demonstrated that the separability from one motion to another was very similar between the two CSP features and could explain why the classification performance of the two CSP features was not significant different in all shift configurations. Comparing (a)–(d) and (e) of Figs. 9, 10, 11 and 12, we found that the misclassifications of one motion vs. another of Variog feature were pretty different from that of the other four features. We suggested that this phenomenon was induced by the different type of classifier that the Variog feature used.
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig9_HTML.gif
Fig. 9
Average confusion matrix of the five features across all subjects in ST1. a CSP-OvO. b CSP-OvR. c TD. d TDAR. e Variog
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig10_HTML.gif
Fig. 10
Average confusion matrix of the five features across all subjects in ST2. a CSP-OvO. b CSP-OvR. c TD. d TDAR. e Variog
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig11_HTML.gif
Fig. 11
Average confusion matrix of the five features across all subjects in SL1. a CSP-OvO. b CSP-OvR. c TD. d TDAR. e Variog
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig12_HTML.gif
Fig. 12
Average confusion matrix of the five features across all subjects in SL2. a CSP-OvO. b CSP-OvR. c TD. d TDAR. e Variog
Shift in EMG feature space
Figure 13 shows the average RCS of the four features (CSP-OvO, CSP-OvR, TD and TDAR) across all subjects in the different shift configurations (ST1, ST2, SL1 and SL2). The average RCS of CSP-OvO changed from 1.15 to 1.25. The average RCS of CSP-OvR changed from 1.18 to 1.27. The average RCS of TD changed from 1.77 to 2.04. The average RCS of TDAR changed from 1.77 to 2.04. The average RCS of CSP features (CSP-OvO and CSP-OvR) was about two thirds of that of classic features (TD and TDAR) in all shift configurations (ST1, ST2, SL1 and SL2).
http://static-content.springer.com/image/art%3A10.1186%2Fs12984-015-0102-9/MediaObjects/12984_2015_102_Fig13_HTML.gif
Fig. 13
Average RCS across all subjects for four features: CSP-OvO, CSP-OvR, TD, and TDAR. Error bars represented the standard deviation
The two-way ANOVA revealed a statistically significant main effect of Feature (p<0.001). No other significant two-way interaction or main effect was revealed. For the factor of Feature, Tukey comparison showed that the RCS of CSP-OvO was not significantly different with that of CSP-OvR (p=1.0), but significant smaller than that of TD (p<0.001) and TDAR (p<0.001). It also showed that the RCS of CSP-OvR was significant smaller than that of TD (p<0.001) and TDAR (p<0.001). However, the RCS of TD was not significantly different with that of TDAR (p=1.0). The results demonstrated that the significant improvement in CA of the CSP features (CSP-OvO and CSP-OvR) was induced by the significantly smaller RCS in the feature space compared with the classic features (TD and TDAR). Unlike the CSP features (CSP-OvO and CSP-OvR), the significant difference in CA between TDAR and TD was not reflected on the RCS. The main reason was likely that the difference in CA between TDAR and TD was much smaller compared with the difference in CA between the CSP features and the classic features.

Interpretation of improvement in patterns of CSP features
Figures 4 and 5 show the last CSP patterns of motion 1 and motion 4 for CSP-OvO extension scheme in the transversal direction shift (ST1) and longitudinal direction shift (SL1) respectively. Figures 6 and 7 show the first CSP pattern of each active motion and rest motions for CSP-OvR extension scheme in the transversal direction shift (ST1) and longitudinal direction shift (SL1) respectively. We found that the locations emphasized by the CSP patterns before and after the shift were very similar. We believed that this was due to the underlying electrophysiology processes were not changed even in the presence of electrode shift. Thus, the CSP patterns of the EMG signals before electrode shift could emphasize the most discriminative locations after electrode shift and improve the CA in all electrode shift configurations (ST1, ST2, SL1 and SL2).

Discussion

As shown in Fig. 8, CSP features (CSP-OvO and CSP-OvR) significantly improved the CA over 10 % with respect to TD features in all shift configurations (ST1, ST2, SL1 and SL2) (p<0.05). The CSP-OvO feature achieved the highest average CA in all electrode configurations (ST, SL, ST1, ST2, SL1 and SL2) and significantly improved the average CA over 5 % with respect to TDAR features in all shift configurations (p<0.05). Except shift configuration ST2, the CSP-OvR feature significantly improved the average CA with respect to TDAR features in shift configurations ST1, SL1 and SL2 (p<0.05). Except transversal shift configurations (ST1 and ST2), the CSP features (CSP-OvO and CSP-OvR) significantly improved the average CA with respect to Variog features in longitudinal shift configurations (SL1 and SL2) (p<0.05). The CSP features (CSP-OvO and CSP-OvR) could achieve the average CA of ∼80 % in transversal direction shift (ST1 and ST2) and ∼95 % in longitudinal direction shift (SL1 and SL2). Thus, the CSP features could improve robustness against electrode shift for myoelectric control with respect to classic features.

Although there was no significant difference between the CA of CSP-OvO feature and that of CSP-OvR feature in all electrode configurations in Fig. 8, the average CA of CSP-OvO feature was slightly higher than that of CSP-OvR feature in all electrode configurations. We attributed this to the fact that the number of features extracted in CSP-OvO scheme was much more than the number of features extracted in CSP-OvR scheme. Therefore, the CSP-OvO feature extracted more helpful information for classification from the HD EMG signals with respect to the CSP-OvR feature. Furthermore, the results showed that the CSP features (CSP-OvO and CSP-OvR) performed best in longitudinal shift configurations (SL1 and SL2). We believed that this was presumably due to the fact that the electrode configuration was shifted in this case along the muscle fiber direction.

The results also showed that TDAR features significantly improved the CA with respect to TD features in shift configurations ST2, SL1 and SL2 (p<0.05). These confirmed the result that TDAR features significantly reduced sensitivity to electrode shift compared with TD features of a previous study [11]. Furthermore, the results showed that the Variog features significantly improved the average CA with respect to TD features in shift configurations ST1 and ST2 (p<0.05) and improved the average CA with respect to TDAR features in shift configuration ST1. However, the CA of the Variog features was not significantly different with that of TDAR features in shift configuration ST2 (p=0.3) and significantly lower than that of TDAR features in longitudinal direction shift (SL1 and SL2) (p<0.05). These results were partially consistent with the results of previous study [18]. Since parameters choosing was very important when using the Variog features and SVM classifiers, we suggested this might be due to that we could not find the optimized parameters in the current study. Moreover, this might be due to the number of motions considered in the current study was eleven which was larger than seven in that previous study.

As shown in Fig. 13, the average RCS of CSP features (CSP-OvO and CSP-OvR) across all subjects was significantly smaller than that of classic features (TD and TDAR) in all shift configurations (p<0.001). Since the average value of the feature vector of each motion and the covariance of all EMG data determined the parameters of the LDA classifier, the smaller RCS indicated that the LDA classifier trained before the electrode shift was more suitable for identifying the EMG data after the electrode shift. Thus, the CA of the features with smaller RCS should be greater than that with larger RCS. Here, we attributed that the improvement of CSP features (CSP-OvO and CSP-OvR) with respect to classic features (TD and TDAR) was induced by the relatively smaller RCS compared with the classic features.

For noise investigations, Hahne et al. have evaluated the performance of CSP features with a high baseline noise of individual channels and proved that the CSP features outperformed the classic features [17]. Thus, we did not test this effect, but only concentrated on the electrode shift in the current study.

The results showed that the proposed CSP features could improve the robustness against electrode shift for myoelectric control compared with the commonly used features. However, a limitation existed in the current study was that the proposed CSP features were not suitable for LD EMG. Geng et al. used CSP method to select LD channels from HD EMG, but they targeted channel selection and did not consider the problem of electrode shift [19]. Huang et al. also used an improved CSP for EMG classification, but they targeted LD EMG and did not consider the problem of electrode shift [24]. For LD EMG application, the proposed CSP features should be modified to common spatio-spectral pattern (CSSP) features and then evaluate their performance against electrode shift. In CSSP, several finite impulse response (FIR) spectral filters were embedded into CSP to constitute a spatio-spectral filter [20]. Since the embedded FIR filters would improve the number of channels for CSP, it could make the CSP suitable for LD EMG. We will investigate the performance against electrode shift of CSSP features for LD EMG application in the future.

As this work is an off-line analysis, an on-line study should be taken into account. In the future, the CSP features (CSP-OvO and CSP-OvR) will be tested in real-time experiments measured by three performance metrics, i.e. motion completion rate, motion completion time and motion selection time [32, 33]. There is a limitation in the current study that the subjects are intact-limb subjects. Although Scheme et al. showed that the results from intact-limb subjects could be generalized to amputees [34], the CSP features (CSP-OvO and CSP-OvR) should be tested on amputees in future work. To test the applicability of CSP features in practice, whether the computation capability of current micro-controller is enough for the analysis of HD EMG signals in myoelectric control should be investigated in the future.

Conclusion

This study evaluated whether the CSP of HD EMG signals could improve the myoelectric control performance under electrode shift for eleven classes of hand and wrist motions. Compared with the TD features, the CSP features significantly improved the CA over 10 % in all shift configurations (ST1, ST2, SL1 and SL2). Compared with the TDAR features, a. the CSP-OvO feature significantly improved the average CA over 5 % in all shift configurations; b. the CSP-OvR feature significantly improved the average CA in shift configurations ST1, SL1 and SL2. Compared with the Variog features, the CSP features significantly improved the average CA in longitudinal shift configurations (SL1 and SL2). It demonstrated that CSP of HD EMG signals could improve robustness against electrode shift for myoelectric control with respect to the commonly used features.

Declarations

Acknowledgements

This work was supported by the National Basic Research Program (973 Program) of China (No. 2011CB013305), the National Natural Science Foundation of China (No. 51375296, 51475292).

Open Access
This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://​creativecommons.​org/​licenses/​by/​4.​0/​), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://​creativecommons.​org/​publicdomain/​zero/​1.​0/​) applies to the data made available in this article, unless otherwise stated.

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

Conceived and designed the experiments: LZP XYZ. Performed the experiments: LZP DGZ XJS. Analyzed the data: LZP DGZ NJ. Contributed reagents/materials/analysis tools: LZP NJ XJS. Wrote the paper: LZP.

References

Farina D, Jiang N, Rehbaum H, Holobar A, Graimann B, Dietl H, et al.The extraction of neural information from the surface EMG for the control of upper-limb prostheses: emerging avenues and challenges. IEEE Trans. Neural Syst Rehabil Eng. 2014; 22(4):797–809.
View ArticlePubMed
Graupe D, Cline WK. Functional separation of EMG signals via ARMA identification methods for prosthesis control purposes. IEEE Trans Syst Man Cybern. 1975; SMC-5(2):252–9.
View Article
Doerschuk PC, Gustafon DE, Willsky A. Upper extremity limb function discrimination using EMG signal analysis. IEEE Trans. Biomed. Eng. 1983; BME-30(1):18–29.
View Article
Fougner A, Stavdahl O, Kyberd P, Losier Y, Parker P. Control of upper limb prostheses: Terminology and proportional myoelectric control-a review. IEEE Trans Neural Syst Rehabil Eng. 2012; 20(5):663–77.
View ArticlePubMed
Hudgins B, Parker P, Scott RN. A new strategy for multifunction myoelectric control. IEEE Trans Biomed Eng. 1993; 40(1):82–94.
View ArticlePubMed
Englehart K, Hudgins B. A robust, real-time control scheme for multifunction myoelectric control. IEEE Trans Biomed Eng. 2003; 50(7):848–54.
View ArticlePubMed
Young AJ, Hargrove LJ, Kuiken TA. Improving myoelectric pattern recognition robustness to electrode shift by changing interelectrode distance and electrode configuration. IEEE Trans Biomed Eng. 2012; 59(3):645–52.
PubMed CentralView ArticlePubMed
Liu J, Zhang D, Sheng X, Zhu X. Quantification and solutions of arm movements effect on semg pattern recognition. Biomedical Signal Process and Control. 2014; 13:189–97.
View Article
Pan L, Zhang D, Sheng X, Zhu X. Improving myoelectric control for amputees through transcranial direct current stimulation. IEEE Trans Biomed Eng. 2015; 62(8):1927–1936.
View ArticlePubMed
Graupe D, Salahi J, Kohn KH. Multifunctional prosthesis and orthosis control via microcomputer identification of temporal pattern differences in single-site myoelectric signals. J Biomed Eng. 1982; 4(1):17–22.
View ArticlePubMed
Young AJ, Hargrove LJ, Kuiken TA. The effects of electrode size and orientation on the sensitivity of myoelectric pattern recognition systems to electrode shift. IEEE Trans Biomed Eng. 2011; 58(9):2537–544.
PubMed CentralView ArticlePubMed
Hargrove L, Englehart K, Hudgins B. A training strategy to reduce classification degradation due to electrode displacements in pattern recognition based myoelectric control. Biomed signal process control. 2008; 3(2):175–80.
View Article
Østlie K, Lesjø IM, Franklin RJ, Garfelt B, Skjeldal OH, Magnus P. Prosthesis rejection in acquired major upper-limb amputees: a population-based survey. Disabil Rehabil Assist Technol. 2012; 7(4):294–303.
View ArticlePubMed
Huang H, Zhou P, Li G, Kuiken T. Spatial filtering improves emg classification accuracy following targeted muscle reinnervation. Ann Biomed Eng. 2009; 37(9):1849–1857.
PubMed CentralView ArticlePubMed
Daley H, Englehart K, Hargrove L, Kuruganti U. High density electromyography data of normally limbed and transradial amputee subjects for multifunction prosthetic control. J Electromyogr Kinesiol. 2012; 22(3):478–84.
View ArticlePubMed
Muceli S, Farina D. Simultaneous and proportional estimation of hand kinematics from emg during mirrored movements at multiple degrees-of-freedom. IEEE Trans Neural Syst Rehabil Eng. 2012; 20(3):371–8.
View ArticlePubMed
Hahne JM, Graimann B, Muller K. Spatial filtering for robust myoelectric control. IEEE Trans Biomed Eng. 2012; 59(5):1436–1443.
View ArticlePubMed
Stango A, Negro F, Farina D. Spatial correlation of high density EMG signals provides features robust to electrode number and shift in pattern recognition for myocontrol. IEEE Trans Neural Syst Rehabil Eng. 2015; 23(2):189–198.
View ArticlePubMed
Geng Y, Zhang X, Zhang YT, Li G, et al. A novel channel selection method for multiple motion classification using high-density electromyography. Biomed eng online. 2014; 13(1):102.
PubMed CentralView ArticlePubMed
Chen Q, Gong P. Automatic variogram parameter extraction for textural classification of the panchromatic ikonos imagery. IEEE Trans Geosci Remote Sensing. 2004; 42(5):1106–1115.
View Article
Tonye E, Fotsing J, Zobo BE, Tankam NT, Kanaa TF, Rudant JP. Contribution of variogram and feature vector of texture for the classification of big size sar images. In: Proc. Seventh Int. Conf. on Signal Image Technology and Internet-Based Systems. New Jersey: IEEE: 2011. p. 382–9.
Ramoser H, Muller-Gerking J, Pfurtscheller G. Optimal spatial filtering of single trial eeg during imagined hand movement. IEEE Trans Rehabil Eng. 2000; 8(4):441–6.
View ArticlePubMed
Blankertz B, Dornhege G, Krauledat M, Müller KR, Curio G. The non-invasive berlin brain–computer interface: fast acquisition of effective performance in untrained subjects. NeuroImage. 2007; 37(2):539–50.
View ArticlePubMed
Huang G, Zhang Z, Zhang D, Zhu X. Spatio-spectral filters for low-density surface electromyographic signal classification. Medical & biological engineering & computing. 2013; 51(5):547–55.
View Article
Dornhege G. Toward Brain-computer Interfacing. Massachusetts: MIT press; 2007.
Huang Y, Englehart KB, Hudgins B, Chan AD. A gaussian mixture model based classification scheme for myoelectric control of powered upper limb prostheses. IEEE Trans Biomed Eng. 2005; 52(11):1801–1811.
View ArticlePubMed
Farrell TR, Weir RF. A comparison of the effects of electrode implantation and targeting on pattern classification accuracy for prosthesis control. IEEE Trans Biomed Eng. 2008; 55(9):2198–211.
PubMed CentralView ArticlePubMed
Chen X, Zhang D, Zhu X. Application of a self-enhancing classification method to electromyography pattern recognition for multifunctional prosthesis control. J Neuroengineering and Rehabil. 2013; 10(1):44.
View Article
Englehart K, Hudgins B, Parker PA, Stevenson M. Classification of the myoelectric signal using time-frequency based representations. Medical Engineering & Physics. 1999; 21(6):431–8.
View Article
Chang CC, Lin CJ. Libsvm: A library for support vector machines. ACM Trans Intell Syst Technol (TIST). 2011; 2(3):27.
Duda RO, Hart PE, et al, Vol. 3. Pattern Classification and Scene Analysis. New York: Wiley; 1973.
Kuiken TA, Li G, Lock BA, Lipschutz RD, Miller LA, Stubblefield KA, Englehart KB. Targeted muscle reinnervation for real-time myoelectric control of multifunction artificial arms. JAMA. 2009; 301(6):619–28.
PubMed CentralView ArticlePubMed
Li G, Schultz AE, Kuiken TA. Quantifying pattern recognition-based myoelectric control of multifunctional transradial prostheses. IEEE Trans Neural Syst Rehabil Eng. 2010; 18(2):185–92.
PubMed CentralView ArticlePubMed
Scheme E, Englehart K. Electromyogram pattern recognition for control of powered upper-limb prostheses: State of the art and challenges for clinical use. J Rehabil Res Dev. 2011; 48(6):643.
View ArticlePubMed

Drosophila Imp iCLIP identifies an RNA assemblage coordinating F-actin formation

Heidi Theil Hansen1, Simon Horskjær Rasmussen1, Sidsel Kramshøj Adolph1, Mireya Plass1, Anders Krogh1, Jeremy Sanford2, Finn Cilius Nielsen3 and Jan Christiansen1*

  • Corresponding author: Jan Christiansen janchr@bio.ku.dk

Author Affiliations
1 Department of Biology, Center for Computational and Applied Transcriptomics, University of Copenhagen, Ole Maaloes Vej 5, Copenhagen, 2200, Denmark

2 MCD Biology, University of California, Santa Cruz 95064, CA, USA

3 Center for Genomic Medicine, Rigshospitalet, University of Copenhagen, Copenhagen, 2100, Denmark

For all author emails, please log on.

Genome Biology 2015, 16:123 doi:10.1186/s13059-015-0687-0

The electronic version of this article is the complete one and can be found online at: http://genomebiology.com/2015/16/1/123

Received: 10 February 2015
Accepted: 29 May 2015
Published: 9 June 2015
© 2015 Hansen et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Formula display:
Abstract
Background
Post-transcriptional RNA regulons ensure coordinated expression of monocistronic mRNAs encoding functionally related proteins. In this study, we employ a combination of RIP-seq and short- and long-wave individual-nucleotide resolution crosslinking and immunoprecipitation (iCLIP) technologies in Drosophila cells to identify transcripts associated with cytoplasmic ribonucleoproteins (RNPs) containing the RNA-binding protein Imp.

Results
We find extensive binding of Imp to 3′ UTRs of transcripts that are involved in F-actin formation. A common denominator of the RNA–protein interface is the presence of multiple motifs with a central UA-rich element flanked by CA-rich elements. Experiments in single cells and intact flies reveal compromised actin cytoskeletal dynamics associated with low Imp levels. The former shows reduced F-actin formation and the latter exhibits abnormal neuronal patterning. This demonstrates a physiological significance of the defined RNA regulon.

Conclusions
Our data imply that Drosophila Imp RNPs may function as cytoplasmic mRNA assemblages that encode proteins which participate in actin cytoskeletal remodeling. Thus, they may facilitate coordinated protein expression in sub-cytoplasmic locations such as growth cones.

Background
To establish a neuronal network, protrusions, in the form of filopodia and lamellipodia, project from the developing axons into their surroundings in search of their appropriate postsynaptic targets. These protrusions are formed at the edge of the outermost part of the axon, the growth cone. This highly motile domain, consisting of bundles of filamentous actin, microtubules and their associated proteins, is an exquisite sensor of chemical gradients and is constantly undergoing dynamic structural changes in response to the surrounding guidance cues (reviewed in [1]). As the continuous remodeling of the growth cone requires the localization of proteins to the tip, synchronous on-site translation of mRNAs makes it possible to swiftly produce a subset of proteins needed for accurate responses to the guidance cues (reviewed in [2]).

The post-transcriptional RNA regulon/operon hypothesis posits that functionally related monocistronic mRNAs are co-regulated by common trans-acting RNA-binding proteins (RBPs) [3]. The distinct binding profiles of yeast RNA-binding Puf isomers provide a striking example of how this type of regulation works with regard to common function and mRNA localization [4]. In situ hybridization studies in Drosophila blastoderms have estimated that 71 % of 3370 examined mRNAs are localized in diverse patterns [5], so the interplay between RBPs and target mRNAs is pervasive. However, most localization studies have focused on a single mRNA species and a corresponding RBP, exemplified by the β-actin mRNA/ZBP1 paradigm in fibroblasts and growth cones [6], [7]. ZBP1 belongs to a vertebrate family of RNA-binding proteins (other acronyms include IGF2BP, VICKZ, IMP) with a domain architecture of two RNA recognition motifs and four K homology (KH) domains, the latter providing the RNA attachment platform [8], [9]. The majority of vertebrate genomes contain three paralogs of this particular RBP family, whereas a single locus on the X chromosome encodes the Drosophila homolog Imp (insulin-like growth factor II mRNA-binding protein), containing the four conserved KH domains, but lacking the N-terminal two RNA recognition motifs. In addition, the Drosophila homolog encompasses a C-terminal Gln-rich extension that is not observed among vertebrates (Fig. 1a). Imp is expressed in the oocyte, the blastoderm embryo as well as in the developing nervous system and pole cells [10]–[12]. During mid-oogenesis, maternal Imp co-localizes with both gurken and oskar mRNAs, which are critical for dorsoventral and anteriorposterior axis formation, respectively [11], [13]. However, the functional significance of Imp in axis formation is uncertain since localization of both gurken and oskar mRNAs is unaffected in an Imp-deficient background, whereas Imp overexpression disrupts the localization and translational regulation of both transcripts. At later stages of development, increased pharate adult lethality is observed in Imp mutants, which is a phenotype often associated with defective synaptic transmission. More specifically, Imp mutant larvae exhibited smaller synaptic junctions [14], and a recent study revealed that extensive remodeling of γ-neurons in the mushroom body during pupation is jeopardized in the absence of Imp. It is not the initial growth of the axons that is compromised but their remodeling in terms of both length and directionality, and overexpression of the profilin homolog Chickadee is able to partially rescue the remodeling defect [15]. Imp is also expressed in the tail end of cyst cells surrounding elongating spermatids, and both spermatids and cyst cells elongate to an extreme length in an F-actin dependent fashion [12].

thumbnailFig. 1. Drosophila Imp is present in cytoplasmic RNPs. a Protein domain architecture of Drosophila Imp and homologous vertebrate IMPs. The four KH domains are conserved whereas the glutamine (Q)-rich region and the two RNA recognition motifs (RRM) are structural features present in arthropods and chordates, respectively. b Immunocytochemistry of S2 cell with DAPI staining (blue), anti-Imp antibody (red) and phalloidin (green). c Correlation between the enrichment of immunoprecipitated transcripts in Imp RNPs from two biological RNA-binding protein immunoprecipitation (RIP)-seq replicates. The scatter plot shows the natural logarithm of RIP enrichment values per nucleotide in the whole transcript for each RIP-seq replicate. d Venn diagram of the 200 most enriched transcripts in Imp RNPs from two biological RIP-seq replicates
In recent years, high-throughput cross-linking and immunoprecipitation (CLIP) procedures have been introduced to obtain information regarding direct binding sites for individual RBPs on a transcriptome-wide scale in vivo [16]–[18]. In this way, novel details of DGCR8 (DiGeorge syndrome critical region 8) functions [19], regulation of splicing [16], [20], and alternative polyadenylation [21] have been elucidated. The salient feature of these procedures is the ability to identify the exact cross-linked position on a global scale by using stringent washing regimens to minimize non-covalent associations.

In the present study, we employed three different high-throughput procedures to identify the full targetome of Drosophila Imp and thus assess the post-transcriptional RNA regulon hypothesis. The data revealed extensive 3′ untranslated region (UTR) attachments in mRNAs encoding components crucial for actin cytoskeletal remodeling, and subsequent Imp knock-down analysis showed decreased F-actin levels in S2 cells. Moreover, an Imp-deficient fly exhibited a phenotype with difficulties reaching the pharate adult stage, mainly due to defects in the peripheral nervous system. Taken together, our study identifies an Imp-mediated RNA assemblage containing transcripts that encode proteins coordinating F-actin formation.

Results
Drosophila Imp localizes to cytoplasmic granules containing multiple mRNA species
To address the subcellular localization and potential of forming RNP granules in the Drosophila S2 cell line, immunostaining of endogenous Imp was performed with co-staining with phalloidin (Fig. 1b). Immunostaining revealed punctate fluorescence with an optical diameter of about 200 nm dispersed throughout the cytoplasm with limited nuclear Imp immunoreactivity, suggesting Imp localization in predominantly cytoplasmic RNP granules.

To characterize the repertoire of RNA targets that are associated with the Imp granules in Drosophila S2 cell lysates, we sequenced Imp-associated mRNAs following immunoprecipitation from cytoplasmic extracts using anti-Imp antibody-coated Protein A Dynabeads. Mapping statistics are listed in Figure S1a in Additional file 1. The enrichment factor (the ratio of immunoprecipitated RNA:input RNA) for each transcript was calculated, and the correlation of the values between the biological replicates was plotted. As shown in Fig. 1c, the replicate experiments are strongly correlated (R = 0.85), underscoring the precision of our measurements. Of the top 200 ranked transcripts in each biological replicate, 158 were present in both replicates (Fig. 1d). Gene ontology (GO) term analysis was performed on the 158 transcripts and showed enrichment in GO terms involved in neuronal development, reproduction, and cytoskeletal organization (Figure S1b in Additional file 1). This result is in agreement with earlier observations of Imp being implicated in ageing in the testis stem-cell niche [22], in spermatid cyst cell elongation [12], and in axonal remodeling in mushroom body neurons [15].

iCLIP and PAR-iCLIP identify direct Imp target sites in 3′ UTRs
Since RIP-seq identifies both direct and indirect RNA targets without information regarding the position of interaction, we performed individual-nucleotide resolution cross-linking (iCLIP) [23] as well as a modified version of photoactivatable ribonucleoside-CLIP (PAR-CLIP) [17], which we have called PAR-iCLIP, on cytoplasmic extracts of S2 cells to pinpoint the positions of direct binding. We generated two biological replicate iCLIP and two biological replicate PAR-iCLIP sequencing libraries for Imp. A representative autoradiograph of both types of RNA–Imp cross-linking complexes used for library generation is depicted in Figure S2a in Additional file 1, and the mapping statistics of the sequenced reads are summarized in Figure S2b in Additional file 1. We confidently mapped 37–50 % of the reads (2.1–4.7 × 10 6 ) to the Drosophila genome, demonstrating the high quality of the produced libraries.

In order to quantify the strength of Imp binding and eliminate the contribution from RNA expression, the Imp iCLIP and PAR-iCLIP tag clusters were normalized to transcript levels determined by RNA-seq. To provide a global view of Imp RNA-binding specificity we analyzed the average distribution of tag clusters across mRNA regions, including the 5′ UTR, coding sequence (CDS) and 3′ UTR (Fig. 2a). The CLIP tag clusters predominantly map to the 3′ UTR, showing a striking preference for Imp binding to this region regardless of the chosen procedure. In addition, the average distribution profile revealed that a higher 3′ UTR:CDS ratio in terms of normalized Imp CLIP tags was observed with the long wavelength PAR-iCLIP approach, reflecting the lesser background observed in the autoradiograph in Figure S2a in Additional file 1.

thumbnailFig. 2. Imp CLIP identifies CA-rich binding motifs in 3′ UTRs. a Standardized transcript profile showing the CLIP enrichment from short-wave (left panel) and long-wave (right panel) cross-linking approaches. The y-axis displays the mean enrichment per nucleotide in bins across the 5′ UTRs (20), coding regions (CDS; 50) and 3′ UTRs (50) on the x-axis. b UCSC genome browser tracks of normalized CLIP-tags on squid (sqd), chickadee (chic), pAbp, and visgun (vsg) primary transcripts (red depicts tags derived from minus-strand transcripts and green those from plus-strand transcripts). The blue boxes correspond to the coding region, whereas UTRs and introns are depicted with thick blue lines and thin blue lines, respectively. c Electrophoretic mobility-shift assay of RNA segments from the coding region and 3′ UTR of the sqd transcript (upper panel) and the pAbp transcript (lower panel) with recombinant Drosophila Imp protein (12.5 nM, 25 nM, 50 nM, and 100 nM). d Scatter plot displaying the Z-score of all tetramers in the cross-link region (x-axis) and flanking regions (y-axis) in the 3000 most enriched iCLIP clusters compared with background clusters. The words with the highest enrichment either in the cross-link region or in the flanking regions are highlighted in red (UAUA, AUAU, AUUU, UAUU, UUAU, UUUU) and blue (ACAC, CACA, AACA, ACAA, CAAA, AAAC, CAAC), respectively. e Plot of the running mean (over five nucleotides) of positional enrichment, showing the distribution of the CA-rich (blue line) and UA-rich (red line) motifs identified in panel (d) in highly enriched iCLIP clusters. The shaded area highlights the cross-link region. Enrichment of the same motifs in randomly selected (background) clusters from 3′ UTRs is depicted by the dashed light blue line (CA-rich motifs) and pink line (UA-rich motifs)
Imp binding frequently covered the entire 3′ UTR with peaks and valleys in the tag clusters, as shown here for squid, chickadee, pAbp and visgun transcripts (Fig. 2b). To assess the reproducibility of the experiments, the correlation of the total enrichment values, based on CLIP tags per nucleotide in 3′ UTRs normalized to the average RNA-seq expression, was evaluated between the two iCLIP libraries and between the two PAR-iCLIP libraries (Figure S2c in Additional file 1). There was a high correlation (R = 0.74 and R = 0.74 among iCLIP and PAR-iCLIP libraries, respectively) and, more importantly, a similar high correlation (R = 0.76) between the enrichment values obtained using iCLIP and PAR-iCLIP procedures (Figure S2c in Additional file 1).

Electrophoretic mobility-shift assays of squid and pAbp 3′ UTR segments, which exhibited efficient UV cross-linking in vivo, confirmed the direct binding of recombinant Imp protein to target RNAs in vitro (Fig. 2c; Figure S2i in Additional file 1). Moreover, the absence of Imp interaction with coding regions in vivo was confirmed by electrophoretic mobility-shift assay in vitro. At increasing concentrations of Imp we observed an oligomerization pattern of recombinant Imp on individual 3′ UTR segments, suggesting formation of higher-order RNP complexes. In contrast to what was observed for the predominantly 3′ UTR-binding human antigen R (HuR) RBP, the number of Imp iCLIP tags per nucleotide correlated with 3′ UTR length (Figure S2d in Additional file 1), thus corroborating the oligomerization behavior of Imp on its RNA targets on a global scale. Taken together, the in vivo cross-linking and in vitro binding experiments imply that 3′ UTRs provide efficient Imp binding platforms, whereas this is not the case for coding regions.

In order to obtain insight into the motif recognized by Imp, we merged the two replicates for each iCLIP protocol and carried out motif analyses using cWords [24] on 25,000 clusters, ranked by the maximal CLIP enrichment of each cluster. The top 10 tetramers obtained from these analyses are depicted in Figure S2e in Additional file 1. We found a significant enrichment of CA-rich motifs, mainly as CA and CAA repeats, and UA-rich motifs in both datasets. The co-occurrence analysis of MACA motifs (M designates C or A) in highly enriched CLIP clusters showed fluctuations following a periodic pattern, confirming the repetitive structure of these motifs in the CLIP clusters (Figure S2f and “Supplementary Materials and methods” in Additional file 1).

Crystallographic studies of single KH domain–nucleic acid interactions such as Nova-2 KH3 [25] and hnRNP K KH3 [26] suggest that KH domains recognize tetranucleotides situated in a molecular vise between the GxxG signature loop and the variable loop. Therefore, we analyzed the positional enrichment of all tetramers in the cross-link region defined by a high occurrence of iCLIP read starts [18]. Figure 2d shows the Z-score of tetramers in the 3000 most enriched iCLIP clusters in 3′ UTRs compared with randomly selected clusters from 3′ UTRs (background) in the cross-link region (x-axis) and the flanking regions (y-axis). The scatter plot identifies a group of UA-rich words that are highly enriched in the cross-link region whereas the CA-rich words are mainly enriched in the flanking regions. The positional enrichment of these motifs in the vicinity of iCLIP clusters is depicted in Fig. 2e. We observe a strong enrichment of the UA-rich words around the beginning of clusters (grey shaded area) whereas CA-rich motifs appear mainly in the flanking regions. Similar results were also found for PAR-iCLIP (Figure S2g, h in Additional file 1), although a 4-thiouridine-mediated bias towards U-rich motifs was observed (Figure S2e in Additional file 1). We conclude that Drosophila Imp has a specific positional sequence preference that extends over distances longer than a single tetramer motif.

Imp binds transcripts important for correct neural development and reproduction
In order to assess the transcriptome-wide Imp association, we examined the CLIP and RIP-seq enrichment of all expressed transcripts in the S2 cell transcriptome (Fig. 3a; “Supplementary Materials and methods” in Additional file 1). In both cases, the profiles showed that many transcripts exhibited some degree of association with Imp. However, a subpopulation of about 200 transcripts displayed a more than three-fold enrichment compared with RNA-seq.

thumbnailFig. 3. Imp RNPs contain transcripts associated with neural development and reproduction. a Enrichment of combined iCLIP and PAR-iCLIP clusters (red curve) and RIP-seq data (blue curve) across the Imp targetome. On the x-axis, the transcripts are ranked according to the enrichment of the individual transcript shown on the y-axis. The horizontal dotted line at y-value 1 indicates equal relative amounts of CLIP clusters or RIP-seq reads and RNA-seq levels. b Venn diagram of highly ranked transcripts derived from six individual experiments utilizing three distinct experimental approaches. In each experiment the 200 most highly ranked transcripts are identified, and the common pool for each experimental approach is represented by the three ellipses (157 for RIP-seq, 154 for iCLIP, and 168 for PAR-iCLIP) c GO term analyses of the 86 transcripts present among the 200 most highly ranked transcripts in each of six individual experiments utilizing three distinct experimental approaches and a P value cutoff of 0.01. Figure S3b in Additional file 1 lists the protein family or conserved domains encoded by each of the 86 Drosophila transcripts. d Standardized transcript profile showing the average number of motifs per nucleotide predicted with the hidden Markov model. The y-axis represents the average motif counts per nucleotide and the x-axis shows the position of the motif along the standardized transcript. The red line indicates the number of matches in the top 86 transcripts identified by iCLIP, PAR-iCLIP and RIP-seq datasets. The blue line displays the average motifs in all transcripts with a least 30 % RNA-seq coverage. Vertical dashed lines indicate borders between UTRs and CDS
To define enriched Imp targets in a conservative manner, the intersection between the top 200 targets identified by iCLIP, PAR-iCLIP and RIP-seq was selected, resulting in 86 transcripts common to all six experiments (Fig. 3b). Enrichment of both iCLIP and PAR-iCLIP clusters among the subpopulation of highly ranked transcripts (red curves in Figure S3a in Additional file 1) was above that exhibited by the whole transcriptome (blue curves in Figure S3a in Additional file 1).

The 86 shared transcripts, as well as their protein family association, are ranked in Figure S3b in Additional file 1 according to their average enrichment, and GO term enrichment analysis was performed on the 86 top targets found in all datasets. Significant enrichment for GO categories representing biological processes such as reproduction, neural development and cell polarity was observed (Fig. 3c).

To examine the enrichment of the previously identified Imp binding motif among the 86 top ranked transcripts, we designed a hidden Markov model (HMM) that models the CA-rich repeat regions and the UA-rich cross-link motif to identify Imp binding sites (see “Materials and methods”; Figure S3c in Additional file 1). As expected, we observe a two- to threefold higher enrichment of Imp binding motifs in 76 top targets (Fig. 3d, red line) compared with the S2 cell line transcriptome (Fig. 3d, blue line). An example of these binding motifs can be found in the highly enriched clusters present in Pendulin and pAbp 3′ UTRs (Figure S3d in Additional file 1).

We conclude that Imp binds to a UA-rich motif flanked by CA-rich motifs in 3′ UTRs of transcripts important for neural development and reproduction. By examining the list of the shared 86 transcripts as well as their GO term categories, we can also conclude that Imp has a strong preference for 3′ UTRs in transcripts encoding crucial components of cellular protrusion dynamics, such as Moesin, Chickadee (profilin homolog), Rho GTPases (Rac1 and Cdc42), the exchange inhibitor GDI, the 14-3-3epsilon adaptor, and Arf79f (Arf1 homolog), to name a few.

Knockdown of Imp affects the F-actin level
The RNA interaction network described in Fig. 3c suggests that Imp may play a role in remodeling the actin cytoskeleton. To test this hypothesis we examined the effect of Imp depletion on F-actin levels and organization using phalloidin, which binds to the interface of actin subunits in F-actin. Double-stranded RNA interference (dsRNAi)-mediated knockdown of Imp was carried out in parallel with dsRNAi-mediated knockdown of firefly luciferase, the latter acting as a negative control. Since inhibition of the Rho-associated protein kinase (ROCK) with Y-27632 has been shown to result in the suppression of cofilin phosphorylation and increased severing of F-actin [27], we also examined the effect of Imp knockdown in the absence and presence of the ROCK inhibitor in S2 cells seeded on concanavalin A [28].

Imp immunoreactivity and phalloidin staining in an S2 cell are depicted in Fig. 4a, whereas visual fields from cell dishes subjected to various treatments are shown in Fig. 4b. Knockdown of Imp resulted in an expected decrease in Imp immunoreactivity, and the phalloidin staining was apparently also decreased in the Imp-deficient cells. In order to quantify the effect on F-actin levels, ten fields of each cell dish were analyzed and the red Alexa 568 (anti-Imp) and green Alexa 488 (phalloidin) pixels were measured using the HISTO functionality of the ZEN 2011 software package. The mean intensities in each field, after thresholding (Fig. 4c), are listed in Figure S4a in Additional file 1. When counting the phalloidin and anti-Imp stained pixels it emerged that a reduction of endogenous Imp levels resulted in significantly less F-actin staining compared with S2 cells treated with control luciferase dsRNA (21 % reduction, P value 0.04; Fig. 4d). The reduction was more pronounced in a ROCK inhibitor-stressed background (36 % reduction, P value 0.002), in spite of a lack of any major changes in F-actin organization upon treatment with the ROCK inhibitor in wild-type S2 cells.

thumbnailFig. 4. Knockdown of Imp leads to diminished F-actin levels. a Structured illumination of Alexa 488-phalloidin (green) and Alexa 568 (red) Imp in Drosophila S2 cells. b Selected visual fields of untreated, luciferase dsRNA-treated, or imp dsRNA-treated S2 cells, in the presence or absence of the ROCK inhibitor Y-27632. c The absolute frequencies of Alexa 488 and Alexa 568 pixels in a visual field of untreated S2 cells and cells with Imp knock-down, which were used to derive the average mean intensities shown in panel (d). d The pixel intensities of ten frames from each cell dish were measured, and the boxplots display the average mean intensity of Imp immunoreactivity (left panel) and phalloidin staining (right panel). P values were calculated using Student’s t-test. *P ≤ 0.05, **P ≤ 0.01 and ****P ≤ 0.0001
Diminished F-actin staining results from an inability to assemble filamentous actin from either a reduced level of actin monomers or a lack of other factors involved in the process. In order to determine if the reduced level of Imp caused down-regulation of mRNAs participating in F-actin formation, RNA-seq from imp dsRNA- and luciferase dsRNA-treated cells were carried out. Imp knockdown resulted in a reduction of the Imp protein level to 11–21 % of the endogenous level in S2 cells (Figure S4b in Additional file 1). The biological triplicate analyses of the steady-state mRNA levels were unable to identify statistically significant changes when correcting for multiple hypotheses testing (for output of differential expression sequencing (DESeq) analysis see Gene Expression Omnibus accession number GSE62997). Since knock-down of Imp did not result in any changes at the transcript level, we investigated whether reduction of Imp caused diminished levels of actin monomers. A western blot analysis of imp dsRNA- and luciferase dsRNA-treated cell lysates was unable to detect a change of actin monomer concentration in an Imp-deficient setting (Figure S4c in Additional file 1). Therefore, we infer that the level of Imp affects the formation of F-actin rather than individual transcripts or actin monomer concentration.

Knockdown of Imp results in abnormal neuronal patterning in embryos
Since reduction of Imp levels resulted in compromised F-actin formation in S2 cells, we hypothesized that an Imp-deficient embryo would exhibit developmental defects. Therefore, we examined the imp G0072 mutant strain, which exhibits decreased imp transcript levels and is semi-lethal due to a P-element insertion in the imp locus [29]. Quantitative RT-PCR analysis of a collection of pupae showed that the imp transcript level was reduced to 8 % of the wild-type level (data not shown). To corroborate that the semi-lethal phenotype of imp G0072 flies was due to a disturbance at the imp locus, excision of the P-element was induced and both wild-type sequence and phenotype were restored, as previously described by Geng and Macdonald [13] (Figure S5 and “Supplementary Materials and methods” in Additional file 1). Increased mortality of the imp G0072 progeny compared with both wild-type and an FM7c balanced/wt was observed during all stages of life. Closer examination of the eclosing flies revealed that only 12 % (n = 383) were hemizygous males compared with the expected 25 %. Moreover, we found 30 hemizygous males struggling to emerge from their pupal cases, and among the 45 hemizygous males that did manage to eclose, all drowned shortly after in the medium.

The observation that the imp G0072 mutants were either incapable of emerging from their pupal cases or drowned in the medium suggests locomotion problems that could be ascribed to defects in the nervous system. This led us to investigate the neuronal patterning in imp G0072 mutant embryos. To identify the 25 % of the imp G0072 progeny that exhibited no imp wild-type allele, a new line from the imp G0072 strain was created, in which a β-galactosidase (β-gal) marker was present on the X-chromosome balancer FM7c, so that expression of β-gal would be linked to the presence of an FM7c balancer with a wild-type imp allele. In order to record the fraction of embryos with an abnormal neuronal pattern, the nervous system of wild-type embryos, and embryos of crosses between either wt/FM7c or imp G0072 /FM7c females and FM7c/Y males, was visualized by enzyme-linked immunohistochemistry. Twenty-nine percent of the embryos showed an abnormal phenotype, which suggested that aside from the hemizygous males (imp G0072 /Y), some heterozygotes (imp G0072 /FM7c) were also affected by the P-element insertion. The binominal P values of imp G0072 embryos compared with either wild-type or wt/FM7c embryos were calculated, showing significantly increased rates of embryos with abnormal morphology and nervous systems (Fig. 5a).

thumbnailFig. 5. Reduced Imp levels result in aberrant neuronal patterning. a Theoretical genotypes and observed abnormal embryos in the F1 generation. FM7c designates an X-chromosome balancer containing a wild-type (wt) imp gene and β-gal gene, whereas imp G0072 contains a P-element in the imp locus on the X chromosome. Wild type and FM7c balanced wild type exhibited equal rates of abnormal embryos. A binominal P value of <0.001 was calculated for the imp G0072 mutant compared with either wild type or FM7c balanced wild type. Number of embryos counted: wt/wt + wt/Y cross 575 embryos, wt/FM7c + FM7c/Y cross 712 embryos, imp G0072 / FM7c + FM7c/Y cross 558 embryos. b Whole-mount immunohistochemistry with the monoclonal antibody 22C10 staining neurofilamental Futsch (red), and β-gal (green) marking the X-chromosome balancer. Top left: an embryo with multiple wild-type Imp alleles and a nervous system that appears normal (scale bar 50 μm). Bottom left: an embryo with one Imp allele which is morphologically abnormal and an aberrant nervous system (scale bar 50 μm). Top right: an imp G0072 hemizygous male embryo with a relatively normal morphology and a distinguishable central nervous system exhibiting pathfinding defects (scale bar 100 μm). Bottom right: an imp G0072 hemizygous abnormal male embryo revealing extensive patterning defects (scale bar 50 μm)
To examine the neuronal patterning of mutant embryos, whole-mount immunostaining of the imp G0072 /FM7c + FM7c/Y resulting embryos, using antibodies against β-gal and the MAP1B-like protein Futsch (mAb 22C10), was carried out. Staining with 22C10 showed a large variation among the mutants, ranging from a morphologically normal looking embryo with a nervous system that at least superficially appears normal (Fig. 5b, genotype FM7c/FM7c or FM7c/Y), to a morphologically normal embryo with an underdeveloped and abnormal nervous system (Fig. 5b, genotypes imp G0072 /FM7c and imp G0072 /Y). We conclude that embryos with reduced Imp levels exhibit abnormal neuronal patterning in the developing nervous system, and that even survivors at the pharate adult stage have profound locomotion problems.

Discussion
In this study, we have identified the RNA targets of the endogenous Drosophila Imp protein on a transcriptome-wide scale, utilizing a highly specific polyclonal antibody in six biologically independent experiments with three different methods. In general, both iCLIP and PAR-iCLIP data revealed a striking preference for direct Imp attachment to 3′ UTRs at the expense of especially the coding region. The distinct binding profile in Fig. 2a provides comprehensive evidence that the CLIP tags are not the result of random unspecific cross-linking of an abundant RBP and is also in sharp contrast to the only other published CLIP study in the S2 cell line, namely an iCLIP analysis of Ago2 in nuclear extracts that exhibited a narrower peak profile [30]. An analysis of motifs associated with Imp iCLIP clusters showed that repeated CA-rich segments provide the most likely RNA–protein interface of a composite binding site consisting of a UA-rich cross-link motif, similar to the UUUAY motif previously identified for Drosophila Imp using SELEX [11], flanked by CA-rich regions in the 3′ UTR, which are enriched considerably among the top 86 targets. The CA-rich motif is reminiscent of the CAUH motif reported to be involved in binding of Drosophila Imp to the unpaired 3′ UTR [22]; the CAUH motif is also recognized by human IGF2BP1–3 in HEK293 cells, and in 30 % of cases is repeated within 3–5 nucleotides [17]. Moreover, a bipartite recognition motif consisting of CGGAC and MCAY within 10–25 nucleotides of each other has been suggested for the human KH34 di-domain [31], although the RNA-protein interface of the full-length protein appears to be considerably more complex spatially [9].

Although both Drosophila Imp and its vertebrate homologs contain the characteristic four KH domains, the former oligomerizes on targets of merely 200 nucleotides (Fig. 2c) whereas the human homologs only dimerize on 100–250 nucleotide target RNAs [32]. The multimerization of the Drosophila isoform, presumably mediated by the C-terminal low-complexity Gln-rich domain, may have a bearing on the extensive 3′ UTR cross-linking pattern reported here. In fact, there is a positive correlation between the average CLIP enrichment per nucleotide and the length of the 3′ UTR, which was not observed in CLIP experiments with other RBPs that also show a strong 3′ UTR enrichment, such as HuR [33] (Figure S2f in Additional file 1), implying that Drosophila Imp preferentially binds transcripts with long 3′ UTRs. This is a hallmark of cooperative multimerization but additional explanations are definitely feasible in a competitive in vivo scenario. Interestingly, hydrogels derived from low-complexity RBPs exhibit a preponderance to incorporate mRNAs with long 3′ UTRs [34], and since Imp encompasses a low-complexity Gln-rich C-terminus, the biochemically identified RNA assemblage that appears “granular” in Fig. 1b may actually be hydrogels, since a biotinylated isoxazole derivative is able to precipitate Imp from S2 cells [35].

The most striking result of our study is the identification of transcripts mediating actin cytoskeletal dynamics. An analysis of the RIP-seq, iCLIP and PAR-iCLIP data revealed 86 transcripts being top Imp targets based on their enrichment in both immunoprecipitated RNA and CLIP tags over transcript levels in the S2 cell line. Forty of these transcripts can be grouped into cellular processes linked to growth cone steering and axonal branching during neuronal development as illustrated in Fig. 6. The Imp RNP contains transcripts encoding the transmembrane protein Prominin-like, which is able to respond to external signals, and Cnx99A (calnexin), which is able to alter the internal Ca 2+ environment, vital for regulation of growth cone outgrowth [36]. The GTPases (Arf79f, Cdc42, Rac1, R (Rap1) and Ran) are needed to initiate intracellular events, resulting in a reorganization of the actin cytoskeleton, while Moesin and alpha-Spectrin provide a link between the membrane and the cytoskeleton. The main instigator of the polymerization of G-actin into F-actin (Chickadee/profilin), as well as the microtubule-associated protein Mapmodulin, are also present. Transcripts encoding proteins necessary for intracellular trafficking in the growth cone (Rab5, Rab11, Arf51F and Surf4) are also found in the Imp RNP. Additional references to Fig. 6 are listed in Figure S6 in Additional file 1. Less conservative exclusion criteria reveal enabled, Myosin light chain cytoplasmic, and Act42a transcripts as additional Imp targets. Interestingly, the major actin isoform in S2 cells encoded by Act42a mRNA was present among the top 140 transcripts in all four CLIP experiments, but exhibited a low enrichment factor in the RIP experiments. This indicates that Imp binds with modest affinity to Act42a mRNA and, therefore, is washed away, at least to some extent, during the immunoprecipitation.

thumbnailFig. 6. A model of the Imp post-transcriptional RNA assemblage in growth cone dynamics. Forty out of the 86 transcripts identified by the three different high-throughput analyses as being associated with Imp (Fig. 3; Figure S3b in Additional file 1) have been divided into five categories (a–e) participating in growth cone dynamics. References for each of the categorized transcripts and their involvement in neuronal development and growth cone biology are in Figure S6 in Additional file 1
The concept of Imp homologs being involved in growth cone turning has been shown previously, but focus has been on the β-actin/ZBP1 paradigm and not on the growth cone transcriptome [37], [38]. However, a recent study of γ-neurons in mushroom bodies of Drosophila Imp knockout flies has shown that Chickadee/profilin overexpression is able to partially rescue a neurite branching defect [15], providing a glimpse of a more composite Imp-regulated growth cone targetome. An intriguing result of the Imp top target analysis was the absence of transcripts encoding Twinstar (the cofilin homolog responsible for severing F-actin), Lim kinase 1 (which inactivates Twinstar), and the serine-3 phosphatase Slingshot (which activates Twinstar) [39]. In an Imp-deficient background, there was a down-regulation of F-actin levels, and this phenotype was more pronounced in the presence of a ROCK inhibitor, lending support to a scenario where the Imp RNA assemblage may be coordinating F-actin polymerization. Inhibiting the function of ROCK causes increased severing of F-actin filaments, but F-actin staining was similar to that seen in untreated cells. This is most likely due to the increased presence of barbed ends and an increased potential for addition of actin monomers to the filaments. Knockdown of Imp diminished the level of F-actin, and this phenotype was more pronounced in the presence of ROCK inhibitor, suggesting that Imp is involved in the addition of actin monomers to the filaments. Since knockdown of Imp did not affect the level of actin monomers but caused a decrease in F-actin staining, we infer that the presence of the Imp assemblage facilitates F-actin polymerization.

One could speculate that the Imp RNP encapsulates transcripts needed for coordinated on-site (de novo) protein synthesis of filopodia components in response to directional guidance cues. This interpretation is supported by experiments performed in mouse primary cortical neurons [40], Drosophila mushroom body γ-neurons [15], and Xenopus retinal ganglion cells [41] that all find Imp responsible for correct branching events rather than being essential for default neurite extension.

The embryo phenotype of Imp knockdown we report in this study exhibited a broad range of penetrance — almost to the point of being stochastic — and the few survivors reaching the pharate adult stage showed severe locomotion defects. In a more general fashion, the broad range of Imp knockout phenotypes suggests that Imp provides robustness to cytoplasmic post-transcriptional regulation by supplying RNA assemblages coordinating spatiotemporal protein production essential for cellular extensions. In the same vein, during spermatogenesis Imp exhibits abundant expression in the tail cyst cell encasing the extraordinarily long spermatids [12], and its elongation is dependent on F-actin formation, so the top-ranked transcripts identified in the present study may therefore also participate in cyst cell elongation via an Imp-facilitated mechanism. Moreover, the clustering of transcripts encoding RNA-binding proteins such as Hrb27c, Squid, Ypsilon schachtel and Pabp2, combined with a critical role of the cortical actin-binding protein Moesin in oskar mRNA anchoring during oogenesis [42], suggests a coordinating role of Imp in the interplay between the localization apparatus and F-actin at the posterior oocyte cortex.

Conclusions
Neurogenesis and gametogenesis are characterized by a requirement for plasticity in terms of local cellular elongation, yet preserving a differentiated or totipotent state, respectively. Therefore, coordination of gene expression has to some extent become an RNA-centric task, and the physical manifestation of the post-transcriptional RNA regulon is provided by cytoplasmic RNP granules/hydrogels encompassing functionally related mRNAs associated with similar RBPs, as shown for Imp in the present study. These RNA assemblages facilitate fast sub-cytoplasmic recruitment of genetic information to the translational apparatus, thereby ensuring synchronous on-site translation in response to extracellular cues.

Materials and methods
Drosophila cell culture
S2 cells (Invitrogen) were cultured at 26 °C in Schneider’s Drosophila medium (Biowest) containing 10 % heat-inactivated fetal bovine serum, 100 U/ml penicillin and 100 μg/ml streptomycin.

Immunofluoresence of S2 cells
S2 cells were fixed on concanavalin A (Sigma-Aldrich) coated glass bottom dishes with 10 % formaldehyde for 10 min and permeabilized with 0.1 % Triton X-100 in phosphate-buffered saline (PBS) for 3 min [43]. After blocking with 1 % normal goat serum in PBS for 1 h at room temperature, cells were incubated overnight at 4 °C with polyclonal antibody raised against Imp [10]. Following washes in PBS, cells were stained with Alexa Fluor® 568 anti-rabbit antibodies (Life Technologies) and Alexa Fluor® 488 phalloidin (Life Technologies) in 1 % normal goat serum in PBS for 1.5 h at room temperature.

Immunofluoresence of Imp dsRNAi-mediated knockdown in S2 cells
In a 6-well plate, 1 × 10 6 cells/ml were soaked with 40 μg/ml dsRNA corresponding to the imp or luciferase genes. Each day, cells received a boost of 40 μg/ml dsRNA. After the boost on the third day, cells were treated with 10 μM ROCK inhibitor (Y-27632, Sigma Aldrich) for 20 h. The medium was removed and cells were fixed and immunostained as previously described. Cells were examined by structured illumination (SIM) and by confocal microscopy (Zeiss LSM 780) to obtain high resolution pictures and quantification of Imp and F-actin, respectively. Employing a 40× objective, Alexa 568 and Alexa 488 pixels were quantified in ten visual fields randomly selected using the automatic controller of the microscope. Pixels were quantified employing the HISTO functionality of the ZEN 2011 software package (Carl Zeiss). Channels were evaluated separately, and the mean intensity in each field was determined after thresholding. Significant differences were analyzed by a Student’s t-test. For more information see “Supplementary Materials and methods” in Additional file 1.

RIP-seq
S2 cells were lysed in 50 mM Tris–HCl (pH 8.0), 100 mM NaCl, 1 % NP-40, 1.5 mM EDTA, complete EDTA-free protease inhibitor cocktail (Roche), 1 U/μl RiboLock (Thermo Scientific) and immunoprecipitated using polyclonal anti-Imp antibody coated Protein A Dynabeads (Life Technologies). Following repeated washes in lysis buffer, TRI Reagent (Sigma-Aldrich) was added to the beads and the RNA was isolated according to the manufacturer’s specifications. RNA from two biologically independent RIP-seq experiments as well as poly(A)-enriched RNA from cells isolated on the same day as cells used for RIP-seq, were fractionated, and library preparation was performed as described by [44], with the final library being amplified with 28 cycles of PCR.

iCLIP and PAR-iCLIP sequencing
Imp iCLIP and Imp PAR-iCLIP were performed in biological replicates essentially as described in [17], [23]. Briefly, 5 × 10 7Drosophila S2 cells were used for each iCLIP and PAR-iCLIP experiment. Cells for iCLIP were irradiated with 250 mJ at 254 nm UV light, whereas cells were treated with 100 μM 4-thiouridine for 16 h prior to UV cross-linking with 250 mJ at 365 nm in the PAR-iCLIP procedure. Cells were lysed with lysis buffer (50 mM Tris–HCl, pH 8.0, 100 mM NaCl, 1 % NP-40, complete EDTA-free protease inhibitor cocktail (Roche)), followed by ribonuclease T1 digestion (Thermo Scientific; 0.005 U/μl for iCLIP and 0.05 U/μl for PAR-iCLIP) in the presence of 0.0025 U/μl DNase I (Thermo Scientific) for 10 min at 37 °C. Imp–RNA complexes were immunoprecipitated using rabbit anti-Imp antibody-coated Protein A Dynabeads (Life Technologies) for 2 h at 4 °C. Beads were washed three times with high-salt buffer (50 mM Tris–HCl, pH 7.4, 500 mM NaCl, 1 % NP-40, 1 mM EDTA, 0.125 % SDS) and twice with wash buffer (20 mM Tris–HCl, pH 7.4, 10 mM MgCl 2 , 0.5 % NP-40). The RNA purification and library generation were performed as described in [23] with 21–34 cycles of PCR amplification. E-gel size selection was done in order to remove the reverse transcription and Illumina primer annealing by-product.

RNA-seq for normalizing iCLIP data
In order to normalize iCLIP and PAR-iCLIP data, RNA-seq was performed on cells harvested on the same day as cells used in the iCLIP procedure. For each replicate, total RNA was isolated from 5 × 10 7 cells using the standard TRI Reagent protocol (Sigma Aldrich). Poly(A) RNA was enriched from 100 μg total RNA using Poly(A) Purist MAG kit (Ambion) according to the manufacturer’s specifications. The poly(A) RNA was fragmented into approximately 280 nucleotides by heating at 95 °C for 3.5 min in the presence of 50 mM Tris–HCl (pH 8.0) and 5 mM MgCl 2 . Treatment of the RNA with T4 polynucleotide kinase (Thermo Scientific), according to the manufacturer’s specifications, resulted in dephosphorylation of the 3′ ends of the RNA. The remaining part of the RNA-seq library preparation was performed as described in [23] with 17 cycles of PCR amplification.

Gene ontology terms
The transcripts present in a top 200 ranking of normalized RIP-seq datasets were used as input for the term enrichment tool AmiGO [45] using a P value cutoff of 0.0001. As background sample, a list was made of all poly(A) RNAs present in the S2 cell line according to the RNA-seq.

Whole mount immunostaining of embryos
Imp G0072 was obtained from the Bloomington Drosophila Stock Center (stock number 11798, w67c23 P{lacW}ImpG0072/FM7c). The strain was generated by a P-element insertion in the imp locus in a region corresponding to an intron upstream of the initiation codons [29]. Embryos were dechorionized for two minutes in 2 % sodium hypochlorit and fixed in 4 % paraformaldehyde. In order to record the fraction of embryos with an abnormal neuronal pattern, the nervous system of wild-type embryos, and embryos of crosses between either wt/FM7c or imp G0072 /FM7c females and FM7c/Y males, was visualized by enzyme-linked immunohistochemistry. The monoclonal neuronal marker 22C10, which recognizes the MAP1B-like protein Futsch and labels both the central and peripheral nervous systems [46], was used as the primary antibody, followed by anti-mouse horse radish peroxidase-catalyzed diaminobenzidine oxidation (Cell Signaling). The frequency of embryos with an abnormal morphology and nervous system was recorded, and the percentage of abnormal embryos was determined. In the co-staining with anti-β-gal antibody (Rockland; 1:10,000) and monoclonal anti-Futsch 22C10 antibody (1:100; Developmental Studies Hybridoma Center, University of Iowa), the secondary antibodies were anti-rabbit Alexa Fluor® 488 and anti-mouse Alexa Fluor® 555 from Molecular Probes (1:1000), respectively. Fluorescent immunostainings were visualized with a confocal Zeiss LSM510 microscope.

Read mapping and processing
All sequencing data were produced by The Danish National High-Throughput DNA Sequencing Centre using Illumina HiSeq and 50–100 bp single-end reads. The mapping procedure for iCLIP, PAR-iCLIP, RIP and RNA-seq controls is described in “Supplementary Materials and methods” in Additional file 1.

CLIP and RIP normalization
The iCLIP and RIP libraries were prepared from the same biological sample as that used for the RNA-seq data sets. Therefore, we normalized each RIP and iCLIP replicate with its corresponding RNA-seq replicate. The PAR-iCLIP replicates were normalized to a pool of two iCLIP RNA-seq controls. The enrichment e of CLIP in a position i of a particular cluster k is calculated as:

ei,k=M.Lg.ciN∑Lgj=1rj (1)
where c i is the count of CLIP base calls in position i, and N the total amount of confidently mapped CLIP reads. As r j is the count of RNA-seq reads, 1MLg(∑j=1Lgrj) is the sum of RNA-seq base calls across the transcript g, normalized to the total amount of confidently mapped RNA-seq reads M and the length of the transcript L g .

Standardized transcript profiles
To make standardized profiles, the longest protein-coding transcript of each gene with at least 30 % RNA-seq coverage in the S2 cell line was used. The regions (5′ UTR, CDS, 3′ UTR) of those transcripts were divided into a fixed number of bins as follows: 20 bins for 5′ UTRs, 50 bins for CDS and 50 bins for 3′ UTRs. The calculation of mean enrichment is described in “Supplementary Materials and methods” in Additional file 1.

Transcriptome-wide Imp binding and transcript ranking
To quantify Drosophila Imp binding across the whole transcriptome (Fig. 3a), we pooled CLIP replicates and RIP-seq replicates separately, and the CLIP was normalized by a pool of RNA-seq replicates associated with the iCLIP while the RIP was normalized to the RNA-seq associated with the RIP-seq. Transcripts were then ordered according to decreasing enrichment. Identification of transcripts bound by Imp was accomplished by calculating a normalized enrichment per nucleotide in 3′ UTRs for the iCLIP and PAR-iCLIP data and RIP-seq data. Each replicate was normalized by a pool of RNA-seq replicates associated with the iCLIP and rankings were summarized and merged. Only the 1157 most expressed transcripts with an expression level higher than the average expression of a given dataset were considered. This was done to avoid a high degree of variation due to low gene expression and thereby inconsistent rankings across replicates. This normalized coverage was then used to rank transcripts according to enrichment (Fig. 3b).

Word enrichment analysis
Imp CLIP-seq data form very large clusters that cover the entire 3′ UTRs in some cases. Thus, upon normalization of clusters from pooled iCLIP and pooled PAR-iCLIP replicates, they were divided into smaller regions based on the enrichment changes across the clusters. In this way, segments that were highly covered by CLIP-data, i.e., highly enriched, and segments less enriched could be separated. We extracted the sequences and ranked them by maximal normalized enrichment in the associated cluster (described in Eq. 1) after discarding small clusters (less than six reads) and genes with low expression levels (average of less than five reads per nucleotide), and selected the 25,000 most enriched clusters in 3′ UTRs. These sequences and the enrichment ranking were subjected to the motif discovery method cWords [24] in two separate analyses, one for each iCLIP protocol.

Hidden Markov model
To assess the presence of putative Drosophila Imp binding sites, we specified a HMM in the anHMM framework [47]. A state diagram of the HMM is shown in Figure S3c in Additional file 1. In the state diagram, all transitions are uniformly assigned, except the transition marked by the t1 label, which is 0.95. All states have emission probabilities of the letter shown in the diagram set to 0.9. The HMM was used to identify independent instances of the CA-rich and UA-rich motifs and post-processed such that alternating occurrences of the two motifs were merged into one composite motif if they were spaced a maximum of 50 nucleotides apart. All other single motif occurrences were filtered out. We annotated the entire genome on both strands with matches using Viterbi decoding. We summarized binding sites in the top 86 transcripts bound by Imp and in all expressed transcripts with at least 30 % RNA-seq coverage in S2 cells. We found matches in 3′ UTRs for 76 of the top 86 transcripts.

Data access
All the data reported in this paper can be found in the Gene Expression Omnibus database with accession number GSE62997.

Abbreviations
β-gal: β-galactosidase

CDS: coding sequence

CLIP: cross-linking and immunoprecipitation

dsRNAi: double-stranded RNA interference

GO: gene ontology

HMM: hidden Markov model

HuR: human antigen R

iCLIP: individual-nucleotide resolution cross-linking and immunoprecipitation

KH: K homology

PAR-CLIP: photoactivatable ribonucleoside cross-linking and immunoprecipitation

PBS: phosphate-buffered saline

RBP: RNA-binding protein

RIP: RNA-binding protein immunoprecipitation

RNA-seq: high-throughput sequencing of cDNAs

RNP: ribonucleoprotein

ROCK: Rho-associated protein kinase

UTR: untranslated region

Competing interests
The authors declare that they have no competing interests.

Authors’ contribution
HTH, SKA and JC designed the project. HTH and SKA performed the experiments. HTH, SKA, SHR, MP and JC wrote the manuscript. FCN performed microscopy and quantification of the pictures in Fig. 4. SHR, MP and AK designed the bioinformatics analyses. SHR performed the bioinformatics analyses. JS contributed with CLIP expertise. All authors read and approved the final manuscript.

Authors’ information
Information and request for reagents: hhansen@bio.ku.dk and janchr@bio.ku.dk

Additional file
Additional file 1:. Supplemental Figures S1–S6 and Supplemental Materials and methods and Supplemental References. Figure S1 is related to Fig. 1, Figure S2 is related to Fig. 2, Figure S3 is related to Fig. 3, Figure S4 is related to Fig. 4, Figure S5 is related to Fig. 5, and Figure S6 is related to Fig. 6.
Format: PDF Size: 4.3MB Download file
This file can be viewed with: Adobe Acrobat ReaderOpen Data
Acknowledgements
We are grateful to Lena Bjørn Johansson for skillful technical assistance. This work was supported by the Danish Council for Strategic Research (Center for Computational and Applied Transcriptomics, DSF-10-092320), the Novo Nordisk Foundation, and the Carlsberg Foundation.

References
Gomez TM, Letourneau PC. Actin dynamics in growth cone motility and navigation. J Neurochem. 2014; 129:221-34. PubMed Abstract | Publisher Full Text OpenURL
Jung H, Gkogkas CG, Sonenberg N, Holt CE. Remote control of gene function by local translation. Cell. 2014; 157:26-40. PubMed Abstract | Publisher Full Text OpenURL
Keene JD. RNA regulons: coordination of post-transcriptional events. Nat Rev Genet. 2007; 8:533-43. PubMed Abstract | Publisher Full Text OpenURL
Gerber AP, Herschlag D, Brown PO. Extensive association of functionally and cytotopically related mRNAs with Puf family RNA-binding proteins in yeast. PLoS Biol. 2004; 2: Article ID E79 PubMed Abstract | Publisher Full Text OpenURL
Lecuyer E, Yoshida H, Parthasarathy N, Alm C, Babak T, Cerovina T et al.. Global analysis of mRNA localization reveals a prominent role in organizing cellular architecture and function. Cell. 2007; 131:174-87. PubMed Abstract | Publisher Full Text OpenURL
Oleynikov Y, Singer RH. Real-time visualization of ZBP1 association with beta-actin mRNA during transcription and localization. Curr Biol. 2003; 13:199-207. PubMed Abstract | Publisher Full Text OpenURL
Zhang HL, Eom T, Oleynikov Y, Shenoy SM, Liebelt DA, Dictenberg JB et al.. Neurotrophin-induced transport of a beta-actin mRNP complex increases beta-actin levels and stimulates growth cone motility. Neuron. 2001; 31:261-75. PubMed Abstract | Publisher Full Text OpenURL
Nielsen FC, Nielsen J, Kristensen MA, Koch G, Christiansen J. Cytoplasmic trafficking of IGF-II mRNA-binding protein by conserved KH domains. J Cell Sci. 2002; 115:2087-97. PubMed Abstract | Publisher Full Text OpenURL
Wachter K, Kohn M, Stohr N, Huttelmaier S. Subcellular localization and RNP formation of IGF2BPs (IGF2 mRNA-binding proteins) is modulated by distinct RNA-binding domains. Biol Chem. 2013; 394:1077-90. PubMed Abstract | Publisher Full Text OpenURL
Adolph SK, DeLotto R, Nielsen FC, Christiansen J. Embryonic expression of Drosophila IMP in the developing CNS and PNS. Gene Expr Patterns. 2009; 9:138-43. PubMed Abstract | Publisher Full Text OpenURL
Munro TP, Kwon S, Schnapp BJ, St Johnston D. A repeated IMP-binding motif controls oskar mRNA translation and anchoring independently of Drosophila melanogaster IMP. J Cell Biol. 2006; 172:577-88. PubMed Abstract | Publisher Full Text OpenURL
Fabrizio JJ, Hickey CA, Stabrawa C, Meytes V, Hutter JA, Talbert C et al.. Imp (IGF-II mRNA-binding protein) is expressed during spermatogenesis in Drosophila melanogaster. Fly (Austin). 2008; 2:47-52. Publisher Full Text OpenURL
Geng C, Macdonald PM. Imp associates with squid and Hrp48 and contributes to localized expression of gurken in the oocyte. Mol Cell Biol. 2006; 26:9508-16. PubMed Abstract | Publisher Full Text OpenURL
Boylan KL, Mische S, Li M, Marques G, Morin X, Chia W et al.. Motility screen identifies Drosophila IGF-II mRNA-binding protein–zipcode-binding protein acting in oogenesis and synaptogenesis. PLoS Genet. 2008; 4: Article ID e36 PubMed Abstract | Publisher Full Text OpenURL
Medioni C, Ramialison M, Ephrussi A, Besse F. Imp promotes axonal remodeling by regulating profilin mRNA during brain development. Curr Biol. 2014; 24:793-800. PubMed Abstract | Publisher Full Text OpenURL
Licatalosi DD, Mele A, Fak JJ, Ule J, Kayikci M, Chi SW et al.. HITS-CLIP yields genome-wide insights into brain alternative RNA processing. Nature. 2008; 456:464-9. PubMed Abstract | Publisher Full Text OpenURL
Hafner M, Landthaler M, Burger L, Khorshid M, Hausser J, Berninger P et al.. Transcriptome-wide identification of RNA-binding protein and microRNA target sites by PAR-CLIP. Cell. 2010; 141:129-41. PubMed Abstract | Publisher Full Text OpenURL
Konig J, Zarnack K, Rot G, Curk T, Kayikci M, Zupan B et al.. iCLIP reveals the function of hnRNP particles in splicing at individual nucleotide resolution. Nat Struct Mol Biol. 2010; 17:909-15. PubMed Abstract | Publisher Full Text OpenURL
Macias S, Plass M, Stajuda A, Michlewski G, Eyras E, Caceres JF. DGCR8 HITS-CLIP reveals novel functions for the Microprocessor. Nat Struct Mol Biol. 2012; 19:760-6. PubMed Abstract | Publisher Full Text OpenURL
Sanford JR, Wang X, Mort M, Vanduyn N, Cooper DN, Mooney SD et al.. Splicing factor SFRS1 recognizes a functionally diverse landscape of RNA transcripts. Genome Res. 2009; 19:381-94. PubMed Abstract | Publisher Full Text OpenURL
Martin G, Gruber AR, Keller W, Zavolan M. Genome-wide analysis of pre-mRNA 3′ end processing reveals a decisive role of human cleavage factor I in the regulation of 3′ UTR length. Cell Rep. 2012; 1:753-63. PubMed Abstract | Publisher Full Text OpenURL
Toledano H, D’Alterio C, Czech B, Levine E, Jones DL. The let-7-Imp axis regulates ageing of the Drosophila testis stem-cell niche. Nature. 2012; 485:605-10. PubMed Abstract | Publisher Full Text OpenURL
Konig J, Zarnack K, Rot G, Curk T, Kayikci M, Zupan B et al.. iCLIP–transcriptome-wide mapping of protein-RNA interactions with individual nucleotide resolution. J Vis Exp. 2011; 50:2638. PubMed Abstract | Publisher Full Text OpenURL
Rasmussen SH, Jacobsen A, Krogh A. cWords – systematic microRNA regulatory motif discovery from mRNA expression data. Silence. 2013; 4:2. PubMed Abstract | BioMed Central Full Text OpenURL
Lewis HA, Musunuru K, Jensen KB, Edo C, Chen H, Darnell RB et al.. Sequence-specific RNA binding by a Nova KH domain: implications for paraneoplastic disease and the fragile X syndrome. Cell. 2000; 100:323-32. PubMed Abstract | Publisher Full Text OpenURL
Backe PH, Messias AC, Ravelli RB, Sattler M, Cusack S. X-ray crystallographic and NMR studies of the third KH domain of hnRNP K in complex with single-stranded nucleic acids. Structure. 2005; 13:1055-67. PubMed Abstract | Publisher Full Text OpenURL
Song X, Chen X, Yamaguchi H, Mouneimne G, Condeelis JS, Eddy RJ. Initiation of cofilin activity in response to EGF is uncoupled from cofilin phosphorylation and dephosphorylation in carcinoma cells. J Cell Sci. 2006; 119:2871-81. PubMed Abstract | Publisher Full Text OpenURL
Rogers SL, Wiedemann U, Stuurman N, Vale RD. Molecular requirements for actin-based lamella formation in Drosophila S2 cells. J Cell Biol. 2003; 162:1079-88. PubMed Abstract | Publisher Full Text OpenURL
Peter A, Schottler P, Werner M, Beinert N, Dowe G, Burkert P et al.. Mapping and identification of essential gene functions on the X chromosome of Drosophila. EMBO Rep. 2002; 3:34-8. PubMed Abstract | Publisher Full Text OpenURL
Taliaferro JM, Aspden JL, Bradley T, Marwha D, Blanchette M, Rio DC. Two new and distinct roles for Drosophila Argonaute-2 in the nucleus: alternative pre-mRNA splicing and transcriptional repression. Genes Dev. 2013; 27:378-89. PubMed Abstract | Publisher Full Text OpenURL
Patel VL, Mitra S, Harris R, Buxbaum AR, Lionnet T, Brenowitz M et al.. Spatial arrangement of an RNA zipcode identifies mRNAs under post-transcriptional control. Genes Dev. 2012; 26:43-53. PubMed Abstract | Publisher Full Text OpenURL
Nielsen J, Kristensen MA, Willemoes M, Nielsen FC, Christiansen J. Sequential dimerization of human zipcode-binding protein IMP1 on RNA: a cooperative mechanism providing RNP stability. Nucleic Acids Res. 2004; 32:4368-76. PubMed Abstract | Publisher Full Text OpenURL
Mukherjee N, Corcoran DL, Nusbaum JD, Reid DW, Georgiev S, Hafner M et al.. Integrative regulatory mapping indicates that the RNA-binding protein HuR couples pre-mRNA processing and mRNA stability. Mol Cell. 2011; 43:327-39. PubMed Abstract | Publisher Full Text OpenURL
Han TW, Kato M, Xie S, Wu LC, Mirzaei H, Pei J et al.. Cell-free formation of RNA granules: bound RNAs identify features and components of cellular assemblies. Cell. 2012; 149:768-79. PubMed Abstract | Publisher Full Text OpenURL
Kato M, Han TW, Xie S, Shi K, Du X, Wu LC et al.. Cell-free formation of RNA granules: low complexity sequence domains form dynamic fibers within hydrogels. Cell. 2012; 149:753-67. PubMed Abstract | Publisher Full Text OpenURL
Tang F, Dent EW, Kalil K. Spontaneous calcium transients in developing cortical neurons regulate axon outgrowth. J Neurosci. 2003; 23:927-36. PubMed Abstract | Publisher Full Text OpenURL
Leung KM, van Horck FP, Lin AC, Allison R, Standart N, Holt CE. Asymmetrical beta-actin mRNA translation in growth cones mediates attractive turning to netrin-1. Nat Neurosci. 2006; 9:1247-56. PubMed Abstract | Publisher Full Text OpenURL
Yao J, Sasaki Y, Wen Z, Bassell GJ, Zheng JQ. An essential role for beta-actin mRNA localization and translation in Ca2 + −dependent growth cone guidance. Nat Neurosci. 2006; 9:1265-73. PubMed Abstract | Publisher Full Text OpenURL
Niwa R, Nagata-Ohashi K, Takeichi M, Mizuno K, Uemura T. Control of actin reorganization by Slingshot, a family of phosphatases that dephosphorylate ADF/cofilin. Cell. 2002; 108:233-46. PubMed Abstract | Publisher Full Text OpenURL
Welshhans K, Bassell GJ. Netrin-1-induced local beta-actin synthesis and growth cone guidance requires zipcode binding protein 1. J Neurosci. 2011; 31:9800-13. PubMed Abstract | Publisher Full Text OpenURL
Kalous A, Stake JI, Yisraeli JK, Holt CE. RNA-binding protein Vg1RBP regulates terminal arbor formation but not long-range axon navigation in the developing visual system. Dev Neurobiol. 2014; 74:303-18. PubMed Abstract | Publisher Full Text OpenURL
Jankovics F, Sinka R, Lukacsovich T, Erdelyi M. MOESIN crosslinks actin and cell membrane in Drosophila oocytes and is required for OSKAR anchoring. Curr Biol. 2002; 12:2060-5. PubMed Abstract | Publisher Full Text OpenURL
Buster DW, Nye J, Klebba JE, Rogers GC. Preparation of Drosophila S2 cells for light microscopy. J Vis Exp. 2010; 40:1982. PubMed Abstract | Publisher Full Text OpenURL
Kielpinski LJ, Boyd M, Sandelin A, Vinther J. Detection of reverse transcriptase termination sites using cDNA ligation and massive parallel sequencing. Methods Mol Biol. 2013; 1038:213-31. PubMed Abstract | Publisher Full Text OpenURL
Boyle EI, Weng S, Gollub J, Jin H, Botstein D, Cherry JM et al.. GO::TermFinder–open source software for accessing Gene Ontology information and finding significantly enriched Gene Ontology terms associated with a list of genes. Bioinformatics. 2004; 20:3710-5. PubMed Abstract | Publisher Full Text OpenURL
Hummel T, Krukkert K, Roos J, Davis G, Klambt C. Drosophila Futsch/22C10 is a MAP1B-like protein required for dendritic and axonal development. Neuron. 2000; 26:357-70. PubMed Abstract | Publisher Full Text OpenURL
Sonnhammer EL, von Heijne G, Krogh A. A hidden Markov model for predicting transmembrane helices in protein sequences. Proc Int Conf Intell Syst Mol Biol. 1998; 6:175-82. PubMed Abstract | Publisher Full Text OpenURL

Transplantation of human oligodendrocyte progenitor cells in an animal model of diffuse traumatic axonal injury: survival and differentiation

Leyan Xu1*†, Jiwon Ryu1†, Hakim Hiel2, Adarsh Menon1, Ayushi Aggarwal1, Elizabeth Rha1, Vasiliki Mahairaki1, Brian J Cummings3 and Vassilis E Koliatsos145

  • Corresponding author: Leyan Xu lxu9@jhmi.edu

† Equal contributors
Author Affiliations
1 Division of Neuropathology, Department of Pathology, The Johns Hopkins University School of Medicine, Baltimore 21205, MD, USA

2 Department of Otolaryngology-Head and Neck Surgery, The Johns Hopkins University School of Medicine, Baltimore 21205, MD, USA

3 Departments of Physical and Medical Rehabilitation, Neurological Surgery, and Anatomy and Neurobiology, Sue and Bill Gross Stem Cell Research Center, Institute for Memory Impairments and Neurological Disorders, University of California at Irvine, Irvine 92697, CA, USA

4 Department of Neurology, The Johns Hopkins University School of Medicine, Baltimore 21205, MD, USA

5 Department of Psychiatry and Behavioral Sciences, The Johns Hopkins University School of Medicine, Baltimore 21205, MD, USA

For all author emails, please log on.

Stem Cell Research & Therapy 2015, 6:93 doi:10.1186/s13287-015-0087-0

Leyan Xu and Jiwon Ryu contributed equally to this work.

The electronic version of this article is the complete one and can be found online at: http://stemcellres.com/content/6/1/93

Received: 19 December 2014
Revisions received: 13 March 2015
Accepted: 1 May 2015
Published: 14 May 2015
© 2015 Xu et al.; licensee BioMed Central.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Abstract
Introduction
Diffuse axonal injury is an extremely common type of traumatic brain injury encountered in motor vehicle crashes, sports injuries, and in combat. Although many cases of diffuse axonal injury result in chronic disability, there are no current treatments for this condition. Its basic lesion, traumatic axonal injury, has been aggressively modeled in primate and rodent animal models. The inexorable axonal and perikaryal degeneration and dysmyelination often encountered in traumatic axonal injury calls for regenerative therapies, including therapies based on stem cells and precursors. Here we explore the proof of concept that treatments based on transplants of human oligodendrocyte progenitor cells can replace or remodel myelin and, eventually, contribute to axonal regeneration in traumatic axonal injury.

Methods
We derived human oligodendrocyte progenitor cells from the human embryonic stem cell line H9, purified and characterized them. We then transplanted these human oligodendrocyte progenitor cells into the deep sensorimotor cortex next to the corpus callosum of nude rats subjected to traumatic axonal injury based on the impact acceleration model of Marmarou. We explored the time course and spatial distribution of differentiation and structural integration of these cells in rat forebrain.

Results
At the time of transplantation, over 90 % of human oligodendrocyte progenitor cells expressed A2B5, PDGFR, NG2, O4, Olig2 and Sox10, a profile consistent with their progenitor or early oligodendrocyte status. After transplantation, these cells survived well and migrated massively via the corpus callosum in both injured and uninjured brains. Human oligodendrocyte progenitor cells displayed a striking preference for white matter tracts and were contained almost exclusively in the corpus callosum and external capsule, the striatopallidal striae, and cortical layer 6. Over 3 months, human oligodendrocyte progenitor cells progressively matured into myelin basic protein(+) and adenomatous polyposis coli protein(+) oligodendrocytes. The injured environment in the corpus callosum of impact acceleration subjects tended to favor maturation of human oligodendrocyte progenitor cells. Electron microscopy revealed that mature transplant-derived oligodendrocytes ensheathed host axons with spiral wraps intimately associated with myelin sheaths.

Conclusions
Our findings suggest that, instead of differentiating locally, human oligodendrocyte progenitor cells migrate massively along white matter tracts and differentiate extensively into ensheathing oligodendrocytes. These features make them appealing candidates for cellular therapies of diffuse axonal injury aiming at myelin remodeling and axonal protection or regeneration.

Introduction
Axonal injury is the defining feature of diffuse axonal injury (DAI), but is also present in blast injuries [1], chronic traumatic encephalopathy [2], and even mild head injuries [3]. Axonal damage in models of DAI is referred to as traumatic axonal injury (TAI), a term often used interchangeably with DAI [4], [5]. In the case of DAI, axonal injury causes disconnection of neural circuits at multiple central nervous system (CNS) sites [6]–[8] and can lead to a number of neurological impairments, including long-term memory problems, emotional disturbances, unconsciousness, and/or a persistent vegetative state. These neurological impairments have no satisfactory treatment besides symptomatic alleviation of various subsyndromes with physical, occupational, speech and language therapy and various categories of CNS-acting drugs including antispasmodics, antidepressants, and mood stabilizers. Although some retraining of circuits is anticipated over time and syndromic pharmacotherapies have some effectiveness, most patients with DAI still remain severely symptomatic years and decades later.

Stem cell therapy presents a promising treatment approach for traumatic brain injury (TBI). Some early success in models of ischemic brain injury [9] has encouraged the use of stem cell or neural precursor (NP) transplantation, primarily in models of focal TBI [10]. Much less is known about the role of stem cell therapies in DAI/TAI. Axonal repair as a target of treatment separate from nerve cell regeneration is not as well established in TBI as in spinal cord injury, and this is especially true with the problem of myelin repair/remyelination [11]. However, demyelination appears to contribute to degeneration of axons in TAI [12], [13] and TAI is associated with active and ongoing attempts at axonal repair [14]. Therefore, adding exogenous oligodendrocyte progenitor cells (OPCs) may furnish competent oligodendrocytes that can assist in remyelination/myelin remodeling and prevent axonal degeneration or help myelinate regenerating axons in TAI.

Animal models are invaluable tools in establishing proof of concept that remyelination by exogenously provided oligodendrocytes is possible in TAI settings. Models of inertial acceleration and impact acceleration (IA) are frequently used for experimental studies of DAI/TAI [5], [15]. In the present study we use the IA model of DAI/TAI [16] and transplant human embryonic stem cell (ESC)-derived OPCs (hOPCs) into the deep sensorimotor cortex next to the corpus callosum. Our findings indicate that exogenous hOPCs differentiate into mature oligodendrocytes, migrate extensively along white matter tracts, and begin to myelinate host axons. Our data are consistent with the view that stem cell grafts may serve as effective myelin remodeling tools in TBI scenarios featured by DAI/TAI.

Materials and methods
Human embryonic stem cell culture and differentiation to human oligodendrocyte progenitor cells
The human ESC line H9 from WiCell (Madison, WI, USA) was maintained according to standard stem cell culture protocols. H9 cells (WA-09; passages 30 to 41) were grown on mitotically inactivated mouse embryonic fibroblasts essentially as described in [17]. hOPCs were generated through extensive passaging as neurospheres based on the method of Hu and colleagues [18], [19] with minor modifications (Additional file 1: Fig. S1). The ventralizing factor sonic hedgehog (SHH; 100 ng/mL) along with the caudalizing factor retinoic acid (0.1 μM) were used to initially pattern neuroepithelial cells; glial differentiation medium (GDM; Dulbecco’s modified Eagle’s medium (DMEM)/F12, B27 without vitamin A, N1, MEM-NEAA, cAMP, biotin, 60 ng/mL triiodothyronine, 10 ng/mL platelet-derived growth factor (PDGF)-AA, insulin-like growth factor (IGF)1 and neurotrophin (NT)3) was used for further differentiation. Cells were trypsinized with TrypLE (Life Technologies, Grand Island, NY, USA) at day 84 after induction of differentiation, counted, and plated on p-L-ornithine- and laminin-coated plates. Cells were grown in GDM supplemented with PDGF-AA, IGF1 and NT3 for 12 days, then trypsinized, counted, and resuspended at high concentration (2.0 × 108 per mL), and finally transplanted on day 98 after induction of differentiation.

Characterization of human oligodendrocyte progenitor cells used for transplantation with immunocytochemistry
Two weeks before transplantation (on day 84, a time point chosen to correspond to the remaining time in differentiation of hOPCs destined for transplantation), hOPC neurospheres were trypsinized with TrypLE and counted. Twenty thousand cells were plated on polyornithine- and laminin-coated coverslips or Matrigel-coated four-well slide chambers and cultured in GDM supplemented with PDGF, IGF and NT3 for 2 weeks. Cultures were then fixed with 4 % paraformaldehyde in phosphate-buffered saline for 20 minutes and then subjected to immunocytochemistry with the oligodendrocytic markers A2B5, platelet-derived growth factor receptor (PDGFR)α, NG2, Sox10, and O4; the neuronal marker type III-tubulin epitope J1 (TUJ1); astrocyte marker glial fibrillary acidic protein (GFAP); and the mesodermal marker bone morphogenetic protein 4 (Table 1).

Table 1. Primary antibodies used for immunocytochemistry, immunohistochemistry and ultrastructural immunohistochemistry
Animals and surgical procedures
Ten-week old male nude rats (Crl:NIH-Foxn1rnu; Charles River, Wilmington, MA, USA) were used for hOPC transplantation. Nude rats were chosen because immunodeficient animals yield greater engraftment and survival of human cells than immunocompetent animals treated with immunosuppressants [20]. All surgical procedures were carried out according to protocols approved by the Animal Care and Use Committee of the Johns Hopkins Medical Institutions using gas anesthesia (isoflurane:oxygen:nitrous oxide = 1:33:66) and aseptic methods. In order to explore the fate of transplanted hOPCs and compare differentiation between injured and uninjured scenarios, animals were separated into IA and sham groups. In the IA group, animals were subjected to injury with full artificial ventilation as described by Marmarou and colleagues [16]. In the present experiments, we employed a severe TBI regimen using a 450 g weight that was freely dropped onto the steel disc through a Plexiglass tube from a height of 2 meters. In the sham group, animals received all aspects of the regimen except the injury itself (weight on the steel disc). One week after injury, a time point that appears to optimize survival and differentiation [21], [22], 200,000 live hOPCs were transplanted into two sites 1 mm apart in the right deep motor cortex next to the corpus callosum (1 mm and 0 mm anterior to bregma, 2 mm lateral to midline and 3 mm ventral to pia) of either injured (n = 10) or sham (n = 5) animals using procedures that have been detailed in our published work [21], [23], [24]. To explore the progress of differentiation of transplanted hOPCs in the TAI environment, animals in the TAI group were allowed to survive for 6 weeks and 3 months. Sham animals with transplanted hOPCs were euthanized at 3 months.

Histology, immunohistochemistry and microscopy
Brain tissues were prepared from animals perfused transcardially with 4 % phosphate-buffered paraformaldehyde. The axonal injury, survival, location and phenotypic fate of hOPC grafts were assessed with ABC peroxidase immunohistochemistry (IHC) and dual-label fluorescent IHC in serial coronal or sagittal sections (40 μm) through the brain as described previously [22], [24], [25]. Axonal injury was studied with well-established TAI markers, including an antibody against the amyloid precursor protein (APP), the monoclonal antibody RMO14 binding to the rod domain of neurofilaments H and M, and a monoclonal antibody against the 68-kDa light chain neurofilament protein. hOPC survival was studied with human-specific nuclei (HNu) or human-specific cytoplasm (SC121) antibody using immunoperoxidase or immunofluorescence labeling. Differentiation was studied with dual-label immunofluorescence combining HNu or SC121 with other oligodendrocyte markers – that is, the progenitor and early marker PDGFRα, the early markers O4 and GalC and late markers myelin basic protein (MBP) and adenomatous polyposis coli protein (APC). Appositions between axons and transplant-derived oligodendrocytes were visualized with the combination of antibodies against the heavy neurofilament subunit (NF-H) and the SC121 epitope as generic axonal and transplant-derived cell markers, respectively. The nuclear mitotic marker Ki67, the early neuronal marker TUJ1 and the astroglial cell marker GFAP were also used in separate co-localization experiments with HNu or SC121 as described elsewhere [22], [24], [26], [27]. All antibody information is listed in Table 1. The Gallyas silver staining method [28] was used to evaluate injured and/or degenerating axons and terminals. For this purpose, sections were processed with a commercially available kit (NeuroSilver kit II; FD Neurotechnologies, Ellicott City, MD, USA) as described previously [29].

Stained sections were studied on a Zeiss Axiophot microscope equipped for epifluorescence (Diagnostic Instruments Inc., Sterling Heights, MI, USA) or a Zeiss LSM 510 inverted confocal microscope (Carl Zeiss Inc., Oberkochen, Germany). Confocal microscopic images were captured with pinhole set at 0.8 μm to ensure co-localization of multiple labels at the same resolution level as semithin sections. Three-dimensional reconstruction by Z-stack scanning through regions of interest was acquired with LSM software (Carl Zeiss Inc., Oberkochen, Germany). Adobe Photoshop 7.0 software (Adobe Systems, San Jose, CA, USA) was used for montaging and image processing. All staining, image collection, and quantification were done in a fashion blind to group assignment.

Ultrastructural immunohistochemistry
Myelin formation by transplanted hOPCs was assessed ultrastructurally with electron microscopy using standard pre-embedding immunoperoxidase-3,3′-diaminobenzidine IHC for the human cytoplasmic antigen SC121 as described previously [23]–[25], [30]. Briefly, brain sections prepared as in the previous section were treated with a solution containing 4 % paraformaldehyde and 0.2 % glutaraldehyde for 24 hours. Sections were then rinsed in 0.1 M phosphate buffer (pH 7.3) for 3 to 10 minutes, immersed in 1 % osmium tetroxide for 15 minutes, dehydrated in graded concentrations of ethanol, embedded in Poly/Bed 812 (Polysciences Inc., Warrington, PA, USA), polymerized at 60 °C for 72 hours, and then finally embedded in BEEM® capsules (Electron Microscopy Sciences, Hatfield, PA, USA). Half the sections were stained en block in uranyl acetate prior to embedding. Serial ultrathin sections were collected on Formvar-coated slotted grids and viewed with a Hitachi H7600 transmission electron microscope equipped with a 2 k×2 k bottom mount AMT XR-100 CCD camera (Hitachi High-Technologies Corporation, Tokyo, Japan). Only sections that were not stained with uranyl acetate were used for studying ensheathment profiles originating in hOPC transplants.

Stereological quantification of human oligodendrocyte progenitor cell survival and differentiation
Numbers of surviving hOPCs were counted in serial, systematically and randomly sampled coronal sections based on the optical fractionator concept with the aid of a motorized stage Axioplan microscope (Carl Zeiss Inc.) equipped with Stereo Investigator (MicroBrightField Bioscience, lliston, VT, USA) as described previously [22]. To evaluate the migration and possible final residing location of differentiating hOPCs, only the contralateral side of transplantation was examined. hOPCs in corpus callosum and cortex were also counted separately for this purpose. Every twelfth serial coronal section through the transplant/injury site was selected for stereological analysis. The counting frame was set at 50×50 μm and the sampling grid and counting depth were 200×200 μm and 10 μm, respectively. Cells around the transplantation site were not counted because of difficulties in discerning individual cells in the densely packed center of the transplant.

Differentiation of survived hOPCs was estimated in a non-stereological fashion as described previously [27]. Briefly, we counted the total number of SC121(+) cells, as well as cells dually labeled with SC121 and the mature oligodendrocyte marker MBP from our immunofluorescent preparations, on randomly selected fields of cortex and corpus callosum using 40× magnification and avoiding the transplantation site. At least three fields in each of four serial sections were used from each animal. Numbers of SC121(+) and double-labeled profiles were pooled from each case and grouped per experimental protocol. Average numbers of single- and double-labeled cells were generated for two TBI groups and one Sham group (n = 5 per group). Differentiation rate was expressed as percentage of SC121 and MBP double-labeled cells in the population of SC121(+) cells.

Migration mapping of oligodendrocyte lineage cells derived from human oligodendrocyte progenitor cell grafts
The positions of all SC121(+) cells were mapped on every twelfth coronal section through brain levels containing the grafted cells and their lineage using Neurolucida software (MicroBrightField Bioscience). Representative cells differentiated from hOPCs and their processes were also traced with Neurolucida software.

Statistical methods
Variance between and across samples of numbers of oligodendrocyte-lineage cells classified by experimental history (IA versus sham), migratory destination in brain (corpus callosum versus neocortex), and time point after transplantation (6 weeks or 3 months) was analyzed with two-way analysis of variance (ANOVA) or t test. In the case of ANOVA, significant differences were further analyzed with post hoc tests to reveal important main effects or interactions. Statistical analyses were performed with STATISTICA 8.0 (StatSoft Inc., Tulsa, OK, USA).

Results
Axonal injury in nude rats using the impact acceleration model
Immunocompromised nude rats were used here to avoid immune rejection of human cell xenografts into rodent brain [20]. Because the original IA model was developed in Sprague-Dawley [16] and Wistar [31] rats, we first explored whether the same IA settings as the ones used in those strains can cause TAI in nude rats. Induction of TAI was studied with IHC strategies routinely used in TBI studies – that is, antibodies against APP, the monoclonal antibody RMO14 binding to the rod domain of neurofilament heavy and medium chains that are exposed after lesion-induced sidearm proteolysis, and a monoclonal antibody against the 68-kDa light chain neurofilament protein (NF68). IHC was used in brain sections from nude rats exposed to a standard severe IA injury (450 g weight drop from a height of 2 meters) [16]. Tissues were also processed with a modification of the Gallyas silver method. Twenty-four hours post-injury, APP, RMO14 and NF68 IHC consistently labeled axonal pathologies such as undulated axons, axonal swellings and bulbs in the corpus callosum and the corticospinal tract as described in several published studies [32]–[35] (data not shown). Argyrophilic axonal degeneration based on Gallyas silver staining became evident 1 week post-injury in the corpus callosum (Fig. 1a), the optic tract and the corticospinal tract (Fig. 1b). Axonal degeneration labeled with Gallyas silver was still present in the corpus callosum (Fig. 1c), corticospinal tract (Fig. 1d) and other white matter tracts 3 months post-injury. These data suggest that the pattern of TAI in nude rates exposed to IA injury is qualitatively similar to the one described for Sprague-Dawley and Wistar rats and, therefore, the nude rat model is suitable for research into hOPC transplantation outcomes in a diffuse TBI background.

thumbnailFig. 1. Traumatic axonal injury in impact acceleration-injured nude rats as demonstrated by Gallyas silver staining. One week after exposure of nude rats to severe impact acceleration injury, argyrophilic axonal degeneration is pronounced in (a) the corpus callosum and (b) the corticospinal tract (arrows). c,d At 3 months after injury, degenerating axons are still evident in these regions (arrows). Axonal bulbs are also present in (d) (arrowhead). Scale bars = 50 μm
Differentiation of human embryonic stem cells to human oligodendrocyte progenitor cells in vitro
As per Hu and colleague’s original description [18], [19], columnar epithelial cells began to appear and organize into rosettes 10 days after induction of differentiation of embryoid bodies with human ESC medium without fibroblast growth factor (FGF)2 (DMEM/F12, Knockout serum replacer, MEM-nonessential amino acid, 0.1 mM β-mercaptoethanol, 4 ng/mL FGF2) for 3 days and then with neural differentiation medium (NDM; DMEM/F12, N2, MEM-nonessential amino acid, 2 ug/mL heparin) for 6 days. Neuroepithelial cells were initially cultured in the presence of 0.1 μM retinoic acid on laminin-coated plates for 4 days as described by Hu and colleagues [18]. The resulting rosette-rich colonies were manually detached and grown into spheres and then continued to be patterned with retinoic acid and SHH for 10 more days. To generate pre-oligodendrocyte progenitors, spheres were passaged by Accutase and cultured in NDM supplemented with B27, SHH (100 ng/mL) and FGF2 (10 ng/mL) for 10 days. For further differentiation into hOPCs, spheres were cultured in GDM (DMEM/F12, N1, B27, MEM-nonessential amino acid, 60 ng/mL T3, 1 μM cAMP, 0.1 μg/mL biotin) supplemented with SHH, PDGF-AA, IGF1 and NT3 for 2 weeks and then dissociated by Accutase and continued to be feed with the same GDM without SHH (day 49). Every 2 or 3 weeks, the spheres were passaged by the same dissociation method using Accutase and cultured in the same GDM with PDGF-AA, IGF1 and NT3. On day 84 after induction of differentiation, spheres were trypsinized into dissociated hOPCs, plated and cultured further as in Materials and methods. On day 98, cells were detached from the plate and resuspended at a high concentration (2.0 × 108 per mL) for transplantation.

As described in Materials and methods, the cell composition of hOPC transplants was analyzed in a representative sample of cells destined for transplantation with immunocytochemistry for protein markers of various neural cell lineages [36]–[39] (Fig. 2). Results show that only a small percentage of hOPCs (less than 10 %) expressed the neuronal marker TUJ1, an even smaller percentage (less than 1 %) were positive for astrocytic markers (GFAP), and no bone morphogenetic protein(+) mesodermal-lineage cells were detected (Fig. 2a-c). In contrast, these hOPC samples were enriched for cells expressing oligodendrocyte-lineage markers including A2B5, PDGFRα, O4, NG2, Sox10 and MBP (Fig. 2d-i).

thumbnailFig. 2. Characterization of human oligodendrocyte progenitor cells used for transplantation 99 days after induction of differentiation of human embryonic stem cells. (a) No mesodermal lineage cells were detected (bone morphogenetic protein; BMP) and very little (b) neural (type III-tubulin epitope J1; TUJ1) and (c) astrocyte (glial fibrillary acidic protein; GFAP) markers were expressed in the oligodendrocyte progenitor cells (OPCs). Most cells (90-95 %) were positive for early and late OPC and pre-oligodendrocyte markers ((d) A2B5, (e) platelet-derived growth factor receptor (PDGFR)α and (f) O4). Fifty percent of cells were (g) NG2 positive and (h) most cells were positive for the transcriptional factor Sox10. (i) About 50 % of cells showed varied expression of the oligodendrocyte marker myelin basic protein (MBP). Insets in (b) and (c) show typical rare neuronal and astrocytic profiles in these human OPC cultures. Inset in (b) is taken from a culture that was double immunostained for NG2 (green) and TUJ1 (red). Insets in (d-i) are magnifications of cells or groups of cells labeled with asterisks to showcase typical cytologies and immunoreactivities, including nuclear (h) and non-nuclear (d,e,f,g,i) localizations. Other than the nuclear localization of the transcription factor Sox10 (H), most immunoreactivities are especially prominent in processes (e,g) or have punctate peripheral localization (f,i). Scale bars = 20 μm
Survival and migration of transplanted human oligodendrocyte progenitor cells in the rat brain
Human OPCs were transplanted into the deep sensorimotor cortex of IA- and sham-injured rats. Transplanted hOPCs survived very well in the brains of injured and uninjured animals and migrated extensively away from the transplantation site in both groups. We mapped and counted migratory profiles at 6 weeks and 3 months post-transplantation. For the time point of 6 weeks we have data only on IA-injured animals. In the case of 3 months, we have data from both IA-injured and sham animals.

At 6 weeks, there was some migration of transplant-derived SC121(+) human cells into the ipsilateral corpus callosum and external capsule and into the contralateral corpus callosum adjacent to cingulate gyrus (Additional file 2: Fig. S2). At 3 months, transplanted cells had migrated much further (Figs. 3 and 4). hOPCs had densely populated the ipsilateral corpus callosum and the entire length of the external capsule and reached further into the corpus callosum and external capsule on the contralateral side; some cells had entered the contralateral neocortex (Fig. 4). In many cases, these cells had migrated 5 mm or more in antero-posterior distance from the transplantation site (Fig. 4). There was little migration into the gray matter, and cortical invasion of hOPCs was limited to layers adjacent to corpus callosum (lower layer 6). In the case of migration into the neostriatum (Fig. 4), hOPCs were localized strictly inside the white matter striae.

thumbnailFig. 3. Extensive migration of transplant-derived cells 3 months post-transplantation. This representative section from an impact acceleration-injured rat shows that SC121(+) cells (brown) migrate extensively from the transplantation site (arrowhead) along the corpus callosum on both sides. Methylene green was used as counterstain. Arrows denote human oligodendrocyte progenitor cell distribution. Insets are enlarged from the frames to convey information on cytology
thumbnailFig. 4. Migration map of human oligodendrocyte progenitor cell transplant-derived cells from a representative case of an impact acceleration-injured rat. Images were acquired and processed from serial coronal sections (40 μm; every 24th) from a case of an impact acceleration-injured rat 3 months post-transplantation using Neurolucida software. Distance between neighboring sections is 0.96 mm. Distance between migrating cells in section 9 and edge of transplantation site (section 4) is 4.8 mm. The maximal migration at 3 months was greater than 5 mm based on the fact that section 9 was not the furthest section containing human cells (other sections are not shown)
Stereological counts of transplant-derived cells contralateral to the injection side provide a good measure of the migratory potential of transplanted hOPCs. Cell counts in IA-injured animals at 6 weeks and 3 months show massive and progressive migration into the corpus callosum and adjacent cortical layer 6 (Fig. 5). For example, the number of oligodendrocyte-lineage cells in the contralateral corpus callosum is 30 times higher at 3 months compared to 6 weeks (Fig. 5a). Two-way ANOVA examining interaction between time and location shows that time tends to advance the position of cells from the corpus callosum into deep cortical layers (Fig. 5a). Experimental history (IA versus sham) shows no effect on numbers of oligodendrocyte-lineage cells in corpus callosum or cortex at 3 months. There are about three to four times as many oligodendrocyte-lineage cells in corpus callosum compared to cortical layer 6 in both IA-injured and sham animals (Fig. 5b). Two-way ANOVA addressing the interaction between experimental history and location shows no significant effect (that is, lesion does not promote more advanced migration into deep cortical layers).

thumbnailFig. 5. Stereological counts of transplanted human oligodendrocyte progenitor cells migrating into the contralateral hemisphere. a Migratory patterns at two time points (6 weeks and 3 months) after transplantation in two brain regions (corpus callosum and deep cortex) in impact acceleration (IA) animals. b Migratory tendencies in the same two locations based on experimental history (IA versus sham) at 3 months post-transplantation. In (A), the difference in cell numbers between 6 weeks and 3 months is significant by t test (*P < 0.05) in both corpus callosum and deep cortex (layer 6); two-way analysis of variance (ANOVA) shows that there is interaction between time and location (that is, time tends to favor deep cortical over callosal location; P < 0.05). In (B), there are no differences in cell numbers between sham and IA-injured subjects in corpus callosum or deep cortex at 3 months post-transplantation. In both groups of subjects, there are more cells in corpus callosum than deep cortex by t test (P < 0.05). Two-way ANOVA shows that there is no interaction between experimental history and location (that is, injury does not seem to influence the location of cells in one site over the other). OPC, oligodendrocyte progenitor cell
At 3 months post-transplantation, a majority of oligodendrocyte-lineage cells around the transplantation site (the triangular region of Fig. 3) had round perikaryal profiles and multiple radial processes consistent with type I morphology (Fig. 6a-c) [40]. On the other hand, the majority of transplant-derived cells in the corpus callosum were spindle-shaped with parallel processes consistent with type II morphology (Fig. 6d-f) [40]. In the gray matter away from the injection site (ipsilateral or contralateral), cytology was mixed.

thumbnailFig. 6. Some cytological features of human oligodendrocyte progenitor transplant-derived cells 3 months post-transplantation. a-c Around the transplantation site, a large number of transplant-derived SC121(+) cells (brown) have round features with extensive and radially arrayed processes. d-f In the corpus callosum (cc), the majority of SC121(+) cells are spindle shaped with long parallel processes. (b,e) Enlargements of bracketed areas in (a) and (d), respectively. (c,f) Neurolucida tracings of representative cells from (b) and (e) indicated with asterisks. Scale bars: (a,d) = 50 μm; (b,e) = 20 μm
Very few (less than 1 %) hOPCs at the transplantation site or within the main migratory domains (corpus callosum and deep neocortex) were positive for the mitotic marker-Ki67 at 6 weeks or 3 months, in injured or sham animals. This pattern suggests that, at the time points studied here, surviving cells are not proliferative at the original transplant site or in their migratory paths and destinations (Fig. 7).

thumbnailFig. 7. Proliferative activity of human oligodendrocyte progenitor cells at 6 weeks and 3 months post-transplantation. Only rare HNu(+) transplant-derived cells (red) are Ki67(+) cycling cells (green) (arrow, the color turns into yellow because of overlapping with red color from HNu) at the transplantation site at 6 weeks (a) or, after migration, in the corpus callosum (cc) at 3 months (b). Images are from a representative impact acceleration-injured rat. Scale bars = 50 μm
Differentiation of transplanted human oligodendrocyte progenitor cells in the rat brain
At 6 weeks and 3 months post-transplantation, under either IA or sham conditions, no neurons and very few astrocytes were derived from transplanted hOPCs (Fig. 8a,b). The majority of transplanted cells were identified as PDGFRα(+) (Fig. 8i-k) or MBP(+) (Fig. 8c-h) profiles in both the transplantation site and migratory pathways/destinations. MBP immunoreactivity was expressed in both round and spindle-shaped oligodendrocyte profiles derived from the transplant (Fig. 8c-h). At 3 months, most transplant-derived cells around the transplantation site, in corpus callosum, and deep cortical layers were also APC(+) (Fig. 8l-n).

thumbnailFig. 8. Differentiation of transplanted human oligodendrocyte progenitor cells at 3 months post-transplantation. All images are from representative impact acceleration-injured animals. a At 3 months, we found no SC121(+) cells expressing type III-tubulin epitope J1 (TUJ1(+)) neuronal phenotypes, and only rare SC121(+) human oligodendrocyte progenitor cells (hOPCs; red) had differentiated into glial fibrillary acidic protein (GFAP(+)) astrocytes (green) at the transplantation site (arrow). Inset in (a) is a magnification of the astrocytic profile indicated by the arrow in the main panel. b In the corpus callosum (cc), no hOPCs (red) are immunoreactive for GFAP (green). c-h Confocal images to show that various types of cells derived from the hOPC transplant (round in (c) or spindle-shaped in (f), red) become mature myelin basic protein (MBP) (+) oligodendrocytes (green in (d) and (g)); both cell bodies (arrows) and processes (arrowheads) of transplant-derived cells are immunoreactive for MBP. i-n Platelet-derived growth factor receptor (PDGFR)α(+) (i-k) and adenomatous polyposis coli protein (APC) (+) (l-n) cell bodies of transplant-derived cells in the corpus callosum. (e,h,k,n) Merged images of panels (c,d), (f,g), (i,j) and (l,m), respectively. Scale bars: (a,b) = 50 μm; (c-n) = 20 μm
In the area surrounding the transplantation site of IA-injured animals, cell counts of PDGFRα(+) profiles show that 74.6 ± 9 % of graft-derived cells are PDGFRα(+) at 6 weeks post-transplantation; this number is significantly reduced to 49.4 ± 11 % at 3 months. Conversely, the percentage of MBP(+) oligodendrocytes derived from hOPCs is significantly higher at 3 months (67.8 ± 12 %) compared to 6 weeks (37.1 ± 9 %) (Fig. 9a, left). A similar pattern is seen in the corpus callosum (Fig. 9a, right), but trends in this case do not reach statistical significance. In the area surrounding the transplantation site, there are no significant differences in PDGFRα(+) or MBP(+) cell rates between sham and injured animals (Fig. 9b, left). In the corpus callosum, the percentage of MBP(+) cells is significantly higher in injured animals compared to shams (Fig. 9b, right).

thumbnailFig. 9. Population sizes of platelet-derived growth factor receptor α- and myelin basic protein-immunoreactive cells derived from human oligodendrocyte progenitor cell transplant. a Differentiation patterns at 6 weeks and 3 months after transplantation in two brain regions (corpus callosum versus transplantation area in cortex) in impact acceleration (IA) animals. b Differentiation patterns in the same two locations based on experimental history (IA versus sham) at 3 months post-transplantation. In (a), IA rats at 3 months have more myelin basic protein (MBP) (+) and fewer platelet-derived growth factor receptor (PDGFR)α(+) transplant-derived oligodendrocytes than at 6 weeks post-transplantation; counts were performed at the transplantation site. In (b), IA rats at 3 months have more transplant-derived MBP(+) oligodendrocytes compared to sham in the corpus callosum (cc). *P < 0.05
Confocal microscopy with three-dimensional image reconstruction for SC121- and NF-H-immunoreactive structures was used to visualize appositions between SC121(+) processes belonging to transplant-derived cells and NF-H(+) SC121(−) host axons. Reconstructed confocal images demonstrated many ensheathing appositions between processes of transplant-derived oligodendrocytes and host axons (Fig. 10).

thumbnailFig. 10. Ensheathment of axons by transplant-derived oligodendrocytes. This three-dimensional reconstructed confocal micrograph depicts SC121(+) process (red) from transplant-derived oligodendrocytes ensheathing neurofilament H(+) axons (NFH, green) in a representative impact acceleration-injured animal at 3 months post-transplantation (arrows on z plane). Ensheathment is confirmed on x and y planes at the corresponding cross-sectional locations (arrow heads). Scale bar = 10 μm
Ultrastructural IHC for the human cytosolic epitope SC121 was used to disclose the involvement of transplant-derived oligodendrocyte processes in the ensheathment of host-derived axons or the formation of myelin. In a pattern similar to the one revealed with confocal microscopy, semi-thin preparations accompanying thin sections showed SC121(+) processes co-localizing with toluidine blue-stained myelin sheaths (Fig. 11a). Using thin sections, we found numerous SC121(+) cytoplasmic projections juxtaposed to or ensheathing unlabeled (host) axons. Ensheathment was featured by complex configurations, including the presence of outer and inner cytoplasmic tongues and close juxtapositions with compact myelin (Fig. 11b,c). It was not possible to ascertain whether compact myelin belonging to the same host axons, as SC121(+) sheaths were continuous with the latter in our preparations because the cytoplasmic human epitope SC121 would not be expected to be present within dense myelin.

thumbnailFig. 11. Ensheathing profiles issued by transplant-derived oligodendrocytes as shown by ultrastructural immunohistochemistry. Preparations are from an injured animal 3 month post-transplantation. a Companion toluidine blue-stained semi-thin section through the corpus callosum shows the co-localization of SC121(+) (brown) processes with blue myelin sheaths in transverse (arrow) axonal profiles. Co-localization profiles are dark brown. Asterisk shows a group of SC121(−) axons. b An SC121(+) process (arrowhead) is shown to ensheath an unlabeled axon. This profile is adjacent to one SC121(+) cell (1) and also one unlabeled cell (2). Cells 1 and 2 have the appearance of oligodendrocytes. c A magnification of the framed area in (b) shows detailed ultrastructural features of ensheathment by transplant-derived oligodendrocytes. SC121(+) tongue processes (arrowheads) are wrapped around a myelinated axon. Myelin sheath on the inside appears to be unlabeled. Scale bars: (a) = 5 μm; (b) = 1 μm; (c) = 500 nm
Discussion
Our findings indicate that the IA model of Marmarou can be effectively replicated in the nude rat background. Using the nude rat IA model, hOPC transplants survive well in the deep sensorimotor cortex and behave in a fashion very different from NPs; that is, they migrate massively and show almost exclusive affinity for white matter tracts, especially the corpus callosum and adjacent white matter in deep cortical layers. The progressive migration of transplanted hOPCs is accompanied by progressive maturation into MBP(+) and APC(+) oligodendrocytes that ensheath host axons. Our findings provide further support to the notion that human ESCs and neural stem cells can be coaxed to specific fates that continue to progress to fully differentiated progenies after transplantation into the adult CNS. These progenies behave in a fashion that is strikingly similar to indigenous differentiated neural cells. Given the very low level of proliferation of transplanted cells as early as 6 weeks post-transplantation and their prompt differentiation into mature oligodendrocytes, the possibility of overgrowth and, hence, tumorigenic risk is very low. We postulate that the high numbers of cells present in the contralateral hemisphere by 3 months had started as pre-existing hOPCs in the dense center of the transplant and did not derive from ongoing cell divisions.

The use of human ESCs such as line H9 was based on a number of considerations including: thorough characterization and inexhaustible supply of the parent line [41]–[44]; great versatility to differentiate to any neural cell type in sufficient quantity for transplantation [18], [43]–[50]; and well-established methods for in vitro manipulation to fate determination prior to transplantation. The choice of human ESCs is based on availability and access considerations, the greater translational value of such cells, and a long experience in our laboratory using human cells as transplants in rodent hosts [21]–[24], [27], [50]–[54].

In adulthood, the sources of usable stem cells or neural progenitors in the CNS are limited to a few forebrain niches, and the yield or repair potential of such niches is low. For example, in mouse models of multiple sclerosis, the limited recruitment of endogenous hOPCs into demyelination sites does not suffice for effective remyelination [55]. Therefore, supplementation of such limited stem cell pools with exogenous progenitors is a reasonable first step for a cellular therapy. Besides providing sources of fully differentiated nerve cells competent to replenish lost cells, transplanted progenitors also release neuroprotective molecules [27], [56] and, importantly, may induce endogenous stem cells/progenitor cells to proliferate and differentiate as auxiliary niches, thereby improving the efficacy of self-repair mechanisms [52].

Generation of human oligodendrocyte progenitor cells from human embryonic stem cells
In the body of in vivo studies reported here, we used hOPCs that were prepared from human ESC line H9 following the methods described by Hu and colleagues with minor modifications [19], [45], [57]. Our experience with culturing and differentiating H9 cells and then characterizing the derived hOPCs in vitro is very similar to the original description of Hu and colleagues; hOPCs derived in this manner express PDGFRα, NG2, O4, and Sox10, a pattern consistent with a classical hOPC identity [36]. Methodological issues concerning hOPC derivation are very important, because the use of highly concentrated hOPCs for grafting is key for achieving the desired outcomes (that is, myelination of host axons) within a limited time period.

Differentiation of human ESCs into hOPCs is a longer and more arduous process than the one leading to neuronal progenitors. In work reported here we initially used two methods [18], [58] for deriving and characterizing hOPCs prior to transplantation. In our hands, the method of Hu and colleagues appeared to be more successful in generating viable transplants, although this observation was not confirmed in a systematic fashion. Major differences between the two methods are: trophic factors used; extracellular matrix used; enzyme used to passage the cells; length of time in three-dimensional culture; and timing of hOPC harvest. In previous published work, the method of Nistor and colleagues generated hOPCs with good viability after transplantation [58]–[62], but the transplantation site of these authors (spinal cord) was different from ours (neocortex). Interestingly, the same team of investigators reported that their hOPCs did not survive past 2 weeks after transplantation in animal models of multiple sclerosis, but these mice were immunocompetent C57B6 mice [63]. Our in vivo outcomes using hOPCs prepared as per Hu and colleagues are consistent with the ones reported by that team on shiverer mice [19].

Issues related to in vivo differentiation of human oligodendrocyte progenitor cell transplants
Two important trends in hOPC maturation were the progressive phenotypic differentiation in the transplantation area and the higher rate of maturation of hOPCs in the corpus callosum of injured subjects. With respect to the former, we have observations only from IA-injured animals, but we speculate that there is a similar differentiation trend in control (sham) animals. We also postulate that differentiation trends in the corpus callosum are not too dissimilar to those in the cortex around the transplantation site and that differences in significance may be caused by the fact that hOPC maturation may be earlier in the white matter [64] compared to grey matter. Regarding the latter, it would appear that injury may contribute to hOPC differentiation; the difference between corpus callosum and cortex around the transplantation site may be due to the fact that cortex is not a primary site of injury in the IA model that preferentially affects white matter tracts. We have previously reported that the injured spinal cord niche influences both the proliferation and differentiation of human CNS stem cells propagated as neurospheres [65].

The role of injured or otherwise pathological environment as a niche for differentiation deserves further commentary. Environments associated with acquired neurological injury such as stroke and spinal cord injury have been shown to promote endogenous stem cell differentiation [9], [66], [67] and this effect may be mediated, in part, by trophic and cytokine signals such as stem cell factor [68], stromal cell-derived factor 1α [69], FGF-2 [70], vascular endothelial growth factor [71], ciliary neurotrophic factor, and CXC chemokine receptor 4 [72]. Some of these factors are known to act as tropic cues for migration or to specify the definitive phenotype of endogenous or exogenous stem cells [68], [73]–[79].

The invasion of migrated hOPCs and their differentiated progenies into deep cortical layers matches native cortical myelination patterns; these patterns show an overall denser myelination in lower cortical layers and variable myelination in superficial layers. It is also interesting that, at least up to 3 months post-transplantation, the differentiated progeny of transplanted hOPCs does not advance to more superficial layers. This distribution is identical to that of native mature oligodendrocytes; in contrast to hOPCs that are radially spread across cortical layers, oligodendrocytes favor deep layers and follow the deep-to-superficial-layer myelin gradient [80].

The presumed intent of OPC transplants as cell therapies is the replenishment of damaged or destroyed myelin sheaths. Although molecular myelin markers such as MBP may be telling of the myelin-forming potential of OPCs and their progenies, they do not directly show that MBP(+) cells are forming myelin. The three-dimensional reconstruction of confocal images indicates a close, ensheathment-like, apposition of host axons with processes of transplant-derived oligodendrocytes, but the presence of structurally mature myelin can only be ascertained ultrastructurally. Using ultrastructural IHC or electron microscopy combined with histochemistry, previous studies have demonstrated the ability of specific progenies of neural stem cells or OPCs to ensheath [21] or increase the thickness of abnormal myelin sheaths in hosts [19], [26], [59], [61], [81], [82], but labeling of exogenously derived myelin is technically difficult. In the present study, ultrastructural IHC divulges that transplant-derived cells ensheath host axons in intimate proximity to myelin sheaths, but the host-versus-transplant identity of the myelin itself is difficult to ascertain. The human cell marker SC121 is a cytosolic marker and, therefore, it is found in oligodendrocyte cytoplasmic projections and ensheathing tongues but not in the membranous myelin sheath itself. Unstained preparations that were used here have not been particularly useful, primarily because of this reason. Furthermore, existing myelin antibodies cannot resolve between human and rodent myelin and this problem limits their usefulness in confocal microscopy or ultrastructural IHC.

Ultimately, the proof of concept that remyelination or myelin remodeling by OPC transplants can be beneficial in TAI/DAI will depend on the demonstration that such exogenous OPCs afford functional benefits. In the case of IA-injured rodents, such benefits can be sought out in a number of behavioral domains. One approach is assessment of motor control, which depends on the intactness of the corticospinal tract. Experiments addressing functional repair with exogenous OPCs have been successfully performed in models of spinal cord injury [10], [11] and demyelination [82]–[85]. In TBI models, motor recovery may occur within the first month after injury in the absence of any therapy, making the assessment of the efficacy of cell therapies more challenging. In such models, the assessment of efficacy or human neural cell therapies may be more straightforward using tasks of cognition or anxious/emotional disposition, faculties that become chronically impaired in rodent TBI. Chronicity of deficits is important because these human cells may take months to proliferate, migrate and terminally differentiate [15].

Stem cell transplantation as experimental therapy for traumatic brain injury
Some success in models of ischemic brain injury [9] has encouraged the use of stem cell/NP transplantation in models of focal TBI [10], [86]. However, because of the complexity of TBI and its animal models, there is a need to identify specific repair targets based on key pathological mechanisms. Such repair tasks include replacing dead neurons, supporting injured neurons, and protecting axons or assisting with axonal repair/regeneration. The problem of neuronal injury/death is encountered both in focal injury [87], [88] and in the course of TAI [89], [90]. Neuronal cell death in focal TBI is acute and has necrotic components, whereas in TAI/DAI it is slow with apoptotic features and may be associated with retrograde and trans-synaptic effects [8], [89], [91]. Although axonal repair/remyelination as a therapeutic target separate from neuronal regeneration is best established in spinal cord injury [11], there is evidence that demyelination may contribute to degeneration of axons in TAI [12], [13]. Therefore, contributing exogenous hOPCs in the case of TAI may assist in remyelination and prevent axonal degeneration and disconnection within brain circuits.

There is very little published work on stem cell-based therapies for models of TAI/DAI. However, the field of TBI and more specifically TAI/DAI can borrow from spinal cord injury that invariably involves trauma in long tracts [59], [60], [92]–[96]. Experimental cell therapies in animal models of spinal cord injury have utilized various stem cell preparations including neurospheres and OPCs [59], [60], [96], [97], and there are several ongoing clinical trials using neural stem cells [98], [99]. In one report, OPCs were found to remyelinate and restore locomotion after contusional spinal cord injury in rodents [59], but functional recovery reported in this study occurred within 12 days of transplant, a time point that is too early with respect to migration and terminal differentiation of OPCs. In contrast to spinal cord injury, where long tracts course in relatively circumscribed areas, DAI involves disparate white matter tracts [100], [101] and it would be difficult to transplant cells into all these sites. Therefore, transplantation route (systemic, ventricular, and parenchymal) and location of transplant (in the case of parenchymal delivery) are critical. The choice of transplantation site may be based on factors such as concentration of axonal pathology or sites of injury responsible for critical symptoms. The choice of transplantation into deep sensorimotor cortex in the present study was based on the expectation that this site would provide oligodendrocytes for both the corpus callosum and the corticospinal tract, which are affected in the IA injury [16], [35]. Our findings indicate that there was little, if any, invasion of cells into the internal capsule by 3 months, but the extensive migration of OPCs via the corpus callosum and descending tracts and their remarkable differentiation into mature oligodendrocytes predicts a broader remyelination potential with longer survival times. Of course, transplantation sites can also be optimized to the desired functional outcome or involve multiple locations as we have shown in models of motor neuron disease [23], [52], [53].

Conclusions
In conclusion, the findings in this study support the idea that hOPCs can serve as a competent source of mature oligodendrocytes that ensheath CNS axons after TAI, and provide proof of concept that regenerative strategies targeting myelin remodeling can be further considered in TBI models in the future. In addition, we demonstrate that the nude rat is a suitable animal model for studying human cell transplants in neurotrauma. In view of the fact that stem cell therapies are being progressively introduced in clinical trials of neurodegenerative and traumatic diseases of the CNS [11], [98], [102]–[108], these timely results should encourage further translational work targeting the problem of axonal degeneration in the context of DAI and, specifically, interventions designed to regenerate or remodel the myelin sheath.

Abbreviations
ANOVA: analysis of variance

APC: adenomatous polyposis coli protein

APP: amyloid precursor protein

CNS: central nervous system

DAI: diffuse axonal injury

DMEM: Dulbecco’s modified Eagle’s medium

ESC: embryonic stem cell

FGF: fibroblast growth factor

GDM: glial differentiation medium

GFAP: glial fibrillary acidic protein

hOPC: human oligodendrocyte progenitor cell

IA: impact acceleration

IGF: insulin-like growth factor

IHC: immunohistochemistry

MBP: myelin basic protein

NDM: neural differentiation medium

NP: neural precursor

NY: neurotrophin

OPC: oligodendrocyte progenitor cell

PDGF: platelet-derived growth factor

PDGFR: platelet-derived growth factor receptor

SHH: sonic hedgehog

TAI: traumatic axonal injury

TBI: traumatic brain injury

TUJ1: type III-tubulin epitope J1

Competing interests
The authors declare that they have no competing interests.

Authors’ contributions
LX, configured the details of experimental plan, performed all animal surgeries and IHC staining, analyzed in vivo data and prepared the first draft of manuscript. JR, established and maintained human OPC cultures and characterized them prior to transplantation. HH, performed ultrastructural IHC. AM, performed stereological counts of PDGFRα and MBD(+) cells derived from the OPC transplant. AA, helped with stereology. ER, performed mapping and stereological counting of migrated OPC-lineage cells. VM, participated in the design of the project and assisted in IHC. BJC, participated in the design of the project and edited the manuscript. VEK, principal investigator, designed the project, troubleshot the experiments and prepared the final manuscript. All authors read and approved the manuscript.

Authors’ information
LX (MD PhD) is an expert on stem cell therapies for animal models of neurological disease and animal models of traumatic brain injury. JR (PhD) is an expert in embryonic and neural stem cell culture and manipulation techniques. HH (PhD) specializes in electron microscope techniques for the nervous system. AM, AA and ER are undergraduate students at Johns Hopkins University. VM (PhD) is an expert in embryonic and neural stem cell culture and manipulation techniques. BJC (PhD) is an expert in stem cell biology and animal models of neurodegeneration and neurotrauma. VEK (MD) is an expert in neural injury and repair, neurodegenerative disease, and traumatic brain injury.

Additional files
Additional file 1: Fig. S1.. A schematic illustration (A) of the method used to prepare hOPCs for transplantation and representative cultures and cells derived (B). Method sketched in A was based on Hu and colleagues [18] with minor modifications in the final stage before transplantation (day 84–99). Panel B shows representative morphologies of hOPCs on days 3, 7, 12, 17, 23, 27, 41, 90 and 97 that roughly correspond to milestones in A. Human ESC H9 colonies were detached by dispase on day 1 to prepare embryoid bodies (EBs) that were subsequently cultured for 3 days in hES medium without FGF (B, Day 3) and then for 3 days in NDM. Embryoid bodies (B, Day 7) were then plated on laminin-coated plates and cultured with NDM for 3 days and NDM with RA for another 5 days (B, Day 12). On day 15, colonies were manually detached and cultured as spheres (B, Day 17) for 10 days in NDM with RA, B27 and SHH. On day 24, big spheres (B, Day 23) were dissociated by Accutase (ACT) into small spheres (B, Day 27) and cultured with NDM containing B27, SHH and FGF for 10 days. On day 35, medium was switched to GDM with SHH, PDGF, IGF and NT3 and cultured for 2 weeks (B, Day 41). On days 49 and 70, cells were passaged with Accutase treatment and treated with GDM with PDGF-AA, IGF1 and NT3 for the remaining of the protocol. On day 84, spheres were trypsinized and plated on p-L-ornithine and laminin-coated plates and coverslips and cultured for 2 weeks in the same medium for transplantation or immunocytochemistry, respectively. On day 99, cells were trypsinized with TrypLE, counted, resuspended in high concentration and used for transplantation (* in A). Scale bars in B: Day 12, 500 μm, Day 3 and 17; 200 μm, all others; 100 μm.
Format: TIFF Size: 6.7MB Download fileOpen Data
Additional file 2: Fig. S2.. Migration of transplant-derived cells at 6 weeks and 3 months post-transplantation. Photographs are taken through the corpus callosum on the side contralateral to transplantation from representative animals. Panel A shows a few scattered SC121(+) profiles at 6 weeks. At 3 months post-transplantation (B), there are numerous SC121(+) cells that had migrated from the transplantation site. Insets are enlargements of framed areas in main panels. Scale bars: 100 μm.
Format: TIFF Size: 18.1MB Download fileOpen Data
Acknowledgement
This work was supported by a Maryland Technology Development Corporation (TEDCO) grant to VEK funded as companion to CIRM grant TR2-01767 to BJC. We would like to thank Ms Devon Hitt who offered great technical help with immunostaining and collection of some quantitative data.

References
Mac Donald CL, Johnson AM, Cooper D, Nelson EC, Werner NJ, Shimony JS, et al.: Detection of blast-related traumatic brain injury in U.S. military personnel.
N Engl J Med 2011, 364:2091-2100. PubMed Abstract | Publisher Full Text OpenURL
Mckee AC, Stern RA, Nowinski CJ, Stein TD, Alvarez VE, Daneshvar DH, et al.: The spectrum of disease in chronic traumatic encephalopathy.
Brain. 2013, 136:43-64. PubMed Abstract | Publisher Full Text OpenURL
Mittl RL, Grossman RI, Hiehle JF, Hurst RW, Kauder DR, Gennarelli TA, et al.: Prevalence of MR evidence of diffuse axonal injury in patients with mild head injury and normal head CT findings.
Am J Neuroradiol. 1994, 15:1583-1589. PubMed Abstract | Publisher Full Text OpenURL
Geddes JF, Whitwell HL, Graham DI: Traumatic axonal injury: practical issues for diagnosis in medicolegal cases.
Neuropathol Appl Neurobiol. 2000, 26:105-116. PubMed Abstract | Publisher Full Text OpenURL
Wang HC, Ma YB: Experimental models of traumatic axonal injury.
J Clin Neurosci. 2010, 17:157-162. PubMed Abstract | Publisher Full Text OpenURL
Adams JH, Doyle D, Graham DI, Lawrence AE, McLellan DR: Microscopic diffuse axonal injury in cases of head injury.
Med Sci Law. 1985, 25:265-269. PubMed Abstract | Publisher Full Text OpenURL
Gennarelli TA, Thibault LE, Adams JH, Graham DI, Thompson CJ, Marcincin RP: Diffuse axonal injury and traumatic coma in the primate.
Ann Neurol. 1982, 12:564-574. PubMed Abstract | Publisher Full Text OpenURL
Buki A, Povlishock JT: All roads lead to disconnection? Traumatic axonal injury revisited.
Acta Neurochir. 2006, 148:181-193. PubMed Abstract | Publisher Full Text OpenURL
Burns TC, Verfaillie CM, Low WC: Stem cells for ischemic brain injury: a critical review.
J Comp Neurol. 2009, 515:125-144. PubMed Abstract | Publisher Full Text OpenURL
Richardson RM, Singh A, Sun D, Fillmore HL, Dietrich DW, Bullock MR: Stem cell biology in traumatic brain injury: effects of injury and strategies for repair.
J Neurosurg. 2010, 112:1125-1138. PubMed Abstract | Publisher Full Text OpenURL
Kwon BK, Sekhon LH, Fehlings MG. Emerging repair, regeneration, and translational research advances for spinal cord injury. Spine (Phila Pa 1976). 2010;35:S263–70.
Maxwell WL: Damage to myelin and oligodendrocytes: a role in chronic outcomes following traumatic brain injury?
Brain Sci. 2013, 3:1374-1394. PubMed Abstract | Publisher Full Text OpenURL
Maxwell WL, Domleo A, McColl G, Jafari SS, Graham DI: Post-acute alterations in the axonal cytoskeleton after traumatic axonal injury.
J Neurotrauma. 2003, 20:151-168. PubMed Abstract | Publisher Full Text OpenURL
Singleton RH, Zhu JP, Stone JR, Povlishock JT: Traumatically induced axotomy adjacent to the soma does not result in acute neuronal death.
J Neurosci. 2002, 22:791-802. PubMed Abstract | Publisher Full Text OpenURL
Gold EM, Su D, Lopez-Velazquez L, Haus DL, Perez H, Lacuesta GA, et al.: Functional assessment of long-term deficits in rodent models of traumatic brain injury.
Regen Med. 2013, 8:483-516. PubMed Abstract | Publisher Full Text OpenURL
Marmarou A, Foda MAA, Vandenbrink W, Campbell J, Kita H, Demetriadou K: A new model of diffuse brain injury in rats. 1. Pathophysiology and biomechanics.
J Neurosurg 1994, 80:291-300. PubMed Abstract | Publisher Full Text OpenURL
Pankratz MT, Li XJ, Lavaute TM, Lyons EA, Chen X, Zhang SC: Directed neural differentiation of human embryonic stem cells via an obligated primitive anterior stage.
Stem Cells. 2007, 25:1511-1520. PubMed Abstract | Publisher Full Text OpenURL
Hu BY, Du ZW, Zhang SC: Differentiation of human oligodendrocytes from pluripotent stem cells.
Nat Protoc. 2009, 4:1614-1622. PubMed Abstract | Publisher Full Text OpenURL
Hu BY, Du ZW, Li XJ, Ayala M, Zhang SC: Human oligodendrocytes from embryonic stem cells: conserved SHH signaling networks and divergent FGF effects.
Development. 2009, 136:1443-1452. PubMed Abstract | Publisher Full Text OpenURL
Anderson AJ, Haus DL, Hooshmand MJ, Perez H, Sontag CJ, Cummings BJ: Achieving stable human stem cell engraftment and survival in the CNS: is the future of regenerative medicine immunodeficient?
Regen Med. 2011, 6:367-406. PubMed Abstract | Publisher Full Text OpenURL
Yan J, Welsh AM, Bora SH, Snyder EY, Koliatsos VE: Differentiation and tropic/trophic effects of exogenous neural precursors in the adult spinal cord.
J Comp Neurol. 2004, 480:101-114. PubMed Abstract | Publisher Full Text OpenURL
Yan J, Xu L, Welsh AM, Hatfield G, Hazel T, Johe K, et al.: Extensive neuronal differentiation of human neural stem cell grafts in adult rat spinal cord.
PLoS Med. 2007., 4
Article ID e39
PubMed Abstract | Publisher Full Text OpenURL
Xu L, Ryugo DK, Pongstaporn T, Johe K, Koliatsos VE: Human neural stem cell grafts in the spinal cord of SOD1 transgenic rats: differentiation and structural integration into the segmental motor circuitry.
J Comp Neurol. 2009, 514:297-309. PubMed Abstract | Publisher Full Text OpenURL
Nasonkin I, Mahairaki V, Xu L, Hatfield G, Cummings BJ, Eberhart C, et al.: Long-term, stable differentiation of human embryonic stem cell-derived neural precursors grafted into the adult mammalian neostriatum.
Stem Cells. 2009, 27:2414-2426. PubMed Abstract | Publisher Full Text OpenURL
Koliatsos VE, Price DL, Clatterbuck RE: Motor neurons in Onuf’s nucleus and its rat homologues express the p75 nerve growth factor receptor: sexual dimorphism and regulation by axotomy.
J Comp Neurol. 1994, 345:510-527. PubMed Abstract | Publisher Full Text OpenURL
Cummings BJ, Uchida N, Tamaki SJ, Salazar DL, Hooshmand M, Summers R, et al.: Human neural stem cells differentiate and promote locomotor recovery in spinal cord-injured mice.
Proc Natl Acad Sci U S A. 2005, 102:14069-14074. PubMed Abstract | Publisher Full Text OpenURL
Xu L, Yan J, Chen D, Welsh AM, Hazel T, Johe K, et al.: Human neural stem cell grafts ameliorate motor neuron disease in SOD-1 transgenic rats.
Transplantation. 2006, 82:865-875. PubMed Abstract | Publisher Full Text OpenURL
Gallyas F: Silver staining of Alzheimer’s neurofibrillary changes by means of physical development.
Acta Morphol Hung. 1971, 19:1-8. OpenURL
Koliatsos VE, Cernak I, Xu L, Song Y, Savonenko A, Crain BJ, et al.: A mouse model of blast injury to brain: initial pathological, neuropathological, and behavioral characterization.
J Neuropathol Exp Neurol. 2011, 70:399-416. PubMed Abstract | Publisher Full Text OpenURL
Hooshmand MJ, Anderson AJ, Cummings BJ. Improved pre-embedded immuno-electron microscopy procedures to preserve myelin integrity in mammalian central nervous tissue. In: Mendez-Vilas A, editor. Microscopy: advances in scientific research and education. Volume 6(1). Badajoz, Spain: Formatex Research Center; 2014. p. 59–65.
Engelborghs K, Verlooy J, Van RJ, Van DB, Van DV, Borgers M: Temporal changes in intracranial pressure in a modified experimental model of closed head injury.
J Neurosurg. 1998, 89:796-806. PubMed Abstract | Publisher Full Text OpenURL
Kallakuri S, Li Y, Zhou R, Bandaru S, Zakaria N, Zhang L, et al.: Impaired axoplasmic transport is the dominant injury induced by an impact acceleration injury device: an analysis of traumatic axonal injury in pyramidal tract and corpus callosum of rats.
Brain Res. 2012, 1452:29-38. PubMed Abstract | Publisher Full Text OpenURL
Zakaria N, Kallakuri S, Bandaru S, Cavanaugh JM: Temporal assessment of traumatic axonal injury in the rat corpus callosum and optic chiasm.
Brain Res. 2012, 1467:81-90. PubMed Abstract | Publisher Full Text OpenURL
Marmarou CR, Walker SA, Davis CL, Povlishock JT: Quantitative analysis of the relationship between intra-axonal neurofilament compaction and impaired axonal transport following diffuse traumatic brain injury.
J Neurotrauma. 2005, 22:1066-1080. PubMed Abstract | Publisher Full Text OpenURL
Foda MAA, Marmarou A: A new model of diffuse brain injury in rats. 2. Morphological characterization.
J Neurosurg 1994, 80:301-313. PubMed Abstract | Publisher Full Text OpenURL
Alsanie WF, Niclis JC, Petratos S: Human embryonic stem cell-derived oligodendrocytes: protocols and perspectives.
Stem Cells Dev. 2013, 22:2459-2476. PubMed Abstract | Publisher Full Text OpenURL
Iannarelli P, Grist M, Wegner M, Richardson WD, Fogarty M, Nicoletta K: Competing waves of oligodendrocytes in the forebrain and postnatal elimination of an embryonic lineage.
Nat Neurosci. 2006, 9:173-179. PubMed Abstract | Publisher Full Text OpenURL
Rowitch DH, Kriegstein AR: Developmental genetics of vertebrate glial-cell specification.
Nature. 2010, 468:214-222. PubMed Abstract | Publisher Full Text OpenURL
Richardson WD, Kessaris N, Pringle N: Oligodendrocyte wars.
Nat Rev Neurosci. 2006, 7:11-18. PubMed Abstract | Publisher Full Text OpenURL
Suchet S. The morphology and ultrastructure of oligodendrocytes and their functional implication. In: Kettenmann H, Ransom BR, editors. Neuroglia. New York: Oxford University Press, Inc; 1995. p. 23–43.
Thomson JA, Itskovitz-Eldor J, Shapiro SS, Waknitz MA, Swiergiel JJ, Marshall VS, et al.: Embryonic stem cell lines derived from human blastocysts.
Science. 1998, 282:1145-1147. PubMed Abstract | Publisher Full Text OpenURL
Xu RH, Peck RM, Li DS, Feng X, Ludwig T, Thomson JA: Basic FGF and suppression of BMP signaling sustain undifferentiated proliferation of human ES cells.
Nat Methods. 2005, 2:185-190. PubMed Abstract | Publisher Full Text OpenURL
Zhang SC, Wernig M, Duncan ID, Brustle O, Thomson JA: In vitro differentiation of transplantable neural precursors from human embryonic stem cells.
Nat Biotechnol. 2001, 19:1129-1133. PubMed Abstract | Publisher Full Text OpenURL
Yan Y, Yang D, Zarnowska ED, Du Z, Werbel B, Valliere C, et al.: Directed differentiation of dopaminergic neuronal subtypes from human embryonic stem cells.
Stem Cells. 2005, 23:781-790. PubMed Abstract | Publisher Full Text OpenURL
Hu BY, Zhang SC: Differentiation of spinal motor neurons from pluripotent human stem cells.
Nat Protoc. 2009, 4:1295-1304. PubMed Abstract | Publisher Full Text OpenURL
Li XJ, Du ZW, Zarnowska ED, Pankratz M, Hansen LO, Pearce RA, et al.: Specification of motoneurons from human embryonic stem cells.
Nat Biotechnol. 2005, 23:215-221. PubMed Abstract | Publisher Full Text OpenURL
Weick JP, Liu Y, Zhang SC: Human embryonic stem cell-derived neurons adopt and regulate the activity of an established neural network.
Proc Natl Acad Sci U S A. 2011, 108:20189-20194. PubMed Abstract | Publisher Full Text OpenURL
Zhang SC, Lipsitz D, Duncan ID: Self-renewing canine oligodendroglial progenitor expanded as oligospheres.
J Neurosci Res. 1998, 54:181-190. PubMed Abstract | Publisher Full Text OpenURL
Nguyen HX, Nekanti U, Haus DL, Funes G, Moreno D, Kamei N, et al.: Induction of early neural precursors and derivation of tripotent neural stem cells from human pluripotent stem cells under xeno-free conditions.
J Comp Neurol. 2014, 522:2767-2783. PubMed Abstract | Publisher Full Text OpenURL
Haus DL, Nguyen HX, Gold EM, Kamei N, Perez H, Moore HD, et al.: CD133-enriched Xeno-free human embryonic-derived neural stem cells expand rapidly in culture and do not form teratomas in immunodeficient mice.
Stem Cell Res. 2014, 13:214-226. PubMed Abstract | Publisher Full Text OpenURL
Nasonkin IO, Koliatsos VE: Nonhuman sialic acid Neu5Gc is very low in human embryonic stem cell-derived neural precursors differentiated with B27/N2 and noggin: implications for transplantation.
Exp Neurol. 2006, 201:525-529. PubMed Abstract | Publisher Full Text OpenURL
Xu L, Mahairaki V, Koliatsos VE: Host induction by transplanted neural stem cells in the spinal cord: further evidence for an adult spinal cord neurogenic niche.
Regen Med. 2012, 7:785-797. PubMed Abstract | Publisher Full Text OpenURL
Xu LY, Shen PL, Hazel T, Johe K, Koliatsos VE: Dual transplantation of human neural stem cells into cervical and lumbar cord ameliorates motor neuron disease in SOD1 transgenic rats.
Neurosci Lett. 2011, 494:222-226. PubMed Abstract | Publisher Full Text OpenURL
Yan J, Xu L, Welsh AM, Chen D, Hazel T, Johe K, et al.: Combined immunosuppressive agents or CD4 antibodies prolong survival of human neural stem cell grafts and improve disease outcomes in amyotrophic lateral sclerosis transgenic mice.
Stem Cells. 2006, 24:1976-1985. PubMed Abstract | Publisher Full Text OpenURL
Boyd A, Zhang H, Williams A: Insufficient OPC migration into demyelinated lesions is a cause of poor remyelination in MS and mouse models.
Acta Neuropathol. 2013, 125:841-859. PubMed Abstract | Publisher Full Text OpenURL
Koliatsos VE, Xu LY, Yan J: Human stem cell grafts as therapies for motor neuron disease.
Expert Opin Biol Ther. 2008, 8:137-141. PubMed Abstract | Publisher Full Text OpenURL
Hu BY, Weick JP, Yu J, Ma LX, Zhang XQ, Thomson JA, et al.: Neural differentiation of human induced pluripotent stem cells follows developmental principles but with variable potency.
Proc Natl Acad Sci U S A. 2010, 107:4335-4340. PubMed Abstract | Publisher Full Text OpenURL
Nistor GI, Totoiu MO, Haque N, Carpenter MK, Keirstead HS: Human embryonic stem cells differentiate into oligodendrocytes in high purity and myelinate after spinal cord transplantation.
Glia. 2005, 49:385-396. PubMed Abstract | Publisher Full Text OpenURL
Keirstead HS, Nistor G, Bernal G, Totoiu M, Cloutier F, Sharp K, et al.: Human embryonic stem cell-derived oligodendrocyte progenitor cell transplants remyelinate and restore locomotion after spinal cord injury.
J Neurosci. 2005, 25:4694-4705. PubMed Abstract | Publisher Full Text OpenURL
Sharp J, Frame J, Siegenthaler M, Nistor G, Keirstead HS: Human embryonic stem cell-derived oligodendrocyte progenitor cell transplants improve recovery after cervical spinal cord injury.
Stem Cells. 2010, 28:152-163. PubMed Abstract | Publisher Full Text OpenURL
Cloutier F, Siegenthaler MM, Nistor G, Keirstead HS: Transplantation of human embryonic stem cell-derived oligodendrocyte progenitors into rat spinal cord injuries does not cause harm.
Regen Med. 2006, 1:469-479. PubMed Abstract | Publisher Full Text OpenURL
Hatch MN, Nistor G, Keirstead HS: Derivation of high-purity oligodendroglial progenitors.
Methods Mol Biol. 2009, 549:59-75. PubMed Abstract | Publisher Full Text OpenURL
Hatch MN, Schaumburg CS, Lane TE, Keirstead HS: Endogenous remyelination is induced by transplant rejection in a viral model of multiple sclerosis.
J Neuroimmunol. 2009, 212:74-81. PubMed Abstract | Publisher Full Text OpenURL
Vigano F, Mobius W, Gotz M, Dimou L: Transplantation reveals regional differences in oligodendrocyte differentiation in the adult brain.
Nat Neurosci. 2013, 16:1370-1372. PubMed Abstract | Publisher Full Text OpenURL
Sontag CJ, Uchida N, Cummings BJ, Anderson AJ: Injury to the spinal cord niche alters the engraftment dynamics of human neural stem cells.
Stem Cell Reports. 2014, 2:620-632. PubMed Abstract | Publisher Full Text OpenURL
Moreno-Manzano V, Rodriguez-Jimenez FJ, Garcia-Rosello M, Lainez S, Erceg S, Calvo MT, et al.: Activated spinal cord ependymal stem cells rescue neurological function.
Stem Cells. 2009, 27:733-743. PubMed Abstract | Publisher Full Text OpenURL
Yang H, Lu P, McKay HM, Bernot T, Keirstead H, Steward O, et al.: Endogenous neurogenesis replaces oligodendrocytes and astrocytes after primate spinal cord injury.
J Neurosci. 2006, 26:2157-2166. PubMed Abstract | Publisher Full Text OpenURL
Sun L, Lee J, Fine HA: Neuronally expressed stem cell factor induces neural stem cell migration to areas of brain injury.
J Clin Invest. 2004, 113:1364-1374. PubMed Abstract | Publisher Full Text OpenURL
Robin AM, Zhang ZG, Wang L, Zhang RL, Katakowski M, Zhang L, et al.: Stromal cell-derived factor 1alpha mediates neural progenitor cell motility after focal cerebral ischemia.
J Cereb Blood Flow Metab. 2006, 26:125-134. PubMed Abstract | Publisher Full Text OpenURL
Tripathi RB, McTigue DM: Chronically increased ciliary neurotrophic factor and fibroblast growth factor-2 expression after spinal contusion in rats.
J Comp Neurol. 2008, 510:129-144. PubMed Abstract | Publisher Full Text OpenURL
Vaquero J, Zurita M, De OS, Coca S, Morales C, Salas C: Expression of vascular permeability factor in craniopharyngioma.
J Neurosurg. 1999, 91:831-834. PubMed Abstract | Publisher Full Text OpenURL
Chi L, Ke Y, Luo C, Li B, Gozal D, Kalyanaraman B, et al.: Motor neuron degeneration promotes neural progenitor cell proliferation, migration, and neurogenesis in the spinal cords of amyotrophic lateral sclerosis mice.
Stem Cells. 2006, 24:34-43. PubMed Abstract | Publisher Full Text OpenURL
Fricker RA, Carpenter MK, Winkler C, Greco C, Gates MA, Bjorklund A: Site-specific migration and neuronal differentiation of human neural progenitor cells after transplantation in the adult rat brain.
J Neurosci. 1999, 19:5990-6005. PubMed Abstract | Publisher Full Text OpenURL
Flax JD, Aurora S, Yang C, Simonin C, Wills AM, Billinghurst LL, et al.: Engraftable human neural stem cells respond to developmental cues, replace neurons, and express foreign genes.
Nat Biotechnol. 1998, 16:1033-1039. PubMed Abstract | Publisher Full Text OpenURL
Kim SU: Human neural stem cells genetically modified for brain repair in neurological disorders.
Neuropathology. 2004, 24:159-171. PubMed Abstract | Publisher Full Text OpenURL
Erlandsson A, Larsson J, Forsberg-Nilsson K: Stem cell factor is a chemoattractant and a survival factor for CNS stem cells.
Exp Cell Res. 2004, 301:201-210. PubMed Abstract | Publisher Full Text OpenURL
Widera D, Holtkamp W, Entschladen F, Niggemann B, Zanker K, Kaltschmidt B, et al.: MCP-1 induces migration of adult neural stem cells.
Eur J Cell Biol. 2004, 83:381-387. PubMed Abstract | Publisher Full Text OpenURL
Imitola J, Raddassi K, Park KI, Mueller FJ, Nieto M, Teng YD, et al.: Directed migration of neural stem cells to sites of CNS injury by the stromal cell-derived factor 1alpha/CXC chemokine receptor 4 pathway.
Proc Natl Acad Sci U S A. 2004, 101:18117-18122. PubMed Abstract | Publisher Full Text OpenURL
Zhang H, Vutskits L, Pepper MS, Kiss JZ: VEGF is a chemoattractant for FGF-2-stimulated neural progenitors.
J Cell Biol. 2003, 163:1375-1384. PubMed Abstract | Publisher Full Text OpenURL
Tomassy GS, Berger DR, Chen HH, Kasthuri N, Hayworth KJ, Vercelli A, et al.: Distinct profiles of myelin distribution along single axons of pyramidal neurons in the neocortex.
Science. 2014, 344:319-324. PubMed Abstract | Publisher Full Text OpenURL
Tirotta E, Carbajal KS, Schaumburg CS, Whitman L, Lane TE: Cell replacement therapies to promote remyelination in a viral model of demyelination.
J Neuroimmunol. 2010, 224:101-107. PubMed Abstract | Publisher Full Text OpenURL
Wang S, Bates J, Li X, Schanz S, Chandler-Militello D, Levine C, et al.: Human iPSC-derived oligodendrocyte progenitor cells can myelinate and rescue a mouse model of congenital hypomyelination.
Cell Stem Cell. 2013, 12:252-264. PubMed Abstract | Publisher Full Text OpenURL
Givogri MI, Galbiati F, Fasano S, Amadio S, Perani L, Superchi D, et al.: Oligodendroglial progenitor cell therapy limits central neurological deficits in mice with metachromatic leukodystrophy.
J Neurosci. 2006, 26:3109-3119. PubMed Abstract | Publisher Full Text OpenURL
Windrem MS, Schanz SJ, Guo M, Tian GF, Washco V, Stanwood N, et al.: Neonatal chimerization with human glial progenitor cells can both remyelinate and rescue the otherwise lethally hypomyelinated shiverer mouse.
Cell Stem Cell. 2008, 2:553-565. PubMed Abstract | Publisher Full Text OpenURL
Pluchino S, Quattrini A, Brambilla E, Gritti A, Salani G, Dina G, et al.: Injection of adult neurospheres induces recovery in a chronic model of multiple sclerosis.
Nature. 2003, 422:688-694. PubMed Abstract | Publisher Full Text OpenURL
Shindo T, Matsumoto Y, Wang Q, Kawai N, Tamiya T, Nagao S: Differences in the neuronal stem cells survival, neuronal differentiation and neurological improvement after transplantation of neural stem cells between mild and severe experimental traumatic brain injury.
J Med Invest. 2006, 53:42-51. PubMed Abstract | Publisher Full Text OpenURL
Clark RSB, Kochanek PM, Watkins SC, Chen MZ, Dixon CE, Seidberg NA, et al.: Caspase-3 mediated neuronal death after traumatic brain injury in rats.
J Neurochem. 2000, 74:740-753. PubMed Abstract | Publisher Full Text OpenURL
Smith FM, Raghupathi R, MacKinnon MA, McIntosh TK, Saatman KE, Meaney DF, et al.: TUNEL-positive staining of surface contusions after fatal head injury in man.
Acta Neuropathol. 2000, 100:537-545. PubMed Abstract | Publisher Full Text OpenURL
Povlishock JT, Katz DI: Update of neuropathology and neurological recovery after traumatic brain injury.
J Head Trauma Rehabilitation. 2005, 20:76-94. Publisher Full Text OpenURL
Singleton RH, Povlishock JT: Identification and characterization of heterogeneous neuronal injury and death in regions of diffuse brain injury: evidence for multiple independent injury phenotypes.
J Neurosci. 2004, 24:3543-3553. PubMed Abstract | Publisher Full Text OpenURL
Greer JE, McGinn MJ, Povlishock JT: Diffuse traumatic axonal injury in the mouse induces atrophy, c-Jun activation, and axonal outgrowth in the axotomized neuronal population.
J Neurosci. 2011, 31:5089-5105. PubMed Abstract | Publisher Full Text OpenURL
Iwanami A, Kaneko S, Nakamura M, Kanemura Y, Mori H, Kobayashi S, et al.: Transplantation of human neural stem cells for spinal cord injury in primates.
J Neurosci Res. 2005, 80:182-190. PubMed Abstract | Publisher Full Text OpenURL
Neirinckx V, Cantinieaux D, Coste C, Rogister B, Franzen R, Wislet-Gendebien S: Concise review: Spinal cord injuries: how could adult mesenchymal and neural crest stem cells take up the challenge?
Stem Cells. 2014, 32:829-843. PubMed Abstract | Publisher Full Text OpenURL
Fujimoto Y, Abematsu M, Falk A, Tsujimura K, Sanosaka T, Juliandi B, et al.: Treatment of a mouse model of spinal cord injury by transplantation of human induced pluripotent stem cell-derived long-term self-renewing neuroepithelial-like stem cells.
Stem Cells. 2012, 30:1163-1173. PubMed Abstract | Publisher Full Text OpenURL
Ogawa Y, Sawamoto K, Miyata T, Miyao S, Watanabe M, Nakamura M, et al.: Transplantation of in vitro-expanded fetal neural progenitor cells results in neurogenesis and functional recovery after spinal cord contusion injury in adult rats.
J Neurosci Res. 2002, 69:925-933. PubMed Abstract | Publisher Full Text OpenURL
Nori S, Okada Y, Yasuda A, Tsuji O, Takahashi Y, Kobayashi Y, et al.: Grafted human-induced pluripotent stem-cell-derived neurospheres promote motor functional recovery after spinal cord injury in mice.
Proc Natl Acad Sci U S A. 2011, 108:16825-16830. PubMed Abstract | Publisher Full Text OpenURL
Sun Y, Xu CC, Li J, Guan XY, Gao L, Ma LX, et al.: Transplantation of oligodendrocyte precursor cells improves locomotion deficits in rats with spinal cord irradiation injury.
PLoS One. 2013., 8
Article ID e57534
PubMed Abstract | Publisher Full Text OpenURL
NCT01772810. Safety study of human spinal cord-derived neural stem cell transplantation for the treatment of chronic SCI. http://clinicaltrials.gov/ct2/show/NCT01772810?term=Neural+Stem+Cell%2C+spinal+cord+injury&rank=1. 2014.
NCT01321333. Study of human central nervous system stem cells (HuCNS-SC) in patients with thoracic spinal cord injury. https://clinicaltrials.gov/ct2/show/NCT01321333?term=NCT01321333&rank=1. 2014.
Povlishock JT: Traumatically induced axonal injury – pathogenesis and pathobiological implications.
Brain Pathol. 1992, 2:1-12. PubMed Abstract | Publisher Full Text OpenURL
Smith DH, Meaney DF, Shull WH: Diffuse axonal injury in head trauma.
J Head Trauma Rehabilitation. 2003, 18:307-316. Publisher Full Text OpenURL
NCT01494480. The clinical trial on the use of umbilical cord mesenchymal stem cells in amyotrophic lateral sclerosis. http://clinicaltrials.gov/ct2/show/NCT01494480?term=NCT01494480&rank=1. 2014.
NCT01217008. Safety study of GRNOPC1 in spinal cord injury. http://clinicaltrials.gov/ct2/show/NCT01217008?term=NCT01217008&rank=1. 2014.
NCT01895439. Safety and efficacy study of autologus bone marrow mesenchymal stem cells in multiple sclerosis. http://clinicaltrials.gov/ct2/show/NCT01895439?term=NCT01895439&rank=1. 2014.
NCT02254863. UCB transplant of inherited metabolic diseases with administration of intrathecal UCB derived oligodendrocyte-like cells (DUOC-01). http://clinicaltrials.gov/ct2/show/NCT02254863?term=NCT02254863&rank=1. 2014.
Brazzini A, Cantella R, De la Cruz A, Yupanqui J, Leon C, Jorquiera T, et al.: Intraarterial autologous implantation of adult stem cells for patients with Parkinson disease.
J Vasc Interv Radiol. 2010, 21:443-451. PubMed Abstract | Publisher Full Text OpenURL
Hess DC, Sila CA, Furlan AJ, Wechsler LR, Switzer JA, Mays RW: A double-blind placebo-controlled clinical evaluation of MultiStem for the treatment of ischemic stroke.
Int J Stroke. 2014, 9:381-386. PubMed Abstract | Publisher Full Text OpenURL
Riley J, Glass J, Feldman EL, Polak M, Bordeau J, Federici T, Johe K, Boulis NM: Intraspinal stem cell transplantation in als: a phase i trial, cervical microinjection and final surgical safety outcomes.
Neurosurgery. 2014, 74:77-87. PubMed Abstract | Publisher Full Text OpenURL

close

Regulatory functions and pathological relevance of the MECP2 3′UTR in the central nervous system

Heather McGowan and Zhiping P. Pang

Corresponding authors: Heather McGowan mcgowahe@rwjms.rutgers.edu – Zhiping P Pang zhiping.pang@rutgers.edu

Author Affiliations
Department of Neuroscience and Cell Biology, Child Health Institute of New Jersey, Rutgers University Robert Wood Johnson Medical School, 89 French Street, Room 3277, New Brunswick 08901, NJ, USA

Cell Regeneration 2015, 4:9 doi:10.1186/s13619-015-0023-x

The electronic version of this article is the complete one and can be found online at: http://www.cellregenerationjournal.com/content/4/1/9

Received: 1 May 2015
Accepted: 18 September 2015
Published: 28 October 2015
© 2015 McGowan and Pang.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Abstract
Methyl-CpG-binding protein 2 (MeCP2), encoded by the gene MECP2, is a transcriptional regulator and chromatin-remodeling protein, which is ubiquitously expressed and plays an essential role in the development and maintenance of the central nervous system (CNS). Highly enriched in post-migratory neurons, MeCP2 is needed for neuronal maturation, including dendritic arborization and the development of synapses. Loss-of-function mutations in MECP2 cause Rett syndrome (RTT), a debilitating neurodevelopmental disorder characterized by a phase of normal development, followed by the progressive loss of milestones and cognitive disability. While a great deal has been discovered about the structure, function, and regulation of MeCP2 in the time since its discovery as the genetic cause of RTT, including its involvement in a number of RTT-related syndromes that have come to be known as MeCP2-spectrum disorders, much about this multifunctional protein remains enigmatic. One unequivocal fact that has become apparent is the importance of maintaining MeCP2 protein levels within a narrow range, the limits of which may depend upon the cell type and developmental time point. As such, MeCP2 is amenable to complex, multifactorial regulation. Here, we summarize the role of the MECP2 3′ untranslated region (UTR) in the regulation of MeCP2 protein levels and how mutations in this region contribute to autism and other non-RTT neuropsychiatric disorders.

Keywords: Methyl-CpG-binding protein 2; 3′ untranslated region; Autism; Rett syndrome
Introduction
In 1999, Huda Zoghbi and her colleagues discovered that MECP2, which codes for methyl-CpG-binding protein 2 (MeCP2), is the gene that is mutated in Rett syndrome (RTT) [1]. It is now known from a decade and a half of study that MeCP2 is a multifunctional protein that plays a complex, yet essential role, in the development and maintenance of the central nervous system (CNS). The diversity of MeCP2 function includes, but may not be limited to: transcription regulation [2]–[4], chromatin-remodeling and histone modification [5]–[7], and regulation of messenger RNA (mRNA)-splicing [8], [9] and microRNA (miRNA)-processing [10]. These molecular functions manifest themselves on a cellular level in ways that are not completely understood but ultimately result in proper neural cell differentiation [11], neuronal maturation [12]–[14], dendritic arborization and spine formation [15]–[17], and adequate production of synaptic proteins and receptors [18], [19]. Moreover, MeCP2 exerts both cell autonomous and non-cell autonomous effects on neurons [20]. Once thought to be exclusive to neurons in the CNS, we now also know that glial cells, including astrocytes and oligodendrocytes, express MeCP2 and require it to adequately support the morphological and functional development of neurons. In turn, glial cells have been implicated as key players in the pathophysiology of RTT and MeCP2-related disorders [21]–[25].

In most cases, loss of MeCP2 function in females results in classic RTT [26]. RTT is an X-linked neurodevelopmental disorder that is characterized by 6–18 months of normal development, followed by a stagnation and eventual regression of developmental milestones. Affected individuals exhibit a myriad of characteristic signs and debilitating symptoms, including microcephaly, intellectual disability, autistic features, overall growth retardation and weight loss, hypotonia, loss of motor coordination, autonomic dysfunction, breathing irregularities and apneas, and replacement of purposeful hand movements with stereotypies such as wringing, clapping, or flapping [27], [28]. In males, RTT-causing mutations most often result in severe neonatal encephalopathy [29], [30]; however, in rare cases, these same mutations can cause classic RTT in males with Klinefelter syndrome (47, XXY) or somatic mosaicism [31], [32]. Loss-of-function mutations that do not cause RTT produce a host of neuropsychiatric abnormalities in both males (e.g., mental retardation, bipolar disorder, schizophrenia, PPM-X syndrome) [30], [33]–[36] and females (e.g., atypical RTT, mental retardation, Angelman-like syndrome, autism) [37]–[41]. In addition, increases in MeCP2 dosage also lead to profound dysfunction. Duplications of the gene locus results in MeCP2 duplication syndrome, which is a progressive neurodevelopmental disorder with RTT-like features that occurs most often in males [42], [43].

While it is intriguing that dramatic losses and gains in the MECP2 gene dosage both result in a similar, debilitating phenotype, it has also been demonstrated that a hypomorphic MECP2 allele that expresses 50 % of the wild type gene level also produces an RTT-like syndrome in mice [44], thus demonstrating a need for MeCP2 protein levels to be maintained within a narrow margin to ensure proper neurological development. Accordingly, there is also evidence that MeCP2 levels are reduced in other neurodevelopmental disorders, including autism, trisomy 21, fragile X syndrome, Angelman syndrome, and Prader-Willi syndrome [37], [45]. Therefore, it is not surprising that MECP2 expression is subject to intense, complex regulation at virtually every level from DNA to protein. While the overall expression patterns for MeCP2 in the developing nervous system has been elucidated, there is not always a clear correlation between mRNA and protein expression, and the various and complex mechanisms governing the transcript-protein balance are not completely understood [12], [13], [37], [46]–[50]. The clinical relevance for elucidating these processes is particularly crucial, not only for providing a more complete understanding of the heterogeneity in the clinical presentation and treatment potential of RTT and related MeCP2-spectrum disorders, but also for those above-mentioned developmental disorders in which MeCP2 expression is dysregulated in the absence of an obvious mutation. In this review, we will summarize the current understanding of the spatio-temporal expression patterns of MeCP2 throughout brain development, as well as the regulatory mechanisms that control the specificity of these patterns, focusing on the regulatory potential of the highly conserved 3′ untranslated region (3′UTR). We will seek to establish the 3′UTR as an important potential contributor in establishing and maintaining homeostatic levels of MeCP2 expression in a developmentally appropriate manner. Finally, we will discuss the clinical relevance of these regulatory mechanisms as they are currently understood and their implications for potential therapeutic strategies.

MECP2 expression
During development, MeCP2 protein is expressed at low levels throughout the brain prenatally but progressively increases with neurogenesis and reaches its peak in mature, post-migratory neurons [12], [13]. The timing of MeCP2 expression varies by brain region and cell type and correlates closely with the maturation of the given cell type and matches the overall ontogenetic maturation pattern of the CNS [49], i.e., older structures, such as the spinal cord and brain stem, become MeCP2 positive before newer structures, such as the cerebral cortex, hippocampus, and cerebellum [50]. This temporal pattern holds within individual brain regions—e.g., in the cortex, the timing of MeCP2 expression follows the inside-out lamination sequence, with Cajal-Retzius cells becoming positive first, followed by early-born deep-layer neurons, and finally later-born superficial-layer neurons [12]. This suggests that the MeCP2 protein only becomes detectable once the individual neuron reaches a certain point of maturity and is consistent with the hypothesis that MeCP2 plays an important role in maintaining neuronal maturation. Additionally, in the cortex, hippocampus, and the granule cells of the cerebellum, MeCP2 expression is most closely correlated with synaptogenesis, consistent with its proposed role in the formation and maintenance of synapses [50].

In addition to the spatio-temporal dependence of the MECP2 expression, the brain contains a heterogeneous cell population, including both low MeCP2-expressing cells (MeCP2 lo ) and high MeCP2-expressing cells (MeCP2 hi ), displaying a defined distribution pattern. MeCP2 lo cells are present in the highest proportion in the granular layer of the cerebellum, whereas layer IV of the cerebral cortex and the molecular layer of the cerebellum exhibit a higher proportion of MeCP2 hi cells. The remaining layers of both regions contain roughly equal proportions of each cell type. The proportion of MeCP2-positive neurons increases throughout postnatal life, and it correlates with the percentage of MeCP2 hi cells, indicating a possibly increasing need for MeCP2 function as the nervous system matures [46].

Alternative splicing generates two MeCP2 isoforms, MeCP2_e1 and MeCP2_e2, which only differ at their N-termini. Isoform E1 is 498 amino acids long and is translated from exons 1, 3, and 4, whereas isoform E2 is 486 amino acids and translated from exons 2, 3, and 4 [51], [52] (Fig. 1a). These isoforms are also spatio-temporally regulated. The e1 isoform is much more abundant in the brain and also demonstrates more widespread expression throughout development. Early in postnatal life, the e2 transcript is widely expressed but with time becomes largely restricted to the dorsal thalamus and cortical layer V [53]. Consistent with this, a more recent report utilized an MeCP2E1-specific antibody to demonstrate that the E1 protein is also widely expressed, with its highest expression in the cortex and cerebellum, and it is also expressed in high amounts in neurons as compared to astrocytes [54].

thumbnailFig. 1. Regulation via the MECP2 3’UTR. a Schematic depicting the cis-acting regulatory elements in the genomic sequence of the MECP2 3’UTR. Open boxes indicate the silencer and enhancer described by Liu and Francke [59]. Closed, colored boxes indicate auxiliary elements involved in polyadenylation, as described by Newnham et al. [58]. E exon, GRS G-rich element, USE upstream sequence element. b Four unique MECP2 transcripts produced by alternative polyadenylation. c Relative position of putative miRNA binding sites in the 3’UTR of the 10.2-kb species of MECP2 mRNA, as defined by Target Scan. This representation is limited to predicted sites for miRNAs that are broadly conserved in vertebrates (with the exception of the human-specific site for miR-483-5p)
It is evident that very intricate regulation is required to achieve the temporal, regional, and cell-population specificity of these expression patterns, and we will now turn our attention to the evidence implicating the 3′UTR as an important contributor to the fine-tuning of homeostatic MeCP2 expression levels throughout development.

3′ untranslated region (3′UTR) of MECP2
Structure and conservation
The MECP2 gene contains a remarkably large, highly conserved 3′UTR. Coy et al. [55] screened cosmid clones with radioactively labeled cDNA clones they had isolated from human tissue samples and identified a long contig with no open reading frame or introns. They mapped the contig to the 3′UTR of MECP2, and further investigation revealed that alternative polyadenylation (poly A) signals could originate transcripts of multiple lengths. Reichwald et al. [56] similarly found by comparative sequence analysis the existence of this additional stretch of sequence 3′ to the previously reported poly(A) site, which terminated in an alternative poly(A) site. In addition, Coy et al. [55] compared the entire human 3′UTR sequence with the mouse sequence and found that homology of this region of the MECP2 gene (~52 %) is lower than the average for 3′UTRs. However, they discovered blocks of sequence within the 3′UTR that were highly conserved, not only in mouse but also kangaroo, rat, hamster, macaque, and chimpanzee sequences. These regions were also predicted to have high minimum free energy, suggesting a weak secondary structure. Conversely, the regions of the 3′UTR that were not highly similar in primary sequence were predicted to have low minimum free energy, suggesting tight folding. Indeed, the free-energy distribution over the entire sequence was highly similar between human and mouse, suggesting that the overall secondary structure of the 3′UTR was conserved throughout evolution regardless of divergent primary structures between different species and therefore may confer important regulatory functions. Likewise, the loose conformation of the highly conserved regions of sequence suggests that these may represent important binding sites for transregulatory elements. Taken together, such evolutionary analysis emphasizes the importance of the 3′UTR on a very fundamental level.

Alternative polyadenylation of MECP2
Alternative poly(A) sites within the 3′UTR can be used to generate four mRNA transcripts of varying length: ~1.8, ~5.4, ~7.5, and ~10.2 kb (Fig. 1b) [57]. Studies have revealed that the distribution and abundance of each transcript varies by tissue and is developmentally regulated. The 10.2-kb transcript is the most abundant in the brain, and in sharp contrast to MeCP2 protein, the expression of this transcript is highly enriched in the embryonic brain but then progressively declines during postnatal development [37], [47], only to be upregulated again in the adult brain [37]. It has been suggested that this decrease in the use of the long 3′UTR transcript accounts for the corresponding increase in MeCP2 protein levels via possible differential regulation conferred by the individual transcripts. However, Samaco et al. [37] found that, while total MECP2 transcript expression levels across neuronal populations decreased from the fetal to the postnatal stage, the increase in MeCP2 protein expression, along with an increase in the percentage of MeCP2 hi cells with age, correlates with an increase in the expression of both the total and the long variant MECP2 transcripts within the MeCP2 hi population. Additionally, they noted a significant decrease in expression of the long MECP2 transcript within the MeCP2 lo population in postnatal brains versus fetal brain and of the ratio of long transcript to total transcripts, which later increased in adult brains. However, the study found no significant correlation between transcript length and MeCP2 protein levels on a single cell level. This highlights how the alternative polyadenylation of the MECP2 3′UTR serves as a dynamic source of brain region and even cell-type-specific regulation of MeCP2 expression. While we do not currently fully understand how each individual transcript relates to protein expression, and how this may change with spatio-temporal context, the heterogeneity of the levels of transcript type among cell populations suggests that the 3′UTR may be important for fine-tuning MeCP2 protein expression to meet the homeostatic needs of individual microenvironments in the brain.

Alternative poly(A) is determined both by the presence of pairs of cis-acting core sequences that specify the site of poly(A) as well as auxiliary regulatory elements that can facilitate or repress poly(A) at a designated site. Newnham et al. [58] discovered such cis regulatory elements both upstream and downstream of the binding sites for cleavage and polyadenylation specificity factor (CPSF) in the MECP2 3′UTR. They discovered a G-rich element (GRS) downstream of the most proximal poly(A) core sequence (Fig. 1a), the mutation of which resulted in significantly reduced efficiency of polyadenylation. They also showed that this site is specifically bound by hnRNP F, a protein involved in the 3′ end formation, indicating it likely plays an important regulatory role in the production of alternative MECP2 transcripts. They also found an element upstream of the most distal poly(A) signal, which was very similar to upstream sequence elements (USEs; Fig. 1a) found in human collagen genes and human COX-2. Mutation at this site also reduced polyadenylation efficiency, albeit to a lesser extent. Interestingly, the DNA sequence of the 3′UTR also harbors enhancer and repression elements that act directly on the MECP2 core promoter (Fig. 1a). These regulatory elements were shown by gel shift assays to bind nuclear proteins, presumably transcription factors [59]. Additional evidence is needed in order to tease out the mechanisms by which these cis elements regulate the transcription and post-transcriptional modification of MECP2. Insight into how alternative polyadenylation of MECP2 is regulated, for example, may shed light on the circumstances under which one transcript is required over another in order to meet the homeostatic needs of the cell.

The complicated nature of the relationship between the length of the MECP2 3′UTR and the expression of the MeCP2 protein is likely due to the complex and multifactorial impact of 3′UTR on gene expression. The 3′UTR can play a role in translation efficiency, localization, and the folding and stability of the mRNA [60]. As such, alternative poly(A) offers the ability of the cell to “customize” a transcript to meet its needs. That is, cleaving the transcript at varying poly(A) sites varies the regulatory elements at the level of both the primary and secondary structure, which in turn will affect the stability, localization, translatability, etc. of the mRNA. As such, mutations in the 3′UTR certainly have the potential to affect the stability of MECP2 transcripts, and indeed, autistic patients carrying non-RTT-causing mutations in conserved sequences of the MECP2 3′UTR displayed reduced levels of MECP2 mRNA compared to controls [61]. Given the fine balance of the MeCP2 expression needed for brain development, this is potentially of high clinical significance.

microRNAs (miRNAs) and post-transcriptional regulation of MeCP2
One feature of 3′UTRs that is often of particular interest is the presence of targeting sequences for trans-acting regulatory elements. Among these regulatory elements are miRNAs, which are short, non-coding RNA molecules that are involved in post-transcriptional gene regulation. miRNAs function by base-pairing with a complementary sequence on target messenger RNA (mRNA). This base-pairing is most often imperfect, with exact matching occurring only between nucleotides 2 and 8 (known as the “seed region”) of the miRNA and a complementary sequence in the 3′UTR of the target mRNA [62]–[64]. This type of partial-binding of miRNA to its target usually results in either translational repression [65], deadenylation [66], or, more rarely, cleavage of the mRNA [67]. miRNAs have also been shown to establish mRNA threshold levels, below which protein translation is highly reduced, thus allowing for a fine-tuning of gene expression levels in addition to overt gene-silencing [68]. miRNAs regulate gene expression in the developing nervous system [69], with roles in regulating neurogenesis [70], [71], neuronal maturation, spinogenesis, dendritic arborization [72], synaptogenesis [73], and neuronal survival [74]. This has profound implications for how miRNA misexpression may affect cognitive capability. For example, Hansen et al. [75] showed that moderate increases in miR-132 in the hippocampus enhanced cognitive capacity, while supra-physiological expression resulted in impaired cognition and an increase in dendritic spines, implying that the decreased capacity for learning and memory resulted from alterations in the structure of synaptic connections.

In line with this, MECP2 is regulated by several miRNAs (Table 1), and its 3′UTR contains putative target sequences for many more (Fig. 1c); however, the role of miRNAs in regulating MECP2 expression is diverse and complex. These promiscuous molecules serve as important potential modifiers for local, tissue-, or possibly even cell-type-specific fine-tuning of MeCP2 protein levels, either developmentally or in response to cellular activity. In the nervous system, miRNA-132 has been demonstrated to regulate MeCP2 expression in a homeostatic feedback loop with brain-derived neurotrophic factor (BDNF) [76]. BDNF is a known target of MeCP2, which interacts directly with its promoter to enhance its expression [77]. Klein et al. [76] demonstrated that activation of the cAMP response element-binding protein (CREB) pathway in cortical neurons stimulated an increase in miR-132 expression and subsequent reduction in MeCP2 expression. Blocking miR-132 increased the expression of both MeCP2 and BDNF, while siRNA-mediated knockdown of MeCP2 reduced both BDNF and miR-132 levels, suggesting a potential CREB-mediated regulatory feedback loop (Fig. 2). miR-212, which is closely related to and genetically arranged in tandem with miR-132, and which has been shown to reduce MeCP2 levels in gastric carcinoma cell lines [78], participates in a similar negative feedback loop with MeCP2 in the dorsal striatum [79]. This miR-212-MeCP2 relationship was also shown to have regulatory implications for BDNF, specifically under the conditions of extended cocaine exposure [79]. Chen and colleagues [80] also demonstrated that MeCP2 is a functional target of miR-7b, which is in turn targeted and silenced by MeCP2 as postnatal neurons mature, suggesting another homeostatic regulatory loop. In addition, Han et al. [81] identified a novel, human-specific targeting site for miR-483-5p in the long MECP2 3′UTR. They postulated a potential role for this miRNA in fetal development by demonstrating an inverse correlation between elevated miR-483-5p levels and decreased MeCP2 expression levels in human fetal brains. Furthermore, regulation of MeCP2 by miR-124a in the spinal cord may modulate nociception. Kynast et al. [82] demonstrated that peripheral noxious stimulation in mice led to a decrease in miR-124a expression in neurons of the dorsal horn, accompanied by an increase in MeCP2 and pro-inflammatory genes, as well as nociceptive behavior.

Table 1. miRNAs that are known to target the MECP2 3’UTR and the reported biological significance for each
thumbnailFig. 2. MeCP2 participates in homeostatic feedback loops involving regulation by miRNAs. An example of several feedback mechanisms involving BDNF is depicted here
The relevance of studying the role of miRNAs in regulating MECP2 is highlighted further by evidence that miRNAs modulate MECP2 expression in physiological processes other than neurogenesis. For example, targeting of MECP2 by miR-22 has been shown to promote smooth muscle cell differentiation [83] and to reduce apoptosis in ischemic cardiomyocytes [84]. Future studies investigating brain-region-specific miRNAs and their potential to target MECP2 may provide additional insight into the enigmatic relationship between transcript variants and MeCP2 protein expression in different parts of the brain.

Clinical implications of the MECP2 3′UTR beyond RTT
In addition to those causing RTT and related disorders, MECP2 mutations have been described in other neurodevelopmental disorders, such as autism, and X-linked mental retardation [33], [34], [61], [85], [86]. Some of these mutations occur in the 3′UTR rather than in the coding sequence, in accordance with a need for finely tuned regulation of MeCP2 protein levels for proper neurological function. Coutinho et al. [61] found 21 variations in the MECP2 3′UTR of 46 out of 172 autistic patients. Of these variations, 12 did not occur in controls. They also found that MECP2 mRNA levels in peripheral blood mononuclear cells of four patients with variations in conserved sequences were significantly lower than that of controls, suggesting that changes in at least these conserved sequences may alter mRNA stability and, thus, protein expression levels. Shibayama et al. [86] also found two 3′UTR variations in autistic patients, as well as one variation in a patient with ADHD. While these alterations seem to be most common to autism [61], [86]–[88], they have also been found in patients with spontaneous intellectual disability [88]–[90], as well as individuals with atypical RTT, or classic RTT with no detectable pathological mutation in the MECP2 coding sequence [88], [91]. Mutations in the 3′UTR of MECP2 were also found in a rare number of patients with RTT in a study by Santos and colleagues [88]; however, none of the five variants they discovered are found in the putative protein-targeting sites of the UTR; additionally, two had been previously described as polymorphisms, one was present in an unaffected father and another was present in an unaffected mother, suggesting that they are not pathogenic, particularly in the case of the father. A summary of these MECP2 3′UTR variants and their pathological relevance can be found in Table 2. Taken together, these findings suggest that mutations in the 3′UTR, which would purportedly impact the expression of MeCP2 rather than its function, may not cause full-blown RTT, except in rare cases, but could still impair neurological function given a context allowing for MeCP2 dysregulation.

Table 2. Sequence variations in the MECP2 3’UTR that have been reported in non-RTT neurological disorders, atypical RTT, or RTT without a detectable pathogenic coding region mutation
In line with this, Hanchard et al. [92] reported an adult male harboring a partial MECP2 duplication, who was high functioning and able to live independently despite suffering from epilepsy and cognitive impairment. The duplication included all four exons but excluded almost the entire 3′UTR. Given the importance of the 3′UTR for the stability and activity of the MeCP2 protein, the loss of the 3′UTR in the duplicated segment of the MECP2 gene more likely mitigates the MeCP2 overexpression and, by extension, the severity of symptoms. Additionally, Samaco et al. [37] demonstrated differences in MeCP2 expression levels in autism, pervasive developmental disorder, Prader-Willi syndrome, and Angelman syndrome; by laser-scanning cytometry, the disturbances in MeCP2 expression were determined as due to differential transcriptional and post-transcriptional mechanisms.

miRNA-targeting of the 3′UTR of MECP2 has also been implicated in neurological dysfunction. Human-specific miR-483-5p, which has been shown to decrease MeCP2 expression, is transcribed from the second intron of the genetically imprinted gene insulin-like growth factor 2 (IGF2). IGF2 expression occurs almost exclusively from the paternal allele, and imprinting defects that lead to expression from the maternal allele cause Beckwith-Wiedemann syndrome (BWS). The study demonstrates that BWS patients with bi-allelic expression of IGF2 also overexpress miR-483-5p and underexpress MeCP2. This finding may have implications for the etiology of higher prevalence of autism in these patients [81]. Additionally, a recent study by Tantra et al. showed that a 50 % overexpression of MeCP2 influences aggression levels in opposing ways in two different strains of mice. To test this interaction between genetic background and expression level in humans, they demonstrated that miR-511, which is expressed in the brain, binds selectively to MECP2 mRNA transcripts that carry a C > T SNP in the 3′UTR. As a result, T carriers have a ~50 % reduction in peripheral MeCP2 expression. The C allele at this locus is associated with increased aggression in schizophrenia, thus implicating the interaction between altered MeCP2 expression and genetic background as a potential mechanism [93]. Increased expression of MeCP2 resulting from reduction of miR-132-mediated repression has also been implicated in the neuroprotective response to pending ischemic injury [94].

These examples highlight how the 3′UTR may serve as a potential source for pathogenic misregulation of MeCP2, and as such, it may be worthwhile to conduct additional studies investigating the potential pathogenicity of mutations and polymorphisms in this critical region.

Conclusions
In summary, regulation of MeCP2 expression is complex, multifactorial, and crucial for the proper maintenance and function of the CNS. In this review, we have focused our attention on the role of the MECP2 3′UTR in this process. Evidence suggests that the 3′UTR confers multiple levels of regulation on MeCP2 expression that are significant for neurological function. This is consistent with the role of the 3′UTR as seen in other important neural proteins. For example, poly(A) produces two BDNF transcripts, with a long or short 3′UTR. The long 3′UTR represses translation at rest, while the short transcript is actively translated. However, upon neuronal activation, the long transcript, but not the short, undergoes rapid translational activation [95]. Likewise, BDNF is also regulated by miRNAs via its 3′UTR [96].

Post-transcriptional regulation by the 3′UTR offers the advantage of rapid and precise homeostatic control over protein levels in response to a cell’s individual needs. However, the complexity of this single avenue of gene regulation highlights the need for intense study, especially given the fact that a vast array of additional mechanisms exist that manage the expression and function of MeCP2 at every level. For example, the MECP2 core promoter is located within a CpG island [56], and gene expression levels have been shown to inversely correlate with methylation. In fact, hypermethylation of the MECP2 core promoter has been observed in autistic patients [97]. Additionally, Liu and Francke [59] identified four enhancers and two silencers within the gene; these enhancers contain predicted binding sites for brain-specific transcription factors, and three of them were able to act in cis with the core promoter. MeCP2 also exists in two isoforms, the alternate expression of which appears to be dependent upon differential methylation of regulatory elements within the MECP2 promoter, which can be manipulated with decitabine [96]. In addition to tight control over expression, mechanisms such as post-translational modification also contribute to the regulation of MeCP2 activity. For example, activity-dependent phosphorylation of threonine 308 blocks the interaction of MeCP2’s repressor domain with NCoR, which attenuates its transcriptional repression [98]. The advances in understanding the complex and interactive nature of MECP2 and its regulatory elements are pivotal to unraveling the mechanisms underpinning the development and progression of complex and poorly understood neurodevelopmental disorders and to devise novel therapeutic approaches. miRNAs in particular are attractive therapeutic targets, as there is evidence that their expression can be modified by small molecules [99]. Likewise, methods are being developed to deliver miRNA mimics and/or inhibitors directly into the CNS (e.g., lipid-, polyethylenimine-, and dendrimer-based methods) [100].

Finally, the bulk of the literature exploring the expression, regulation, and function of MeCP2 has focused almost exclusively on neurons, as early studies were not able to show any expression in glial cells. Recently, increasing attention has been pointed at the role played by glial cells in conditioning the morphological and functional development of neurons, and several studies are deciphering the specific roles of glia in neurodevelopment, as well as in RTT and related disorders. However, the relative paucity of information regarding how MeCP2 is regulated in non-neuronal brain cells is a gap that needs to be filled by further investigations.

Abbreviations
MeCP2: methyl-CpG-binding protein 2

CNS: Central Nervous System

RTT: Rett Syndrome

3’UTR: 3′ untranslated region

mRNA: messenger RNA

miRNA: microRNA

poly(A): Polyadenylation

CPSF: Cleavage and polyadenylation specificity factor

GRS: G-rich element

USE: Upstream sequence element

BDNF: Brain derived neurotrophic factor

CREB: cAMP response element-binding protein

IGF2: Insulin-like growth factor 2

BWS: Beckwith-Wiedemann syndrome

Competing interests
None to report.

Authors’ contributions
HM researched and wrote the manuscript. ZPP edited and critically evaluated the manuscript. Both authors read and approved the final manuscript.

Acknowledgements
We want to thank the support from NIH-NINDS F31NS084551 NRSA predoctoral fellowship and the Jérôme LeJeune Foundation.

References
Amir RE, Van den Veyver IB, Wan M, Tran CQ, Francke U, Zoghbi HY: Rett syndrome is caused by mutations in X-linked MECP2, encoding methyl-CpG-binding protein 2.
Nat Genet 1999, 23(2):185-8. PubMed Abstract | Publisher Full Text OpenURL
Nan X, Campoy FJ, Bird A: MeCP2 is a transcriptional repressor with abundant binding sites in genomic chromatin.
Cell 1997, 88(4):471-81. PubMed Abstract | Publisher Full Text OpenURL
Chahrour M, Jung SY, Shaw C, Zhou X, Wong ST, Qin J, et al.: MeCP2, a key contributor to neurological disease, activates and represses transcription.
Science 2008, 320(5880):1224-9. PubMed Abstract | Publisher Full Text OpenURL
Li Y, Wang H, Muffat J, Cheng AW, Orlando DA, Loven J, et al.: Global transcriptional and translational repression in human-embryonic-stem-cell-derived Rett syndrome neurons.
Cell Stem Cell 2013, 13(4):446-58. PubMed Abstract | Publisher Full Text OpenURL
Jones PL, Veenstra GJ, Wade PA, Vermaak D, Kass SU, Landsberger N, et al.: Methylated DNA and MeCP2 recruit histone deacetylase to repress transcription.
Nat Genet 1998, 19(2):187-91. PubMed Abstract | Publisher Full Text OpenURL
Nan X, Ng HH, Johnson CA, Laherty CD, Turner BM, Eisenman RN, et al.: Transcriptional repression by the methyl-CpG-binding protein MeCP2 involves a histone deacetylase complex.
Nature 1998, 393(6683):386-9. PubMed Abstract | Publisher Full Text OpenURL
Nikitina T, Ghosh RP, Horowitz-Scherer RA, Hansen JC, Grigoryev SA, Woodcock CL: MeCP2-chromatin interactions include the formation of chromatosome-like structures and are altered in mutations causing Rett syndrome.
J Biol Chem 2007, 282(38):28237-45. PubMed Abstract | Publisher Full Text OpenURL
Young JI, Hong EP, Castle JC, Crespo-Barreto J, Bowman AB, Rose MF, et al.: Regulation of RNA splicing by the methylation-dependent transcriptional repressor methyl-CpG binding protein 2.
Proc Natl Acad Sci U S A 2005, 102(49):17551-8. PubMed Abstract | Publisher Full Text OpenURL
Maunakea AK, Chepelev I, Cui K, Zhao K: Intragenic DNA methylation modulates alternative splicing by recruiting MeCP2 to promote exon recognition.
Cell Res 2013, 23(11):1256-69. PubMed Abstract | Publisher Full Text OpenURL
Cheng TL, Wang Z, Liao Q, Zhu Y, Zhou WH, Xu W, et al.: MeCP2 suppresses nuclear microRNA processing and dendritic growth by regulating the DGCR8/Drosha complex.
Dev Cell 2014, 28(5):547-60. PubMed Abstract | Publisher Full Text OpenURL
Gao H, Bu Y, Wu Q, Wang X, Chang N, Lei L, et al.: Mecp2 regulates neural cell differentiation by suppressing the Id1 to Her2 axis in zebrafish.
J Cell Sci 2015, 128(12):2340-50. PubMed Abstract | Publisher Full Text OpenURL
Shahbazian MD, Antalffy B, Armstrong DL, Zoghbi HY: Insight into Rett syndrome: MeCP2 levels display tissue- and cell-specific differences and correlate with neuronal maturation.
Hum Mol Genet 2002, 11(2):115-24. PubMed Abstract | Publisher Full Text OpenURL
Kishi N, Macklis JD: MECP2 is progressively expressed in post-migratory neurons and is involved in neuronal maturation rather than cell fate decisions.
Mol Cell Neurosci 2004, 27(3):306-21. PubMed Abstract | Publisher Full Text OpenURL
Matarazzo V, Cohen D, Palmer AM, Simpson PJ, Khokhar B, Pan SJ, et al.: The transcriptional repressor Mecp2 regulates terminal neuronal differentiation.
Mol Cell Neurosci 2004, 27(1):44-58. PubMed Abstract | Publisher Full Text OpenURL
Chapleau CA, Calfa GD, Lane MC, Albertson AJ, Larimore JL, Kudo S, et al.: Dendritic spine pathologies in hippocampal pyramidal neurons from Rett syndrome brain and after expression of Rett-associated MECP2 mutations.
Neurobiol Dis 2009, 35(2):219-33. PubMed Abstract | Publisher Full Text OpenURL
Landi S, Putignano E, Boggio EM, Giustetto M, Pizzorusso T, Ratto GM: The short-time structural plasticity of dendritic spines is altered in a model of Rett syndrome.
Sci Rep 2011, 1:45. PubMed Abstract | Publisher Full Text OpenURL
Stuss DP, Boyd JD, Levin DB, Delaney KR: MeCP2 mutation results in compartment-specific reductions in dendritic branching and spine density in layer 5 motor cortical neurons of YFP-H mice.
PLoS One 2012., 7(3)
Article ID e31896
PubMed Abstract | Publisher Full Text OpenURL
Blue ME, Naidu S, Johnston MV: Development of amino acid receptors in frontal cortex from girls with Rett syndrome.
Ann Neurol 1999, 45(4):541-5. PubMed Abstract | Publisher Full Text OpenURL
Colantuoni C, Jeon OH, Hyder K, Chenchik A, Khimani AH, Narayanan V, et al.: Gene expression profiling in postmortem Rett Syndrome brain: differential gene expression and patient classification.
Neurobiol Dis 2001, 8(5):847-65. PubMed Abstract | Publisher Full Text OpenURL
Kishi N, Macklis JD: MeCP2 functions largely cell-autonomously, but also non-cell-autonomously, in neuronal maturation and dendritic arborization of cortical pyramidal neurons.
Exp Neurol 2010, 222(1):51-8. PubMed Abstract | Publisher Full Text OpenURL
Ballas N, Lioy DT, Grunseich C, Mandel G: Non-cell autonomous influence of MeCP2-deficient glia on neuronal dendritic morphology.
Nat Neurosci 2009, 12(3):311-7. PubMed Abstract | Publisher Full Text OpenURL
Maezawa I, Swanberg S, Harvey D, LaSalle JM, Jin LW: Rett syndrome astrocytes are abnormal and spread MeCP2 deficiency through gap junctions.
J Neurosci 2009, 29(16):5051-61. PubMed Abstract | Publisher Full Text OpenURL
Lioy DT, Garg SK, Monaghan CE, Raber J, Foust KD, Kaspar BK, et al.: A role for glia in the progression of Rett’s syndrome.
Nature 2011, 475(7357):497-500. PubMed Abstract | Publisher Full Text OpenURL
Derecki NC, Cronk JC, Lu Z, Xu E, Abbott SB, Guyenet PG, et al.: Wild-type microglia arrest pathology in a mouse model of Rett syndrome.
Nature 2012, 484(7392):105-9. PubMed Abstract | Publisher Full Text OpenURL
Nguyen MV, Felice CA, Du F, Covey MV, Robinson JK, Mandel G, et al.: Oligodendrocyte lineage cells contribute unique features to Rett syndrome neuropathology.
J Neurosci 2013, 33(48):18764-74. PubMed Abstract | Publisher Full Text OpenURL
Wan M, Lee SS, Zhang X, Houwink-Manville I, Song HR, Amir RE, et al.: Rett syndrome and beyond: recurrent spontaneous and familial MECP2 mutations at CpG hotspots.
Am J Hum Genet 1999, 65(6):1520-9. PubMed Abstract | Publisher Full Text OpenURL
Hagberg B, Aicardi J, Dias K, Ramos O: A progressive syndrome of autism, dementia, ataxia, and loss of purposeful hand use in girls: Rett’s syndrome: report of 35 cases.
Ann Neurol 1983, 14(4):471-9. PubMed Abstract | Publisher Full Text OpenURL
Nomura Y: Early behavior characteristics and sleep disturbance in Rett syndrome.
Brain Dev 2005, 27(Suppl 1):S35-42. PubMed Abstract | Publisher Full Text OpenURL
Imessaoudene B, Bonnefont JP, Royer G, Cormier-Daire V, Lyonnet S, Lyon G, et al.: MECP2 mutation in non-fatal, non-progressive encephalopathy in a male.
J Med Genet 2001, 38(3):171-4. PubMed Abstract | Publisher Full Text OpenURL
Zeev BB, Yaron Y, Schanen NC, Wolf H, Brandt N, Ginot N, et al.: Rett syndrome: clinical manifestations in males with MECP2 mutations.
J Child Neurol 2002, 17(1):20-4. PubMed Abstract | Publisher Full Text OpenURL
Budden SS, Dorsey HC, Steiner RD: Clinical profile of a male with Rett syndrome.
Brain Dev 2005, 27(Suppl 1):S69-71. PubMed Abstract | Publisher Full Text OpenURL
Maiwald R, Bonte A, Jung H, Bitter P, Storm Z, Laccone F, et al.: De novo MECP2 mutation in a 46, XX male patient with Rett syndrome.
Neurogenetics 2002, 4(2):107-8. PubMed Abstract | Publisher Full Text OpenURL
Meloni I, Bruttini M, Longo I, Mari F, Rizzolio F, D’Adamo P, et al.: A mutation in the Rett syndrome gene, MECP2, causes X-linked mental retardation and progressive spasticity in males.
Am J Hum Genet 2000, 67(4):982-5. PubMed Abstract | Publisher Full Text OpenURL
Orrico A, Lam C, Galli L, Dotti MT, Hayek G, Tong SF, et al.: MECP2 mutation in male patients with non-specific X-linked mental retardation.
FEBS Lett 2000, 481(3):285-8. PubMed Abstract | Publisher Full Text OpenURL
Cohen D, Lazar G, Couvert P, Desportes V, Lippe D, Mazet P, et al.: MECP2 mutation in a boy with language disorder and schizophrenia.
Am J Psychiatry 2002, 159(1):148-9. PubMed Abstract | Publisher Full Text OpenURL
Klauck SM, Lindsay S, Beyer KS, Splitt M, Burn J, Poustka A: A mutation hot spot for nonspecific X-linked mental retardation in the MECP2 gene causes the PPM-X syndrome.
Am J Hum Genet 2002, 70(4):1034-7. PubMed Abstract | Publisher Full Text OpenURL
Samaco RC, Nagarajan RP, Braunschweig D, LaSalle JM: Multiple pathways regulate MeCP2 expression in normal brain development and exhibit defects in autism-spectrum disorders.
Hum Mol Genet 2004, 13(6):629-39. PubMed Abstract | Publisher Full Text OpenURL
Lam CW, Yeung WL, Ko CH, Poon PM, Tong SF, Chan KY, et al.: Spectrum of mutations in the MECP2 gene in patients with infantile autism and Rett syndrome.
J Med Genet 2000, 37(12):E41. PubMed Abstract | Publisher Full Text OpenURL
Carney RM, Wolpert CM, Ravan SA, Shahbazian M, Ashley-Koch A, Cuccaro ML, et al.: Identification of MeCP2 mutations in a series of females with autistic disorder.
Pediatr Neurol 2003, 28(3):205-11. PubMed Abstract | Publisher Full Text OpenURL
Watson P, Black G, Ramsden S, Barrow M, Super M, Kerr B, et al.: Angelman syndrome phenotype associated with mutations in MECP2, a gene encoding a methyl CpG binding protein.
J Med Genet 2001, 38(4):224-8. PubMed Abstract | Publisher Full Text OpenURL
Milani D, Pantaleoni C, D’Arrigo S, Selicorni A, Riva D: Another patient with MECP2 mutation without classic Rett syndrome phenotype.
Pediatr Neurol 2005, 32(5):355-7. PubMed Abstract | Publisher Full Text OpenURL
Van Esch H, Bauters M, Ignatius J, Jansen M, Raynaud M, Hollanders K, et al.: Duplication of the MECP2 region is a frequent cause of severe mental retardation and progressive neurological symptoms in males.
Am J Hum Genet 2005, 77(3):442-53. PubMed Abstract | Publisher Full Text OpenURL
del Gaudio D, Fang P, Scaglia F, Ward PA, Craigen WJ, Glaze DG, et al.: Increased MECP2 gene copy number as the result of genomic duplication in neurodevelopmentally delayed males.
Genet Med 2006, 8(12):784-92. OpenURL
Samaco RC, Fryer JD, Ren J, Fyffe S, Chao HT, Sun Y, et al.: A partial loss of function allele of methyl-CpG-binding protein 2 predicts a human neurodevelopmental syndrome.
Hum Mol Genet 2008, 17(12):1718-27. PubMed Abstract | Publisher Full Text OpenURL
Zhang A, Shen CH, Ma SY, Ke Y, El Idrissi A: Altered expression of autism-associated genes in the brain of Fragile X mouse model.
Biochem Biophys Res Commun 2009, 379(4):920-3. PubMed Abstract | Publisher Full Text OpenURL
LaSalle JM, Goldstine J, Balmer D, Greco CM: Quantitative localization of heterogeneous methyl-CpG-binding protein 2 (MeCP2) expression phenotypes in normal and Rett syndrome brain by laser scanning cytometry.
Hum Mol Genet 2001, 10(17):1729-40. PubMed Abstract | Publisher Full Text OpenURL
Balmer D, Goldstine J, Rao YM, LaSalle JM: Elevated methyl-CpG-binding protein 2 expression is acquired during postnatal human brain development and is correlated with alternative polyadenylation.
J Mol Med 2003, 81(1):61-8. PubMed Abstract | Publisher Full Text OpenURL
Jung BP, Jugloff DG, Zhang G, Logan R, Brown S, Eubanks JH: The expression of methyl CpG binding factor MeCP2 correlates with cellular differentiation in the developing rat brain and in cultured cells.
J Neurobiol 2003, 55(1):86-96. PubMed Abstract | Publisher Full Text OpenURL
Cassel S, Revel MO, Kelche C, Zwiller J: Expression of the methyl-CpG-binding protein MeCP2 in rat brain. An ontogenetic study.
Neurobiol Dis 2004, 15(2):206-11. PubMed Abstract | Publisher Full Text OpenURL
Mullaney BC, Johnston MV, Blue ME: Developmental expression of methyl-CpG binding protein 2 is dynamically regulated in the rodent brain.
Neuroscience 2004, 123(4):939-49. PubMed Abstract | Publisher Full Text OpenURL
Mnatzakanian GN, Lohi H, Munteanu I, Alfred SE, Yamada T, MacLeod PJ, et al.: A previously unidentified MECP2 open reading frame defines a new protein isoform relevant to Rett syndrome.
Nat Genet 2004, 36(4):339-41. PubMed Abstract | Publisher Full Text OpenURL
Kriaucionis S, Bird A: The major form of MeCP2 has a novel N-terminus generated by alternative splicing.
Nucleic Acids Res 2004, 32(5):1818-23. PubMed Abstract | Publisher Full Text OpenURL
Dragich JM, Kim YH, Arnold AP, Schanen NC: Differential distribution of the MeCP2 splice variants in the postnatal mouse brain.
J Comp Neurol 2007, 501(4):526-42. PubMed Abstract | Publisher Full Text OpenURL
Zachariah RM, Olson CO, Ezeonwuka C, Rastegar M: Novel MeCP2 isoform-specific antibody reveals the endogenous MeCP2E1 expression in murine brain, primary neurons and astrocytes.
PLoS One 2012, 7(11):e49763. PubMed Abstract | Publisher Full Text OpenURL
Coy JF, Sedlacek Z, Bachner D, Delius H, Poustka A: A complex pattern of evolutionary conservation and alternative polyadenylation within the long 3”-untranslated region of the methyl-CpG-binding protein 2 gene (MeCP2) suggests a regulatory role in gene expression.
Hum Mol Genet 1999, 8(7):1253-62. PubMed Abstract | Publisher Full Text OpenURL
Reichwald K, Thiesen J, Wiehe T, Weitzel J, Poustka WA, Rosenthal A, et al.: Comparative sequence analysis of the MECP2-locus in human and mouse reveals new transcribed regions.
Mamm Genome 2000, 11(3):182-90. PubMed Abstract | Publisher Full Text OpenURL
Pelka GJ, Watson CM, Christodoulou J, Tam PP: Distinct expression profiles of Mecp2 transcripts with different lengths of 3’UTR in the brain and visceral organs during mouse development.
Genomics 2005, 85(4):441-52. PubMed Abstract | Publisher Full Text OpenURL
Newnham CM, Hall-Pogar T, Liang S, Wu J, Tian B, Hu J, et al.: Alternative polyadenylation of MeCP2: influence of cis-acting elements and trans-acting factors.
RNA Biol 2010, 7(3):361-72. PubMed Abstract | Publisher Full Text OpenURL
Liu J, Francke U: Identification of cis-regulatory elements for MECP2 expression.
Hum Mol Genet 2006, 15(11):1769-82. PubMed Abstract | Publisher Full Text OpenURL
Decker CJ, Parker R: Diversity of cytoplasmic functions for the 3’ untranslated region of eukaryotic transcripts.
Curr Opin Cell Biol 1995, 7(3):386-92. PubMed Abstract | Publisher Full Text OpenURL
Coutinho AM, Oliveira G, Katz C, Feng J, Yan J, Yang C, et al.: MECP2 coding sequence and 3’UTR variation in 172 unrelated autistic patients.
Am J Med Genet B Neuropsychiatr Genet 2007, 144B(4):475-83. PubMed Abstract | Publisher Full Text OpenURL
Lewis BP, Shih IH, Jones-Rhoades MW, Bartel DP, Burge CB: Prediction of mammalian microRNA targets.
Cell 2003, 115(7):787-98. PubMed Abstract | Publisher Full Text OpenURL
Lewis BP, Burge CB, Bartel DP: Conserved seed pairing, often flanked by adenosines, indicates that thousands of human genes are microRNA targets.
Cell 2005, 120(1):15-20. PubMed Abstract | Publisher Full Text OpenURL
Wang XJ, Reyes JL, Chua NH, Gaasterland T: Prediction and identification of Arabidopsis thaliana microRNAs and their mRNA targets.
Genome Biol 2004, 5(9):R65. PubMed Abstract | BioMed Central Full Text OpenURL
Williams AE: Functional aspects of animal microRNAs.
Cell Mol Life Sci 2008, 65(4):545-62. PubMed Abstract | Publisher Full Text OpenURL
Bazzini AA, Lee MT, Giraldez AJ: Ribosome profiling shows that miR-430 reduces translation before causing mRNA decay in zebrafish.
Science 2012, 336(6078):233-7. PubMed Abstract | Publisher Full Text OpenURL
Kawasaki H, Taira K: MicroRNA-196 inhibits HOXB8 expression in myeloid differentiation of HL60 cells.
Nucleic Acids Symp Ser (Oxf) 2004, 48:211-2. Publisher Full Text OpenURL
Mukherji S, Ebert MS, Zheng GX, Tsang JS, Sharp PA, van Oudenaarden A: MicroRNAs can generate thresholds in target gene expression.
Nat Genet 2011, 43(9):854-9. PubMed Abstract | Publisher Full Text OpenURL
Yoo AS, Staahl BT, Chen L, Crabtree GR: MicroRNA-mediated switching of chromatin-remodelling complexes in neural development.
Nature 2009, 460(7255):642-6. PubMed Abstract | Publisher Full Text OpenURL
Liu C, Zhao X: MicroRNAs in adult and embryonic neurogenesis.
Neuromolecular Med 2009, 11(3):141-52. PubMed Abstract | Publisher Full Text OpenURL
Volvert ML, Rogister F, Moonen G, Malgrange B, Nguyen L: MicroRNAs tune cerebral cortical neurogenesis.
Cell Death Differ 2012, 19(10):1573-81. PubMed Abstract | Publisher Full Text OpenURL
Schratt GM, Tuebing F, Nigh EA, Kane CG, Sabatini ME, Kiebler M, et al.: A brain-specific microRNA regulates dendritic spine development.
Nature 2006, 439(7074):283-9. PubMed Abstract | Publisher Full Text OpenURL
Edbauer D, Neilson JR, Foster KA, Wang CF, Seeburg DP, Batterton MN, et al.: Regulation of synaptic structure and function by FMRP-associated microRNAs miR-125b and miR-132.
Neuron 2010, 65(3):373-84. PubMed Abstract | Publisher Full Text OpenURL
Vo NK, Cambronne XA, Goodman RH: MicroRNA pathways in neural development and plasticity.
Curr Opin Neurobiol 2010, 20(4):457-65. PubMed Abstract | Publisher Full Text OpenURL
Hansen KF, Karelina K, Sakamoto K, Wayman GA, Impey S, Obrietan K: miRNA-132: a dynamic regulator of cognitive capacity.
Brain Struct Funct 2013, 218(3):817-31. PubMed Abstract | Publisher Full Text OpenURL
Klein ME, Lioy DT, Ma L, Impey S, Mandel G, Goodman RH: Homeostatic regulation of MeCP2 expression by a CREB-induced microRNA.
Nat Neurosci 2007, 10(12):1513-4. PubMed Abstract | Publisher Full Text OpenURL
Chen WG, Chang Q, Lin Y, Meissner A, West AE, Griffith EC, et al.: Derepression of BDNF transcription involves calcium-dependent phosphorylation of MeCP2.
Science 2003, 302(5646):885-9. PubMed Abstract | Publisher Full Text OpenURL
Wada R, Akiyama Y, Hashimoto Y, Fukamachi H, Yuasa Y: miR-212 is downregulated and suppresses methyl-CpG-binding protein MeCP2 in human gastric cancer.
Int J Cancer 2010, 127(5):1106-14. PubMed Abstract | Publisher Full Text OpenURL
Im HI, Hollander JA, Bali P, Kenny PJ: MeCP2 controls BDNF expression and cocaine intake through homeostatic interactions with microRNA-212.
Nat Neurosci 2010, 13(9):1120-7. PubMed Abstract | Publisher Full Text OpenURL
Chen Y, Shin BC, Thamotharan S, Devaskar SU: Differential methylation of the micro-RNA 7b gene targets postnatal maturation of murine neuronal Mecp2 gene expression.
Dev Neurobiol 2014, 74(4):407-25. PubMed Abstract | Publisher Full Text OpenURL
Han K, Gennarino VA, Lee Y, Pang K, Hashimoto-Torii K, Choufani S, et al.: Human-specific regulation of MeCP2 levels in fetal brains by microRNA miR-483-5p.
Genes Dev 2013, 27(5):485-90. PubMed Abstract | Publisher Full Text OpenURL
Kynast KL, Russe OQ, Moser CV, Geisslinger G, Niederberger E: Modulation of central nervous system-specific microRNA-124a alters the inflammatory response in the formalin test in mice.
Pain 2013, 154(3):368-76. PubMed Abstract | Publisher Full Text OpenURL
Zhao H, Wen G, Huang Y, Yu X, Chen Q, Afzal TA, et al.: MicroRNA-22 regulates smooth muscle cell differentiation from stem cells by targeting methyl CpG-binding protein 2.
Arterioscler Thromb Vasc Biol 2015, 35(4):918-29. PubMed Abstract | Publisher Full Text OpenURL
Feng Y, Huang W, Wani M, Yu X, Ashraf M: Ischemic preconditioning potentiates the protective effect of stem cells through secretion of exosomes by targeting Mecp2 via miR-22.
PLoS One 2014, 9(2):e88685. PubMed Abstract | Publisher Full Text OpenURL
Dotti MT, Orrico A, De Stefano N, Battisti C, Sicurelli F, Severi S, et al.: A Rett syndrome MECP2 mutation that causes mental retardation in men.
Neurology 2002, 58(2):226-30. PubMed Abstract | Publisher Full Text OpenURL
Shibayama A, Cook EH Jr, Feng J, Glanzmann C, Yan J, Craddock N, et al.: MECP2 structural and 3’-UTR variants in schizophrenia, autism and other psychiatric diseases: a possible association with autism.
Am J Med Genet B Neuropsychiatr Genet 2004, 128B(1):50-3. PubMed Abstract | Publisher Full Text OpenURL
Xi CY, Ma HW, Lu Y, Zhao YJ, Hua TY, Zhao Y, et al.: MeCP2 gene mutation analysis in autistic boys with developmental regression.
Psychiatr Genet 2007, 17(2):113-6. PubMed Abstract | Publisher Full Text OpenURL
Santos M, Yan J, Temudo T, Oliveira G, Vieira JP, Fen J, et al.: Analysis of highly conserved regions of the 3’UTR of MECP2 gene in patients with clinical diagnosis of Rett syndrome and other disorders associated with mental retardation.
Dis Markers 2008, 24(6):319-24. PubMed Abstract | Publisher Full Text OpenURL
Tejada MI, Penagarikano O, Rodriguez-Revenga L, Martinez-Bouzas C, Garcia B, Badenas C, et al.: Screening for MECP2 mutations in Spanish patients with an unexplained mental retardation.
Clin Genet 2006, 70(2):140-4. PubMed Abstract | Publisher Full Text OpenURL
Ylisaukko-Oja T, Rehnstrom K, Vanhala R, Kempas E, von Koskull H, Tengstrom C, et al.: MECP2 mutation analysis in patients with mental retardation.
Am J Med Genet A 2005, 132A(2):121-4. PubMed Abstract | Publisher Full Text OpenURL
Fendri-Kriaa N, Mkaouar-Rebai E, Moalla D, Belguith N, Louhichi N, Zemni R, et al.: Mutational analysis of the MECP2 gene in Tunisian patients with Rett syndrome: a novel double mutation.
J Child Neurol 2010, 25(8):1042-6. PubMed Abstract | Publisher Full Text OpenURL
Hanchard NA, Carvalho CM, Bader P, Thome A, Omo-Griffith L, del Gaudio D, et al.: A partial MECP2 duplication in a mildly affected adult male: a putative role for the 3’ untranslated region in the MECP2 duplication phenotype.
BMC Med Genet 2012, 13:71. PubMed Abstract | BioMed Central Full Text OpenURL
Tantra M, Hammer C, Kastner A, Dahm L, Begemann M, Bodda C, et al.: Mild expression differences of MECP2 influencing aggressive social behavior.
EMBO Mol Med 2014, 6(5):662-84. PubMed Abstract | Publisher Full Text OpenURL
Lusardi TA, Farr CD, Faulkner CL, Pignataro G, Yang T, Lan J, et al.: Ischemic preconditioning regulates expression of microRNAs and a predicted target, MeCP2, in mouse cortex.
J Cereb Blood Flow Metab 2010, 30(4):744-56. PubMed Abstract | Publisher Full Text OpenURL
Lau AG, Irier HA, Gu J, Tian D, Ku L, Liu G, et al.: Distinct 3’UTRs differentially regulate activity-dependent translation of brain-derived neurotrophic factor (BDNF).
Proc Natl Acad Sci U S A 2010, 107(36):15945-50. PubMed Abstract | Publisher Full Text OpenURL
Varendi K, Matlik K, Andressoo JO: From microRNA target validation to therapy: lessons learned from studies on BDNF.
Cell Mol Life Sci 2015, 72(9):1779-94. PubMed Abstract | Publisher Full Text OpenURL
Nagarajan RP, Hogart AR, Gwye Y, Martin MR, LaSalle JM: Reduced MeCP2 expression is frequent in autism frontal cortex and correlates with aberrant MECP2 promoter methylation.
Epigenetics 2006, 1(4):e1-11. PubMed Abstract | Publisher Full Text OpenURL
Ebert DH, Gabel HW, Robinson ND, Kastan NR, Hu LS, Cohen S, et al.: Activity-dependent phosphorylation of MeCP2 threonine 308 regulates interaction with NCoR.
Nature 2013, 499(7458):341-5. PubMed Abstract | Publisher Full Text OpenURL
Broderick JA, Zamore PD: MicroRNA therapeutics.
Gene Ther 2011, 18(12):1104-10. PubMed Abstract | Publisher Full Text OpenURL
Zhang Y, Wang Z, Gemeinhart RA: Progress in microRNA delivery.
J Controlled Release 2013, 172(3):962-74. Publisher Full Text OpenURL

Neuron anatomy structure reconstruction based on a sliding filter

Gongning Luo1, Dong Sui1, Kuanquan Wang1* and Jinseok Chae2*

Author Affiliations

1Research Center of Perception and Computing, School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China

2Department of Computer Science and Engineering, Incheon National University, Incheon, Korea

For all author emails, please log on.

BMC Bioinformatics 2015, 16:342  doi:10.1186/s12859-015-0780-0

The electronic version of this article is the complete one and can be found online at:http://www.biomedcentral.com/1471-2105/16/342

Received: 14 May 2015
Accepted: 16 October 2015
Published: 24 October 2015

© 2015 Luo et al.

Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Abstract

Background

Reconstruction of neuron anatomy structure is a challenging and important task in neuroscience. However, few algorithms can automatically reconstruct the full structure well without manual assistance, making it essential to develop new methods for this task.

Methods

This paper introduces a new pipeline for reconstructing neuron anatomy structure from 3-D microscopy image stacks. This pipeline is initialized with a set of seeds that were detected by our proposed Sliding Volume Filter (SVF), given a non-circular cross-section of a neuron cell. Then, an improved open curve snake model combined with a SVF external force is applied to trace the full skeleton of the neuron cell. A radius estimation method based on a 2D sliding band filter is developed to fit the real edge of the cross-section of the neuron cell. Finally, a surface reconstruction method based on non-parallel curve networks is used to generate the neuron cell surface to finish this pipeline.

Results

The proposed pipeline has been evaluated using publicly available datasets. The results show that the proposed method achieves promising results in some datasets from the DIgital reconstruction of Axonal and DEndritic Morphology (DIADEM) challenge and new BigNeuron project.

Conclusion

The new pipeline works well in neuron tracing and reconstruction. It can achieve higher efficiency, stability and robustness in neuron skeleton tracing. Furthermore, the proposed radius estimation method and applied surface reconstruction method can obtain more accurate neuron anatomy structures.

Keywords:

Neuron anatomy structure reconstruction; Radius estimation; Sliding filter; Open curve snake model

Background

Neuron morphology and structure information is critical for neuroscience research. Hence, reconstructing the entire anatomy structure of a neuron is an essential task in the field of neuron informatics [1], [2]. However, reconstructing the anatomy structure of a neuron artificially is labor intensive. Efficient, advanced methods for anatomy structure reconstruction of neurons are greatly demanded in this field. Specifically, with the rapid development of microscopic imaging technology, a wide range of scales of bio-images can be obtained, which is helpful for us to develop new methods and algorithms to meet the needs in neuroscience research [3], [4]. The reconstructed digital neuron structure, including axons and dendrites as well as thickness information, can be used in conjunction with electrophysiological simulations to determine the complex mechanisms of the nervous system [5], [6].

The computer-aided manual neuron reconstruction method was first proposed in 1965 and was achieved by a biologist using a microscope [7]. Following this milestone, numerous algorithms and open softwares were introduced to reduce manual labor consisting of the boring task of tracing and analysis [8]–[11], but most of them were still limited to semi-automation and required manual validation by experts to achieve accurate reconstruction of whole neurons. Hence, the lack of powerful and effective computational tools for automatically reconstructing neuron cells has emerged as a major technical bottleneck in neuroscience research. This problem motivated the DIgital reconstruction of Axonal and DEndritic Morphology (DIADEM) challenge [12] and BigNeuron project [13], [14], which began in 2010 and 2015 respectively. They provided an open-source platform for researchers from all over the world and aimed to promote the development of computer algorithms for reconstructing the full anatomy structure of neurons. The data sets from DIADEM are most widely used in the domain of neuron reconstruction to date. However, the BigNeuron proposed some new challenges for the further research in the field of neuron reconstruction.

Generally speaking, before the DIADEM project, the neuron tracing methods were categorized into several types: shortest path methods [15], [16], minimum spanning tree methods [17], [18], sequential tracing methods [19], [20], skeletonization methods [21], [22], neuromuscular projection fiber tracing methods [23]–[25] and active contour-based tracing methods [26]–[28]. Based on these methods, some new improved methods were proposed [29]. The DIADEM final listed five well-performed algorithms: the model-based method [30], geometry-based method[31], probabilistic approach-based method [32], open snake-based method [33] and cost minimization trees-based approach [34]. In the model-based method, Myers’s team employed the idea of shortest paths to refine local tracing, which is based on the model of Al-Kofahi [19] and a formal tube model. This pipeline can reconstruct the neuron from raw or preprocessed images[30]. In the geometry-based method, Erdogmus’s team introduced a principal curve to represent the skeleton of axons, and they then extracted the topology information using a recursive principal curve tracing method [31]. In the probabilistic approach-based method, Gonzalez’s team built a set of candidate trees to choose the best one by a global objective function, which combined geometric priors from image evidence [32]. In the open snake-based method, Roysam’s team proposed a three dimensional open curve snake model that was initiated automatically by a set of skeletons from binary images generated by the 2-D graph cut pre-segmentation method, and the snake curve could be stretched bi-directionally along the centerline to trace the neuron cell structure [33]. Stepanyants’s team proposed trees-based method, which can merge individual branches into trees based on a cost minimization strategy [34]. After the DIADEM final, Liu’s group proposed a 3D neuronal morphology reconstruction method based on the augmented ray burst sampling method [35]. This method consisted of a single step to achieve the tracing and reconstruction, in which the centerline extraction or the extra radius estimation was unnecessary but the first seed must be set artificially. Peng’s team proposed series of efficient methods for neuron reconstruction, such as an anisotropic path searching method [36], an all-path pruning method [37], a hierarchical-path pruning method based on a gray-weighted image distance-tree[38], an automatic distance-field neuron tracing method based on global threshold foreground extraction [39], a smart tracing method based on machine learning [40] and a method based on reverse mapping and assembling of 2D projections [41]. These methods can work well with the neuron center lines tracing under the complex and noisy background. Kakadiaris’s team proposed a learning 3D tubular models-based method, which can use a morphology-guided deformable model to extract the dendritic centerline and use minimum shape-cost tree to represent the neuron morphology [42]. In addition, to achieve more accurate neuron tracing results, some open source softwares have been developed, such as flNeuronTool [35], FarSight [33], V3D [10], and Vaa3D[43], [44]. Along with all the existing algorithms, these open source softwares also promote the development of neuron reconstruction.

Despite the large number of proposed neuron tracing algorithms mentioned above, few methods can automatically reconstruct the complete and detailed neuron morphology, including complex dendritic and axonal arbors and variable thickness information. Moreover, because of the limited computer power, the automatic and accurate reconstruction of neuron anatomy structure is still a significant challenge.

In this paper, we propose a new 3D seed detection method based on Sliding Volume Filter (SVF) to initialize our framework, and we designed an open curve snake model combined with a SVF external force for centerline extraction and tracing. This open curve snake model has higher efficiency in the convergence of endpoints and detection of branch collision. In addition, radius estimation is another critical problem in neuron reconstruction, and accurate radius estimation can benefit simulation and functional research. Hence, this paper also proposes a new radius estimation method based on a 2D sliding band to estimate the radius of a neuron. The proposed radius estimation method can fit the real edges of neuron non-circular cross-sections better than previous methods. Finally, a surface reconstruction method based on contour lines is adopted to reconstruct detailed neuron morphology.

Methods

As shown in Fig. 1, some critical steps, such as seeding, tracing, radius estimating and surface reconstruction, are included in the pipeline of our protocol. The details of every critical step will be explained.

thumbnailFig. 1. Pipeline of neuron anatomy structure reconstruction

Seed detection

Seed detection is a critical procedure in the open snake-based tracing protocol, and an ideal seed list can ensure tracing accuracy. The proposed seeding method includes the following two stages:

(1) We used the proposed SVF based method to select coarse seeding points in the interior of neuron cells.

(2) The ridge criterion was used to achieve the further filter to obtain better seeding points, which are always near the center of the neuron cell.

A. SVF-Based seeding

In the field of computer vision and image processing, the convex region is defined as follows:

a) A rounded convex region is a region with higher intensity in the center than the edge, and the gradient vectors of this region point to its center.

b) A tube-like convex region is a region with higher intensity along its centerline than the edge, and the gradient vectors point to the centerline from the edge.

Quelhas’s group proposed a 2D Sliding Band Filter (SBF) for cell nucleus detection based on the characteristic of a rounded convex region [45]. In the data sets of microscopic imaging, a 3D neuron cell has not only a tube-like convex region but also a non-circular cross-section. Given these two characteristics, we extended the SBF into 3D space and designed a Sliding Volume Filter (SVF) to enhance the tube-like convex region for seed detection of neuron volume data.

To explain the calculation of SVF, we first explained the Voxel Convergence Index (VCI). As shown in Fig. 2a, O is an interested voxel in 3D volume datasets with its coordinate located at (x, y, z). A sphere support region R is located around the center O, and P are the voxels in the support regionR except at O, whose coordinate is (i, j, k). ϕ(i, j, k) is the angle between PO and the gradient vector direction. The VCI of P is defined as follows:

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M1','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M1">View MathML</a>

(1)

thumbnailFig. 2. Scheme of Spatial Convergence Index. a The model of 3D spatial convergence index. b The model of 3D sliding volume filter in y-z plate section. c the discretization calculation of SVF using the polar coordinates

Figure 2b and c show the calculation scheme of the sliding volume filter in a support region Rwhose radius is rad. To finish the discretization computation efficiently, the polar coordinate is introduced into this scheme, and the SVF is calculated as

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M2','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M2">View MathML</a>

(2)

with

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M3','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M3">View MathML</a>

(3)

where M is the number of support region lines radiating from the center pixel O(x, y, z), ρ denotes the radial coordinate, a and b stand for the angular coordinates, d is the thickness of sliding volume, r is the center position of the sliding volume in the support region line ranging from R minto R max , Q is the points between [r−d/2,r + d/2], and φ(qx ρ , qy ρ , qz ρ ) is the angle between the gradient vector at Q and the direction of QO. Additionally, the angles a ∈ [0, 2π] and b ∈ [0, π] are divided into 2 L parts and L parts, respectively. Thus, M = 2 L 2 . Specially, the number of parts of L determines the accuracy and efficiency of computation.

After the SVF was applied to the neuron volume data for seed detection voxel by voxel, we selected the voxels as the raw seeds whose SVF response values are higher than the threshold T. Notably, there are more gradient vectors that point to the center of a tube-like structure in the marginal regions than in the other regions [45]. Hence, the sliding volumes of support regions of interior points are more likely to converge in the marginal regions. As shown in Fig. 3a, the voxelA in the interior of the nerve is more likely to be selected as a raw seed than the external voxel B. Because A has a higher SVF response value than voxel B, the orientations of gradient vectors in the sliding volumes of support region of A are more likely to point to A. However, the orientations of gradient vectors in the sliding volumes of the support region of B are not consistent and sometimes point away from B. Moreover, a nerve cell is not a uniform tube-like structure but instead has variable thickness. Therefore, SVF is the proper filter for raw seed selection.

thumbnailFig. 3. Scheme of computation of Sliding Volume Filter, as well as the selection of seed points. a The procedure of seed points filtrate, after the SVF and ridge criterion, proper seed points are chosen. b The seeding points selection after step1 and step2

B. Ridge criterion

The Aylward’s ridge criterion method was applied to the raw seeds for the final seed choice [20]. As shown in Fig. 3a, I is the volume data set, ∇I(p) is the gradient vector at voxel p with its coordinate (x, y, z) in I, and ev1 , ev2 and ev3 are the eigenvectors computed from Hessian matrix of I. ev1 (p) is the principle direction along the center lines of the tube-like structure, and ev2 (p) and ev3 (p) are the other two orthogonal eigenvectors. The seeds near the center of the tube-like structcure meet the condition of Eq. 4.

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M4','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M4">View MathML</a>

(4)

The raw seeding points were further chosen according to the ridge criterion Eq. 4. As shown in Fig. 3b, after the steps of SVF and ridge criterion, the proper seeds near the center line of the nerve are chosen and included in the seed list, in which the seed points are sorted by the response values. Simultaneously, response values from SVF are used to enhance the intensity of voxels in the original data, which benefits the deformation of the open curve model in the following step. The SVF volume enhancement method is denoted as

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M5','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M5">View MathML</a>

(5)

where I SVF (p) is the intensity of point p after SVF enhancement, I(p) is the intensity of point pbefore SVF enhancement, and SVF(p) is the SVF value of point p.

SEF-OCS Neuron tracing

Tracing the full neuron skeleton is still a challenging task in neuron science, although many methods have been proposed. In this section, a new tracing model is proposed called an SVF external force open curve snake (SEF-OCS, SEF-Open Curve Snake). The open curve snake model was initially applied to automated actin filament segmentation and tracking [33], [46]. Extended the application of the open curve snake model to neuron tracing. However, the computation was tedious in the tracing framework of [33], especially in the step of branch detection. The proposed SEF-OCS includes three parts: open curve deformation, curve extension, and collision detection.

A. Open curve deformation

This model is a parametric open curve model, and the total snake energy can be defined as

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M6','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M6">View MathML</a>

(6)

E Total is the total image energy combined with internal energy and external energy. This model is a traditional deformable model, which resembles previous work in [16]. The open snake model is a parametric curve, c(s) = (x(s), y(s), z(s)), s∈[0, 1], and the snake internal and external energy are defined as follows:

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M7','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M7">View MathML</a>

(7)

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M8','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M8">View MathML</a>

(8)

with

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M9','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M9">View MathML</a>

In Eq. 7, α and β are the “elasticity coefficient” and “stiffness coefficient”, respectively, in internal energy, and they can control the regularity of the curve in the process of evolution. In Eq. 8, the external energy term is used to make the snake deform near the center line of the neuron and stretch the endpoints to the tail of the neuron. ∇E im is the negative normalized Gradient Vector Flow (GVF) of the volume data enhanced by SVF, p signifies point (x(s), y(s), z(s)) on the open curve, and I SVF is the volume after SVF enhancement in this paper. Instead of the original 3D image GVF, we calculated the GVF of I SVF . The SVF can enhance the tube-like convex region to smooth the GVF. As shown in Fig. 4a, the blue arrows show examples of gradient vectors from the volume enhanced by SVF. The vectors point toward the centers of neurons, which can make the seed points (the yellow points in Fig. 4a) move to the center position (the position of the red points in Fig. 4a). Specifically, the stretching force ∇E str (c(s)) is only implemented to the final endpointsc (0) and c (1). The cs (s)/||cs (s)|| denotes the direction of the stretching force. The value <a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M10','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M10">View MathML</a> is used to measure the tube-like level around the end point. When a curve reaches the end of a neuron, the end points will lose the tube-like characteristic. Hence, ∇Estr (c(s)) approaches zero, and the open active curve converges. According to a large number of experiments, this strategy is not only efficient and reliable but also can avoid the leakage of the neuron boundary. To minimize the energy function E Total , the points on the snake curves are updated as:

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M11','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M11">View MathML</a>

(9)

thumbnailFig. 4. Scheme of SEF open curve snake model. a The open curve is driven to the center of neuron by external force in the volume after SVF.b The procedure of open snake curve extension and collision detection in the branching region

where the parameters t and γ control the iteration numbers and size of the step at each iteration, respectively. The iterations are stopped when t reaches the threshold of the max iteration number.

B. Curve extension

The initial open snake curve is formed by three points (fewer than three points will not be traced as a branch of neuron). The first point p has the best response value in the seed list, and the other two points are generated by extending along the first principal direction to ev1 (p) and −ev1 (p). As shown in Fig. 4b, along with the open snake curve moving to the center of neuron, it also extends toward the two inverse tangential directions, cs (p0 ) and −cs (p1 ), in which the p0 and p1 are the two temporary endpoints. During the procedure of extension, the seed points belonging to one curve were labelled with new values (the default value is zero) in accord with the ID of the curve. For example, in Fig. 4b the yellow points and green points belong to different curves.

C. Collision detection

Neurons have many branches, especially in the dendrite region. Hence, detecting branching points and handling collision are essential. In the proposed scheme, two types of collision exist in the collision region and are shown in Fig. 4b. The first collision is branching point collision, which occurs when the open snakes reach a seed point whose value is not zero, and this point is recorded as the branching point (pink point in Fig. 4b). This branching point detection strategy is based on labelling seeds and is highly efficient. It also can handle the second type of collision, contour lines collision. The contour lines coming from the following step of radius estimation are the foundation of neuroanatomy reconstruction. However, due to the ambiguity of radius estimation in the collision region, the contour lines from two curves easily intersect. In Fig. 4b, this situation is illustrated in the imaginary pink circle and the embedded image, which is an experimental result in the branching region. This collision will influence the accuracy of the following reconstruction algorithm. Hence, a backoff strategy is proposed to avoid contour line collision. First, radius estimation in the branching points will not be executed. Second, if an extending curve reaches the branching point, it will be cut back the length of D, which is usually set as double the average estimated radius of the current curve.

In other words, the imaginary pink circle is not necessary in radius estimation because the following reconstruction algorithm would interpolate the information using triangular meshes automatically. Finally, the tracing algorithm ends when all the seed points are traversed.

The entire tracing algorithm procedure is shown as follows:

In summary, compared to the open snake method in [33], we improved this model in the following three aspects. First, the volume after SVF enhancement has more straightforward gradient vectors, which point to the center line of the neuron and can be used in driving the initial lines to the center of the neuron. Second, the proposed method can cut down the computation of the stretching force of end nodes. Third, in the step of collision detection, compared to the method based on labelling neighbor voxels, the method based on labelling seeds has higher detection efficiency and benefits the following reconstruction procedure.

Radius estimation

Radius estimation is another critical task in neuron anatomy reconstruction, and it can provide more quantitative information for neuroscience research. Peng, Aylward, and Wang had proposed some radius estimation methods [16], [20], [33], but most of them are based on the assumption that the neuron have a uniform tube-like structure, whose cross-sections are regular circles. However, the real cross-sections are not regular circles, as shown in the embedded image of Fig. 5. To reconstruct the neuron morphological structure more accurately, fitting the real edge of the neuron cell is achieved by a new proposed radius estimation method based on a 2D Sliding Band Filter (SBF) [45]. The SBF can converge on the real edge of a neuron cross-section that has the rounded convex region in [45].

thumbnailFig. 5. Illustration of radius estimation of the neuron cross-section. The left embedded image shows the real cross-section of neuron, and the estimation result with different parameters. And in the right image v 1 is the tangential direction in S i , v 2 and v 3 are the orthogonal vectors which define the cross-section

Figure 5 shows the scheme of the radius estimation method based on a 2D sliding band filter. We could obtain the cross-section according to the normal vector v 1 , which points to the tangential direction of the open curve. Additionally, the v 2 and v 3 are the orthogonals in the cross-section. To obtain an accurate estimation of neuron cross-section, n radiuses radiating from the center point S on the snake curve are estimated as different lengths. The radius lengths are equal to r in the Eq. 10 with the maximum SBF response value.

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M12','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M12">View MathML</a>

(10)

with,

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M13','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M13">View MathML</a>

where B is the boundary points on the cross-section, which are at the centers of sliding bands and will be used to fit the real edge. (x rn , y rn , z rn ) are the spatial coordinates of S. The computation method of SCI in point P, which is in the range of [r−d/2, r + d/2], has been introduced in Eq. 4. dis the width of the sliding band, r is the distance between B and the center point S, and it can slide in the range of [R min , R max ] to obtain the optimal position of B with the maximum SBF response value. The boundary points B can be connected clockwise to fit the edge of the neuron cell.

In the proposed method, the parameter n is related to the accuracy of radius estimation. As shown in the embedded image of Fig. 5, the larger n is, the more accurate the cross-section fitting will be. However, considering the efficiency and accuracy in the actual application, the parameter n should be adjusted flexibly.

Neuron surface reconstruction

Most of the traditional neuron reconstruction methods were based on the fast marching method and some supplemental processes for connecting different fragments [33], [47]. However, in this paper, Liu’s non-parallel contour lines surface reconstruction method is employed for surface reconstruction [48], considering the non-parallel characteristic of circles generated from previous steps. On the premise of an accurate description of the entire neuron anatomy structure, this method is efficient. Although this method had been widely used in other biological models, it has rarely been used in neuron model reconstruction. The generated mesh model of the neuron can benefit the future finite element mesh subdivision and simulation.

Figure 6 shows the scheme of Liu’s method. First, it constructs medial axes (MA) between adjacent contour lines (Fig. 6a). Second, the points and lines from different contours are projected on the MA (Fig. 6b). Third, triangular meshes are used to connect the curve networks to their projection points on the MA (Fig. 6c). Finally, the surface meshes, which are connected with different contour lines, are formed as the boundaries between neighboring compartments [48]. To obtain a smoother neuron surface, we use a surface diffusion smoothing algorithm to minimize the curvature of the model surface to obtain a smooth 3D model [49]. As shown in Fig. 6, the initial neuron surface (Fig. 6e) is formed by contour lines (Fig. 6d) which are obtained by radius estimation, and the final smooth neuron surface is shown in Fig. 6f. In addition, the most outstanding advantage of Liu’s method is that it can automatically handle branch reconstruction, especially of circles without intersections in the branching region (the intersection problem was resolved through removing the collisions of circles in the SEF-OCS neuron tracing step). As shown in Fig. 7, in the branching region, two branches could be automatically reconstructed with different label colors.

thumbnailFig. 6. Process of contour reconstruction. a Construction of projective plate MA [48]. b Projection of points and lines on MA [48]. c Triangulation of adjacent contour lines [48]. d The contour lines of neuron cell. e The initial surface from triangulation of adjacent contour lines. f The final surface model after smoothing

thumbnailFig. 7. Process of neuron reconstruction in branching region. a The input branching data from [48]. b The reconstruction result of input data of (a).c The input data of neuron contour lines. d The reconstruction result of (c), in which different branches are labelled by different colors

Results and discussion

Parameters

We validated and evaluated the steps of seeding, tracing, radius estimation and neuron reconstruction in the proposed method using synthetic data and real data from the DIADEM challenge [12] and parts of datasets from BigNeuron project [13], [14]. All of the experiments were performed on an ordinary computer (Intel Core i5 3.2 CPU, NVIDIA GeForce GTX 960, 8 GB RAM, Windows 7). The proposed algorithm was developed using C++ language. In addition, to compare the other methods equally, we did not adopt any manual interactive operations shown in the Fig. 1, such as preprocessing, picking and expending seeds, checking and validating data, tracing editing, branch refining, and rooting setting, although these operation can improve the result of neuron tracing.

According to the image data sets (DIADEM challenge and BigNeuron project) chosen in this paper, Table 1 shows the parameters selected for the following experiments. Some experimental parameters such as d, T, L, γ, ɑ, and β remain constant for the following experiments. Other parameters such as rad, R max , R min , n and t could be adjusted for optimal results.

Table 1. Parameter selection

Seeding

Generally speaking, an ideal seed point is located in the neuron body of the foreground, and its position is near the center of the neuron cell’s cross-section. To quantify the performance of SVF seed detection, we evaluated the seeding method using an artificial helix dataset and the real dataset of the DIADEM challenge [12], according to following point deviation measurement principle:

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M14','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M14">View MathML</a>

(11)

where P g denotes point sets in the gold standard, P s denotes point sets generated by tested methods and N is the number of points in P s . d min is the distance between a seeding point and its nearest point in the gold standard.

We compared the proposed seeding methods with the two most widely used seeding point detection methods, the global threshold method [18] and the LoG threshold method [50]. We set the parameters rad = 20, R max= 16, and R min= 5 in this experiments.

In the proposed SVF filter method, compared with two other methods, most of the seed points are in the interior of the neuron body. Table 2 shows the seed deviation results of the artificial helix body and OP_1, in which the SVF seeding method can achieve the lowest deviation.

Table 2. Comparisons of different seeding methods on test datasets

As shown in Fig. 8, some seeds fall outside of the artificial helix body, which are detected by traditional threshold methods and highlighted with arrows. Figure 9 shows the seeding results in drosophila olfactory axonal datasets (OP_1 of DIADEM challenge), which are generated by the three mentioned methods. These results also suggest that our method is better than the two other seed detection methods.

thumbnailFig. 8. Comparison of seeding method on test dataset; a The seeds detection result of global threshold method. b The seeds detection result of LoG threshold method. c The seeds detection result of SVF method

thumbnailFig. 9. Comparison of seeding method on drosophila olfactory axonal data sets (OP_1). a The seeds detection result of global threshold method. b The seeds detection result of LoG threshold method. c The seeds detection result of SVF method and the enlargement image in intensive region of dendrites

We also compared the enhancement results from the three methods, and we chose the same cross-section of the neuron volume image to demonstrate that the SVF seeding method can enhance the region around the center line and simultaneously save the contrast information of the tube-like volume. The results are shown in Fig. 10, in which the red part has a higher response value than the blue part. This result suggests that the SVF method can extract the center region better than the other two methods and can enhance the original volume data properly. Furthermore, the better results in both seed detection and SVF volume enhancement can benefit neuron tracing.

thumbnailFig. 10. Comparisons among seeding methods for image enhancement. The results of global threshold method lose the contrast information between center and edge; The results of LoG threshold method extend the center region exorbitantly; The results of SVF seeding method can enhance the original volume properly

Skeleton tracing

A. Tracing accuracy

We adopted the drosophila olfactory axonal datasets (OP_1 to OP_9 from DIADEM challenge) and neocortical layer 1 axons subset 1 datasets (NC_1 to NC_6 from DIADEM challenge) to evaluate the performance of the proposed neuron skeleton tracing method in the term of accuracy. Meanwhile, we compared the SEF-OCS tracing method with some start-of-the-art algorithms, such as the Open Curve Snake tracing method (OCS) [33], Neural Circuit Tracer method (NCT) [34], all-path pruning method (APP) [37], improved all-path pruning method (APP2) [38], distance-field based method (DF) [39], and 3D tubular models based method (TM) [42].

To compare with these methods fairly, we conducted the experiments without any manual interactions and corrections, using the widely used accuracy principle to measure the test results. Similarly, we set the parameters rad = 20, R max= 16, R min= 5, and t = 10 in the proposed method, and we chose the better parameters for other six methods according to the features of different datasets. The measured principle is defined as:

<a onClick="popup('http://www.biomedcentral.com/1471-2105/16/342/mathml/M15','MathML',630,470);return false;" target="_blank" href="http://www.biomedcentral.com/1471-2105/16/342/mathml/M15">View MathML</a>

(12)

where Precision is measured as the proportion of the length of correct traces to the total length of the traces generated by the tested methods, and Recall is the proportion of the length of correct traces to the length of the gold standard of adopted datasets.

Figure 11a and b show the reconstruction results of OP_1, and Fig. 11c and d show the results of OP_4. All of these results were generated by our proposed SEF-Open curve snake method. To illustrate the higher performance of our proposed method, we choose different colors to indicate the differences; the blue lines are the gold standard, and the green lines are the skeleton reconstructed by our method. Additionally, more tracing results of the other datasets are shown in the Additional file 1.

thumbnailFig. 11. The tracing results of OP_1 and OP_4. a The tracing result of OP_1 from multi-view. b The magnified result of branch parts of OP_1. cThe tracing result of OP_4 from multi-view; d The magnified result of branch parts of OP_4

Table 3 summarizes the comparisons between the OCS and other six methods in terms of precision and recall. We can see that SEF-OCS is far better than other six methods in most datasets in terms of accuracy and recall. In addition, the SEF-OCS outperforms other six methods in average accuracy and recall. We also conducted the DIADEM score test [51] to evaluate the proposed method in the precision of reconstructed neuron topology and compare with the other methods. To the best of our knowledge, due to the various features of different datasets, no methods can get higher DIADEM score in all the datasets automatically. Hence a box plot is adopted to show the DIADEM score distribution of different methods tested in the different datasets. As we can see from Fig. 12, our method can achieve higher DIADEM score in most of the tested datasets and outperforms other six methods in average value (0.87 ± 0.001), median (0.86) and minimum (0.81). This results also proved that the proposed method has higher stability. In order to evaluate the automaticity of the proposed method, we used fixed parameters to test our method in this paper. Actually, The changed parameters can also be tried to get more meaningful tracing results. For instance, when the neuron data includes a big cell body, the bigger radparameter is needed. Additionally, some other methods also can be tried to trace neuron according to the features of different neuron cells. For example, the APP, APP2 and DF methods can achieve better results sometimes.

Table 3. Comparisons among different methods with different image datasets in tracing accuracy

thumbnailFig. 12. The box plot of DIADEM scores of the different methods for different datasets (OP_1 to OP_9 and NC_1 to NC_6 datasets). The median is the middle pink bar. The box indicates the lower quartile (splits 25 % of lowest data) and the upper quartile (splits 75 % of highest data). The red bar and blue bar are the maximum and minimum values. The blue diamond denotes the mean value of the scores

B. Tracing robustness

To verify the robustness of our method, we designed three kinds of experiments. Firstly, the datasets with different levels of signal attenuation were tested. Secondly, the datasets with deferent levels of Gaussian noise were tested. Thirdly, the datasets (checked6_frog_scrippts, checked6_human_culturedcell_Cambridge_in_vitro_confocal_GFP,checked6_human_allen_confocal and checked6_fruitfly_larvae_gmu) from BigNeuron project were also tested using the proposed method.

Firstly, we compared the length of the traced skeleton with OCS in handling image signal reduction. Compared with the traditional robustness test method, which always added Gaussian noise to the original volume, this paper’s test method has a special meaning. Unbalanced light will lead to different levels of signal attenuation in the process of capturing images from a microscope. The OP_1 data set is chosen as an example, and we reduced the image signal from 10 % to 40 %. The content of the neuron images with different degrees of reduction is shown in Fig. 13. In Fig. 14, we list the lengths of the neuron skeleton traced by the two methods for comparison. With the image information reducing from 10 % to 40 %, our method can trace more information than the open curve snake in skeleton length. This result conveys that our method has higher robustness upon image signal reduction. All these results suggest that the proposed method performs better than the OCS method.

thumbnailFig. 13. Comparison in signal removed image datasets. To achieve more clear comparison, we follow the same experiment design in [16]

thumbnailFig. 14. Changes of skeleton length with signal reduction (Unit: Volex)

Secondly, we tested the robustness of our method using the datasets with different levels of Gaussian noise (The mean is 0 and the variances are 0.01, 0.02, 0.03 and 0.04 respectively.). As we can see from Fig. 15, the blue lines represent the tracing results of the proposed method. The tracing results are tolerable even when the variance is 0.04 and the major branches of neurons are not missed.

thumbnailFig. 15. The tracing results of NC_2 dataset with different levels of Gaussian noise. a The dataset with a variance of 0.01. b The dataset with a variance of 0.02. c The dataset with a variance of 0.03. d The dataset with a variance of 0.04. To prove the robustness of our method clearly, we follow the similar design of experiment in [39]

Thirdly, the datasets from the BigNeuron project were also tested. To my best knowledge, the BigNeuron will be a new trend in this field and most of datasets are challenging. The Fig. 16 shows some tracing results of the datasets of the BigNeuron project. Similarly, the blue lines represent the tracing results of our method. The results are tolerable even these datasets are complex and sometimes include a big cell body (In the Fig. 16, the big cell body is highlighted using the red rectangle). However, rad parameter must be set bigger (we set the parameters rad = 35, Rmax = 31, Rmin = 5, and t = 10 in the proposed method) to get better results when the datasets contain a big cell body. Additionally, the APP2 from Vaa3D [43], [44] can also achieve good results automatically when a big cell body exists in the datasets.

thumbnailFig. 16. The tracing results of the datasets from the BigNeuron project. aThe tracing result of the done_err_Recon112012no2-2 data of checked6_frog_scrippts. b The tracing result of image 7 data of checked6_human_culturedcell_Cambridge_in_vitro_confocal_GFP. c The tracing result of in_house1 data of checked6_human_allen_confocal. dThe tracing result of done_1_CL-I_X_OREGON_R_ddaD_membrane-GFP data of checked6_fruitfly_larvae_gmu

Radius estimation and surface reconstruction

Adaptive radius estimation and surface reconstruction methods can improve the neuron model, which has branches of varying widths. In radius estimation, the proposed method can fit the edge of the cross-section of the neuron cell. Furthermore, the credibility of radius estimation can be adjusted by the parameter n. As we can see from Fig. 5, the higher parameter n, the greater the credibility of radius estimation and the higher the computation intensity. Though a higher credibility of radius estimation must be achieved by higher computation intensity, the proposed method has solved the non-circular radius estimation problem mentioned in [33]. Generally, the parameter n = 16 is sufficient for most applications. Hence, we set n = 16 in the efficiency comparison experiments. Figure 17b shows the radius estimation results of the entire neuron using the proposed method based on the skeleton of the original volume, which is shown in Fig. 17a.

thumbnailFig. 17. Radius estimation and anatomical reconstruction of OP_1. a The original result by volume rendering method. b The result of radius estimation by 2D sliding band method, in which the blue contour lines are used to fit the edges of neuron cross-sections. c The anatomical reconstruction result based on the contour lines, in which the different branches are labelled with different colors

The adopted reconstruction method can interpolate meshes based on contour lines from radius estimation and can also handle branching reconstruction problems efficiently. To illustrate the performance of the proposed framework, the reconstruction result of the most complex data (OP_1) in the OP series data sets is shown in Fig. 17. The morphology of a reconstructed neuron cell of OP_1 was obtained, and the different branches were labelled with different colors during reconstruction.

Computational efficiency

In the term of automatic computation efficiency, we tested every step of the proposed pipeline and compared with other methods to evaluate the 3D neuron reconstruction efficiency. In addition, we set the parameters rad = 20, Rmax = 16, Rmin = 5, and t = 10 in the proposed method.

As we can see from Table 4, the proposed method is more efficient than the OCS framework, especially in the seeding step because seeding in the OCS framework is based on a complex procedure including graph-cut segmentation and skeleton and seeding point selection. In contrast, our seeding method is more concise and efficient. The higher tracing efficiency also demonstrated the improvements in stretching force in the open snake model and the collision point detection strategy. Additionally, Table 4 also shows that the radius estimation and reconstruction are more efficient in the SEF-OCS framework than in the OCS framework. These results prove that the proposed radius estimation and surface reconstruction methods outperform the corresponding methods in the OCS framework.

Table 4. Comparisons with OCS method in the efficiency of the proposed pipeline

To test the ability of parallel computation, we also developed our CUDA implementation for the four main steps. The computation time is shown in the Table 4. We can see that the proposed method is faster than OCS method in the same parallel environment. In addition, the average speedup ratio of SEF-OCS can achieve 63.94, which is higher than OCS’s average speedup ratio 48.8. This also demonstrates that the proposed method has higher parallel computation ability.

Table 5 shows the efficiency comparison results with some other methods from Table 3. Actually, most of neuron reconstruction methods don’t contain the surface reconstruction procedure. Hence, we conducted the comparison experiments in another way. We cut down the computational cost of the surface reconstruction step in order to carry out the comparison among different methods fairly. What’s more, the experiments are conducted three times for every method corresponding to every data set to avoid the errors from the operation system environment. The average results of three times are shown in Table 5. The comparison results show that the proposed method achieve the lowest average computational cost. We also can see that the computational cost of our method mainly depend on the size of the data sets and will realize higher efficiency with the development of computation parallel capacity.

Table 5. Comparisons in the efficiency of neuron tracing

Conclusions

Neuron cell anatomy structure reconstruction plays a very important role in the field of neurology. In this paper, we have developed a new neuron tracing framework, which is based on a sliding filter. We improved every step of the traditional framework compared to the OCS framework. First, given a non-circular cross-section of a neuron, the sliding filter method was introduced to the proposed seeding method (SVF) and radius estimation method (SBF), which is critical for accurately tracing skeletons and reconstructing real morphology. Second, on the basis of better seeding results, the traditional open curve snake model was improved by introducing a new external force to aid the curve evolution for neuron skeleton tracing and a new strategy for collision detection. Finally, a surface reconstruction method based on contour lines was used to generate whole neuron morphology.

A series of experiments have proved that the proposed framework has higher efficiency, stability and robustness in tracing accuracy. In addition, the proposed estimation method and adopted neuron reconstruction method can obtain more accurate neuron morphology, which is meaningful for future works such as simulation and analysis of neuron function in the field of neuroscience research.

Availability of supporting data

The source code can be available in the website [52]. The OP and NC datasets come from DIADEM challenge project, it can be downloaded from [53]. The checked6_frog_scrippts datasets, the checked6_human_culturedcell_Cambridge_in_vitro_confocal_GFP datasets, the checked6_human_allen_confocal datasets and the checked6_fruitfly_larvae_gmu datasets are available in the BigNeuron project whose website is [54].

Abbreviations

VCI: Voxel Convergence Index

SBF: Sliding Band Filter

SVF: Sliding Volume Filter

SEF-OCS: SVF external force open curve snake

OCS: Open Curve Snake

MA: Medial axes

NCT: Neural Circuit Tracer method

APP: All-path pruning method

APP2: Improved all-path pruning method

DF: Distance-field based method

TM: 3D tubular models based method

S: Seeding

T: Tracing

RE: Radius estimation

SR: Surface reconstruction

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

GL and DS carried out the program of the neuron tracing and reconstruction and finish the work of writing the draft, KW designed the workflow of neuron cell tracing and reconstruction pipeline. JC provided the suggestion of manuscript revision and pipeline optimization. All authors read and approved the final manuscript.

Additional file

Additional file 1:. This document includes additional figures not included in the paper. Some other tracing results are shown in the supplementary material (12 figures). S1-S7 are some tracing results of BigNeuron datasets. S8-S9 are some other tracing results of NC datasets. S10-S12 are some other tracing results of OP datasets. (PDF 1439 kb)

Format: PDF Size: 1.4MB Download file

This file can be viewed with: Adobe Acrobat ReaderOpen Data

Acknowledgements

This work was supported by the Incheon National University International Cooperative Research Grant in 2013. Our deepest gratitude goes to the anonymous reviewers for their careful work and constructive suggestions that have helped us to improve this paper substantially. We also appreciate the support of softwares and datasets from the team of the BigNeuron project.

References

  1. Roysam B, Shain W, Ascoli GA. The central role of neuroinformatics in the national academy of engineering’s grandest challenge: reverse engineer the brain.Neuroinformatics. 2009; 7(1):1-5. PubMed Abstract | Publisher Full Text OpenURL
  2. van Pelt J, van Ooyen A, Uylings HBM. The need for integrating neuronal morphology databases and computational environments in exploring neuronal structure and function. Anat Embryol. 2001; 204(4):255-65. PubMed Abstract | Publisher Full Text OpenURL
  3. Peng HC. Bioimage informatics: a new area of engineering biology. Bioinformatics. 2008; 24(17):1827-36. PubMed Abstract | Publisher Full Text OpenURL
  4. Peng HC, Bateman A, Valencia A, Wren JD. Bioimage informatics: a new category in bioinformatics. Bioinformatics. 2012; 28(8):1057-57. PubMed Abstract |Publisher Full Text OpenURL
  5. Arbib MA, Bonaiuto JJ, Bornkessel-Schlesewsky I, Kemmerer D, MacWhinney B, Nielsen FA et al.. Action and language mechanisms in the brain: data, models and neuroinformatics. Neuroinformatics. 2014; 12(1):209-25. PubMed Abstract |Publisher Full Text OpenURL
  6. Peng HC, Roysam B, Ascoli GA. Automated image computing reshapes computational neuroscience. BMC Bioinformatics. 2013;14:293. doi:10.1186/1471-2105-14-293.
  7. Glaser EM, Vanderloos H. A semi-automatic computer-microscope for the analysis of neuronal morphology. IEEE Trans Biomed Eng. 1965; 12:22-31. PubMed Abstract |Publisher Full Text OpenURL
  8. Fordholevinski TS, Dahlberg TA, Agranoff BW. A microcomputer-based image analyzer for quantitating neurite outgrowth. Brain Res. 1986; 368(2):339-46. Publisher Full TextOpenURL
  9. Ascoli GA. Neuroinformatics grand challenges. Neuroinformatics. 2008; 6(1):1-3.PubMed Abstract | Publisher Full Text OpenURL
  10. Peng HC, Ruan ZC, Long FH, Simpson JH, Myers EW. V3D enables real-time 3D visualization and quantitative analysis of large-scale biological image data sets. Nat Biotechnol. 2010; 28(4):348-53. PubMed Abstract | Publisher Full Text OpenURL
  11. Peng HC, Long FH, Myers EW. VANO: a volume-object image annotation system.Bioinformatics. 2009; 25(5):695-7. PubMed Abstract | Publisher Full Text OpenURL
  12. Brown KM, Barrionuevo G, Canty AJ, De Paola V, Hirsch JA, Jefferis GSXE et al.. The DIADEM data Sets: representative light microscopy images of neuronal morphology to advance automation of digital reconstructions. Neuroinformatics. 2011; 9(2-3):143-57. PubMed Abstract | Publisher Full Text OpenURL
  13. Peng HC, Meijering E, Ascoli GA. From DIADEM to BigNeuron. Neuroinformatics. 2015;13(3):259-60. PubMed Abstract | Publisher Full Text OpenURL
  14. Peng HC, Hawrylycz M, Roskams J, Hill S, Spruston N, Meijering E et al.. BigNeuron: large-scale 3D neuron reconstruction from optical microscopy images. Neuron. 2015;87(2):252-6. PubMed Abstract | Publisher Full Text OpenURL
  15. Meijering E, Jacob M, Sarria JCF, Steiner P, Hirling H, Unser M. Design and validation of a tool for neurite tracing and analysis in fluorescence microscopy images. Cytom Part A. 2004; 58A(2):167-76. Publisher Full Text OpenURL
  16. Peng HC, Ruan ZC, Atasoy D, Sternson S. Automatic reconstruction of 3D neuron structures using a graph-augmented deformable model. Bioinformatics. 2010;26(12):i38-46. PubMed Abstract | Publisher Full Text OpenURL
  17. Yuan XS, Trachtenberg JT, Potter SM, Roysam B. MDL constrained 3-D grayscale skeletonization algorithm for automated extraction of dendrites and spines from fluorescence confocal images. Neuroinformatics. 2009; 7(4):213-32. PubMed Abstract |Publisher Full Text OpenURL
  18. Gonzalez G, Turetken E, Fleuret F, Fua P. Delineating trees in noisy 2D images and 3D image-stacks. Proc Cvpr IEEE. 2010:2799-806. doi:10.1109/CVPR.2010.5540010.
  19. Al-Kofahi KA, Lasek S, Szarowski DH, Pace CJ, Nagy G, Turner JN et al.. Rapid automated three-dimensional tracing of neurons from confocal image stacks. IEEE T Inf Technol B. 2002; 6(2):171-87. Publisher Full Text OpenURL
  20. Aylward SR, Bullitt E. Initialization, noise, singularities, and scale in height ridge traversal for tubular object centerline extraction. IEEE Trans Med Imaging. 2002;21(2):61-75. PubMed Abstract | Publisher Full Text OpenURL
  21. Cohen AR, Roysam B, Turner JN. Automated tracing and volume measurements of neurons from 3-D confocal fluorescence microscopy data. J Microsc-Oxford. 1994;173:103-14. Publisher Full Text OpenURL
  22. He W, Hamilton TA, Cohen AR, Holmes TJ, Pace C, Szarowski DH et al.. Automated three-dimensional tracing of neurons in confocal and brightfield images. Microsc Microanal. 2003; 9(4):296-310. PubMed Abstract | Publisher Full Text OpenURL
  23. Cai HM, Xu XY, Lu J, Lichtman J, Yung SP, Wong STC. Using nonlinear diffusion and mean shift to detect and connect cross-sections of axons in 3D optical microscopy images. Med Image Anal. 2008; 12(6):666-75. PubMed Abstract | Publisher Full Text OpenURL
  24. Lu J, Fiala JC, Lichtman JW. Semi-automated reconstruction of neural processes from large numbers of fluorescence images. Plos One. 2009; 4(5):e5655. PubMed Abstract |Publisher Full Text OpenURL
  25. Srinivasan R, Li Q, Zhou XB, Lu J, Lichtman J, Wong STC. Reconstruction of the neuromuscular junction connectome. Bioinformatics. 2010; 26(12):i64-70.PubMed Abstract | Publisher Full Text OpenURL
  26. Schmitt S, Evers JF, Duch C, Scholz M, Obermayer K. New methods for the computer-assisted 3-D reconstruction of neurons from confocal image stacks. Neuroimage. 2004; 23(4):1283-98. PubMed Abstract | Publisher Full Text OpenURL
  27. Cai HM, Xu XY, Lu J, Lichtman JW, Yung SP, Wong STC. Repulsive force based snake model to segment and track neuronal axons in 3D microscopy image stacks.Neuroimage. 2006; 32(4):1608-20. PubMed Abstract | Publisher Full Text OpenURL
  28. Vasilkoski Z, Stepanyants A. Detection of the optimal neuron traces in confocal microscopy images. J Neurosci Meth. 2009;178(1):197–204.
  29. Halavi M, Hamilton KA, Parekh R, Ascoli GA. Digital reconstructions of neuronal morphology: three decades of research trends. Front Neurosci. 2012; 6:49.PubMed Abstract | Publisher Full Text OpenURL
  30. Zhao T, Xie J, Amat F, Clack N, Ahammad P, Peng HC et al.. Automated reconstruction of neuronal morphology based on local geometrical and global structural models.Neuroinformatics. 2011; 9(2-3):247-61. PubMed Abstract | Publisher Full Text OpenURL
  31. Bas E, Erdogmus D. Principal curves as skeletons of tubular objects. Neuroinformatics. 2011; 9(2-3):181-91. PubMed Abstract | Publisher Full Text OpenURL
  32. Turetken E, Gonzalez G, Blum C, Fua P. Automated reconstruction of dendritic and axonal trees by global optimization with geometric priors. Neuroinformatics. 2011;9(2-3):279-302. PubMed Abstract | Publisher Full Text OpenURL
  33. Wang Y, Narayanaswamy A, Tsai CL, Roysam B. A broadly applicable 3-D neuron tracing method based on open-curve snake. Neuroinformatics. 2011; 9(2-3):193-217.PubMed Abstract | Publisher Full Text OpenURL
  34. Chothani P, Mehta V, Stepanyants A. Automated tracing of neurites from light microscopy stacks of images. Neuroinformatics. 2011; 9(2-3):263-78. PubMed Abstract |Publisher Full Text OpenURL
  35. Ming X, Li AA, Wu JP, Yan C, Ding WX, Gong H et al.. Rapid reconstruction of 3D neuronal morphology from light microscopy images with augmented rayburst sampling. Plos One. 2013; 8(12):e84557. PubMed Abstract | Publisher Full Text OpenURL
  36. Xie J, Zhao T, Lee T, Myers E, Peng HC. Anisotropic path searching for automatic neuron reconstruction. Med Image Anal. 2011;15(5):680–9.
  37. Peng HC, Long FH, Myers G. Automatic 3D neuron tracing using all-path pruning.Bioinformatics. 2011; 27(13):I239-47. PubMed Abstract | Publisher Full Text OpenURL
  38. Xiao H, Peng HC. APP2: automatic tracing of 3D neuron morphology based on hierarchical pruning of a gray-weighted image distance-tree. Bioinformatics. 2013;29(11):1448-54. PubMed Abstract | Publisher Full Text OpenURL
  39. Yang JZ, Gonzalez-Bellido PT, Peng HC. A distance-field based automatic neuron tracing method. BMC Bioinformatics. 2013; 14(1):93. PubMed Abstract |BioMed Central Full Text OpenURL
  40. Chen HB, Xiao H, Liu TM, Peng HC. SmartTracing: self-learning-based neuron reconstruction. Brain Inform. 2015; 1(1):1. OpenURL
  41. Zhou Z, Liu XX, Long B, Peng HC. TReMAP: Automatic 3D neuron reconstruction based on tracing, reverse mapping and assembling of 2D projections. Neuroinformatics. 2015. doi:10.1007/s12021-015-9278-1.
  42. Santamaria-Pang A, Hernandez-Herrera P, Papadakis M, Saggau P, Kakadiaris I.Automatic morphological reconstruction of neurons from multiphoton and confocal microscopy images using 3D tubular models. Neuroinformatics. 2015; 13(3):297-320.PubMed Abstract | Publisher Full Text OpenURL
  43. Peng HC, Bria A, Zhou Z, Iannello G, Long FH. Extensible visualization and analysis for multidimensional images using Vaa3D. Nat Protoc. 2014; 9(1):193-208.PubMed Abstract | Publisher Full Text OpenURL
  44. Peng HC, Tang JY, Xiao H, Bria A, Zhou JL, Butler V et al. Virtual finger boosts three-dimensional imaging and microsurgery as well as terabyte volume image visualization and analysis. Nat Commun. 2014;5:4342. doi:10.1038/ncomms5342.
  45. Quelhas P, Marcuzzo M, Mendonca AM, Campilho A. Cell nuclei and cytoplasm joint segmentation using the sliding band filter. IEEE Trans Med Imaging. 2010;29(8):1463-73. PubMed Abstract | Publisher Full Text OpenURL
  46. Li HS, Shen T, Smith MB, Fujiwara I, Vavylonis D, Huang XL. Automated actin filament segmentation, tracking and tip elongation measurements based on open active contour models. 2009 IEEE Int Symp Biomed Imaging. 2009; 1 And 2:1302-5. OpenURL
  47. Sethian JA. A review of level set and fast marching methods for image processing.Nato Sci Ser Ii Math. 2002; 75:365-96. OpenURL
  48. Liu L, Bajaj C, Deasy JO, Low DA, Ju T. Surface reconstruction from non-parallel curve networks. Comput Graph Forum. 2008; 27(2):155-63. Publisher Full Text OpenURL
  49. Xu GL, Pan Q, Bajaj CL. Discrete surface modelling using partial differential equations. Comput Aided Geom D. 2006; 23(2):125-45. Publisher Full Text OpenURL
  50. Byun JY, Verardo MR, Sumengen B, Lewis GP, Manjunath BS, Fisher SK. Automated tool for the detection of cell nuclei in digital microscopic images: application to retinal images. Mol Vis. 2006; 12(105-07):949-60. PubMed Abstract | Publisher Full Text OpenURL
  51. Gillette TA, Brown KM, Ascoli GA. The DIADEM metric: comparing multiple reconstructions of the Same Neuron. Neuroinformatics. 2011; 9(2-3):233-45.PubMed Abstract | Publisher Full Text OpenURL
  52. Neuron reconstruction project. http://biometrics.hit.edu.cn/projects/neuron-reconstruction. Accessed 10 Oct 2015.
  53. DIADEM challenge datasets. http://diademchallenge.org/data_sets.html. Accessed 1 Feb 2012.
  54. BigNeuron project datasets. http://alleninstitute.org/bigneuron/data/. Accessed 1 Jan 2015.

 

Development of a novel imaging system for cell therapy in the brain

Maria-Adelaide Micci1*, Debbie R. Boone1, Margaret A. Parsley1, Jingna Wei2, Igor Patrikeev2, Massoud Motamedi2 and Helen L. Hellmich1

Author Affiliations

1Department of Anesthesiology, University of Texas Medical Branch, 301 University Blvd., Galveston 77555, TX, USA

2Center for Biomedical Engineering, University of Texas Medical Branch, 301 University Blvd., Galveston TX 77555, USA

For all author emails, please log on.

Stem Cell Research & Therapy 2015, 6:131  doi:10.1186/s13287-015-0129-7

The electronic version of this article is the complete one and can be found online at:http://stemcellres.com/content/6/1/131

Received: 9 December 2014
Revisions received: 19 May 2015
Accepted: 9 July 2015
Published: 21 July 2015

© 2015 Micci et al.

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Abstract

Introduction

Stem cells have been evaluated as a potential therapeutic approach for several neurological disorders of the central and peripheral nervous system as well as for traumatic brain and spinal cord injury. Currently, the lack of a reliable and safe method to accurately and non-invasively locate the site of implantation and track the migration of stem cells in vivo hampers the development of stem cell therapy and its clinical application.

In this report, we present data that demonstrate the feasibility of using the human sodium iodide symporter (hNIS) as a reporter gene for tracking neural stem cells (NSCs) after transplantation in the brain by using single-photon emission tomography/computed tomography (SPECT/CT) imaging.

Methods

NSCs were isolated from the hippocampus of adult rats (Hipp-NSCs) and transduced with a lentiviral vector containing the hNIS gene. Hipp-NSCs expressing the hNIS (NIS-Hipp-NSCs) were characterized in vitro and in vivo after transplantation in the rat brain and imaged by using technetium-99m ( 99m Tc) and a small rodent SPECT/CT apparatus. Comparisons were made between Hipp-NSCs and NIS-Hipp-NSCs, and statistical analysis was performed by using two-tailed Student’s t test.

Results

Our results show that the expression of the hNIS allows the repeated visualization of NSCs in vivoin the brain by using SPECT/CT imaging and does not affect the ability of Hipp-NSCs to generate neuronal and glial cells in vitro and in vivo.

Conclusions

These data support the use of the hNIS as a reporter gene for non-invasive imaging of NSCs in the brain. The repeated, non-invasive tracking of implanted cells will accelerate the development of effective stem cell therapies for traumatic brain injury and other types of central nervous system injury.

Introduction

During the past decade, several reports have explored the therapeutic potential of neural stem cells (NSCs) for the treatment of neurological disorders of the central nervous system (CNS) (i.e., macular degeneration, Parkinson’s disease, Alzheimer’s disease, and multiple sclerosis) and for traumatic brain and spinal cord injury [1]–[8]. However, the successful development of stem cell therapy and its translation to the clinical setting is currently hampered by the lack of a reliable and safe method to accurately monitor the location, migration, and phenotypical differentiation of transplanted cells. So far, the majority of preclinical studies on stem cell fate have relied on ex vivo histological analysis of green fluorescent protein (GFP) or β-galactosidase expression in the grafted cells. These analyses require the euthanasia of the animals at each time point studied and therefore are laborious and time-consuming. Recently, the development of in vivo non-invasive imaging technologies has provided the means to monitor the delivery, grafting, and survival of stem cells. Current imaging modalities to monitor cells in the brain include magnetic resonance imaging (MRI), single-photon emission tomography (SPECT), positron emission tomography (PET), bioluminescence, and fluorescence imaging [9].

The use of fluorescence and luciferase for bioluminescence imaging is an excellent tool to monitor grafted cells in small animals but is not translatable to the human patient. MRI and SPECT/PET are non-invasive imaging modalities that are suited for human use. Although MRI has a higher spatial resolution than SPECT/PET, the sensitivity of detection is higher for SPECT/PET than MRI (SPECT/PET: 10 −10 –10 −12 M levels of probe; MRI: 10 −3 –10 −5 M levels of probe). Additionally, PET has the ability to detect reporter genes [10], [11].

The use of reporter genes to track stem cell fate in vivo is particularly appealing, as it is the only method that allows studying stem cell survival (only viable cells will be able to express the reporter protein), proliferation (the reporter gene will be passed on to daughter cells, and the corresponding imaging signals will increase in intensity), and death (cells that are apoptotic or dead will not be able to express the reporter protein). Moreover, a reporter gene can be placed under cell-specific promoters (e.g., neuron-specific or glia-specific), thus allowing monitoring of the fate of the transplanted cells within the host tissue [12], [13].

When clinicians decide on which reporter gene to use for imaging stem cells, important factors to be considered are (1) biological distribution of the gene, (2) availability of the probes, and (3) effect of the expression of the reporter gene on the physiology of the cells.

In this respect, the sodium iodide symporter (NIS) reporter gene represents a good choice for imaging stem cells in the brain because (1) it is not expressed in the brain, (2) the radio probes for NIS are readily available in most nuclear medicine clinics and no radio synthesis is required, (3) the metabolism and clearance in the body of both radiodiodide and technetium-99m ( 99m Tc) are well known, and (4) the imaging potentials of NIS have been shown in vitro and in vivo[14], [15].

The use of the NIS to monitor in vivo the delivery, grafting, and phenotypical differentiation of cells after transplantation has recently been investigated in particular in cardiovascular research[16]–[18]. The NIS has also been used to monitor trafficking of immune cells in vivo. Seo et al.reported how transfecting immortalized macrophage cell lines with the hNIS allowed monitoring of their migration toward areas of inflammation in vivo in nude mice by using PET imaging [19].

Up until now, however, no studies have been reported on the use of the NIS for imaging NSCs in the brain in vivo. Here, we report on the development of the use of the NIS for imaging NSCs in the brain.

Methods

Animals

Male Sprague–Dawley rats (200–350 g) were used in all of the experiments described. Experimental protocols were approved by the Institutional Animal Care and Use Committee at the University of Texas Medical Branch, Galveston, in accordance with the guidelines provided by the National Institutes of Health.

Isolation and in vitro expansion of rat hippocampal neural stem cells

Cell culture reagents were obtained from Invitrogen (Carlsbad, CA, USA), except where noted. Adult male Sprague–Dawley rats (200–250 g) were anesthetized with isofluorane (4 % by inhalation) and euthanized by decapitation. The brain was rapidly removed, and the hippocampi were identified and dissected out. The hippocampi from three or four rats were collected into a 50-ml Falcon tube containing sterile Dulbecco’s modified Eagle’s medium/F12 (DMEM/F12) medium with antibiotics (penicillin and streptomycin) and kept on ice. The tissue was minced into small pieces in cold Hanks’ Balanced Salt Solution (HBSS) medium containing 1 mM EGTA in a sterile Petri dish. The tissue pieces were transferred into a sterile 50-ml tube containing HBSS (without Ca2+ and Mg 2+) with 0.1 % collagenase/dispase, 0.01 % DNase I; 1 ml for 100 mg of tissue) and incubated for 30 min at 37 °C. The tissue was triturated every 10 min by using a sterile disposable 5-ml pipette. At the end of the incubation, the cell suspension was centrifuged at 200×g for 5 min at room temperature (RT). The pellet was resuspended in HBSS containing 0.025 % trypsin/ethylenediaminetetraacetic acid and incubated for 10 min at 37 °C. An equal volume of DMEM/F12, 2 mM L-glutamine, 10 % fetal bovine serum (FBS), and antibiotics was added to stop the digestion and the cell suspension was centrifuged at 200×g for 5 min. The pellet was resuspended in sterile DMEM/F12 with L-glutamine, 10 % FBS, and antibiotics; triturated by using a 1-ml tip; and centrifuged. This step was repeated two to four times until a single cell suspension was obtained. The cell suspension was filtered through a 70-μm Falcon filter, centrifuged at 200×gfor 5 min, and resuspended in complete growth medium: neurobasal A medium containing the serum-free supplement B27 (without retinoic acid), 2 mM L-glutamine, 20 ng/ml of epidermal growth factor (EGF), 20 ng/ml of basic fibroblast growth factor-2 (FGF-2), and penicillin/streptomycin. The cells were plated in an uncoated T25 flask at approximately 1×10 6cells per 10 ml of complete growth medium. The medium was changed after 24 h and every other day after that.

Coating of plates with poly-ornithine and laminin

Stock solutions were prepared as follows: poly-L-ornithine (Sigma-Aldrich, St. Louis, MO, USA, catalog number P3655) was reconstituted with sterile distilled water at a concentration of 10 mg/ml, and aliquots were stored at −20 °C; laminin (1 mg/ml; Invitrogen) was stored at −20 °C.

The poly-L-ornithine stock solution was diluted 1:1000 in sterile distilled water to a final working concentration of 10 μg/ml and added to the plates as follow: 1.5 ml/well of six-well plates; 0.2 ml/well of eight-well chambered slides (Lab-Tek II Chambered Slide System; Nalgene Nunc International, Naperville, IL, USA). Plates and slides were incubated overnight at RT and then washed twice with distilled water, incubated with laminin overnight at RT (10 μg/ml in Dulbecco’s phosphate-buffered saline (D-PBS) without Ca 2+ and Mg 2+ ; 1 ml/well of six-well plates and 0.2 ml/well of eight-well chambered slides), and stored at −20 °C until ready to use. The plates were washed with D-PBS once before use.

Induction of differentiation

Hipp-NSCs were dissociated into single cells by using StemPro Accutase (Invitrogen) in accordance with the instructions of the manufacturer and counted on a hemocytometer. The cells were plated onto poly-ornithine/laminin-coated plates as follows: 125,000 cells per well into a six-well plate and 10,000 cells per well into eight-well chambered slides. To induce differentiation, the cells were cultured for 8–10 days in neurobasal A medium containing B27 with retinoic acid, 0.5 mM L-glutamine, 1 % FBS, and penicillin/streptomycin. The medium was changed every other day.

Lentivirus packaging

The full-length human sodium iodide symporter cDNA (hNIS, a kind gift from Sissy M. Jhiang) was cloned into the pCDH-CMV-MCS-EF1-copGFP cDNA cloning and expression lentivector (System Biosciences, Mountain View, CA, USA). The lentivector product was packaged into VSV-G pseudotyped viral particles by System Biosciences (titer 3.12×10 9 infectious units/ml).

Transduction of Hipp-NSCs

Adult rat Hipp-NSCs were dissociated into single cells by using StemPro Accutase (Invitrogen) in accordance with the instructions of the manufacturer and counted on a hemocytometer. The cells were divided into sterile 1.5-ml Eppendorf tubes (100,000 cells per tube in 250 μl of complete growth medium). Pseudovirus particles packaged with the lentivector containing the hNIS were added to each tube at 5, 10, 20, and 40 multiplicity of infection, and the tubes were incubated at 37 °C in a 5 % CO 2 incubator for 30 min. The cells were then transferred to a 24-well plate (250 μl of cells was added to 750 μl of complete growth medium per well) and placed in a 5 % CO 2incubator at 37 °C. After 48 h, the cells were collected into a sterile 15-ml conical tube, centrifuged at 200×g for 5 min, resuspended in complete growth medium, and plated out into a sterile T75 flask. The medium was changed every other day. Expression of GFP (indicative of transduction efficiency) was monitored by using an inverted fluorescent microscope equipped with a 488-nm excitation filter (Nikon Eclipse TS100; Nikon, Tokyo, Japan).

Fluorescence-activated cell sorting

Transduced Hipp-NSCs were dissociated into single cells by using StemPro Accutase, resuspended in sterile D-PBS containing 1 % bovine serum albumin at a concentration of 10–20×10 6 cells/ml, and passed through a 70-μm Falcon filter. The cells were kept on ice and brought to the University of Texas Medical Branch (UTMB) Flow Cytometry Core Facility, where they were run through a FACSAria cell sorter (BD Biosciences, San Jose, CA, USA). The cells were sorted on the basis of the expression of GFP and collected into a 15-ml tube containing 1 ml of complete growth medium with 10 mM HEPES. The collected cells were centrifuged and replated in fresh complete growth medium.

In vitro uptake of technetium-99m

99m Tc was obtained from the Department of Nuclear Medicine at UTMB. Cells (Hipp-NSCs expressing the hNIS and naïve Hipp-NSCs not transduced) were plated out into a six-well plate at a density of 500,000 cells/ml (1.5×10 6 cells per well) in complete growth medium. The cells were incubated with 99m Tc (3–5 μCi/well) for 30 min at 37 °C in a 5 % CO 2 incubator and then washed with ice-cold PBS two times by collection and centrifugation at 200×g for 5 min. At the end of the last centrifugation, the cells were lysed in 500 μl of NaOH (0.33 M) containing 1 % SDS. Counts per minute (CPMs) were read on a gamma counter. Baseline readings were obtained from tubes containing NaOH-1 % SDS and no cells. Data were corrected by subtracting the baseline reading averaged from three separate measurements from the sample CPMs.

Cell proliferation assay

Cell proliferation was measured by using the CellTiter 96 Aqueous non-radioactive cell proliferation assay (Promega Corporation, Madison, WI, USA). This assay is based on the principle that a tetrazolium compound ([3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium, inner salt] (MTS)) is converted by viable cells into a soluble formazan product that can be measured by reading the absorbance at 490 nm. Briefly, Hipp-NSCs and Hipp-NSCs transduced with the hNIS were dissociated into single cells by using StemPro Accutase, plated out into a 96-well plate (10,000 and 20,000 cells per well) in complete growth medium (100 μl/well), and cultured for 24–48 h. At the desired time point, 20 μl of MTS/PMS solution (prepared in accordance with the instructions of the manufacturer) was added to each well and the plate was incubated for 3–4 h at 37 °C in a humidified incubator, 5 % CO 2 atmosphere. The plate was read in an enzyme-linked immunosorbent assay plate reader (Molecular Devices, Sunnyvale, CA, USA) at 490 nm. Background absorbance was measured from wells that were loaded with complete growth medium only and to which MTS/PMS was added.

Total protein extraction and Western Blot Analysis

Cells growing as neurospheres in complete growth medium were collected in a 15-ml conical tube and centrifuged at 200×g for 5 min at RT. The cells were resuspended in D-PBS and centrifuged again. RIPA lysis buffer containing protease inhibitor cocktail, phosphatase inhibitors cocktail, and 100 μM phenylmethylsulfonyl fluoride (PMSF) (all from Sigma-Aldrich) was added to the cell pellet (150–250 μl of lysis buffer for 2–5×10 6 cells). The lysed cells were transferred to 1.5-ml Eppendorf tubes, vortexed, and incubated on ice for 5 min and then quickly frozen in liquid nitrogen and stored at −80 °C. Cells growing on poly-ornithine/laminin-coated plates were detached by using StemPro Accutase, collected into 15-ml tubes, and centrifuged at 200×g for 5 min. The cell pellet was washed in D-PBS and lysed in RIPA lysis buffer as described above. Total protein content was measured by using a Pierce BCA protein assay kit (Thermo Scientific, Rockford, IL, USA).

The protein samples were processed for SDS-polyacrylamide gel electrophoresis (PAGE) with XCell SureLock® Mini-Cell (Invitrogen). After electrophoresis, proteins were transferred to polyvinylidene difluoride (PVDF) membranes (Bio-Rad Laboratories, Hercules, CA, USA) overnight at 4 °C. Blots were incubated in blocking buffer (5 % non-fat dry milk in Tris-buffered saline containing 0.1 % Tween 20; TBS-T) for 1 h at RT and probed with primary antibodies diluted in blocking buffer overnight at 4 °C. After three washes in TBS-T, the blots were incubated with horseradish peroxidase-conjugated secondary antibodies (Cell Signaling Technology, Inc., Danvers, MA, USA; 1:200 dilution in blocking buffer) for 1 h at RT.

Detection was performed using ECL plus kit (Amersham, GE Healthcare, Little Chalfont, UK) followed by exposure of x-ray films (Bioexpress, Kaysville, UT, USA) that were developed with M35A-X-OMAT Processor (Kodak, Rochester, NY, USA). Band intensities were quantified by using ImageJ software, and data were normalized to the expression of the housekeeping protein GAPDH (glyceraldehyde 3-phosphate dehydrogenase).

Immunofluorescence analysis

Cells growing on poly-ornithine/laminin-coated chambered slides were washed in PBS (pH 7.4) and fixed in 4 % paraformaldehyde (pH 7.4 in PBS) for 15 min at RT. After two washes in PBS, the cells were blocked and permeabilized in PBS containing 5 % normal goat serum and 0.3 % triton X-100 for 30 min at RT. The cells were incubated with primary antibodies (Table 1) diluted in 1.5 % normal goat serum in PBS overnight at 4 °C in a humid chamber. After three washes in PBS, the cells were incubated with secondary antibodies (1:400; Alexa-conjugated anti-mouse and anti-rabbit; Invitrogen) in 1.5 % normal goat serum in PBS for 1 h at RT in a humid chamber. The cells were washed two times in PBS and once in tap water before being coverslipped with mounting medium containing DAPI (4′,6-diamidino-2-phenylindole), a fluorescent stain that binds strongly to A-T-rich regions in DNA and is used to label cell nuclei (Vector laboratories, Burlingame, CA, USA). The slides were viewed with an Olympus BX51 fluorescence microscope equipped with a cooled charge-coupled device camera (Microfire; Optronics, Goleta, CA, USA) and image acquisition software (PictureFrame; Optronics).

Table 1. List of antibodies used for the phenotypical characterization of Hipp-NSCs and hNIS-Hipp-NSCs

Injection of Hipp-NSCs in the hippocampus

Hipp-NSCs were labeled with the cell tracker CMFDA (Invitrogen), suspended at a concentration of 50,000 cells/μl in D-PBS, and kept on ice until ready to inject. Experimental recipient adult male rats were deeply anesthetized with isoflurane (2–4 %) and intubated with an endotracheal tube (6.5-French tubing attached to 16-gauge needle adapter) attached to a ventilator (tidal volume set at 12–15 cc/kg and breath rate set at approximately 28–30 breaths per minute). The anesthetized rats were placed at the flat-skull position on a small animal stereotaxic apparatus, and a craniotomy was performed. The cells were injected by using a 33-gauge cannulae supported by the outside 26-gauge guide cannulae connected with a Hamilton syringe. The cells were injected into the hippocampal CA1—anterior-posterior (AP): −3.6, medial-lateral (ML): +2.0, dorsal-ventral (DV): −2.6—and CA3 (AP: −3.6, ML: +3.6, DV: −3.6) regions (1 μl/site at a speed of 0.2 μl/min). The number of cells injected was chosen on the basis of the work by Gao et al.[20].

Injection of Hipp-NSCs in the lateral ventricles

Hipp-NSCs were suspended in D-PBS at a concentration of 2×10 5 cells/μl. Experimental recipient adult male rats were deeply anesthetized with isoflurane (2–4 %) and intubated with an endotracheal tube (6.5-French tubing attached to 16-gauge needle adapter) attached to a ventilator (tidal volume set at 12–15 cc/kg and breath rate set at approximately 28–30 breaths per minute). The anesthetized rats were placed at the flat-skull position on a small animal stereotaxic apparatus, and a craniotomy was performed. The injection site was drilled by using a Dremel drill 1.5 mm posterior to bregma and 1.2 mm to lateral right of sagittal suture for ventricular pressure verification, and injection of cells was performed (5 μl; 0.2 μl/min).

SPECT/CT imaging

Imaging was performed by using an Inveon™ PET-SPECT-CT system (Siemens Medical Solutions, Knoxville, TN, USA) equipped with two gamma cameras (for SPECT modality) and x-ray source and camera for CT imaging mounted on a rotating stage. The PET detector is mounted separately and was not used in this study. This scanner combines PET, SPECT, and CT modalities in one system under the control of a single workstation that allows integrated data acquisition and processing. CT imaging provides mostly anatomical reference information (where the gamma signal comes from), and SPECT provides functional information showing the intensity and the distribution of the isotope in the body. Injected in the animal, a gamma-emitted isotope serves as a source of signal, and the signal is detectable by the SPECT cameras. Since the isotope (in our study, 99m Tc) is specifically attaching to the cells with NIS (naturally, to the thyroid cells; in our study, to the modified stem cells) and is not attaching to the other cells, it provides the imaging contrast between the signal (coming from the cells of interest) and background. Cameras are rotated around the animal and the signal collected for 3D reconstruction. In general, we followed the same CT/SPECT imaging as described by Terrovitis et al.[16].

In vivo SPECT imaging

The rats were anesthetized with isofluorane and positioned on the bed of the SPECT/CT module with their nose inside a mask connected to the isofluorane dispenser. High-resolution CT scans were performed before each SPECT session. The 99m Tc signal was co-registered to the CT scan image by using data processing software Inveon Research Workspace (Siemens Medical Solutions). CT parameters were 2048×3072 pixels, 70 kV, 500 μA, and 520 steps, giving a 3D isotropic resolution of approximately 0.1 mm. SPECT parameters were 40 projections and 9-degree step, for a total of about 16 min for one scan.

Intracerebral cannulation for delivery of 99m Tc in the brain

The rats were anesthetized with isoflurane (2–4 %) and intubated with an endotracheal tube (6.5-French tubing attached to 16-gauge needle adapter) attached to a ventilator (tidal volume set at 12–15 cc/kg and breath rate set at approximately 28–30 breaths per minute). The anesthetized rats were placed at the flat-skull position on a small animal stereotaxic apparatus, and a craniotomy was performed. A microdialysis CMA guide cannula/probe (BASi, West Lafayette, IN, USA) was inserted into the hippocampus (Bregma: −4.3 mm; ML: −3.5 mm; Dura: −3.2 mm).

Tissue processing and immunofluorescence analysis

At the desired time point after cell transplantation, the rats were anesthetized with pentobarbital and cardially perfused with heparinized saline followed by ice-cold 4 % paraformaldehyde (pH 7.4). The brains were removed and embedded in 20 % sucrose in PBS (pH 7.4). Sections (20 μm) were cut on a cryostat, collected on glass microscope slides (Superfrost Plus; Thermo Fisher Scientific Inc., Marietta, OH, USA), and stored at −20 °C. For immunofluorescence analysis of grafted cells, the sections were hydrated in PBS and incubated in PBS containing 10 % normal goat serum and 0.3 % Triton X-100 for 30 min at RT. The sections were incubated with primary antibodies diluted in PBS containing 1.5 % normal goat serum, overnight at 4 °C and with secondary antibodies (594 Alexa-conjugated, Invitrogen; 1:400 dilution in PBS with 1.5 % normal goat serum) for 1 h at RT. After washing in PBS, the sections were rinsed in tap water and coverslipped with mounting medium with DAPI (Vector Laboratories).

Statistical analysis

Data were expressed as the mean ± standard error. Statistical analysis between two independent experimental groups was performed by using two-tailed Student’s t test. Results were considered significant for P values of less than 0.05.

Results

Expansion and characterization of Hipp-NSCs

NSCs were isolated from the hippocampus of adult rats and maintained in culture in serum-free neurobasal A medium supplemented with B27 and the growth factors EGF and FGF-2 (complete growth medium). Under these conditions, the cells actively proliferate and form a characteristic spheroid group of cells that express the NSC marker nestin (Fig. 1a, b). NSCs are defined not only by their ability to proliferate and generate daughter cells that are unspecialized and express specific stem cell markers but also by their ability to generate neurons and glia. When Hipp-NSCs were plated onto poly-ornithine/laminin-coated plates in serum-free medium containing B27, retinoic acid, and no growth factors (EGF and FGF-2), they attached to the substrate and underwent rapid morphological changes (Fig. 1c). Immuncytochemical analysis of cells grown for 8 days under differentiating conditions confirmed that both neurons and glia were generated from Hipp-NSCs (Fig. 1d).

thumbnailFig. 1. Hipp-NSCs form neurospheres and generate neurons and glia in vitro. a A typical neurosphere formed by Hipp-NSCs growing in serum-free medium with EGF and FGF-2. b Immunofluorescence analysis of nestin expression (red) in Hipp-NSC neurospheres. Nuclei are stained blue with DAPI. c Phase-contrast photograph of Hipp-NSCs growing on poly-ornithine-coated plates in differentiation medium for 8 days. dImmunofluorescence analysis of Hipp-NSCs after 8 days of differentiation showing the expression of the neuronal marker βIII-tubulin (red) and of the glial marker GFAP (green). Nuclei are stained blue with DAPI. Scale bars = 50 μm. DAPI4′,6-diamidino-2-phenylindole, EGF epidermal growth factor, FGF-2 fibroblast growth factor-2,GFAP glial fibrillary acidic protein, Hipp-NSC hippocampus-derived neural stem cell

Generation of Hipp-NSCs expressing the hNIS

Hipp-NSCs were transduced by using pseudoviral particles packaged with a lentivector containing the hNIS and the reporter gene GFP (Fig. 2a, b). The expression of GFP and fluorescence-activated cell sorting was used to select out Hipp-NSCs expressing the hNIS (Fig. 2c, d). The expression of the hNIS in the sorted cells was confirmed by immunocytochemical analysis (Fig. 2e). We also determined whether the expression of the hNIS would be retained after differentiation. Hipp-NSCs transduced with the hNIS were cultured onto poly-ornithine/laminin-coated plates in differentiation medium for 7 days. The hNIS was maintained in the differentiated cells as shown by the expression of GFP and by immunofluorescence analysis by using a specific antibody against hNIS (Fig. 2f, g).

thumbnailFig. 2. Generation of Hipp-NSCs expressing the hNIS. a Map of the HIV-based lentivector CD11B-1 (System Biosciences). The hNIS cDNA was cloned in the multiple cloning site (MSC) located downstream of the CMV promoter. Downstream of the expression cassette for the hNIS, an EF1 promoter drives the expression of GFP. b Expression of GFP in Hipp-NSCs transduced with pseudoviral particles for 48 h. The fluorescent image is superimposed on the phase-contrast image. c Flow cytometric analysis of Hipp-NSCs transduced with the hNIS showing the expression of GFP (indicative of the expression of hNIS) in 62 % of the cells before sorting and in 96 % of the cells after sorting (d). e Immunofluorescence analysis of the expression of the hNIS (in red) in Hipp-NSCs transduced with the lentivector and selected by sorting for GFP. Nuclei are stained blue with DAPI. f Expression of GFP in differentiated Hipp-NSCs transduced with the hNIS. g Immunoreactivity for the hNIS in differentiated NIS-Hipp-NSCs. Nuclei are counterstained blue with DAPI. h Uptake of 99m Tc in NIS-Hipp-NSCs was measured as counts per minute (CPM) in a gamma counter and corrected by subtracting background CPM. N = 3; *P < 0.01 by Student’s t test. Scale bars = 100 μm. CMVcytomegalovirus, DAPI 4′,6-diamidino-2-phenylindole, GFP green fluorescent protein, Hipp-NSC hippocampus-derived neural stem cell, hNIS human sodium iodide symporter, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter,99m Tc technetium-99m

To determine whether the hNIS expressed by Hipp-NSCs is functional, we tested the ability of the transduced cells to transport 99m Tc, a gamma-emitting radioisotope that is used by the NIS instead of endogenous iodide. A significant increase in CPM was measured in Hipp-NSCs expressing the hNIS as compared with control Hipp-NSCs not expressing the hNIS and exposed to99m Tc (Fig. 2h).

Characterization of Hipp-NSCs expressing the hNIS in vitro

To assess the effect of the hNIS on cell viability, Hipp-NSCs were plated in a 96-well plate in neurobasal A medium containing EGF and FGF (complete growth medium) at two different densities (10,000 and 20,000 cells per well) and cultured for 24 h. To measure proliferation, the cells were plated out at a density of 10,000 cells per well and cultured for 24 and 48 h. Cell viability was assessed by using the MTS assay (CellTiter 96 Aqueous non-radioactive cell proliferation assay; Promega Corporation). No significant differences were detected between Hipp-NSCs expressing the hNIS (NIS-Hipp-NSCs) and Hipp-NSCs in their naïve form (not expressing the hNIS) (Fig. 3a, b). To test the effect of 99m Tc incorporation on cell viability, Hipp-NSCs and NIS-Hipp-NSCs were exposed to 99m Tc (100 μCi added to the cell culture medium) for 30 min. The cells were washed, resuspended in proliferation medium, plated out in a 96-well plate (10,000 cells per well), and cultured for 48 h. Cell viability was assessed 48 h later by using the MTS assay (CellTiter 96 Aqueous non-radioactive cell proliferation assay). No significant differences were detected between Hipp-NSCs and NIS-Hipp-NSCs, indicating that the incorporation of 99m Tc in NIS-Hipp-NSCs does not reduce cell viability. In addition to studying the effect of the hNIS on Hipp-NSC viability and proliferation, we studied whether NIS-Hipp-NSCs retain their characteristic stem cell signature. This was done by analyzing the expression of NSC markers by Western blot analysis. There was no difference in the expression of nestin and SOX2 between Hipp-NSCs transduced with the hNIS (NIS-Hipp-NSCs) and naïve Hipp-NSCs (Fig. 3c, d).

thumbnailFig. 3. NIS-Hipp-NSCs are viable, proliferate, and express NSC markers. MTS-based analysis of cell viability (a) and proliferation (b) of Hipp-NSCs transduced with the hNIS (NIS-Hipp-NSCs) compared with naïve Hipp-NSCs. N = 3; P > 0.05 by Student’s t test. c Hipp-NSCs and NIS-Hipp-NSCs were exposed to 99m Tc for 30 min. MTS-based analysis of cell viability was performed 48 h later. N = 3; P > 0.05 by Student’s t test. dWestern blot analysis of total protein extracts from NIS-Hipp-NSCs and naïve Hipp-NSCs grown under proliferating conditions. e Densitometric analysis. The expression of GAPDH was used to normalize the densitometry value index (DVI). N = 3; P > 0.05 by Student’s t test. GAPDH glyceraldehyde 3-phosphate dehydrogenase, Hipp-NSC hippocampus-derived neural stem cell, hNIS human sodium iodide symporter, MTS 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium, inner salt, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter, NSC neural stem cell, 99m Tc technetium-99m

Differentiation potential of Hipp-NSCs expressing the hNIS

To test the effect of hNIS on the ability of Hipp-NSCs to generate neurons and glia, the cells were plated out onto poly-ornithine/laminin-coated plates and cultured for 8 days in neurobasal medium containing B27 with retinoic acid and 1 % FBS (differentiation medium). Neuronal and glial differentiation was assessed by immunofluorescence and Western blot analysis. The results show that the expression of hNIS does not affect the ability of Hipp-NSCs to generate neurons and glia in vitro (Fig. 4).

thumbnailFig. 4. Differentiation potential of NIS-Hipp-NSCs. a Immunofluorescence analysis of Hipp-NSCs transduced with the hNIS (NIS-Hipp-NSCs) and of naïve Hipp-NSCs cultured in differentiation medium for 7 days. Nuclei are counterstained blue with DAPI. b Western blot analysis of total protein extracts from differentiated NIS-Hipp-NSCs and naïve Hipp-NSCs. The expression of GAPDH was used to normalize the densitometry value index (DVI). N = 3; *P < 0.05 by Student’s t test. DAPI 4′,6-diamidino-2-phenylindole, GAPDH glyceraldehyde 3-phosphate dehydrogenase, Hipp-NSC hippocampus-derived neural stem cell, hNIS human sodium iodide symporter, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter

SPECT/CT imaging of Hipp-NSCs expressing the hNIS

In vitro imaging

Hippocampus-derived NSCs can be visualized by using SPECT imaging by providing 99m Tc to the cells and measuring the gamma-ray emission from the cells that have taken up the 99m Tc via the NIS. Hipp-NSCs stably expressing the NIS were incubated in vitro with 99m Tc and imaged by placing the cells in an Eppendorf tube inside the SPECT/CT imaging module. Control cells (Hipp-NSCs not expressing the NIS) were also incubated with 99m Tc and imaged (Fig. 5a).

thumbnailFig. 5. SPECT imaging of NIS-Hipp-NSCs. a SPECT signal from hNIS-Hipp-NSCs and Hipp-NSCs that were incubated with 99m Tc for 30 min and washed in phosphate-buffered saline before imaging. The SPECT signal is shown superimposed on the CT scan image. b SPECT signal recorded from NIS-Hipp-NSCs that were incubated with 99m Tc in vitro before transplantation in the right lateral ventricle. Transverse, coronal, and sagittal orientations of the brain are shown. c Analysis of the brain shows grafted cells (identified by the expression of GFP) in the ventricle. The expression of the NIS in the grafted cells was confirmed by immunofluorescence by using an antibody against hNIS (inset). CTcomputed tomography, GFP green fluorescent protein, Hipp-NSC hippocampus-derived neural stem cell, hNIS human sodium iodide symporter, NIS sodium iodide symporter, NIS-Hipp-NSChippocampus-derived neural stem cell expressing the human sodium iodide symporter,SPECT single-photon emission tomography, 99m Tc technetium-99m

In vivo imaging

To assess whether grafted cells can be visualized by using SPECT/CT imaging in the live animal, in a first set of experiments we incubated NIS-Hipp-NSCs in vitro with 99m Tc before transplantation in the rat brain. The cells were injected in the left lateral ventricle (1×10 6 cells) (Fig. 5b) or in the hippocampus (two injection sites per hippocampus; amount of cells per site ranging from 440,000 to 55,000 cells) (Fig. 6a), and the rats were subjected to SPECT imaging. The SPECT signal was superimposed to the CT scan to localize the cells within the brain. Quantification of the signal in the hippocampus was performed by measuring the average number of pixels in the region of interest by using image analysis software (Image J) and correlated to the number of cells injected (Fig. 6b). Post-imaging analysis of the rat brains confirmed the presence of the cells by immunofluorescence analysis (Fig. 6c).

thumbnailFig. 6. In vivo SPECT imaging of NIS-Hipp-NSCs injected in the brain hippocampus. a SPECT signal recorded from NIS-Hipp-NSCs that were incubated with 99m Tc in vitro before transplantation in the hippocampus (two sites/hippocampus). Transverse, coronal, and sagittal orientations of the brain are shown. b Correlation between the number of cells injected in the hippocampus and the signal detected by SPECT imaging. Correlation coefficient r = 0.9566; two-tailed P value = 0.0108. c Analysis of the brain shows grafted cells (identified by the expression of GFP) in the hippocampus. The expression of the NIS in the grafted cells was confirmed by immunofluorescence by using an antibody against hNIS (inset). GFP green fluorescent protein, Hipp-NSC hippocampus-derived neural stem cell, hNIS human sodium iodide symporter, NIS sodium iodide symporter, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter, SPECT single-photon emission tomography, 99m Tc technetium-99m

In a second set of experiments, NIS-Hipp-NSCs were injected through an intracranial cannula implanted in the rat brain. For these experiments, NIS-Hipp-NSCs were resuspended in PBS and injected in the brain (1–2×10 5 cells in 5 μl at 0.2 μl/min). At various time points after cell injection (2 and 24 h), 99m Tc was delivered to the brain via the cannula (0.5 mCi in 5 μl at a speed of 0.5 μl/min), and the rats were subjected to SPECT/CT imaging 4 h later (Fig. 7).

thumbnailFig. 7. Repetitive in vivo SPECT imaging of NIS-Hipp-NSCs. a CT scan of the rat head showing the placement of the cannula used to deliver 99m Tc into the brain. b SPECT imaging of the brain of a rat that had received99m Tc and no cells. c SPECT signal recorded from the rat brain 2 and 24 h after NIS-Hipp-NSC intracranial injection following delivering of 99m Tc. Transverse and coronal slice orientations are shown. CT computed tomography, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter, SPECT single-photon emission tomography, 99m Tc technetium-99m

These data demonstrate that Hipp-NSCs expressing the hNIS can be visualized in vivo in the rat brain. Specifically, these data show adequate spatial resolution and signal detection sensitivity of the SPECT imaging system and the feasibility to repetitively image grafted NSCs in the rat brain.

Characterization of Hipp-NSCs expressing the hNIS after transplantation in the rat brain

To determine whether the expression of the hNIS would alter the ability of Hipp-NSCs to graft and survive after transplantation in the intact rat brain, Hipp-NSCs expressing the hNIS were injected in the hippocampus of adult rats and the animals were euthanized 2 weeks later. Control groups consisted of rats that were injected with Hipp-NSCs without the hNIS. Grafted cells in the brain were localized on the basis of the presence of CMFDA (a cell tracker dye that was loaded in the cells before transplantation). Both Hipp-NSCs and Hipp-NSCs expressing the hNIS were found grafted in the hippocampus 2 weeks after transplantation. Grafted cells were localized in the pyramidal layer (CA1) and some cells were also visible between the CA1 and granular layer. Phenotypical characterization of the grafted cells performed by immunofluorescence analysis revealed that some of the grafted cells had differentiated into mature neuronal cells expressing NeuN (Fig. 8).

thumbnailFig. 8. In vivo SPECT imaging of NIS-Hipp-NSCs injected in the brain hippocampus. Representative cross-sections of rat brain 2 weeks after transplantation of NIS-Hipp-NSCs or Hipp-NSCs in the hippocampus (CA1 region). Grafted cells are located on the basis of the expression of the cell tracker dye CMFDA. Mature neurons are identified on the basis of the expression of NeuN. Merged images show that some of the grafted cells express the neuronal marker NeuN (arrows). Cell nuclei are stained blue with DAPI. Scale bar = 50 μm. DAPI 4′,6-diamidino-2-phenylindole, Hipp-NSC hippocampus-derived neural stem cell, NIS-Hipp-NSC hippocampus-derived neural stem cell expressing the human sodium iodide symporter, SPECT single-photon emission tomography

Discussion

NIS expression in NSCs

The first objective of this work was to determine the effect of NIS expression on the biology of NSCs. The analysis of the effect of the expression of NIS on Hipp-NSCs was directed to both the basic biological properties as well as to their ability to respond to the host brain tissue in vivo. Our data show that the expression of NIS in Hipp-NSCs does not change their proliferation rate in vitro. The ability to proliferate is a key functional signature of NSCs and is of particular importance in clinical applications because in vitro NSC expansion is necessary to generate a clinically useful number of cells for transplantation. Biochemical expression of proteins that are characteristically found in NSCs (SOX2 and nestin) was also not affected by the NIS. Although the expression of SOX2 and nestin is commonly used to identify NSCs, a more complete analysis of their biochemical profile and of their karyotype should be performed in future studies. Nonetheless, the results shown in this work demonstrate that the exogenous expression of NIS in Hipp-NSCs does not alter the basic biological properties of the cells. This is in agreement with other studies that have characterized the use of the NIS in other types of stem cells.

Possibly the most important feature of NSCs, and one that makes these cells so appealing for use in replacement therapy of the nervous system, is their ability to generate neurons and glia. This important feature is maintained in Hipp-NSCs that express the NIS, as demonstrated by the expression of proteins that are known markers of differentiated phenotypes—βIII-tubulin and NeuN for neurons and glial fibrillary acidic protein (GFAP) and oligodendrocyte-specific protein for glia—analyzed by using both immunofluorescence and Western blot techniques. Additionally, no differences were found between Hipp-NSCs expressing the NIS and naïve Hipp-NSCs when comparing the expression of neuronal and glial markers by Western blot analysis. This is true bothin vitro as well as in vivo in the rat brain, where Hipp-NSCs expressing the NIS where found grafted in the hippocampus 2 weeks after transplantation with at least some cells showing expression of neuronal markers. Although the studies presented in this report are limited to a 2-week follow-up, a large body of literature testifies to the safety of 99m Tc accumulation in cells in vivo[16],[21]–[23]. Specifically, NIS-based imaging has been validated and used in humans [24], [25], thus supporting the translational potential of this reporter system.

Molecular imaging of NSCs using the NIS

The main goal of this work was to develop an imaging system that will allow tracking NSCs after transplantation in the brain. We have provided several pieces of evidence, in vitro and in vivo, that the expression of the reporter gene hNIS allows the visualization of live and viable Hipp-NSCs by using SPECT imaging. We have first demonstrated that the expression of the hNIS in Hipp-NSCs is functional in that it allows for the selective uptake of 99m Tc inside the cells. 99m Tc is a gamma-emitting radioisotope that is taken up by the NIS in place of iodide (the endogenous, biological ligand of the NIS). Moreover, 99m Tc is widely available and currently approved by the US Food and Drug Administration in combination with SPECT imaging for diagnostic purposes.

In the in vivo experiments, 99m Tc was delivered directly in the brain parenchyma where the cells were grafted. The results demonstrate that the Hipp-NSCs grafted in the brain are viable, can take up 99m Tc via the NIS, and can be visualized by SPECT imaging.

A successful imaging system needs to have sufficient signal sensitivity to detect grafted cells and sufficient spatial resolution to determine their location within the tissue/organ. The data presented in this article demonstrate that NIS combined with SPECT/CT imaging possesses enough spatial resolution and sensitivity to detect viable grafted cells in the brain. Specifically, when Hipp-NSCs were injected at different sites within the hippocampus and at different densities, discrete regions of positive signals were detected and a correlation between the signal intensity and the number of cells injected could be derived. Another important result of this work is the demonstration that Hipp-NSCs can be visualized repetitively over time on the same rat. This is a critical aspect of a reporter gene imaging system because, by allowing the visualization of grafted cells over time and for as long as they are viable, it sets this system apart from other imaging methodologies.

Limitations

One limitation of using the hNIS reporter system for imaging cells in the brain parenchyma is the relatively low permeability of 99m Tc through the blood–brain barrier (BBB). In this report, therefore, we opted to use an intracranial cannula to deliver 99m Tc directly to the brain in order to bypass any potential issue with 99m Tc penetration through the BBB. This allowed us to provide proof-of-principle data demonstrating the feasibility of tracking grafted NSCs in the brain by using the NIS reporter system and SPECT detection. More clinically relevant and non-invasive delivery routes for 99m Tc have been reported, including intravenous injections to image intracranial gliomas in experimental animals [26], [27], injection into the olfactory route [28], [29], encapsulation of 99m Tc into liposomes, or nanoparticles targeted to the CNS via specific BBB transporters or receptor systems [30]–[32]. In future studies, we will test these routes of delivery; however, even as it stands, our data show that delivery of 99m Tc via implanted intracranial cannulas is a valuable method for tracking NSCs in the brain in preclinical animal models of disease that will allow researchers to follow cell grafts over time in the same animals. This not only will provide valuable information on the location and viability of the cells and their biological activity but also will significantly shorten the time and reduce the number of experimental animals needed to perform the studies, thus expediting the translation of stem cell therapy to the clinical setting.

Although here we used NSCs isolated from the hippocampus of adult rats (Hipp-NSCs), it has to be noted that more easily accessible sources of NSCs (i.e., induced pluripotent stem cells, embryonic stem cell-derived NSCs, and NSCs derived from mesenchymal stem cells isolated from bone marrow, fat, or umbilical cord) have been described and are clinically attractive sources of cells as they allow autologous transplantation, an appealing approach that eliminates the need for immunosuppressive therapies. Because NSCs share common basic biological properties, the results of this work are significant, as they lay the groundwork for the use of the NIS to image NSCs and can very likely be extrapolated to other sources of NSCs.

Additionally, it is important to notice that in this study the cells were transplanted in naïve/uninjured rats. The response of Hipp-NSCs to an injured (i.e., after traumatic brain injury) or diseased brain will likely differ. However, a detailed analysis of grafting and differentiation potentials of Hipp-NSCs in the injured and uninjured brain goes beyond the scope of this work and will be the focus of future studies.

Conclusions

In this report, we have characterized the NIS as a reporter gene for imaging NSCs in the brain. Developing non-invasive imaging of grafted cells has emerged as a fundamental tool to advance the field of stem cell therapy. The ability to track cells in vivo after transplantation on the same animal over time will allow a reduction in the number of experimental animals needed to determine the optimal source of cells as well as the best route and site of delivery that will produce a functional recovery in animal models of disease. This will significantly shorten the time between experimental animal work and translation to the clinical setting.

Abbreviations

AP: Anterior-posterior

BBB: Blood–brain barrier

CNS: Central nervous system

CPM: Counts per minute

CT: Computed tomography

DAPI: 4′,6-diamidino-2-phenylindole

DMEM: Dulbecco’s modified Eagle’s medium

D-PBS: Dulbecco’s phosphate-buffered saline

DV: Dorsal-ventral

EGF: Epidermal growth factor

FBS: Fetal bovine serum

FGF-2: Fibroblast growth factor-2

GFP: Green fluorescent protein

HBSS: Hanks’ Balanced Salt Solution

Hipp-NSC: Hippocampus-derived neural stem cell

hNIS: Human sodium iodide symporter

ML: Medial-lateral

MRI: Magnetic resonance imaging

MTS: 3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium, inner salt

NIS: Sodium iodide symporter

NIS-Hipp-NSC: Hippocampus-derived neural stem cell expressing the human sodium iodide symporter

NSC: Neural stem cell

PBS: Phosphate-buffered saline

PET: Positron emission tomography

PMS: Phenazine methosulfate

SPECT: Single-photon emission tomography

TBS-T: Tris-buffered saline containing 0.1 % Tween 20

Tc: Technetium-99m

UTMB: University of Texas Medical Branch

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

M-AM designed and planned the experiments, isolated and cultured Hipp-NSCs, performed immunofluorescence and Western blotting analysis, analyzed the data, and wrote the manuscript. DRB performed the cloning of the hNIS into the lentivector. MAP performed the cell transplantation procedures and cannulated the rats. JW operated the SPECT/CT imaging system. IP operated the SPECT/CT imaging system and calculated required 99m Tc dose. MM participated in the design of the imaging studies. HLH participated in the design of the study and in the cloning of the hNIS into the lentivector. All authors read and approved the final manuscript.

Acknowledgments

This study was supported by grant 1UL1RR029876-01 from the National Center for Research Resources, National Institutes of Health and by the Moody Project for Translational Traumatic Brain Injury Research. We thank Sissy M. Jhiang for providing the full-length hNIS cDNA, Lori Follis for preparing 99m Tc, Elizabeth Bishop for helping with the cell culture of Hipp-NSCs, Donald J. Deyo for providing valuable expertise and assistance with the intraventricular injection procedure, and Christine Courteau-Butler for editorial support.

References

  1. Cummings BJ, Uchida N, Tamaki SJ, Salazar DL, Hooshmand M, Summers R, et al.: Human neural stem cells differentiate and promote locomotor recovery in spinal cord-injured mice.

    Proc Natl Acad Sci U S A 2005, 102:14069-74. PubMed Abstract | Publisher Full Text OpenURL

  2. Jain KK: Cell therapy for CNS trauma.

    Mol Biotechnol 2009, 42:367-76. PubMed Abstract | Publisher Full Text OpenURL

  3. Pluchino S, Zanotti L, Deleidi M, Martino G: Neural stem cells and their use as therapeutic tool in neurological disorders. Brain research.

    Brain Res Rev 2005, 48:211-9. PubMed Abstract | Publisher Full Text OpenURL

  4. Martino G, Pluchino S: The therapeutic potential of neural stem cells.

    Nat Rev Neurosci 2006, 7:395-406. PubMed Abstract | Publisher Full Text OpenURL

  5. Ourednik J, Ourednik V: Graft-induced plasticity in the mammalian host CNS.

    Cell Transplant 2004, 13:307-18. PubMed Abstract | Publisher Full Text OpenURL

  6. Ourednik V, Ourednik J: Graft/host relationships in the developing and regenerating CNS of mammals.

    Ann N Y Acad Sci 2005, 1049:172-84. PubMed Abstract | Publisher Full Text OpenURL

  7. Conti L, Reitano E, Cattaneo E: Neural stem cell systems: diversities and properties after transplantation in animal models of diseases.

    Brain Pathol 2006, 16:143-54. PubMed Abstract | Publisher Full Text OpenURL

  8. Baghbaderani BA, Mukhida K, Hong M, Mendez I, Behie LA: A review of bioreactor protocols for human neural precursor cell expansion in preparation for clinical trials.

    Curr Stem Cell Res Ther 2011, 6:229-54. PubMed Abstract | Publisher Full Text OpenURL

  9. Zhao C, Tian M, Zhang H: In vivo stem cell imaging.

    Open Nucl Med J 2010, 2:171-7. Publisher Full Text OpenURL

  10. Schaller B: Usefulness of positron emission tomography in diagnosis and treatment follow-up of brain tumors.

    Neurobiol Dis 2004, 15:437-48. PubMed Abstract | Publisher Full Text OpenURL

  11. Spiriev T, Sandu N, Schaller B: Molecular imaging and tracking stem cells in neurosciences.

    Methods Mol Biol 2013, 1052:195-201. PubMed Abstract | Publisher Full Text OpenURL

  12. Brader P, Serganova I, Blasberg RG: Noninvasive molecular imaging using reporter genes.

    J Nucl Med 2013, 54:167-72. PubMed Abstract | Publisher Full Text OpenURL

  13. Hwang-do W, Kang JH, Jeong JM, Chung JK, Lee MC, Kim S, et al.: Noninvasive in vivo monitoring of neuronal differentiation using reporter driven by a neuronal promoter.

    Eur J Nucl Med Mol Imaging 2008, 35:135-45. PubMed Abstract | Publisher Full Text OpenURL

  14. Penheiter AR, Russell SJ, Carlson SK: The sodium iodide symporter (NIS) as an imaging reporter for gene, viral, and cell-based therapies.

    Curr Gene Ther 2012, 12:33-47. PubMed Abstract | Publisher Full Text OpenURL

  15. Portulano C, Paroder-Belenitsky M, Carrasco N: The Na+/I- symporter (NIS): mechanism and medical impact.

    Endocr Rev 2014, 35:106-49. PubMed Abstract | Publisher Full Text OpenURL

  16. Terrovitis J, Kwok KF, Lautamaki R, Engles JM, Barth AS, Kizana E, et al.: Ectopic expression of the sodium iodide symporter enables imaging of transplanted cardiac stem cells in vivo by single-photon emission computed tomography or positron emission tomography.

    J Am Coll Cardiol 2008, 52:1652-60. PubMed Abstract | Publisher Full Text OpenURL

  17. Higuchi T, Anton M, Saraste A, Dumler K, Pelisek J, Nekolla SG, et al.: Reporter gene PET for monitoring survival of transplanted endothelial progenitor cells in the rat heart after pretreatment with VEGF and atorvastatin.

    J Nucl Med 2009, 50:1881-6. PubMed Abstract | Publisher Full Text OpenURL

  18. Templin C, Zweigerdt R, Schwanke K, Olmer R, Ghadri JR, Emmert MY, et al.:Transplantation and tracking of human-induced pluripotent stem cells in a pig model of myocardial infarction: assessment of cell survival, engraftment, and distribution by hybrid single photon emission computed tomography/computed tomography of sodium iodide symporter transgene expression.

    Circulation 2012, 126:430-9. PubMed Abstract | Publisher Full Text OpenURL

  19. Seo JH, Jeon YH, Lee YJ, Yoon GS, Won DI, Ha JH, et al.: Trafficking macrophage migration using reporter gene imaging with human sodium iodide symporter in animal models of inflammation.

    J Nucl Med 2010, 51:1637-43. PubMed Abstract | Publisher Full Text OpenURL

  20. Gao J, Prough DS, McAdoo DJ, Grady JJ, Parsley MO, Ma L, et al.: Transplantation of primed human fetal neural stem cells improves cognitive function in rats after traumatic brain injury.

    Exp Neurol 2006, 201:281-92. PubMed Abstract | Publisher Full Text OpenURL

  21. Beasley TM, Palmer HE, Nelp WB: Distribution and excretion of technetium in humans.

    Health Phys 1966, 12:1425-35. PubMed Abstract | Publisher Full Text OpenURL

  22. Miyagawa M, Beyer M, Wagner B, Anton M, Spitzweg C, Gansbacher B, et al.: Cardiac reporter gene imaging using the human sodium/iodide symporter gene.

    Cardiovasc Res 2005, 65:195-202. PubMed Abstract | Publisher Full Text OpenURL

  23. Shin JH, Chung JK, Kang JH, Lee YJ, Kim KI, Kim CW, et al.: Feasibility of sodium/iodide symporter gene as a new imaging reporter gene: comparison with HSV1-tk.

    Eur J Nucl Med Mol Imaging 2004, 31:425-32. PubMed Abstract | Publisher Full Text OpenURL

  24. Barton KN, Stricker H, Brown SL, Elshaikh M, Aref I, Lu M, et al.: Phase I study of noninvasive imaging of adenovirus-mediated gene expression in the human prostate.

    Mol Ther 2008, 16:1761-9. PubMed Abstract | Publisher Full Text OpenURL

  25. Barton KN, Stricker H, Elshaikh MA, Pegg J, Cheng J, Zhang Y, et al.: Feasibility of adenovirus-mediated hNIS gene transfer and 131I radioiodine therapy as a definitive treatment for localized prostate cancer.

    Mol Ther 2011, 19:1353-9. PubMed Abstract | Publisher Full Text OpenURL

  26. Cho JY, Shen DH, Yang W, Williams B, Buckwalter TL, La Perle KM, et al.: In vivo imaging and radioiodine therapy following sodium iodide symporter gene transfer in animal model of intracerebral gliomas.

    Gene Ther 2002, 9:1139-45. PubMed Abstract | Publisher Full Text OpenURL

  27. Varma NR, Barton KN, Janic B, Shankar A, Iskander A, Ali MM, et al.: Monitoring adenoviral based gene delivery in rat glioma by molecular imaging.

    World J Clin Oncol 2013, 4:91-101. PubMed Abstract | Publisher Full Text OpenURL

  28. Illum L: Nasal drug delivery–possibilities, problems and solutions.

    J Control Release 2003, 87:187-98. PubMed Abstract | Publisher Full Text OpenURL

  29. Okuyama S: The first attempt at radioisotopic evaluation of the integrity of the nose-brain barrier.

    Life Sci 1997, 60:1881-4. PubMed Abstract | Publisher Full Text OpenURL

  30. Singh R, Lillard JW Jr: Nanoparticle-based targeted drug delivery.

    Exp Mol Pathol 2009, 86:215-23. PubMed Abstract | Publisher Full Text OpenURL

  31. Olivier JC, Huertas R, Lee HJ, Calon F, Pardridge WM: Synthesis of pegylated immunonanoparticles.

    Pharm Res 2002, 19:1137-43. PubMed Abstract | Publisher Full Text OpenURL

  32. Begley DJ: Delivery of therapeutic agents to the central nervous system: the problems and the possibilities.

    Pharmacol Ther 2004, 104:29-45. PubMed Abstract | Publisher Full Text OpenURL

 

White matter injury restoration after stem cell administration in subcortical ischemic stroke

Laura Otero-Ortega1, María Gutiérrez-Fernández1*, Jaime Ramos-Cejudo1, Berta Rodríguez-Frutos1, Blanca Fuentes1, Tomás Sobrino2, Teresa Navarro Hernanz3, Francisco Campos2, Juan Antonio López4, Sebastián Cerdán3, Jesús Vázquez4 and Exuperio Díez-Tejedor1*

Author Affiliations

1Department of Neurology and Stroke Center, Neuroscience and Cerebrovascular Research Laboratory, La Paz University Hospital, Neuroscience Area of IdiPAZ Health Research Institute, Autónoma University of Madrid, Paseo de la Castellana 261, Madrid, 28046, Spain

2Department of Neurology, Clinical Neurosciences Research Laboratory, Hospital Clínico Universitario, Health Research Institute of Santiago de Compostela (IDIS), University of Santiago de Compostela, Travesía de Choupana, Santiago de Compostela, s/n, 15706, Spain

3Laboratory for Imaging and Spectroscopy by Magnetic Resonance (LISMAR), Institute of Biomedical Research Alberto Sols, CSIC-UAM, Arturo Duperier, Madrid, 4, 28029, Spain

4Cardiovascular Proteomics Laboratory & Proteomics Unit, Centro Nacional de Investigaciones Cardiovasculares, CNIC, Melchor Fernández, Almagro 3, 28029, Madrid, Spain

For all author emails, please log on.

Stem Cell Research & Therapy 2015, 6:121  doi:10.1186/s13287-015-0111-4
Laura Otero-Ortega, María Gutiérrez-Fernández and Jaime Ramos-Cejudo contributed equally to this work.

The electronic version of this article is the complete one and can be found online at:http://stemcellres.com/content/6/1/121

Received: 8 January 2015
Revisions received: 12 January 2015
Accepted: 10 June 2015
Published: 19 June 2015

© 2015 Otero-Ortega et al.

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Abstract

Introduction

Despite its high incidence, nerve fiber (axon and myelin) damage after cerebral infarct has not yet been extensively investigated. The aim of this study was to investigate white matter repair after adipose-derived mesenchymal stem cell (ADMSC) administration in an experimental model of subcortical stroke. Furthermore, we aimed to analyze the ADMSC secretome and whether this could be implicated in this repair function.

Methods

An animal model of subcortical ischemic stroke with white matter affectation was induced in rats by injection of endothelin-1. At 24 hours, 2 × 10 6 ADMSC were administered intravenously to the treatment group. Functional evaluation, lesion size, fiber tract integrity, cell death, proliferation, white matter repair markers (Olig-2, NF, and MBP) and NogoA were all studied after sacrifice (7 days and 28 days). ADMSC migration and implantation in the brain as well as proteomics analysis and functions of the secretome were also analyzed.

Results

Neither ADMSC migration nor implantation to the brain was observed after ADMSC administration. In contrast, ADMSC implantation was detected in peripheral organs. The treatment group showed a smaller functional deficit, smaller lesion area, less cell death, more oligodendrocyte proliferation, more white matter connectivity and higher amounts of myelin formation. The treated animals also showed higher levels of white matter-associated markers in the injured area than the control group. Proteomics analysis of the ADMSC secretome identified 2,416 proteins, not all of them previously described to be involved in brain plasticity.

Conclusions

White matter integrity in subcortical stroke is in part restored by ADMSC treatment; this is mediated by repair molecular factors implicated in axonal sprouting, remyelination and oligodendrogenesis. These findings are associated with improved functional recovery after stroke.

Introduction

White matter injury and the mechanisms of nerve fiber (axon and myelin) repair have seldom been investigated in translational stroke research [1], [2], despite the fact that blood supply disruption also compromises whole axons and fibers and therefore brain connectivity. Even though relevant, white matter injury in stroke has not been extensively studied in the past due to the intrinsic difficulties associated with animal models; for instance, the fact that the rodent brain has substantially less white matter than higher mammals or humans [3], [4]. However, not only are up to 25 % of ischemic strokes in humans subcortical [1], but cortical infarcts also produce white matter injury. The high incidence of such damage motivates the search for an effective therapy that would enhance the mechanisms underlying the repair of damaged nerve fibers after any kind of stroke.

Stem cell therapy has demonstrated its efficacy in cortical stroke and may have a positive effect on subcortical lesions. In this regard, preclinical studies indicate that adipose-derived mesenchymal stem cells (ADMSC) are a promising new therapy for subcortical stroke that could promote recovery by improving the global brain repair mechanisms [5]–[7]. Trophic factor release, paracrine interactions and immunomodulatory effects have been suggested as the main functional mechanisms involved in ADMSC therapy [8], [9]. In this regard, stem cells are known to have paracrine effects on neurogenesis, gliogenesis, synaptogenesis, vasculogenesis and immunomodulation. However, there is little evidence whether stem cell administration can promote oligodendrogenesis and white matter fiber repair when axonal tract integrity has been compromised.

Therefore, the aim of this study was to investigate the therapeutic effects (improvement of functional deficits and enhancement of white matter fiber repair) of the intravenous administration of ADMSC in rats submitted to subcortical stroke with white matter injury.

Methods

Ethics statement

The procedure was carried out at our Cerebrovascular and Neuroscience Research Laboratory, La Paz University Hospital, Madrid, Spain. All experiments were designed to minimize animal suffering in compliance with, and approved by, our medical school’s Ethical Committee of La Paz University Hospital for the Care and Use of Animals in Research according to the Spanish and European Union rules (86/609/CEE, 2003/65/CE, 2010/63/EU, RD 1201/2005 and RD53/2013).

Animals and surgery

A total of 72 male Sprague–Dawley rats weighing 200–250 g (Charles River Laboratories, France) were used. In all animals, the femoral artery was cannulated during surgery and induction of cerebral ischemia to allow continuous monitoring of physiological parameters including blood glucose levels, blood gases and blood pressure (Omicron ALTEA Monitor; RGB Medical Devices, Madrid, Spain). Cranial and body temperature were monitored and maintained at 36.5 ± 0.5 °C. Male Sprague–Dawley rats (200–250 g) were anesthetized using 3.5 % isoflurane in 2 L/minute oxygen and given meloxicam 2 mg/kg for analgesia. To provoke white matter injury, subcortical stroke was induced by injection of 1 μL endothelin-1 (ET-1; Calbiochem, Germany) (0.25 μg/μL) with the use of a SYR 5 μL Hamilton syringe (Tecknokroma, Barcelona, Spain) into the striatum using stereotactic references (+0.4 mm AP, +3.5 mm L, + 6 mm DV from bregma) as previously described [10].

After 24 hours, the treatment group received intravenously (i.v.) 2 × 10 6 ADMSC in 1 ml of saline solution (n = 24) by the tail vein. Dose was determined based on previous studies [8], [9]. In the control (n = 24) and sham-operation (n = 24) groups, only saline solution was i.v. administered via the tail vein. Rats were sacrificed at 24 hours (n = 4 in each group for comparative anatomical analysis of the lesion in fresh tissue and to analyze ADMSC distribution), 7 days (n = 10 in each group) or 28 days (n = 10 in each group) after treatment for cell death and proliferation analysis and immunohistochemistry, immunofluorescence and Western blot studies.

Cell culture protocol

ADMSC obtained from allogeneic adipose tissue of Sprague–Dawley rats (250–300 g) were cultured. The adipose tissue was digested with collagenase (Sigma Aldrich, Madrid, Spain) and incubated at 37 °C in 5 % CO 2 . On the third pass, the cell cultures were divided into three groups: 1) 1.0 × 10 5 ADMSC for characterization, 2) 1.5 × 10 6 ADMSC for proteomics analysis of the culture supernatant, and 3) 42 × 10 6 ADMSC for the treatment of rats. For characterization, ADMSC were trypsinized and labeled with fluorescein isothiocyanate (FITC)-, phycoerythrin (PE)- or Alexa 647-conjugated primary antibodies. The cells were incubated for 20 minutes at 4 °C in the dark with the following antibodies: CD90-FITC (AbD Serotec, Oxford, UK), CD29-PE (AbD Serotec), CD45-PE (AbD Serotec) and CD11b-PE (AbD Serotec). Matched isotype controls were purchased from Biolegend (San Diego, CA, USA). Flow cytometry analysis of CD90+/CD29+/CD45–/CD11b– cells was performed using a FACScalibur cytometer and CellQuest Pro software (Becton Dickinson, Madrid, Spain). For ADMSC treatment, ADMSC with >95 % viability were administered i.v. The dose, route and time of administration were based on previously reported data [9], [11].

Proteomics data analysis

For proteomic analysis of cell culture supernants, ADMSC were cultured overnight with a free fetal bovine serum/protein culture medium. After 24 hours, cell supernatants were collected and the protein content was analyzed as follows. Proteins were digested using the filter aided sample preparation (FASP) protocol [12]. Briefly, samples were dissolved in 50 mM Tris–HCl pH 8.5, 4 % SDS and 50 mM DTT, boiled for 10 minutes and then centrifuged. Protein concentration in the supernatant was measured by the Direct Detect® Spectrometer (Millipore, Billerica, MA, USA). Approximately 50 μg of protein was diluted in 8 M urea in 0.1 M Tris–HCl (pH 8.5), and loaded onto 30 kDa centrifugal filter devices (FASP Protein Digestion Kit, Expedeon, Knoxville, TN, USA). The denaturation buffer was replaced by washing three times with UA. Proteins were later alkylated using 50 mM iodoacetamide in UA for 20 minutes in the dark, and the excess alkylation reagents were eliminated by washing three times with UA and three additional times with 50 mM ammonium bicarbonate. Proteins were digested overnight at 37 °C with modified trypsin (Promega, Madison, WI, USA) in 50 mM ammonium bicarbonate at 40:1 protein to trypsin (w/w) ratio. The resulting peptides were eluted by centrifugation with 50 mM ammonium bicarbonate (twice) and 0.5 M sodium chloride. Trifluoroacetic acid (TFA) was added to a final concentration of 1 % and the peptides were finally desalted onto C18 Oasis-HLB cartridges and dried-down for further analysis.

Peptides were loaded into the LC-MS/MS liquid chromatography tandem mass spectrometry system for on-line desalting onto C18 cartridges and analyzing by LC-MS/MS using a C-18 reversed phase nano-column (75 μm internal diameter × 50 cm, 2 μm particle size, Acclaim PepMap RSLC, 100 C18; Thermo Fisher Scientific, Waltham, MA, USA) in a continuous acetonitrile gradient consisting of 0–30 % B in 180 minutes, 50–90 % B in 3 minutes (B = 90 % acetonitrile, 0.5 % formic acid). A flow rate of 200 nL/minute was used to elute peptides from the RP nano-column to an emitter nanospray needle for real time ionization and peptide fragmentation on an Orbitrap Elite mass spectrometer (Thermo Fisher Scientific). An enhanced FT-resolution Fourier- Transform spectrum (resolution = 35,000) followed by the MS/MS spectra from the most intense 15 parent ions were analyzed along the chromatographic run. Dynamic exclusion was set at 30 seconds.

For peptide identification, all spectra were analyzed with Proteome Discoverer (version 1.4.0.29; Thermo Fisher Scientific) using SEQUEST-HT (Thermo Fisher Scientific). For database searching on the Uniprot database containing all sequences from humans (6 March 2013), parameters were selected as follows: trypsin digestion with two maximum missed cleavage sites, precursor and fragment mass tolerances of 600 ppm and 0.02 Da, respectively, carbamidomethyl cysteine as fixed modification and methionine oxidation as dynamic modifications. Peptide identification was validated using the probability ratio method [13] with an additional filtering for precursor mass tolerance of 10 ppm. False discovery rate (FDR) was calculated using inverted databases and the refined method [14]. For the study of the biological functions of identified proteins, gene onthology analysis was performed using the GOrilla (Gene Ontology Enrichment Analysis and Visualization) research tool [15].

Biodistribution analysis

For identification of donor cells, ADMSC were labeled with DiI (Celltracker CM-DiI, Invitrogen, Barcelona, Spain) prior to administration and stained with CD90, and possible migration and implantation were analyzed using immunofluorescence. Biodistribution of labeled ADMSC with DiI 24 hours after i.v. administration was analyzed using immunofluorescence techniques in both control and treated animals. Cryosections (10 μm thick) of brain, kidney, liver, lung and spleen were counterstained with 4′,6-diamino-2-phenylindole (DAPI) and analyzed by immunofluorescence staining (n = 4 per group).

Functional evaluation

Functional evaluation was performed in all animals by a blinded observer before surgery and after 1, 3, 7, 14 and 28 days. Motor performance was evaluated using the beam walking and rotarod tests and Rogers’ functional scale. The beam walking test measured the ability of rats to walk along a wooden beam (2.5 × 2.5 × 80 cm). Scores were assigned as follows: 0, traversed the beam with no foot slip; 1, traversed with grasping of the lateral side of the beam; 2, difficulty crawling along the beam but able to traverse; 3, required >10 seconds to traverse the beam because of difficulty with walking; 4, unable to traverse the beam; 5, unable to move the body or any limb on the beam; and 6, unable to stay on the beam for >10 seconds [16]. The rotarod test measured the latency to fall from a rotating cylinder [9]. A variant of Rogers’ functional scale was used to assign scores as follows: 0, no functional deficit; 1, failure to extend forepaw fully; 2, decreased grip of forelimb while tail gently pulled; 3, spontaneous movement in all directions, contralateral circling only if pulled by the tail; 4, circling; 5, walking only when stimulated; 6, unresponsive to stimulation with a depressed level of consciousness; and 7, dead (n = 10 per group) [17].

In vivo magnetic resonance imaging and tractography

Lesion size was analyzed after 1, 7 and 28 days by magnetic resonance imaging (MRI) using a 7-Tesla horizontal bore magnet (Bruker Pharmascan, Ettlingen, Germany) and T2 maps (RARE 8 T2, 180° flip angle, three averages) as previously described [9]. The lesion area was expressed as a percentage of the contralateral hemisphere, after correcting for brain edema. For tractography, diffusion tensor imaging (DTI) was performed after 1, 7 and 28 days using a spin-echo single-shot echo-planar imaging pulse sequence with the following parameters: TR/TE, 5000/35 ms; signal average, 10; 30 non-collinear diffusion gradients with diffusion weighting of b = 1,000 s/mm 2 and b = 0 s/mm 2 ; and field of view 3.5 × 3.5 cm. A total of 496 slices were evaluated from data acquired in 30 directions. The images were obtained using medInria (Inria, France), a multi-platform medical image processing and visualization software. Zoomed lesion site three-dimensional diffusion tensor images were represented using ParaView 4.1.0 software (Los Álamos National Laboratory, New México, USA) (n = 6 per group, 10 sections per animal) as previously described [10].

Cell death evaluation

Cell death was analyzed in the infarct zone of at least 10 sections from each animal using TUNEL staining (TdT-FragEL DNA Fragmentation Detection Kit, Oncogene Research Products, San Diego, CA). The number of positive cells was counted in a minimum of 10 different microscopical fields based on their nuclear morphology and dark color using a 40× objective lens and image analysis software (Image-Pro Plus 4.1, Media Cybernetics, Rockville, MD, USA) (n = 6 per group, 10 sections per animal).

Cell proliferation analysis

Cell proliferation was analyzed using Ki-67 staining (1:100, Chemicon, Temecula, CA, USA) after 7 and 28 days, in 10 sections corresponding to the infarct area of each animal, selected as previously described [18], [19]. The number of positive cells was counted in a minimum of 10 different random microscopic fields using a 40× objective lens and Image-Pro Plus 4.1 software (n = 6 per group, 10 sections per animal).

The differentiation of the proliferating cells was analyzed using co-staining with Ki-67 and NeuN (1:100, Millipore), Olig-2 (1:400, Millipore) and glial fibrillary acidic protein (GFAP; 1:400, Chemicon) followed by goat anti-mouse Alexa Fluor 488 antibody (1:750, Invitrogen). Images were acquired as a confocal maximum projection using a Leica TCS-SPE confocal microscope (Leica Microsystems, Heidelberg, Germany) and the number of double-positive cells was counted in a minimum of 10 different microscopic fields using a 40× objective lens and Image-Pro Plus 4.1 software.

Immunohistochemical, immunofluorescence and Western blot analyses

Frozen sections were stained using the CryoMyelin Kit (Hitobiotech, Wilmington, USA), which allows sensitive localization and visualization of myelin fibers. The mean intensity of myelin staining in the region of interest (ROI) was quantified using a Nikon Eclipse-Ti inverted microscope and NIS-elements software. The lesion area was studied in detail using immunofluorescence and Western blot analyses. The white matter-associated antibodies used were neurofilament (NF; 1:100, Dako, Glostrup, Denmark), neurite outgrowth inhibitor (NogoA; 1:100, Abcam, Cambridge, UK), myelin basic protein (MBP; 1:100, Abcam) and oligodendrocyte (Olig-2; 1:500, Millipore) followed by goat anti-mouse Alexa Fluor 488 and anti-rabbit Alexa Fluor 594 (1:750, Invitrogen). For Western blot analysis, the units were normalized based on Β-actine (1:400, Sigma-Aldrich). To quantify the expression of white matter-associated markers, the mean fluorescence intensity was evaluated in a minimum of 10 different microscopic fields using a 40× objective lens. The experiments, images and quantification of the samples were performed on the same day using the same microscope configurations, by blinded observers, to eliminate bias due to background normalization (four animals, four sections per animal).

Statistical analysis

Results were expressed as mean ± standard error of the mean (SEM). Data were compared using the Kruskal-Wallis test followed by the Mann–Whitney test. Values of p < 0.05 were considered as statistically significant. The analysis was performed using statistical SPSS 16 and GraphPad software (GraphPad Software Inc, CA, USA).

Results

ADMSC characterization, migration and implantation in the injured brain area

ADMSC showed typical fibroblast-like cell morphology and their phenotype was CD90+/CD29+/CD45–/CD11b– (Fig. 1a).

thumbnailFig. 1. Characterization and biodistribution of ADMSC. a ADMSC characterization by flow cytometry. Rat ADMSC were labeled with CD29, CD90, CD11b, and CD45 and analyzed by flow cytometry. Of the ADMSC population, 95 % expressed CD29 and CD90. Additionally, these cells lacked expression (5 % positive) of CD11b, CD45. b Migration and implantation in the brain and peripheral organs (liver, lung and spleen) of DiI- and CD90-labelled cells at 24 hours after treatment. AD-MSC adipose-derived mesenchymal stem cells, DAPI 4′,6-diamino-2-phenylindole, FITC fluorescein isothiocyanate,PE phycoerythrin

DiI and CD90 co-labeled cells were not observed in the control group. Migration and implantation in the brain were not observed on immunofluorescence images of the injured brain area after intravenous administration of DiI and CD90 co-labeled cells. However, DiI and CD90 co-labeled cells were observed in peripheral organs such as the liver, lung and spleen (Fig. 1b).

Effect of ADMSC treatment on functional recovery

To assess the potential of ADMSC administration to improve functional recovery after subcortical stroke, motor function was assessed before surgery and after 1, 3, 7, 14 and 28 days using the walking beam, rotarod and Roger’s test. There were significant differences in motor deficit scores between the treatment and control group; the walking beam test performance was significantly better in the treatment group than in the control group after 3 days (p < 0.01), 14 days (p < 0.05) and 28 days (p < 0.05). The Rotarod test was significantly better in the ADMSC treatment group than in the control group after 1 day (p < 0.01) and 28 days (p < 0.05). The Rogers’ functional scale score was significantly better in the treatment group than in the control group after 7 and 28 days (both p < 0.05) (Fig. 2).

thumbnailFig. 2. Improved functional outcome after ADMSC administration in subcortical stroke. Beam walking test performance (left) was improved at 3 (p < 0.01), 14 and 28 days (p < 0.05). Rotarod test (middle) showed significant differences between the ADMSC group and control animals at 1 (p < 0.01) and 28 days (p < 0.05). The ADMSC group also showed significantly better scores compared to controls at 7 and 28 days (p < 0.05) on the Roger’s test (right). Data are shown as mean ± SEM; *p < 0.05, **p < 0.005; n = 10 animals per group. AD-MSC adipose-derived mesenchymal stem cells

Effect of ADMSC treatment on lesion size and tract connectivity

MRI analysis showed no significant difference in infarct size between the treatment and control groups after 1 and 7 days. However, the infarct size was significantly smaller in the treatment group than in the control group after 28 days (0.12 ± 0.01 vs. 0.6 ± 0.26, p < 0.05) (Fig. 3).

thumbnailFig. 3. ADMSC treatment reduced infarct size and increased fiber tract and myelin integrity after subcortical white matter damage. aMorphological study by CryoMyelin staining identified the zone of the lesion as an area of white matter injury located in the subcortical zone, showing restored myelinated axons in the ADMSC-treated animals. bQuantification of mean ROI intensity of the CryoMyelin staining. Stroked line indicates ROI; yellow line indicates a representative longitudinal profile of pixel intensity.c Comparative image analysis T2-weighted MRI and tractography at 1, 7 and 28 days showed a progressive reduction in white matter infarct size in controls and treated animals. Detail of tractography image in the lesion is given below showing augmented connectivity of fiber tracts in ADMSC-treated animals at 28 days. d Quantitative analysis of MRI images showed that ADMSC therapy reduced lesion size at 28 days compared to the controls (p < 0.05). Data are shown as mean ± SEM; *p < 0.05; n = 6 animals, 10 sections each per group. ADMSCadipose-derived mesenchymal stem cells, d days, MRI magnetic resonance image, ROIregion of interest

Myelin was stained using the CryoMyelin Kit to identify the area of white matter injury in the subcortical infarct. The mean intensity of staining in the ROI was evaluated in the treatment and control groups, with white indicating absence of myelin and black indicating the presence of myelinated axons. There was higher intensity (indicating absence of myelin) in the control group (210.23 ± 14.30 mean intensity) than in the treatment group (155.71 mean intensity ± 21.23) (Fig. 3).

DTI tractography data showed similar results in axial diffusivity (120.35 ± 12.45 µm 2 /s and 126.78 ± 18.34; p > 0.05) in both the control and treated groups, respectively, at 7 days. However, compared with the control rats, 28 days after treatment the ADMSC-treated rats showed significantly improved axial diffusivity (127.98 ± 9.21 and 162.99 ± 13.65 µm 2 /s, respectively; p< 0.05) compared to controls. These results suggest that there was a significant improvement in white matter thickness (width, breadth, depth) and restoration of tract connectivity in the ADMSC-treated animals compared with controls at 28 days.

Effect of acute ADMSC treatment on cell death and brain cell proliferation

Cell death was analyzed on frozen sections by TUNEL staining after 7 and 28 days. After 7 days, no significant differences were found in TUNEL-positive cells in both the control group (718.5 ± 146.3 cells) and the treatment group (460.33 ± 120.5 cells). After 28 days, there were significantly fewer TUNEL-positive cells in the ischemic area in the treatment group than in the control group (24.5 ± 1.73 vs. 56.0 ± 3.46 cells, p < 0.05) (Fig. 4a).

thumbnailFig. 4. ADMSC administration led to a cell death reduction and improved brain proliferation activity. a Quantitative analysis of cell death by TUNEL technique showed a significant reduction in TUNEL-positive cells after ADMSC therapy compared to the control group (p < 0.05). b At 28 days, Ki-67 staining shows a significant increase in the number of proliferating cells in ADMSC-treated animals compared to the control group (p < 0.05).c At 28 days after treatment, Ki-67 co-labeling with NeuN, GFAP, and Olig-2 showed different cell type proliferation including oligodendrocytes. Data are shown as mean ± SEM; scale bars = 20 μm; n = 6 animals, 10 sections each per group. ADMSC adipose-derived mesenchymal stem cells, DAPI 4′,6-diamino-2-phenylindole,GFAP glial fibrillary acid protein, NeuN neuronal nuclei, Olig-2 oligodendrocite transcription factor 2

Quantitative analysis of proliferative cells was performed using Ki-67 labeling after 7 and 28 days. After 7 days, there were no significant differences in proliferative cells in both the treatment group (83.5 ± 34.65 cells) and the control group (19.5 ± 7.78 cells) (Fig. 4b). After 28 days, the number of Ki-67-positive cells was significantly higher in the treatment group than in the control group (13 ± 1.15 vs. 3 ± 1.15 cells, p < 0.05). The proportions of proliferating cell types observed by co-staining with Ki-67 were not significantly different between the treatment and control groups (NeuN, 15 ± 3 % vs. 16 ± 5 %, Olig-2: 27 ± 3 % vs. 26 ± 4 %; GFAP, 33 ± 15 % vs. 27 ± 13 %) (Fig. 4c). However, although there were no significant differences between the proportions of the cell lines, higher levels of each cell type were found in the ADMSC-treated group.

Effect of ADMSC treatment on white matter-associated marker expression

After finding that ADMSC administration had a beneficial effect on functional outcome, we investigated whether the functional outcome was related to the levels of white matter-associated markers. Western blot analysis found that the level of NF (a marker of axonal sprouting) was not significantly different between the treatment and control groups after 7 days. However, the NF level was significantly higher in the treatment group than in the control group after 28 days (4.27 ± 1.26 vs. 1.00 ± 0.69 Arbitrary Units (AU), p < 0.001). The MBP level was significantly higher in the treatment group than in the control group after 7 days (1.09 ± 0.23 vs. 0.63 ± 0.18 AU, p < 0.05) and 28 days (0.80 ± 0.11 vs. 0.43 ± 0.24 AU, p < 0.05). The Olig-2 level was significantly higher in the treatment group than in the control group after 28 days (1.08 ± 0.10 vs. 0.32 ± 0.08 AU, p < 0.05). The NogoA level was not significantly different between the treatment and control groups after 7 days (0.68 ± 0.35 vs. 0.67 ± 0.17 AU, p > 0.05), and tended to be lower in the treatment group than in the control group after 28 days (0.55 ± 0.35 vs. 1.16 ± 0.34 AU, p > 0.05) (Fig. 5). The NF immunofluorescence intensity was significantly higher in the treatment group than in the control group after 28 days (10,020.28 ± 1,231.19 vs. 3,536.21 ± 643.2 average fluorescence intensity, p < 0.05). The MBP immunofluorescence intensity was significantly higher in the treatment group than in the control group after 7 days (5,714.61 ± 529.59 vs. 3,529.97 ± 1,222.40 AU, p < 0.05) and 28 days (6,920.39 ± 1,134.27 vs. 3,736.34 ± 324.50 AU, p < 0.05). The Olig-2 immunofluorescence intensity was significantly higher in the treatment group than in the control group after 28 days (2,439.00 ± 231.12 vs. 353.40 ± 111.12 AU, p < 0.05). The NogoA immunofluorescence intensity tended to be lower in the treatment group than in the control group after 28 days (1,534.21 ± 767.32 vs. 2,423.88 ± 876.70 AU, p > 0.05).

thumbnailFig. 5. White matter-associated markers are enhanced in striatum after ADMSC therapy in subcortical stroke model. a Immunofluorescence images and b immunofluorescence quantification of white matter repair-associated markers (NF, MBP, Olig-2 and NogoA) at 7 and 28 days after treatment. c Western blot and d Western blot quantification showed increased levels of MBP in the treated group compared to controls at both 7 and 28 days (p < 0.05), as well as augmented levels of Olig-2, NF (p < 0.05) and a trend to decreased levels of NogoA. Data are shown as mean ± SEM; scale bars = 20 μm; *p < 0.05; n = 4 animals, 4 sections each per group. ADMSC adipose-derived mesenchymal stem cells, ddays, MBP myelin basic protein, NF neurofilament, Nogo-A neurite outgrowth inhibitor, Olig-2oligodendrocite transcription factor 2

Proteomics analysis of the ADMSC in vitro secretome

Proteomics analysis of the secretome of cell cultures identified 2,416 proteins in the cell supernatants that are implicated into different cell functions (Fig. 6a), such as protein binding (carbohydrate binding, antigen binding, ion binding, sulfur compound binding, and lipid binding), metabolic processes, single and multicellular organism processes, development, endodermal cell differentiation, skin and cartilage morphogenesis, immune system processes, cellular organization and biogenesis, response to stimulus and biological adhesion processes . All functions in detail are shown in Fig. 6c and d. Full proteomic data is given in Additional file 1. Interestingly, some of these proteins are also implicated in growth factor activity, such as several trophic factors and receptors known to be also involved in brain plasticity. Some of these trophic factors are shown in Fig. 6b.

thumbnailFig. 6. Proteomics analysis of ADMSC secretome reveals multiple biological functions. a Gene onthology (GO) analysis of the 2,416 proteins identified by Orbitrap proteomic study. Bars indicate the number of proteins from the total associated with each biological function. b Trophic factor-related proteins identified in the proteomic analysis; PSM: peptide- spectrum matches c GO functional clustering of proteins associated with binding function and d other biological processes

Discussion

The results of this study showed that ADMSC administration plays a major role in improving the repair of white matter fiber tracts in an experimental model of subcortical stroke. We found that the treated group had better functional recovery and smaller lesion size than the control group. In addition, animals which received ADMSC treatment showed significantly higher number of proliferating cells (including oligodendrocyte progenitors) and significantly less cell death at the lesion region than animals in the control group. Analysis of fiber tract integrity by tractography and CryoMyelin staining showed that white tract thickness had been recovered in the treatment group. The treated group with ADMSC also had higher levels of white matter-associated markers (NF, MBP and Olig-2) than the control group, suggesting that ADMSC administration induced repair of white matter fiber tracts.

Up to 25 % of ischemic strokes in humans are subcortical or lacunar, which are confined to white matter regions such as the striatum and internal capsule [1]. The high frequency of damage to these areas in stroke patients has motivated the search for useful experimental animal models of subcortical stroke with white matter affectation, as well as effective therapies to enhance the mechanisms underlying repair of damaged white matter (axon and myelin). In this regard, endothelin (one of the most potent known vasoconstrictors) was considered the best candidate to induce this injury with white matter affectation [10], [20], [21].

ADMSC administration is considered an appropriate therapeutic strategy because ADMSC enhance the natural repair processes of the brain after injury. However, the mechanisms underlying these repair processes are still unknown. Our proteomics analysis identified thousands of proteins, many of them not previously associated with stem cell properties or stroke repair; for instance, we identified a number of proteins such as hepatoma-derived growth factor, latent-transforming growth factor beta, and connective tissue growth factor. Other proteins previously implicated in stem cell therapy function such as transforming growth factor-beta, fibroblast growth factor, vascular endothelial growth factor or brain-derived neurotrophic factor were also identified. Moreover, gene onthology analysis identified a number of protein functions not previously associated with stem cell therapy function in stroke recovery. In this regard, protein binding (carbohydrate binding, antigen binding, ion binding, sulfur compound binding, lipid binding), metabolic processes, single and multicelular organism processes, development, endodermal cell differentiation, skin and cartilage morphogenesis, immune system processes, cellular organization and biogenesis, response to stimulus and biological adhesion processes were highly represented. Interestingly, growth factor activity was not the main represented function in the cell secretome, indicating that other functions are also relevant. Our findings suggest that the release of the identified proteins by the administered ADMSC could contribute to improve functional recovery when allocated to peripheral organs (spleen, lung and liver). However, futures studies will be needed in order to understand the complex molecular mechanisms involved in stem cell therapy-mediated stroke recovery.

Various tests have been used to evaluate motor function following brain injury. Previous studies by our group have demonstrated improvement in the functional outcome after ADMSC administration in another experimental animal model of cerebral ischemia [9]. The present study showed that functional recovery at 28 days was significantly improved after ADMSC administration than in the control group after subcortical ischemic stroke.

In the present study, MRI studies showed a significantly smaller infarct size in the treated group than in the control group after 7 and 28 days. Our results are consistent with previous studies that reported a reduction in infarct size after ADMSC administration in another animal model of cerebral ischemia [9]. This reduction in infarct size could be related to the increased tract thickness and axonal projections observed by tractography and CryoMyelin staining. These results agree with previously reported findings that neural progenitor cell treatment in an animal model of cortical ischemia results in white matter reorganization shown by fiber tracking maps derived from DTI and by histological staining [22].

On the other hand, we found that the density of TUNEL-positive cells in the ischemic area peaked after 7 days in both treated and control groups. However, the treated group had less focal damage at 28 days, with significantly lower numbers of TUNEL-positive cells than in the control group. These results are consistent with those of previously reported studies, which found that ADMSC administration inhibited cell death in the infarct area [5], [9].

The central nervous system continuously generates new cells in several specific regions of the adult mammalian brain, and this proliferation has been shown to be enhanced by cell therapy. Ki-67 staining showed large numbers of proliferative cells after 7 days in both the treatment and control groups. In this regard, our results are consistent with a previous study [23]. Furthermore, there was greater cell proliferation after 28 days in the treated group than in the control group. These results provide clear evidence that ADMSC administration induces significant cell proliferation at 28 days after subcortical stroke. Furthermore, analysis of the proliferating cell lines showed similar proportions of co-staining with Ki-67 and NeuN, Olig-2 and GFAP in the treatment and control groups, indicating that almost half of the proliferating cells in both groups were white matter-associated cells (neurons and oligodendrocytes). These findings indicate genesis of new astrocytes and neurons as well as oligodendrocyte progenitors and immature oligodendrocytes after cerebral ischemia. Although the effects of ADMSC on neurogenesis, gliogenesis, synaptogenesis and vasculogenesis after stroke have already been described, the ability of ADMSC to promote oligodendrogenesis after subcortical stroke is a novel finding.

It is known that some brain repair mechanisms are quickly activated after cortical ischemia [24], but there is little information available regarding brain repair mechanisms after subcortical white matter stroke. To increase our understanding of these mechanisms, we investigated the levels of white matter-associated markers (NF, MBP, Olig-2 and NogoA) in both the treatment and control groups. NF levels, a marker of axonal sprouting, were not significantly different between the treated and control groups after 7 days, suggesting that the extent of white matter injury was similar in these groups during the acute phase. However, NF levels were higher in the group treated with ADMSC after 28 days, which could be explained by enhancement of axonal sprouting after ADMSC administration. In addition to axonal growth, restoration of the myelin sheath is important for the repair of white matter. Furthermore, Olig-2 levels were significantly higher in the treated group than in the control group after 28 days. These results support the concept that ADMSC enhance oligodendrogenesis to restore their loss due to ischemic injury. Our findings are supported by those of a recent in vitro study, which found that mesenchymal stem cell-conditioned medium promoted oligodendroglial cell maturation [25]. In addition to formation of mature oligodendrocytes, repair of the myelin sheath by oligodendrocytes is essential for achieving propagation of nerve impulses along axons [26]. The present study analyzed the MBP level as a marker of myelination. There were significantly higher levels of MBP in the treatment group compared to the control group after 7 and 28 days. These higher levels of MBP and Olig-2 are consistent with the increased proliferation observed for all phases of oligodendrocyte progenitors. Our results suggest that ADMSC administration increases oligodendrogenesis after white matter stroke.

Finally, among the myelin-associated proteins, NogoA, myelin-associated glycoprotein and oligodendrocyte myelin-associated protein all share a single receptor complex [27]. The present study did not find a difference in the NogoA level between the treated and control groups after 7 days. However, the NogoA level tended to be lower in the treatment group than in the control group after 28 days. These results may indicate that ADMSC administration can enhance axonal growth and plasticity by reducing the level of NogoA. In addition to indicating axonal sprouting, these findings are in accordance with the tractography and CryoMyelin stain findings that suggest restoration of white tract connectivity in the area of ischemia induced by injection of ET-1.

Conclusions

The findings of this study support the concept that ADMSC play an important role in enhancing some of the major mechanisms of remyelination. ADMSC administration resulted in a smaller lesion size and less cell death, as well as increased cell proliferation including oligodendrocyte progenitors, and higher levels of white matter-associated markers (NF, MBP and Olig-2) and restoration of white tract connectivity in the infarct area. All these processes may help to explain the improvement in functional outcome after ADMSC administration. Therapies that enhance remyelination may help to prevent the functional deficits resulting from those strokes affecting the white matter.

Abbreviations

ADMSC: Adipose tissue-derived mesenchymal stem cells

AU: Arbitrary units

DAPI: 4′,6-diamino-2-phenylindole

DiI: Celltracker CM-DiI

DTI: Diffusion tensor imaging

ET-1: Endothelin-1

FASP: Filter aided sample preparation

FITC: Fluorescein isothiocyanate

GFAP: Glial fibrillary acidic protein

i.v.: Intravenously

LC-MS/MS: Liquid chromatography tandem mass spectrometry

MBP: Myelin basic protein

MRI: Magnetic resonance imaging

NF: Neurofilament

NogoA: Neurite outgrowth inhibitor

Olig-2: Oligodendrocyte transcription factor 2

PE: Phycoerythrin

ROI: Region of interest

SEM: Standard error of the mean

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

LOO and JRC designed the experiments, performed animal experiments, and participated in drafting the manuscript. BRF was responsible for the laboratory assays. BF participated in coordination and helped in drafting the manuscript. TS and FC participated in cell culture studies and helped in drafting the manuscript. TNH and SC were responsible for the MRI laboratory assays and tractography. JAL and JV performed and analyzed the proteomics experiments. MGF and EDT designed the experiments, participated in coordination and helped in drafting the manuscript. All authors have read and approved the manuscript for publication.

Additional file

Additional file 1:. rADMSC proteomic analysis of cell supernatants.

Format: XLS Size: 1.3MB Download file

This file can be viewed with: Microsoft Excel ViewerOpen Data

Acknowledgements

This study was supported by research grants PS12/01754, PI11/00909 and INVICTUS (RD12/0014) (Spanish Neurovascular Network), SAF2010-37926, ProteoRed-PT13/0001/0017 and a Sara Borrell postdoctoral fellowship (CD12/00706, to LOO) from Research Institute Carlos III, Ministry of Science and Innovation of Spain. We greatly appreciate advice from Prof. Avendaño and Dr Negredo and we thank ServingMed.com for linguistic assistance. Furthermore, TS (CP12/03121) and FC (CP14/00154) are recipients of a research contract from Miguel Servet Program of Instituto de Salud Carlos III.

References

  1. Matute C, Domercq M, Pérez-Samartin A, Ransom BR: Protecting white matter from stroke injury.

    Stroke. 2013, 44:1204-11. PubMed Abstract | Publisher Full Text OpenURL

  2. Zhang R, Chopp M, Zhang ZG: Oligodendrogenesis after cerebral ischemia.

    Front Cell Neurosci. 2013, 7:201. PubMed Abstract | Publisher Full Text OpenURL

  3. Goldberg MP, Ransom BR: New light on white matter.

    Stroke. 2003, 34:330-2. PubMed Abstract | Publisher Full Text OpenURL

  4. Sozmen EG, Kolekar A, Havton LA, Carmichael ST: A white matter stroke model in the mouse: axonal damage, progenitor responses and MRI correlates.

    J Neurosci Meth. 2009, 180:261-72. Publisher Full Text OpenURL

  5. Leu S, Lin YC, Yuen CM, et al.: Adipose-derived mesenchymal stem cells markedly attenuate brain infarct size and improve neurological function in rats.

    J Transl Med. 2010, 8:63. PubMed Abstract | BioMed Central Full Text OpenURL

  6. Lindvall O, Kokaia Z: Stem cell research in stroke: how far from the clinic?

    Stroke. 2011, 42:2369-75. PubMed Abstract | Publisher Full Text OpenURL

  7. Gutiérrez-Fernández M, Rodríguez-Frutos B, Otero-Ortega L, et al.: Adipose tissue-derived stem cells in stroke treatment: from bench to bedside.

    Discov Med. 2013, 16:37-43. PubMed Abstract | Publisher Full Text OpenURL

  8. Gutiérrez-Fernández M, Rodríguez-Frutos B, Alvarez-Grech J, et al.: Functional recovery after hematic administration of allogenic mesenchymal stem cells in acute ischemic stroke in rats.

    Neuroscience. 2011, 175:394-405. PubMed Abstract | Publisher Full Text OpenURL

  9. Gutiérrez-Fernández M, Rodríguez-Frutos B, Ramos-Cejudo J, et al.: Effects of intravenous administration of allogenic bone marrow- and adipose tissue-derived mesenchymal stem cells on functional recovery and brain repair markers in experimental ischemic stroke.

    Stem Cell Res Ther. 2013, 4:11. PubMed Abstract | BioMed Central Full Text OpenURL

  10. Ramos-Cejudo J, Gutiérrez-Fernández M, Otero-Ortega L, et al.: BDNF-administration mediated oligodendrocyte differentiation and myelin formation in subcortical ischemic stroke.

    Stroke. 2015, 46:221-8. PubMed Abstract | Publisher Full Text OpenURL

  11. Yang B, Migliati E, Parsha K, et al.: Intra-arterial delivery is not superior to intravenous delivery of autologous bone marrow mononuclear cells in acute ischemic stroke.

    Stroke. 2013, 44:3463-72. PubMed Abstract | Publisher Full Text OpenURL

  12. Wiśniewski JR, Zougman A, Nagaraj N, Mann M: Universal sample preparation method for proteome analysis.

    Nat Methods. 2009, 5:359-62. Publisher Full Text OpenURL

  13. Martínez-Bartolomé S, Navarro P, Martín-Maroto F, et al.: Properties of average score distributions of SEQUEST: the probability ratio method.

    Mol Cell Proteomics. 2008, 6:1135-45. Publisher Full Text OpenURL

  14. Navarro P, Trevisan-Herraz M, Bonzon-Kulichenko E, et al.: General statistical framework for quantitative proteomics by stable isotope labeling.

    J Proteome Res. 2014, 3:1234-47. Publisher Full Text OpenURL

  15. Eden E, Navon R, Steinfeld I, et al.: GOrilla: a tool for discovery and visualization of enriched GO terms in ranked gene lists.

    BMC Bioinformatics. 2009, 10:48. PubMed Abstract | BioMed Central Full Text OpenURL

  16. Britton GL, Kim H, Kee PH, et al.: In vivo therapeutic gas delivery for neuroprotection with echogenic liposomes.

    Circulation. 2010, 122:1578-87. PubMed Abstract | Publisher Full Text OpenURL

  17. Rogers DC, Campbell CA, Stretton JL, Mackay KB: Correlation between motor impairment and infarct volume after permanent and transient middle cerebral artery occlusion in the rat.

    Stroke. 1997, 28:2060-5. PubMed Abstract | Publisher Full Text OpenURL

  18. Otero L, Bonilla C, Aguayo C, et al.: Intralesional administration of allogeneic bone marrow stromal cells reduces functional deficits after intracerebral hemorrhage.

    Histol Histopathol. 2010, 25:453-61. PubMed Abstract | Publisher Full Text OpenURL

  19. Otero L, Zurita M, Bonilla C, et al.: Allogeneic bone marrow stromal cell transplantation after cerebral hemorrhage achieves cell transdifferentiation and modulates endogenous neurogenesis.

    Cytotherapy. 2012, 14:34-44. PubMed Abstract | Publisher Full Text OpenURL

  20. Yanagisawa M, Kurihara H, Kimura S, et al.: A novel potent vasoconstrictor peptide produced by vascular endothelial cells.

    Nature. 1988, 332:411-5. PubMed Abstract | Publisher Full Text OpenURL

  21. Bailey EL, McCulloch J, Sudlow C, Wardlaw JM: Potential animal models of lacunar stroke: a systematic review.

    Stroke. 2009, 40:451-8. Publisher Full Text OpenURL

  22. Jiang Q, Zhang ZG, Ding GL, et al.: MRI detects white matter reorganization after neural progenitor cell treatment of stroke.

    Neuroimage. 2006, 32:1080-9. PubMed Abstract | Publisher Full Text OpenURL

  23. Tanaka Y, Imai H, Konno K, et al.: Experimental model of lacunar infarction in the gyrencephalic brain of the miniature pig: neurological assessment and histological, immunohistochemical, and physiological evaluation of dynamic corticospinal tract deformation.

    Stroke. 2008, 39:205-12. PubMed Abstract | Publisher Full Text OpenURL

  24. Ramos-Cejudo J, Gutiérrez-Fernández M, Rodríguez-Frutos B, et al.: Spatial and temporal gene expression differences in core and periinfarct areas in experimental stroke: a microarray analysis.

    PLoS One. 2012., 7

    Article ID e52121

    PubMed Abstract | Publisher Full Text OpenURL

  25. Jadasz JJ, Kremer D, Göttle P, et al.: Mesenchymal stem cell conditioning promotes rat oligodendroglial cell maturation.

    PLoS One. 2013., 8

    Article ID e71814

    PubMed Abstract | Publisher Full Text OpenURL

  26. Crawford AH, Chambers C, Franklin RJ: Remyelination: the true regeneration of the central nervous system.

    J Comp Pathol. 2013, 149:242-54. PubMed Abstract | Publisher Full Text OpenURL

  27. Cheatwood JL, Emerick AJ, Schwab ME, Kartje GL: Nogo-A expression after focal ischemic stroke in the adult rat.

    Stroke. 2008, 39:2091-8. PubMed Abstract | Publisher Full Text OpenURL

 

GAML: genome assembly by maximum likelihood

Vladimír Boža, Broňa Brejová and Tomáš Vinař*

  • Corresponding author: Tomáš Vinař vinar@fmph.uniba.sk

Author Affiliations
Faculty of Mathematics, Physics, and Informatics, Comenius University, Mlynská dolina, Bratislava, 842 48, Slovakia

For all author emails, please log on.

Algorithms for Molecular Biology 2015, 10:18 doi:10.1186/s13015-015-0052-6

The electronic version of this article is the complete one and can be found online at: http://www.almob.org/content/10/1/18

Received: 8 April 2015
Accepted: 7 May 2015
Published: 3 June 2015
© 2015 Boža et al.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Formula display:
Abstract
Background
Resolution of repeats and scaffolding of shorter contigs are critical parts of genome assembly. Modern assemblers usually perform such steps by heuristics, often tailored to a particular technology for producing paired or long reads.

Results
We propose a new framework that allows systematic combination of diverse sequencing datasets into a single assembly. We achieve this by searching for an assembly with the maximum likelihood in a probabilistic model capturing error rate, insert lengths, and other characteristics of the sequencing technology used to produce each dataset. We have implemented a prototype genome assembler GAML that can use any combination of insert sizes with Illumina or 454 reads, as well as PacBio reads. Our experiments show that we can assemble short genomes with N50 sizes and error rates comparable to ALLPATHS-LG or Cerulean. While ALLPATHS-LG and Cerulean require each a specific combination of datasets, GAML works on any combination.

Conclusions
We have introduced a new probabilistic approach to genome assembly and demonstrated that this approach can lead to superior results when used to combine diverse set of datasets from different sequencing technologies. Data and software is available at http://compbio.fmph.uniba.sk/gaml webcite.

Keywords: Genome assembly; Maximum likelihood; Simulated annealing; De Bruijn graphs; Next generation sequencing
Background
The second and third generation sequencing technologies have dramatically decreased the cost of sequencing. Nowadays, we have a surprising variety of sequencing technologies, each with its own strengths and weaknesses. For example, Illumina platforms are characteristic by low cost and high accuracy, but the reads are short. On the other hand, Pacific Biosciences offer long reads at the cost of quality and coverage. In the meantime, the cost of sequencing was brought down to the point, where it is no longer a sole domain of large sequencing centers; even small labs can experiment with cost-effective genome sequencing. As a result, it is not realistic to assume an existence of a single standard protocol for sequencing genomes of a particular size. In this paper, we propose a framework for genome assembly that allows flexible combination of datasets from different technologies in order to harness their individual strengths.

Modern genome assemblers are usually based either on the overlap–layout–consensus framework (e.g. Celera [1], SGA [2]), or on de Bruijn graphs (e.g. Velvet [3], ALLPATHS-LG [4]). Both approaches can be seen as special cases of a string graph [5], in which we represent sequence fragments as vertices, while edges represent possible adjacencies of fragments in the assembly. A genome assembly is simply a set of walks through this graph. The main difference between the two frameworks is how we arrive at a string graph: through detecting long overlaps of reads (overlap–layout–consensus) or through construction of de Bruijn graphs based on k-mers.

However, neither of these frameworks is designed to systematically handle pair-end reads and additional heuristic steps are necessary to build larger scaffolds from assembled contigs. For example, ALLPATHS-LG [4] uses libraries with different insert lengths for scaffolding contigs assembled without the use of paired read information, while Cerulean [6] uses Pacific Biosystems long reads for the same purpose. Recently, the techniques of paired de Bruijn graphs [7] and pathset graphs [8] were developed to address paired reads systematically, however these approaches cannot combine libraries with different insert sizes.

Combination of sequencing technologies with complementary strengths can help to improve assembly quality. However, it is not feasible to design new algorithms for every possible combination of datasets. Often it is possible to supplement previously developed tools with additional heuristics for new types of data. For example, PBJelly [9] uses Pacific Biosystems reads solely to aid gap filling in draft assemblies. Assemblers like PacbioToCa [10] or Cerulean [6] use short reads to improve the quality of Pacific Biosystems reads so that they can be used within traditional assemblers. However, such approaches do not use all information contained within the datasets.

We propose a new framework that allows systematic combination of diverse datasets into a single assembly, without requiring a particular type of data for specific heuristic steps. Recently, probabilistic models have been used very successfully to evaluate the quality of genome assemblers [11]–[13]. In our work, we use likelihood of a genome assembly as an optimization criterion, with the goal of finding the assembly with the highest likelihood. Even though this may not be always feasible, we demonstrate that optimization based on simulated annealing can be very successful at finding high likelihood genome assemblies.

To evaluate the likelihood, we adapted a model by Ghodsi et al. [13], which can capture characteristics of each dataset, such as sequencing error rate, as well as length distribution and expected orientation of paired reads (“Probabilistic model for sequence assembly”). We can thus transparently combine information from multiple diverse datasets into a single score. Previously, there have been several works in this direction in much simpler models without sequencing errors [14], [15]. These papers used likelihood to estimate repeat counts, without considering other problems, such as how exactly are repeats integrated within scaffolds.

To test our framework, we have implemented a prototype genome assembler genome assembly by maximum likelihood (GAML) that can use any combination of insert sizes with Illumina or 454 reads, as well as PacBio reads. The starting point of the assembly are short contigs derived from Velvet [3] with very conservative settings in order to avoid assembly errors. We then use simulated annealing to combine these short contigs into high likelihood assemblies (“Finding a high likelihood assembly”). We compare our assembler to existing tools on benchmark datasets (“Experimental evaluation”), demonstrating that we can assemble genomes of up to 10 MB long with N50 sizes and error rates comparable to ALLPATHS-LG or Cerulean. For larger genomes, we can start from an assembly given by a different tool and improve on the result. While ALLPATHS-LG and Cerulean each require a very specific combination of datasets, GAML works on any combination.

Probabilistic model for sequence assembly
Recently, several probabilistic models were introduced as a measure of the assembly quality [11]–[13]. All of these authors have shown that the likelihood consistently favours higher quality assemblies. In general, the probabilistic model defines the probability View MathML that a set of sequencing reads R is observed assuming that assembly A is the correct assembly of the genome. Since the sequencing itself is a stochastic process, it is very natural to characterize concordance of reads and an assembly by giving a probability of observing a particular read. In our work, instead of evaluating the quality of a single assembly, we use the likelihood as an optimization criterion with the goal of finding high likelihood genome assemblies. We adapt the model of Ghodsi et al. [13], which we describe in this section.

Basics of the likelihood model
The model assumes that individual reads are independently sampled, and thus the overall likelihood is the product of likelihoods of the reads: View MathML To make the resulting value independent of the number of reads in set R, we use as the main assembly score the log average probability of a read computed as follows: View MathML Note that maximizing View MathML is equivalent to maximizing View MathML.

If the reads were error-free and each position in the genome was sequenced equally likely, the probability of observing read r would simply be View MathML, where View MathML is the number of occurrences of the read as a substring of the assembly A, L is the length of A, and thus 2L is the length of the two strands combined [14]. Ghodsi et al. [13] have shown a dynamic programming computation of read probability for more complex models, accounting for sequencing errors. The algorithm marginalizes over all possible alignments of r and A, weighting each by the probability that a certain number of substitution and indel errors would happen during sequencing. In particular, the probability of a single alignment with m matching positions and s errors (substitutions and indels) is defined as View MathML, where View MathML and View MathML is the sequencing error rate.

However, the full dynamic programming is too time consuming, and in practice only several best alignments contribute significantly to the overall probability. We approximate the probability of observing read r with an estimate based on a set Sr of a few best alignments of r to genome A, as obtained by one of the standard fast read alignment tools:

View MathML

(1)
where mj is the number of matches in the jth alignment, and sj is the number of mismatches and indels implied by this alignment. The formula assumes the simplest possible error model, where insertions, deletions, and substitutions have the same probability, and ignores GC content bias. Of course, much more comprehensive read models are possible (see e.g. [12]).

Paired reads
Many technologies provide paired reads produced from the opposite ends of a sequence insert of a certain size. We assume that the insert size distribution in a set of reads R can be modeled by the normal distribution with known mean μ and standard deviation σ. The probability of observing paired reads r1 and r2 can be estimated from the sets of alignments View MathML and View MathML as follows:

View MathML

(2)
As before, View MathML and View MathML are the numbers of matches and sequencing errors in alignment ji respectively, and View MathML is the distance between the two alignments as observed in the assembly. If alignments j1 and j2 are in two different contigs, or on inconsistent strands, View MathML is zero.

Reads that have no good alignment to A
Some reads or read pairs do not align well to A, and as a result, their probability View MathML is very low; our approximation by a set of high-scoring alignments can even yield zero probability if set Sr is empty. Such extremely low probabilities then dominate the log likelihood score. Ghodsi et al. [13] propose a method that assigns such a read a score approximating the situation when the read would be added as a new contig to the assembly. We modify their formulas for variable read length, and use score View MathML for a single read of length View MathML or View MathML for a pair of reads of lengths View MathML and View MathML. Values k and c are scaling constants set similarly as by Ghodsi et al. [13]. These alternative scores are used instead of the read probability View MathML whenever the probability is lower than the score.

Multiple read sets
Our work is specifically targeted at a scenario, where we have multiple read sets obtained from different libraries with different insert lengths or even with different sequencing technologies. We use different model parameters for each set and compute the final score as a weighted combination of log average probabilities for individual read sets View MathML:

View MathML

(3)
In our experiments, we use weight View MathML for most datasets, but we lower the weight for Pacific Biosciences reads, because otherwise they dominate the likelihood value due to their longer length. The user can also increase or decrease weights wi of individual sets based on their reliability.

Penalizing spuriously joined contigs
The model described above does not penalize obvious misassemblies when two contigs are joined together without any evidence in the reads. We have observed that to make the likelihood function applicable as an optimization criterion for the best assembly, we need to introduce a penalty for such spurious connections. We say that a particular base j in the assembly is connected with respect to read set R if there is a read which covers base j and starts at least k bases before j, where k is a constant specific to the read set. In this setting, we treat a pair of reads as one long read. If the assembly contains d disconnected bases with respect to d, penalty View MathML is added to the View MathML score (α is a scaling constant).

Properties of different sequencing technologies
Our model can be applied to different sequencing technologies by appropriate settings of model parameters. For example, Illumina technology typically produces reads of length 75–150 bp with error rate below 1% [16]. For smaller genomes, we often have a high coverage of Illumina reads. Using paired reads or mate pair technologies, it is possible to prepare libraries with different insert sizes ranging up to tens of kilobases, which are instrumental in resolving longer repeats [4]. To align these reads to proposed assemblies, we use Bowtie2 [17]. Similarly, we can process reads by the Roche 454 technology, which are characteristic by higher read lengths (hundreds of bases).

Pacific Biosciences technology produces single reads of variable length, with median length reaching several kilobases, but the error rate exceeds 10% [6], [16]. Their length makes them ideal for resolving ambiguities in assemblies, but the high error rate makes their use challenging. To align these reads, we use BLASR [18]. When we calculate the probability View MathML, we consider not only the best alignments found by BLASR, but for each BLASR alignment, we also add probabilities of similar alignments in its neighborhood. More specifically, we run a banded version of the forward algorithm by [13], considering all alignments in a band of size three around a guide alignment produced by BLASR.

Finding a high likelihood assembly
Complex probabilistic models, like the one described in “Probabilistic model for sequence assembly”, were previously used to compare the quality of several assemblies [11]–[13]. In our work, we instead attempt to find the highest likelihood assembly directly. Of course, the search space is huge, and the objective function too complex to admit exact methods. Here, we describe an effective optimization routine based on the simulated annealing framework [19].

Our algorithm for finding the maximum likelihood assembly consists of three main steps: preprocessing, optimization, and postprocessing. In preprocessing, we decrease the scale of the problem by creating an assembly graph, where vertices correspond to contigs and edges correspond to possible adjacencies between contigs supported by reads. In order to make the search viable, we will restrict our search to assemblies that can be represented as a set of walks in this graph. Therefore, the assembly graph should be built in a conservative way, where the goal is not to produce long contigs, but rather to avoid errors inside them. In the optimization step, we start with an initial assembly (a set of walks in the assembly graph), and iteratively propose changes in order to optimize the assembly likelihood. Finally, postprocessing examines the resulting walks and splits some of them into shorter contigs if there are multiple equally likely possibilities of resolving ambiguities. This happens, for example, when the genome contains long repeats that cannot be resolved by any of the datasets. In the rest of this section, we discuss individual steps in more detail.

Optimization by simulated annealing
To find a high likelihood assembly, we use an iterative simulated annealing scheme. We start from an initial assembly View MathML in the assembly graph. In each iteration, we randomly choose a move that proposes a new assembly View MathML similar to the current assembly A. The next step depends on the likelihoods of the two assemblies A and View MathML as follows:

If View MathML, the new assembly View MathML is accepted and the algorithm continues with the new assembly.

If View MathML, the new assembly View MathML is accepted with probability View MathML; otherwise View MathML is rejected and the algorithm retains the old assembly A for the next step.

Here, parameter T is called the temperature, and it changes over time. In general, the higher the temperature, the more aggressive moves are permitted. We use a simple cooling schedule, where View MathML in the ith iteration. The computation ends when there is no improvement in the likelihood for a certain number of iterations. We select the assembly with the highest LAP score as the result.

To further reduce the complexity of the assembly problem, we classify all contigs as either long (more than 500 bp) or short and concentrate on ordering the long contigs correctly. The short contigs are used to fill the gaps between the long contigs. Recall that each assembly is a set of walks in the assembly graph. A contig can appear in more than one walk or can be present in a single walk multiple times.

thumbnailFigure 1. Examples of proposal moves. a Walk extension joining two walks. b Local improvement by addition of a new loop. c Repeat interchange.
Proposals of new assemblies are created from the current assembly using the following moves:

Walk extension (Figure 1a) We start from one end of an existing walk and randomly walk through the graph, in every step uniformly choosing one of the edges outgoing from the current node. Each time we encounter the end of another walk, the two walks are considered for joining. We randomly (uniformly) decide whether we join the walks, end the current walk without joining, or continue walking.

Local improvement (Figure 1b) We optimize the part of some walk connecting two long contigs s and t. We first sample multiple random walks starting from contig s. In each walk, we only consider nodes from which contig t is reachable. Then we evaluate these random walks and choose the one that increases the likelihood the most. If the gap between contigs s and t is too big, we instead use a greedy strategy where in each step we explore multiple random extensions of the walk of length around 200 bp and pick the one with the highest score.

Repeat optimization We optimize the copy number of short tandem repeats. We do this by removing or adding a loop to some walk. We precompute the list of all short loops (up to five nodes) in the graph and use it for adding loops.

Joining with advice We join two walks that are spanned by long reads or paired reads with long inserts. We first select a starting walk, align all reads to this walk and randomly choose a read which has the other end outside the walk. Then we find to which node this other end belongs to and join appropriate walks. If possible, we fill the gap between the two walks using the same procedure as in the local improvement move. Otherwise we introduce a gap filled with Ns.

Disconnecting We remove a path through short contigs connecting two long contigs in the same walk, resulting in two shorter walks.

Repeat interchange (Figure 1c) If a long contig has several incoming and outgoing walks, we optimize the pairing of incoming and outgoing edges. In particular, we evaluate all moves that exchange parts of two walks through this contig. If one of these changes improves the score, we accept it and repeat this step, until the score cannot be improved at this contig.

At the beginning of each annealing step, the type of the move is chosen randomly; each type of move has its own probability. We also choose randomly the contig at which we attempt to apply the move.

Note that some moves (e.g. local improvement) are very general, while other moves (e.g. joining with advice) are targeted at specific types of data. This does not contradict a general nature of our framework; it is possible to add new moves as new types of data emerge, leading to improvement when using specific datasets, while not affecting the performance when such data is unavailable.

Preprocessing and the initial assembly
To obtain the assembly graph, we use Velvet with basic error correction and unambiguous concatenation of k-mers. These settings will produce very short contigs, but will also give a much lower error rate than a regular Velvet run. GAML with the default settings then uses each long contig as a separate walk in the starting assembly for the simulated annealing procedure.

Postprocessing
The assembly obtained by the simulated annealing procedure may contain walks with no evidence for a particular configuration of incoming and outgoing edges in the assembly graph. This happens for example if a repeat is longer than the span of the longest paired read. In this case, there would be several versions of the assembly with the same or very similar likelihood score. In the postprocessing step, we therefore apply the repeat interchange move at every possible location of the assembly. If the likelihood change resulting from such a move is negligible, we break the corresponding walks into shorter contigs to avoid assembly errors.

Fast likelihood evaluation
The most time consuming step in our algorithm is evaluation of the assembly likelihood, which we perform in each iteration of simulated annealing. This step involves alignment of a large number of reads to the current assembly. However, only a small part of the assembly is changed in each annealing step, which we can use to significantly reduce the running time. Next, we describe three optimizations implemented in our software.

Limiting read alignment to affected regions of the assembly Since only a small portion of the assembly is affected in each step, we can keep most alignments from the previous iterations and only align reads to the regions that changed. To determine these regions, we split walks into overlapping windows, each window containing several adjacent contigs of a walk. Windows should be as short as possible, but adjacent windows should overlap by at least View MathML bases, where View MathML is the length of the longest read. As a result, each alignment is completely contained in at least one window even in the presence of extensive indels.

We determine the window boundaries by a simple greedy strategy, which starts at the first contig of a walk, and then extends the window by at least View MathML bases beyond the boundary of the first contig. The next window always starts at the latest possible location that ensures a sufficient overlap and extends at least View MathML bases beyond the end of the previous window.

For each window, we keep the position and edit distance of all alignments. In each annealing step, we identify which windows of the assembly were changed since the last iteration. We then glue together overlapping windows and align reads against these sequences.

We further improve this heuristics by avoiding repeated alignments of reads to interiors of long contigs, because these parts of the assembly never change. In particular, if some window starts with a long contig, we only realign reads to the last View MathML bases from that contig, and similarly we use only the first View MathML bases from a long contig at the end of a window.

Reducing the number of reads which need to be aligned The first improvement eliminates most of the assembly from read mapping. In contrast, the second improvement reduces the set of reads which need to be realigned, because most of the reads will not align to the changed part of the assembly. We use a prefiltering step to find the reads which are likely to align to the target sequence. In the current implementation, we use the following three options for such filtering.

In the simplest approach, we look for reads which contain some k-mer (usually View MathML) from the target sequence. We store an index of all k-mers from all reads in a hash map. In each annealing step, we iterate over all k-mers in the target portion of the assembly and retrieve reads that contain them. This approach is very memory consuming, because the identifier of each read is stored for each k-mer from this read.

In the second approach, we save memory using min-hashing [20]. Given hash function h, the min-hash of set A is defined as View MathML. For each read R, we calculate min-hash for the set of all its k-mers. Thus, the identifier of each read is stored in the hash table only once. In each annealing step, we calculate the min-hash for each substring of the target sequence of length View MathML and retrieve the reads that have the same min-hash.

An important property of min-hashing is that View MathML, where View MathML is the Jaccard similarity of two sets A and B[21]. The statement holds if the hash function h is randomly chosen from a family with the min-wise independence property, which means that for every subset of elements X, each element in X has the same chance to have the minimum hash.

Note that strings with a very small edit distance have a high Jaccard similarity between their k-mer sets, and therefore a high chance that they will hash to the same value using min-hashing. We can use several min-hashes with different hash functions to improve the sensitivity of our filtering at the cost of additional memory.

In our implementation, we use a simple hash function which maps k-mers into 32-bit integers. We first represent the k-mer as an integer (where each base corresponds to two bits). We then xor this integer with a random number. Finally, we perform mixing similar to the finalization of the Murmur hash function [22]:

We choose this finalizer because the Murmur hash function is fast and results in few collisions. It is not min-hash independent, but we found it to perform well in practice.

To illustrate the specificity and sensitivity of min-hashing, we have compared our min-hashing approach with indexing all k-mers (with View MathML) on evaluating LAP of the Allpaths-LG assembly of Staphylococus aureus (using read set SA1 described in “Experimental evaluation” and aligning it to the whole S. aureus genome). Indexing all k-mers resulted in 3,659,273 alignments found by examining 21,241,474 candidate positions. Using min-hashing with three hash functions, we were able to find 3,639,625 alignments by examining 3,905,595 candidates positions. Since these reads have a low error rate, k-mer indexing retrieves practically all relevant alignments, while the sensitivity of min-hashing is approximately 99.5%. In min-hashing, 93% of examined positions yield an alignment, whereas specificity of k-mer indexing is only 17%. Also min-hashing used 30 times smaller index.

Note that min-hashing was previously used in a similar context by Berlin et al. [23] to find similarities among PacBio reads. However, since PacBio reads have a high error rate, the authors had to use a high number of hash functions, whereas we use only a few hash functions to filter Illumina reads, which have a low error rate.

In GAML, we filter PacBio reads by a completely different approach, which is based on alignments, rather than k-mers. In particular, we take all reasonably long contigs (at least 100 bases) and align them to PacBio reads. Since BLASR can find alignments where a contig and a read overlap by only around 100 bases, we use these alignments as a filter.

Final computation of the likelihood score When all reads are properly aligned to the new version of the assembly, we can combine the alignments to the final score. In the implementation, we need to handle several issues, such as correctly computing likelihood for reads that align to multiple walks, assigning a special likelihood to reads without any good alignment, and avoiding double counting for reads that align to regions covered by two overlapping windows of the same walk.

Again we improve the running time by considering only reads that were influenced by the most recent change. Between consecutive iterations, we keep all alignments for each sequence window of the assembly and recompute only alignments to affected windows, as outlined above. We also keep the likelihood value of each read or a read pair. Recall that the likelihood of a read or a read pair is the sum of likelihoods of individual alignments.

In each iteration, we then identify which walks were removed and added. Then we calculate likelihoods of all read alignments in these walks (using stored or newly computed alignments) and we use these values to adjust the likelihood values of individual reads, subtracting for removed walks and adding for new walks. At this step, we also handle paired reads, identifying pairs of alignments in correct distance and orientation. Finally, we sum likelihoods of all reads in each dataset and compute the total likelihood score.

Experimental evaluation
We have implemented the algorithm proposed in the previous section in a prototype assembler GAML. At this stage, GAML can assemble small genomes (approx. 10 Mbp) in a reasonable amount of time (approximately 4 h on a single CPU and using 10GB of memory).

To evaluate the quality of our assembler, we have adopted the methodology used for Genome Assembly Gold-Standard Evaluation [24], using metrics on scaffolds. We have used the same genomes and libraries as Salzber et al. [24] (the Staphylococus aureus genome and the human chromosome 14) and Deshpande et al. [6] (the Escherichia coli genome). The overview of the datasets is shown in Table 1. An additional dataset EC3 (long insert, low coverage) was simulated using the ART software [25].

We have evaluated GAML in the following scenarios:

  1. combination of fragment and short insert Illumina libraries (SA1, SA2),
  2. combination of a fragment Illumina library and a long-read high-error-rate Pacific Biosciences library (EC1, EC2),

  3. combination of a fragment Illumina library, a long-read high-error-rate Pacific Biosciences library, and a long jump Illumina library (EC1, EC2, EC3),

Table 1. Properties of datasets used
In each scenario, we use the short insert Illumina reads (SA1 or EC1) in Velvet with conservative settings to build the initial contigs and assembly graph. For the LAP score, we give all Illumina datasets weight 1 and the PacBio dataset weight 0.01. The results are summarized in Table 2. Note that none of the assemblers considered here can effectively run in all three of these scenarios, except for GAML.

Table 2. Comparison of assembly accuracy in the first three scenarios
In the first scenario, GAML performance ranks third among zero-error assemblers in the N50 length. The best N50 assembly is given by ALLPATHS-LG [4]. A closer inspection of the assemblies indicates that GAML missed several possible joins. One such miss was caused by a 4.5 kbp repeat, while the longest insert size in this dataset is 3.5 kbp. Even though in such cases it is sometimes possible to reconstruct the correct assembly thanks to small differences in the repeated regions, the difference in likelihood between alternative repeat resolutions may be very small. Another missed join was caused by a sequence coverage gap penalized in our scoring function. Perhaps in both of these cases the manually set constants may have caused GAML to be overly conservative. Otherwise, the GAML assembly is very similar to the one given by ALLPATHS-LG.

In the second scenario, Pacific Biosystems reads were employed instead of jump libraries. These reads pose a significant challenge due to their high error rate, but they are very useful due to their long length. Assemblers such as Cerulean [6] deploy special algorithms taylored to this technology. GAML, even though not explicitly tuned to handle Pacific Biosystems reads, builds an assembly with N50 size and the number of scaffolds very similar to that of Cerulean. In N50, both programs are outperformed by PacbioToCA [10], however, this is again due to a few very long repeats (approx. 5,000 bp) in the reference genome which were not resolved by GAML or Cerulean. (Cerulean also aims to be conservative in repeat resolution.) Note that in this case, simulated annealing failed to give the highest likelihood assembly among those that we examined, so perhaps our results can be improved by tuning the likelihood optimization.

The third scenario shows that the assembly quality can be hugely improved by including a long jump library, even if the coverage is really small (we have used 0.5× coverage in this experiment). This requires a flexible genome assembler; in fact, only Celera [1] can process this data, but GAML assembly is clearly superior. We have attempted to run also ALLPATHS-LG, but the program could not process this combination of libraries. Compared to the previous scenario, GAML N50 size increased approximately sevenfold (or approx. fourfold compared to the best N50 from the second scenario assemblies).

Table 3. Improving existing assemblies of the human chromosome 14 by GAML
Improving previously assembled genomes
For medium and large genomes, it would take GAML too many iterations to arrive at a reasonable assembly starting from the contigs produced by Velvet with conservative settings. However, it is still possible to scale up GAML to larger genomes by using another assembler to provide a more reasonable starting point.

To this end, we have to map such an input assembly to the assembly graph. We first align the assembly contigs to the Velvet contigs using NUCmer [26]. We keep only alignments which cover entire Velvet contigs and have a high sequence identity. If a single input contig is aligned to several Velvet contigs, we connect these Velvet contigs to a walk in the assembly graph. The missing portions of the walk are found by dynamic programming so as to minimize the edit distance between the input contig and the walk. In the dynamic programming, we consider only edit distance of up to 10, and if we do not find a connection within this threshold, we add a corresponding number of Ns to our walk.

If the input assembly differs too much from the Velvet contigs, a good mapping of the contigs to walks in the Velvet assembly graph cannot be found. In such cases, we construct the assembly graph directly from the input assembly. We first build a deBruijn graph from the contigs, and then we concatenate nodes connected by unambiguous connections.

We can now use GAML to improve medium-size genome assemblies (approx. 100 Mbp). In this setting, 10,000 iterations require approximately 2 days time and 50GB of memory.

We have tested this approach by using Illumina reads with three different insert sizes (H1, H2, H3) on the human chromosome 14 (data from [24]; see Table 1). We use the non-conservative Velvet assembly and the ALLPATHS assembly as our starting point. The results are shown in Table 3.

Starting from the Velvet assembly, GAML makes 787 breaks and 234 joins, reducing the error count by more than a thousand. Our joins did not introduce any new errors to the assembly. The ALLPATHS assembly has a much higher quality, and starting from this assembly, GAML decreases the number of errors only by one at the cost of introducing ten breaks. In both cases, we were able to remove some assembly errors, while not decreasing the error-corrected N50 values. Perhaps more corrections could be found if we ran our algorithm for more iterations (especially in the Velvet case).

Since breaks predominate in the changes made by GAML, we have also compared our results to REAPR [27], which is a tool that aligns reads to an existing assembly and then splits contigs at the positions weakly supported or even in conflict with the reads. When it concludes that some place is not a breakpoint, but should instead contain an insertion, it inserts a sequence of Ns. Note that REAPR can only process one jumping library along with an optional fragment library, and it requires the library to have a reasonable coverage (15×). Due to these constraints, we have used REAPR only with short jump library H2. For the Velvet assembly, REAPR removes significantly more errors than GAML, but at the cost of a great increase in the number of contigs and a decrease in the error-corrected N50 value. REAPR also introduces many cuts in the ALLPATHS assembly and the GAGE error checking tools report a high increase in errors. We hypothesize that this due to REAPR adding many regions of Ns in the corrected assembly, which leads to a high number of small contigs which GAGE checker cannot align correctly.

Conclusion
We have presented a new probabilistic approach to genome assembly, maximizing likelihood in a model capturing essential characteristics of individual sequencing technologies. It can be used on any combination of read datasets and can be easily adapted to other technologies arising in the future. We have also adapted our tool to improve existing assemblies after converting a given assembly to a set of walks.

Our work opens several avenues for future research. First, we plan to further improve running time and memory and to allow the use of our tool on larger genomes. Second, the simulated annealing procedure could be improved by optimizing probabilities of individual moves or devising new types of moves. Finally, it would be interesting to explore even more detailed probabilistic models, featuring coverage biases and various sources of experimental error.

Authors’ contributions
TV and VB have conceived the study. All authors have participated in design of algorithms and experiments. VB has implemented the software and conducted the experiments. All authors participated in manuscript preparation. All authors read and approved the final manuscript.

Acknowledgements
This research was funded by VEGA Grants 1/1085/12 (BB) and 1/0719/14 (TV). The authors would like to thank Viraj Deshpande for sharing his research data. An early version of this paper was published in WABI 2014 conference proceedings.

Compliance with ethical guidelines
Competing interests The authors declare that they have no competing interests.

References
Myers EW, Sutton GG, Delcher AL, Dew IM, Fasulo DP, Flanigan MJ, et al.: A whole-genome assembly of Drosophila.
Science 2000, 287(5461):2196-2204. PubMed Abstract | Publisher Full Text OpenURL
Simpson JT, Durbin R: Efficient construction of an assembly string graph using the FM-index.
Bioinformatics 2010, 26(12):367-373. Publisher Full Text OpenURL
Zerbino DR, Birney E: Velvet: algorithms for de novo short read assembly using de Bruijn graphs.
Genome Res 2008, 18(5):821-829. PubMed Abstract | Publisher Full Text OpenURL
Gnerre S, MacCallum I, Przybylski D, Ribeiro FJ, Burton JN, Walker BJ, et al.: High-quality draft assemblies of mammalian genomes from massively parallel sequence data.
Proc Nat Acad Sci 2011, 108(4):1513-1518. PubMed Abstract | Publisher Full Text OpenURL
Myers EW: The fragment assembly string graph.
Bioinformatics 2005, 21(suppl 2):79-85. Publisher Full Text OpenURL
Deshpande V, Fung ED, Pham S, Bafna V (2013) Cerulean: a hybrid assembly using high throughput short and long reads. In: Algorithms in Bioinformatics (WABI). LNCS, vol 8126. Springer, Berlin, pp 349–363
Medvedev P, Pham S, Chaisson M, Tesler G, Pevzner P: Paired de Bruijn graphs: a novel approach for incorporating mate pair information into genome assemblers.
J Comput Biol 2011, 18(11):1625-1634. PubMed Abstract | Publisher Full Text OpenURL
Pham SK, Antipov D, Sirotkin A, Tesler G, Pevzner PA, Alekseyev MA: Pathset graphs: a novel approach for comprehensive utilization of paired reads in genome assembly.
J Comput Biol 2013, 20(4):359-371. PubMed Abstract | Publisher Full Text OpenURL
English AC, Richards S, Han Y, Wang M, Vee V, Qu J, et al.: Mind the gap: upgrading genomes with Pacific Biosciences RS long-read sequencing technology.
PLoS One 2012, 7(11):47768. Publisher Full Text OpenURL
Koren S, Schatz MC, Walenz BP, Martin J, Howard JT, Ganapathy G, et al.: Hybrid error correction and de novo assembly of single-molecule sequencing reads.
Nat Biotechnol 2012, 30(7):693-700. PubMed Abstract | Publisher Full Text OpenURL
Rahman A, Pachter L: CGAL: computing genome assembly likelihoods.
Genome Biol 2013, 14(1):8. BioMed Central Full Text OpenURL
Clark SC, Egan R, Frazier PI, Wang Z: ALE: a generic assembly likelihood evaluation framework for assessing the accuracy of genome and metagenome assemblies.
Bioinformatics 2013, 29(4):435-443. PubMed Abstract | Publisher Full Text OpenURL
Ghodsi M, Hill CM, Astrovskaya I, Lin H, Sommer DD, Koren S, et al.: De novo likelihood-based measures for comparing genome assemblies.
BMC Res Notes 2013, 6(1):334. PubMed Abstract | BioMed Central Full Text OpenURL
Medvedev P, Brudno M: Maximum likelihood genome assembly.
J Comput Biol 2009, 16(8):1101-1116. PubMed Abstract | Publisher Full Text OpenURL
Varma A, Ranade A, Aluru S (2011) An improved maximum likelihood formulation for accurate genome assembly. In: Computational Advances in Bio and Medical Sciences (ICCABS 2011). IEEE, pp 165–170
Quail MA, Smith M, Coupland P, Otto TD, Harris SR, Connor TR, et al.: A tale of three next generation sequencing platforms: comparison of Ion Torrent, Pacific Biosciences and Illumina MiSeq sequencers.
BMC Genomics 2012, 13(1):341. PubMed Abstract | BioMed Central Full Text OpenURL
Langmead B, Salzberg SL: Fast gapped-read alignment with Bowtie 2.
Nat Methods 2012, 9(4):357-359. PubMed Abstract | Publisher Full Text OpenURL
Chaisson MJ, Tesler G: Mapping single molecule sequencing reads using basic local alignment with successive refinement (BLASR): application and theory.
BMC Bioinform 2012, 13(1):238. BioMed Central Full Text OpenURL
Eglese R: Simulated annealing: a tool for operational research.
Eur J Oper Res 1990, 46(3):271-281. Publisher Full Text OpenURL
Broder AZ (1997) On the resemblance and containment of documents. In: Proceedings of the Compression and Complexity of Sequences 1997. IEEE, pp 21–29
Broder AZ, Charikar M, Frieze AM, Mitzenmacher M: Min-wise independent permutations.
J Comput Syst Sci 2000, 60(3):630-659. Publisher Full Text OpenURL
Appleby A (2008) MurmurHash. https://code.google.com/p/smhasher/wiki/MurmurHash
Berlin K, Koren S, Chin CS, Drake J, Landolin JM, Phillippy AM (2014) Assembling large genomes with single-molecule sequencing and locality sensitive hashing. bioRxiv, 008003
Salzberg SL, Phillippy AM, Zimin A, Puiu D, Magoc T, Koren S, et al.: GAGE: a critical evaluation of genome assemblies and assembly algorithms.
Genome Res 2012, 22(3):557-567. PubMed Abstract | Publisher Full Text OpenURL
Huang W, Li L, Myers JR, Marth GT: ART: a next-generation sequencing read simulator.
Bioinformatics 2012, 28(4):593-594. PubMed Abstract | Publisher Full Text OpenURL
Delcher AL, Phillippy A, Carlton J, Salzberg SL: Fast algorithms for large-scale genome alignment and comparison.
Nucleic Acids Res 2002, 30(11):2478-2483. PubMed Abstract | Publisher Full Text OpenURL
Hunt M, Kikuchi T, Sanders M, Newbold C, Berriman M, Otto TD: Reapr: a universal tool for genome assembly evaluation.
Genome Biol 2013, 14(5):47. BioMed Central Full Text OpenURL

Fair evaluation of global network aligners

Joseph Crawford1, Yihan Sun123 and Tijana Milenković1*

  • Corresponding author: Tijana Milenković tmilenko@nd.edu

Author Affiliations
1 Department of Computer Science and Engineering, Interdisciplinary Center for Network Science and Applications (iCeNSA), ECK Institute for Global Health, University of Notre Dame, Notre Dame 46556, IN, USA

2 Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China

3 Department of Computer Science, Carnegie Mellon University, Pittsburgh 15213-3891, PA, USA

For all author emails, please log on.

Algorithms for Molecular Biology 2015, 10:19 doi:10.1186/s13015-015-0050-8

The electronic version of this article is the complete one and can be found online at: http://www.almob.org/content/10/1/19

Received: 30 August 2014
Accepted: 10 May 2015
Published: 9 June 2015
© 2015 Crawford et al.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Formula display:
Abstract
Background
Analogous to genomic sequence alignment, biological network alignment identifies conserved regions between networks of different species. Then, function can be transferred from well- to poorly-annotated species between aligned network regions. Network alignment typically encompasses two algorithmic components: node cost function (NCF), which measures similarities between nodes in different networks, and alignment strategy (AS), which uses these similarities to rapidly identify high-scoring alignments. Different methods use both different NCFs and different ASs. Thus, it is unclear whether the superiority of a method comes from its NCF, its AS, or both. We already showed on state-of-the-art methods, MI-GRAAL and IsoRankN, that combining NCF of one method and AS of another method can give a new superior method. Here, we evaluate MI-GRAAL against a newer approach, GHOST, by mixing-and-matching the methods’ NCFs and ASs to potentially further improve alignment quality. While doing so, we approach important questions that have not been asked systematically thus far. First, we ask how much of the NCF information should come from protein sequence data compared to network topology data. Existing methods determine this parameter more-less arbitrarily, which could affect alignment quality. Second, when topological information is used in NCF, we ask how large the size of the neighborhoods of the compared nodes should be. Existing methods assume that the larger the neighborhood size, the better.

Results
Our findings are as follows. MI-GRAAL’s NCF is superior to GHOST’s NCF, while the performance of the methods’ ASs is data-dependent. Thus, for data on which GHOST’s AS is superior to MI-GRAAL’s AS, the combination of MI-GRAAL’s NCF and GHOST’s AS represents a new superior method. Also, which amount of sequence information is used within NCF does not affect alignment quality, while the inclusion of topological information is crucial for producing good alignments. Finally, larger neighborhood sizes are preferred, but often, it is the second largest size that is superior. Using this size instead of the largest one would decrease computational complexity.

Conclusion
Taken together, our results represent general recommendations for a fair evaluation of network alignment methods and in particular of two-stage NCF-AS approaches.

Keywords: Protein–protein interaction networks; Network alignment ; Network similarity; Across-species protein function prediction
Background
Motivation and related work
Analogous to sequence alignment, which finds regions of similarity that are a likely consequence of functional or evolutionary relationships between the sequences, network (or graph) alignment finds regions of topological and functional similarity between networks of different species [1]. Then, functional (e.g., aging-related [2]–[4]) knowledge can be transferred between species across conserved (aligned) network regions. Thus, just as sequence alignment, network alignment can be used for establishing from biological network data orthologous relationships between different proteins or phylogenetic relationships between different species [5]–[7]. Also, it can be applied to research problems in other domains, such as semantically matching entities in different ontologies [8], or comparing online social networks with impacts on user privacy [9].

Network alignment can be performed locally and globally. Local network alignment (LNA) aims to optimize similarity between local regions of different networks [10]–[19]. As such, LNA often leads to many-to-many node mapping between different networks. However, LNA is generally unable to find large conserved subgraphs. Thus, methods for global network alignment (GNA) have been proposed, which aim to optimize global similarity between different networks and can thus find large conserved subgraphs [2], [3], [5]–[7], [9], [20]–[31]. Unlike LNA, GNA typically results in one-to-one node mapping between different networks (though some exceptions exist that result in one-to-many or many-to-many node mapping [24], [32]). In this study, we focus on one-to-one GNA due to its recent popularity [2], [3], [31], but all concepts and ideas can also be applied to one-to-many or many-to many GNA, as well as to LNA.

More formally, we define GNA as a one-to-one mapping between nodes of two networks that aligns the networks well with respect to a desired topological or functional criterion. GNA is a computationally hard problem to solve due to the underlying subgraph isomorphism problem [33]. This is an NP-complete problem that asks whether a network exists as an exact subgraph of a larger network. GNA is a more general problem which aims to fit well two networks when one network is not necessarily an exact subgraph of another network. Since GNA is computationally hard, heuristic methods need to be sought. Many (though not all) GNA heuristic algorithms typically achieve an alignment via two algorithmic components: node cost function (NCF) and alignment strategy (AS) [5]–[7], [25], [26], [30], [34]–[36]. NCF captures pairwise costs (or equivalently, similarities) of aligning nodes in different networks, and AS uses these costs to identify a good-quality alignment out of all possible alignments with respect to some topological or biological alignment quality measure [2], [3], [5]–[7], [20], [24]–[26], [34].

Different existing two-step GNA methods use both different NCFs and ASs, so it is unclear whether the superiority of a method comes from its NCF, AS, or both. For this reason, in our recent study [2], [3], we combined NCFs and ASs of MI-GRAAL [7] and IsoRankN [24], two state-of-the-art methods at the time, as a proof of concept that it is important to fairly evaluate the contribution of each component to alignment quality. In the process, we showed that NCF of MI-GRAAL is superior to that of IsoRankN, and importantly, we proposed the combination of MI-GRAAL’s NCF and IsoRankN’s AS as a new superior method for multiple GNA, i.e., for GNA of more than two networks at a time [2], [3].

In the meanwhile, a new state-of-the-art method has appeared, called GHOST [25]. When recently tested against many other both previous and newer GNA methods, GHOST was described as still “an excellent performer” [31]. Thus, in this study, we aim to understand whether it is GHOST’s NCF or AS (or both) that leads to its good performance, as well as to explore the possibility of further increasing GHOST’s performance by replacing its current NCF with a different, potentially superior NCF. For these reasons, we fairly evaluate MI-GRAAL against GHOST by mixing and matching their NCFs and ASs. We use MI-GRAAL in this study because we already demonstrated the superiority of its NCF, as discussed above [2], [3]. At the same time, we ask several additional important questions regarding the choice of appropriate GNA parameters, which have surprisingly been neglected thus far.

We note that some of the existing one-to-one GNA methods do not belong to this two-stage NCF-AS method category, and clearly, our study might not directly be applicable to such approaches. However, many of the existing one-to-one GNA methods do belong to the two-stage category, such as two versions of IsoRank [20], [22], GRAAL [5], H-GRAAL [6], MI-GRAAL [7], and GHOST [25]. It is very likely that many new methods will build on top of these well-established state-of-the-art methods, and thus, our study is of importance for future GNA method development.

Also, we note that although we already showed on the example of MI-GRAAL and IsoRankN that combining NCF of one method and AS of another method can lead to a new superior method [2], [3], testing whether the same holds for MI-GRAAL and GHOST, and in particular identifying the superior of the two NCFs, is of importance. First, validating that this also holds for MI-GRAAL and GHOST would only further stress out the need to carefully design a strategy for evaluating a novel approach against existing ones. Simply comparing the approaches, as has typically been done, is not enough. A more advanced evaluation strategy, such as our mix-and-match approach, is more appropriate. Second, MI-GRAAL’s NCF is a graphlet-based measure of topological node similarity [37] that is also used by many other network aligners [2], [3], [5], [6], [38] or even network clustering methods [37], [39], [40] to link network topology with biological function. When a new measure of topological similarity appears that is also argued to successfully capture biological function, such as GHOST’s NCF, it is extremely important to fairly compare it against the graphlet-based node similarity measure (which has not been done to date). In this way, future studies oriented towards learning new biological knowledge from network topology can focus on the most accurate node similarity measure. And this is exactly one of the goals of our study—to determine which of the two NCFs is superior. (We already demonstrated the superiority of MI-GRAAL’s graphlet-based NCF over IsoRankN’s popular PageRank-based NCF [2], [3].)

Our approach and contributions
MI-GRAAL [7] and GHOST [25] are two state-of-the-art global network aligners that injectively map nodes between two networks in a way that preserves topologically or functionally conserved network regions. The two methods are conceptually similar, in the sense that their NCFs assume two nodes from different networks to be similar if their topological neighborhoods are similar. However, the mathematical and implementation details of the two NCFs are different. The same holds for the two methods’ ASs. To evaluate the contribution to the alignment quality of each of the two NCFs and two ASs, we mix and match these, resulting in a total of four different combinations. We then use each combination to produce alignments for synthetic networks with known ground truth node mapping as well as for real-world networks without known ground truth node mapping, and we evaluate the quality of each alignment with respect to five topological and two biological alignment quality measures.

In general, we find that MI-GRAAL’s NCF is superior to GHOST’s NCF, while the superiority of the methods’ ASs is data-dependent. Hence, for those network data on which GHOST’s AS is superior to MI-GRAAL’s AS, we propose the combination of MI-GRAAL’s NCF and GHOST’s AS as a new superior network aligner.

While fairly evaluating MI-GRAAL’s and GHOST’s NCFs and ASs, we approach two additional important research questions that, to our knowledge, have not been asked systematically in the context of network alignment thus far: (1) how much of the node similarity information within the NCF should come from protein sequence data compared to network topology data, and (2) how large the size of the neighborhoods of the compared nodes from different networks should be when generating topological similarity information within the NCF. Current GNA methods generally use a seemingly arbitrary amount of sequence information in their NCF, and also, they assume that the larger the size of a node’s neighborhood, the better the alignment quality. Thus, in this study, we evaluate whether these “state-of-the-art” choices are actually appropriate. We note that the first question has been recognized in some of the existing work [25], [31], [41], but this question has not been systematically addressed to the same extent as in our study. To our knowledge, the second question has not been addressed at all thus far.

In general, we find that which amount of sequence information is used within NCF does not drastically affect neither topological or biological alignment quality, while the effect of topological information is drastic. Namely, using no topological information within NCF results in poor topological and sometimes even biological alignment quality. Hence, topology takes precedence over sequence when it comes to improving alignment quality. Also, we find that using larger network neighborhood sizes within NCF in most cases leads to better alignment quality than using smaller neighborhood sizes. However, it is not always the case that the largest neighborhood size is the best; in many cases, the second largest size is the best. Therefore, using this size instead of the largest one would drastically decrease computational complexity of the given method without decreasing its accuracy.

We note that a recent study [31] performed a valuable survey of a number of GNA methods, focusing in the process on ranking the different methods based on their performance. However, that study did not focus on in-depth understanding why a given aligner performs the way it does, which is what we aim to do in our study. By analyzing a GNA method’s NCF and AS individually, we are able to understand the effect on alignment quality of each of the two algorithmic components. Furthermore, this existing study [31] compared the different methods with respect to a topological alignment quality measure called induced conserved structure (ICS) [25]. However, recently it was shown that ICS is an inappropriate measure of topological alignment quality, and a new superior measure was proposed, called symmetric substructure score (View MathML) [30]. Here, we use the View MathML measure, along with several additional measures, thus increasing the confidence in our results compared to the results reported in Clark and Kalita [31]. In addition, this existing study [31] evaluated the different network aligners only on real-world networks of different species, for which the ground truth node mapping is not known. Here, we do the same, and we also align a high-confidence biological network to its noisy counterparts (“Data sets”). In the latter case, the ground truth node mapping is known and we can thus measure how well each aligner reconstructs the node mapping [corresponding to node correctness (“Network alignment quality measures”)]. This important evaluation cannot be done when the actual node mapping is not known and was thus not carried out in Clark and Kalita [31], despite the fact that measuring node correctness is the most appropriate way of evaluating a network aligner’s accuracy [5]–[7], [25], [30] before applying the aligner to networks of different species to learn new biological knowledge. Moreover, this existing study [31] still arrived to the conclusion that GHOST is “an excellent performer”, despite the fact that many newer methods were involved into the comparison. Thus, our results showing that we can improve GHOST even further by using its AS on top of MI-GRAAL’s NCF are an additional novel contribution of our study. Finally, we note again that in addition to providing comprehensive in-depth evaluation of the two prominent network aligners (rather than simply comparing their performance as in Clark and Kalita [31]), we also study in detail the effect of different parameters (such as the amount of sequence information or neighborhood size considered within NCF) on the alignment quality; this was not done in the recent study [31].

Methods
Data sets
We use two popular benchmark sets of networks in this study: (1) synthetic networks with known ground truth node mapping and (2) real-world protein–protein interaction (PPI) networks without known ground truth node mapping [2], [3], [7], [25], [30].

The synthetic network data with known node mapping consists of a high-confidence yeast PPI network, which has 1,004 proteins and 8,323 PPIs [5]–[7], [25], [30], [42], and five additional networks that add noise to the yeast network. Noise is the addition to the yeast network of low-confidence edges from the same data set [42], and each of the five additional noisy networks adds View MathML noise to the original network, where View MathML varies from 5 to 25% in increments of 5%. In this network set, we align the original yeast network to each of the synthetic networks with View MathML noise, resulting in the total of five network pairs to be aligned.

The real-world PPI network data without known node mapping consists of PPI networks of the following four species: S. cerevisiae (yeast/Y), D. melanogaster (fly/F), C. elegans (worm/W), and H. sapiens (human/H). The yeast, fly, worm, and human networks have 3,321 proteins and 8,021 PPIs, 7,111 proteins and 23,376 PPIs, 2,582 proteins and 4,322 PPIs, and 6,167 proteins and 15,940 PPIs, respectively [43]. In this network set, we align PPI networks for each pair of species, resulting in the total of six network pairs to be aligned.

We note that the synthetic network data is not truly synthetic, as both the original yeast network and the noise in terms of the lower-confidence PPIs come from an actual experimental study [42]. We refer to this network set as synthetic simply because we know the known ground truth node mapping, unlike for the real-world PPI network set. Also, we note that the synthetic network data encompasses “co-complex” PPIs obtained by affinity purification followed by mass-spectrometry (AP/MS), among other PPI types, while the real-world PPI network data consists of “binary” yeast two-hybrid (Y2H) PPIs. Another difference between the two network sets is that for the synthetic data the smaller (original yeast) network is an exact subgraph of the larger (noisy) network, whereas this is not the case for networks of different species in the real-world data.

When evaluating the amount of sequence data that should be used within NCF when generating an alignment, we use protein sequence similarity data. This data set comes from BLAST bit-values from the NCBI database [44].

When evaluating the biological alignment quality with respect to functional enrichment of the aligned nodes, we use Gene Ontology (GO) annotation data from our recent study [2], [3].

Importantly, we note that we use the above data sources and versions of the data because the exact same data have already been used in the existing work, which allows for fair and consistent method evaluation. If the main focus of one’s work was to predict new biological knowledge rather than to conduct fair method evaluation and comparison, then we would recommend using the latest and thus most complete versions of the data.

Existing network aligners and their NCFs and ASs
MI-GRAAL’s NCF
MI-GRAAL improves upon its predecessors, GRAAL [5] and H-GRAAL [6], by using the same NCF (see below) but by combining GRAAL’s and H-GRAAL’s ASs into a new superior AS (see below).

MI-GRAAL’s NCF relies on the concept of small induced subgraphs called graphlets (Figure 1) [37], [39], [40], [45]–[47]. All 2–5-node graphlets are considered. Because of the small-world nature of real-world networks, using larger graphlets would unnecessarily increase the computational complexity needed the count the graphlets [5], [6]. Based on the graphlets, the node graphlet degree vector (node-GDV) is computed for each node in each network, which counts how many times the given node touches each of the 2–5-node graphlets, i.e., each of their 73 node symmetry groups (or automorphism orbits; Figure 1). As such, node-GDV captures up to a four-deep network neighborhood of the node of interest. By comparing node-GDVs of two nodes to compute their node-GDV-similarity, and by doing so between each pair of nodes in different networks, one is able to capture pairwise topological node similarities between the different networks.

thumbnailFigure 1. Illustration of MI-GRAAL’s NCF. To compute topological node similarities, this NCF relies on thirty 2-, 3-, 4-, and 5-node graphlets View MathML and their “node symmetry groups”, also called automorphism orbits, numbered 0, 1, 2,…, 72. In a graphlet View MathML, View MathML, nodes belonging to the same orbit are of the same shade. For details, see the original publication [49].
MI-GRAAL also allows for integration of other node similarity measures into its NCF, such as protein sequence similarity. Thus, MI-GRAAL has the built-in functionality of allowing the user to incorporate their own custom pairwise node scores rather than rely on MI-GRAAL’s NCF, which is exactly how we incorporate GHOST’s NCF as input into MI-GRAAL’s AS.

MI-GRAAL’s AS
GRAAL’s AS utilizes a seed-and-extend approach to greedily maximize the total NCF over all aligned nodes. H-GRAAL, on the other hand, finds optimal alignments with respect to the total NCF by using the Hungarian algorithm to solve the linear assignment problem. MI-GRAAL’s AS combines GRAAL’s greedy seed-and-extend approach with H-GRAAL’s optimal AS into a superior AS.

Specifically, for graphs View MathML and View MathML, MI-GRAAL’s AS selects a pair of nodes View MathML and View MathML, where View MathML and View MathML, which have the highest similarity score among all pairs of nodes from the different networks. It then begins to align these nodes’ neighbors as follows. Let us denote by View MathML and View MathML the sets of neighbors of nodes View MathML and View MathML, respectively. A bipartite graph is constructed using nodes from View MathML and View MathML, where there exists an edge between a node View MathML from View MathML and a node View MathML from View MathML if and only if a neighbor of View MathML has already been aligned to a neighbor of View MathML. A confidence weight (i.e., the NCF-based similarity between two given nodes) is then assigned to each edge. Given the resulting bipartite graph, MI-GRAAL’s AS solves the maximum weight bipartite matching problem to determine which nodes in View MathML and View MathML should be aligned to each other. After MI-GRAAL’s AS is done aligning nodes from View MathML to nodes from View MathML, it then expands to these nodes’ neighbors and repeats the above steps to align them. The expansion continues iteratively until the entire smaller network is exhausted. For more details on MI-GRAAL’s AS, see the original publication [7].

GHOST’s NCF
GHOST’s NCF takes into account a node’s View MathML-hop neighborhood View MathML, which is the induced subgraph on all nodes whose shortest path distance from the node in question is less than or equal to View MathML (Figure 2). Intuitively, GHOST’s NCF computes topological distance (or equivalently similarity) between two nodes from different networks by comparing the nodes’ “spectral signatures”. These signatures are based on the spectrum of the normalized Laplacian for subgraphs of radius View MathML centered around a given node. Essentially, the spectral signature of a node is based on subgraph counts in the node’s k-hop neighborhood [25]. GHOST also allows for the incorporation of sequence information into its NCF, in which the resulting NCF is a linear combination of GHOST’s topological and sequence distance scores. For further details on GHOST’s NCF, refer to the original publication [25]. In our study, we consider View MathML, which allows for a fair comparison of GHOST’s NCF to MI-GRAAL’s NCF when varying the size of network neighborhood that is considered within the NCFs (“Aligners resulting from combining existing NCFs and ASs, and their parameters”).

thumbnailFigure 2. Illustration of GHOST’s NCF. To compute topological node similarities, this NCF compares two nodes in different networks with respect to similarity of each of their View MathML-hop neighborhoods, View MathML. All blue edges and blue nodes are within the given View MathML-hop neighborhood of the red node.
GHOST’s AS
GHOST’s AS is also a seed-and-extend method, but unlike MI-GRAAL’s AS that deals with the linear assignment problem, GHOST’s AS deals with the quadratic assignment problem (Figure 3 illustrates this). GHOST’s AS uses a two-phase seed-and-extend strategy by first selecting nodes View MathML and View MathML, where View MathML and View MathML, which have the highest similarity score among all pairs of nodes from the different networks, and then extending around these nodes to align their neighbors [i.e., nodes from View MathML and View MathML]. To do this, GHOST’s AS considers pairwise similarities between nodes in View MathML and View MathML in addition to similarities between nodes within the same network, and all of these similarities are used to estimate a solution to the quadratic assignment problem, which is the node alignment. For further details on GHOST’s AS, refer to the original publication [25].

thumbnailFigure 3. Intuitive comparison of MI-GRAAL’s and GHOST’s ASs. Let us assume that we are aligning two graphs View MathML and View MathML Let View MathML, let View MathML, and let the NCF distance (equivalently, similarity) between the node pairs be View MathML, as illustrated. MI-GRAAL’s alignment strategy only considers the values View MathML and View MathML when creating an alignment, while GHOST’s AS considers the values View MathML when doing so.
Aligners resulting from combining existing NCFs and ASs, and their parameters
Mixing and matching different NCFs and ASs
To fairly evaluate the two algorithmic components of MI-GRAAL and GHOST, we aim to first compare the two NCFs under the same AS, for each of the two ASs. We then aim to compare the two ASs under the same NCF, for each of the two NCFs. This results in a total of four aligners, i.e., different combinations of the two methods’ NCFs and ASs. However, GHOST does not allow the user to import their own (e.g., MI-GRAAL’s) NCF into its AS, so we are unable to study the combination of MI-GRAAL’s NCF and GHOST’s AS. Thus, in total, we consider three different aligners (Table 1).

Table 1. The three aligners considered in this study
Varying the amount of sequence versus topological information within NCF
An additional goal of this paper is to determine the most appropriate amount of sequence information versus topological information to be included into NCF. Thus, for each aligner, we generate NCFs with varying amounts of sequence and topology information, as View MathML where View MathML represents topological similarity score (e.g. node-GDV-similarity) and View MathML represents sequence similarity score. We vary View MathML from 0 to 1 in increments of 0.1.

Varying the size of network neighborhood within NCF
Further, we aim to determine the most appropriate neighborhood size that should be used within NCF when producing an alignment. Thus, for each aligner (and for each value of View MathML), we also consider four different neighborhood sizes, as described in Table 2. We note that although we have tried to classify under the same neighborhood size label (e.g. T1 in Table 2) graphlet sizes considered within MI-GRAAL’s NCF and View MathML-hop values considered within GHOST’s NCF, it is not necessarily the case that the neighborhood of a node that is covered by graphlets of a given size and the neighborhood of the same node that is covered by the corresponding View MathML-hop value match exactly. That is, for example, 2–3-node graphlets and 2-hop neighborhood (both corresponding to T2 in Table 2) do not necessarily cover exactly the same amount of network topology. Yet, we have aimed to provide as accurate as possible classification in Table 2, in order to allow for as fair as possible comparison of the two methods’ NCFs under varying sizes of network neighborhoods.

Table 2. The four neighborhood sizes that we vary within each aligner
Implementation details
The types of scores that MI-GRAAL and GHOST take in as input are different: MI-GRAAL looks at node similarities (the higher the score, the more similar the nodes), while GHOST looks at node distances (the lower the score, the more similar the nodes). We carefully take this into account to allow for fair method comparison. For example, to ensure that neither NCF has an advantage due to the format of the scores, we normalize all scores. That is, node similarity scores used in MI-GRAAL can exceed the value one, while no scores generated by GHOST are greater than one. To make the two sets of scores comparable, we scale MI-GRAAL’s node similarity scores to the [0–1] range by dividing each of the scores by the maximum similarity score. Because GHOST deals with distances rather than similarities, we take one minus GHOST’s NCF and then plug in the resulting node scores into MI-GRAAL’s AS.

Further, MI-GRAAL’s NCF returns all pairwise node similarity scores between two networks. However, GHOST’s NCF returns only a subset of all pairwise node distance scores, depending on the network size. To complete GHOST’s pairwise node score matrix and thus allow for it to be given as input into MI-GRAAL’s AS, we assign a score equal to the highest distance score returned by GHOST to all node pairs for which GHOST did not return a distance score.

Finally, the current implementation of MI-GRAAL’s AS does not function properly when a large pairwise node similarity matrix is plugged into it. Thus, MI-GRAAL’s AS has had difficulty aligning the two largest networks from our study, the fly and human networks. As a solution, we create a matrix that contains only the top 21 million node similarity scores of the original node similarity matrix, this being the maximum that our computational resources would process. With this adjustment, we are successfully able to generate all fly-human alignments.

Network alignment quality measures
We use well established network alignment quality measures [2], [3], [30]. Let View MathML and View MathML be two graphs such that View MathML. An alignment of View MathML to View MathML is a total injective function View MathML; every element of View MathML is matched uniquely with an element of View MathML. Let us denote by View MathML the set of edges from View MathML that exist between nodes in View MathML that are aligned by View MathML to nodes in View MathML.

Topological evaluation
We use five measures of topological alignment quality:

  1. Node correctness (NC) If View MathML is the correct ground truth node mapping between View MathML and View MathML (when such mapping is known), then NC of alignment View MathML is: View MathML[5]. This measure can be computed only for alignments of the synthetic network set with known ground truth node mapping (“Data sets”). All remaining measures (listed below) can be computed for the real network set with unknown node mapping as well.
  2. Edge correctness (EC) EC is the percentage of edges from View MathML, the smaller network (in terms of the number of nodes), which are aligned to edges from View MathML, the larger network [5]. Formally, View MathML where the numerator is the number of “conserved” edges, i.e., edges that are aligned under the given node mapping. The larger the EC score, the better the alignment.

  3. Induced conserved structure (ICS) View MathML EC might fail to differentiate between alignments that one might intuitively consider to be of different topological quality [25], since it is defined with respect to edges in View MathML. For example, aligning a View MathML-node cycle in View MathML to a View MathML-node cycle in View MathML would result in the same EC as aligning a View MathML-node cycle in View MathML to a View MathML-node clique (complete graph) in View MathML Clearly, the former is intuitively a better alignment than the latter, since no edges that exist between the View MathML nodes in View MathML are left unaligned in the first case, whereas many edges are left unaligned in the second case. Since ICS is defined with respect to edges in View MathML, it would have the maximum value of 100% when aligning a View MathML-node cycle to a View MathML-node cycle, and it would have a lower value when aligning a View MathML-node cycle to a View MathML-node clique [30]. The larger the ICS, the better.

  4. Symmetric substructure score (View MathML) EC penalizes the alignment for having misaligned edges in the smaller network. ICS penalizes the alignment for having misaligned edges in the larger network. SView MathML on the other hand, aims to improve upon EC and ICS by penalizing for misaligned edges in both the smaller and larger network. SView MathML For details, see the original publication [30].

  5. The size of the largest connected common subgraph (LCCS) [5], which we use for the following reason. Of two alignments with similar EC, ICS, or SView MathML scores, one could expose large, contiguous, and topologically complex regions of network similarity, while the other could fail to do so. Thus, in addition to counting aligned edges or nodes that participate in the aligned edges, it is important that the aligned edges cluster together to form large connected subgraphs rather than being isolated. Hence, we define a connected common subgraph (CCS) as a connected subgraph (not necessarily induced) that appears in both networks [6]. We measure the size of the largest CCS (LCCS) in terms of the number of nodes as well as edges. Namely, we compute the LCCS score as in our recent work [30]. First, we count View MathML, the percentage of nodes from View MathML that are in the LCCS. Then, we count View MathML, the percentage of edges that are in the LCCS out of all edges that could have been aligned between the nodes in the LCCS. That is, View MathML is the minimum of the number of edges in the subgraph of View MathML that is induced on the nodes from the LCCS, and the number of edges in the subgraph of View MathML that is induced on the nodes from the LCCS [30]. Finally, we compute their geometric mean as View MathML, in order to penalize alignments that have small View MathML or small View MathML. Large values of this final LCCS score are desirable.

Biological evaluation
Only alignments in which many aligned node pairs perform the same function should be used to transfer function from annotated parts of one network to unannotated parts of another network [30]. Hence, we measure GO [48] enrichment of aligned proteins pairs, i.e., the percentage of protein pairs in which the two proteins share at least one GO term, out of all aligned protein pairs in which both proteins are annotated with at least one GO term. We refer to this percentage as GO correctness (GO). We do this with respect to complete GO annotation data, independent of GO evidence code. Also, since many GO annotations have been obtained via sequence comparison, and since some of the aligners use sequence information, we repeat the analysis considering only GO annotations with experimental evidence codes, in order to avoid the circular argument. In this case, we refer to GO correctness as experimental GO correctness (EXP). The higher the GO and EXP values, the better [30].

Results and discussion
We aim to answer the following three main questions in the context of network alignment: (1) which NCF and AS is superior to the other, and is there perhaps a combination of one existing method’s NCF and another existing method’s AS that is the superior aligner in terms of accuracy as well as time complexity (“What is the best NCF and the best AS?”)? (2) How much sequence versus topological information to use within NCF (“The amount of sequence versus topological information within NCF?”)? (3) How large the size of network neighborhoods of compared nodes to consider within NCF (“The size of nodes’ neighborhoods within NCF?”)? In addition, we comment on relationships between different alignment quality measures (“Relationships between different alignment quality measures”). Finally, we conclude in “Conclusions”.

What is the best NCF and the best AS?
By comparing M-M and G-M aligners, we can fairly compare the two NCFs under the same (MI-GRAAL’s) AS. Also, by comparing G-M and G-G, we can fairly compare the two ASs under the same (GHOST’s) NCF. See “Aligners resulting from combining existing NCFs and ASs, and their parameters” for details on each aligner.

Synthetic networks with known node mapping
Overall, GHOST’s NCF is slightly superior to that of MI-GRAAL (Figure 4a, b). Also, GHOST’s AS is superior to MI-GRAAL’s AS (Figure 4a, b). However, these findings are based on all alignments (with known node mapping) for all values of View MathML, all neighborhood sizes, and all measures of alignment quality combined (“Aligners resulting from combining existing NCFs and ASs, and their parameters”), which might not be fair. Thus, in Figure 5a–c, for each aligner, for each alignment quality measure, we show results for the best alignments over all values of View MathML and all neighborhood sizes, for three out of all five network pairs (for the remaining network pairs, see Additional file 1: Figures S1 and S2). Now, the general trend (and especially with respect to NC as the most accurate ground truth measure of alignment quality) is that the best scores for M-M are either comparable or superior to those of G-M, indicating slight superiority of MI-GRAAL’s NCF over GHOST’s. Nonetheless, G-G still always outperforms G-M, indicating superiority of GHOST’s AS over MI-GRAAL’s AS.

thumbnailFigure 4. The ranking of the three aligners (M-M, G-M, and G-G). The ranking is shown over all alignments for all values of View MathML and all neighborhood sizes, with respect to: a all topological scores of all alignments with known ground truth node mapping, b all biological scores of alignments with known node mapping, c all topological scores of alignments with unknown node mapping, and d all biological scores of alignments with unknown node mapping. Percentages represent the percentage of cases that an aligner achieved a certain ranking.
It is possible to break down the above results and study how the ranking of the different NCFs and ASs changes with the change in the value of View MathML, which corresponds to the amount of topological similarity information used within NCF (Additional file 1: Figures S3–S7). In general, MI-GRAAL’s NCF is comparable to GHOST’s NCF across all View MathML values, as M-M and G-M scores are similar. On the other hand, GHOST’s AS shows superiority over MI-GRAAL’s AS, as G-G consistently results in higher scores than G-M. We note that we show that the value of View MathML does not greatly affect alignment quality (“The amount of sequence versus topological information within NCF?”).

It is also possible to break down the above results even further and study how the ranking of the different NCFs and ASs changes with the change in the neighborhood size that is considered within NCF (Additional file 1: Figures S3–S7). In general, for the smaller neighborhood sizes (T1 and T2), GHOST’s NCF generally produces comparable or superior results to MI-GRAAL’s NCF, as G-M scores are higher than M-M scores. However, for the larger neighborhood sizes (T3 and T4), MI-GRAAL’s NCF is comparable or superior to GHOST’s NCF. And because we show that the larger neighborhood sizes (T3 and T4) are overall superior (“The size of nodes’ neighborhoods within NCF?”), this means that overall MI-GRAAL’s NCF is comparable to or superior to GHOST’s NCF. On the other hand, in general, for all network sizes, GHOST’s AS consistently outperforms MI-GRAAL’s AS, as G-G scores is typically higher than G-M scores.

When comparing the different aligners with respect to computational complexity (rather than accuracy, as above), we find the following. Overall, G-G is the fastest, followed by M-M, followed by G-M (Figure 6a). This implies that since M-M is faster than G-M, MI-GRAAL’s NCF is less computationally intensive than GHOST’s NCF. Also, since G-G is faster than G-M, GHOST’s AS is less computationally intensive than MI-GRAAL’s AS.

thumbnailFigure 5. Alignment quality results of the three aligners (M-M, G-M, and G-G). The results are shown for best alignments over all values of View MathML and all neighborhood sizes, for a–c three network pairs with known node mapping (yeast–yeast 5%, yeast–yeast 10%, and yeast–yeast 15%, respectively) and d–f three network pairs with unknown mapping (human–yeast, human–worm, and worm–yeast, respectively). Percentages represent the scores achieved by an alignment quality measure. For equivalent results for the remaining network pairs, see the Additional file 1: Figures S1–S2 and S8–S10.
We note that in order to fairly compare the running times of all aligners used in this study, we run all aligners using neighborhood size T4 (Table 2). We cannot do this for the other (smaller) neighborhood sizes for the following reasons. While GHOST allows the user to specify any desired neighborhood size as input, MI-GRAAL’s NCF does not. Namely, the current implementation of MI-GRAAL by default computes all up to 5-node graphlets (i.e., T4). Then, to get the information contained in up to 2-, 3-, or 4-node only graphlets, one simply considers the relevant dimensions of the entire up to 5-node graphlet degree vector and discards all other dimensions. Thus, we cannot evaluate the computational complexity of considering 2-, 3-, or 4-node only graphlets, as with the current implementation, each of these options takes the same (longest) amount of time that computing up to 5-node graphlets takes.

Real networks with unknown node mapping
Overall, unlike for the synthetic network data set with known node mapping, on the real network data set with unknown mapping, MI-GRAAL’s NCF is now comparable or superior to that of GHOST (Figure 4c, d). Further, MI-GRAAL’s AS is now comparable or superior to GHOST’s AS (Figure 4c, d). We confirm these findings even when we limit from all alignments (Figure 4c, d) to the best alignments only (just as above) (Figure 5d–f) (Additional file 1: Figures S8–S10).

thumbnailFigure 6. The CPU time needed for M-M, G-M, and G-G (when using neighborhood size T4) to generate alignments of: a the synthetic noisy yeast networks and b the real-world networks of different species. All experiments were run on the same server with 16 2.3 GHz processors and 128 GB of RAM.
When zooming into the results further to observe the effect of the View MathML parameter, in general, for all values of View MathML, MI-GRAAL’s NCF is comparable or superior to GHOST’s NCF and MI-GRAAL’s AS is comparable to GHOST AS across all values of View MathML (Additional file 1: Figures S11–S16). The same holds independent on the neighborhood size that is considered within NCF (Additional file 1: Figures S11–S16).

When comparing the different aligners with respect to computational complexity (rather than accuracy, as above), we find the following. Unlike for the synthetic network data, we now observe that M-M is significantly the fastest, followed by G-M, followed by G-G (Figure 6b). This implies that since M-M is faster than G-M, MI-GRAAL’s NCF is less computationally intensive than GHOST’s NCF. Also, since G-M is faster than G-G, MI-GRAAL’s AS is less computationally intensive than GHOST’s AS.

thumbnailFigure 7. The ranking of the 11 values of View MathML (from 0 to 1 in increments in 0.1). The ranking is shown over all alignments for all aligners and all neighborhood sizes, with respect to: a all topological scores of alignments with known ground truth node mapping, b all biological scores of alignments with known node mapping, c all topological scores of alignments with unknown node mapping, and d all biological scores of alignments with unknown node mapping. Percentages represent the scores achieved by an alignment quality measure.
Summary
Which NCF or AS is the best overall is not easy to determine, as the results are data-dependent. But when we limit analyses of each aligner to the best alignments over all parameters, M-M is comparable or superior to G-M, indicating that MI-GRAAL’s NCF is better than GHOST’s NCF, while the performance of G-M versus G-G, i.e., of MI-GRAAL’s AS versus GHOST’s AS, is still data-dependent. These results hold not just in terms of accuracy but also in terms of computational complexity. We note that the reason why the performance of the two ASs is data-dependent (GHOST’s AS performing better on the synthetic networks, and MI-GRAAL’s AS performing better on the real-world networks) could be due to the differences of the two network data sets. Namely, recall that the synthetic network data encompasses “co-complex” PPIs obtained by AP/MS, among other PPI types, while the real-world network data consists of “binary” Y2H PPIs (“Data sets”).

The above results imply that the graphlet-based measure of topological node similarity [37] that MI-GRAAL uses (along with many other network aligners [2], [3], [5], [6] or even network clustering methods [37], [39], [40]) remains the state-of-the-art, as it is superior to the newer spectral signature-based node similarity measure that GHOST uses (and especially to the PageRank-based node similarity measure that aligners from the IsoRank family use, as we already showed in our recent study [2], [3]). Our results indicate that the slight superiority of GHOST (i.e., G-G) over MI-GRAAL (i.e., M-M) that was claimed in the original GHOST publication [25] seems to come from GHOST’s AS and not its NCF, which is not surprising, since GHOST’s AS deals with the quadratic assignment problem whereas MI-GRAAL’s AS deals only with linear assignment problem. Further, our results indicate that the combination of MI-GRAAL’s NCF and GHOST’s AS (i.e., M-G) could be a new aligner that is superior to the existing MI-GRAAL (i.e., M-M) and GHOST (ie., G-G) aligners on at least some data sets. Unfortunately, explicitly testing this is not possible with the current implementation of GHOST, as per our conversation with the authors of GHOST, the current implementation is too complex to modify to allow for plugging MI-GRAAL’s (or any other method’s) NCF into GHOST’s AS.

The amount of sequence versus topological information within NCF?
Recall that we vary the amount of topological node similarity information within NCF with the View MathML parameter (where View MathML of 0 means that no topology information is used, i.e., that only sequence information is used, whereas View MathML of 1 means that only topology information is used; “Aligners resulting from combining existing NCFs and ASs, and their parameters”). Here, we study the effect of the View MathML parameter on alignment quality.

Synthetic networks with known node mapping
Overall, the value of View MathML does not affect alignment quality, as long as some amount of topological information is used. That is, only View MathML results in completely inferior alignments, especially with respect to topological alignment quality, whereas all other values of alpha are more-less comparable (Figure 7a, b).

It is expected that the larger the value of View MathML, i.e., the more of topological information is used within NCF, the better the topological alignment quality. Again, this is exactly what we observe (Figure 7a). It is also expected that the smaller the value of View MathML, i.e., the more of sequence information is used within NCF, the better the biological alignment quality. Surprisingly, this is not what we observe (Figure 7b): larger values of View MathML (e.g., 0.7) result in more of high-quality alignments than View MathML.

When zooming into the results further to observe the effect of the aligner, in general, we see the same trends as above independent of the aligner (Additional file 1: Figures S3–S7). Namely, the results from Figure 7a, b hold independent on which NCF or AS is used. Further, there is no difference in the results across the two NCFs (Figure 8a, b). There is only a minor difference in the results across the two ASs, in the sense that the results are somewhat more stable across different View MathMLs for GHOST’s AS than for MI-GRAAL’s AS (Figure 8b, c). Also, GHOST’s AS suggests that in addition to not using View MathML (i.e., sequence alone), one should not use View MathML either (i.e., topology alone); but other than that, the choice of View MathML still has no major effect (Figure 8c).

thumbnailFigure 8. Detailed illustration of the effect of the View MathML parameter for a M-M, b G-M, and c G-G aligners. In particular, results are shown for the yeast-yeast 5% alignment and for the neighborhood size T4. Percentages represent the scores achieved by an alignment quality measure. For other network pairs and other neighborhood sizes, see Additional file 1: Figures S3–S7 for synthetic network data and see see Additional file 1: Figures S11–S16 for real-world PPI network data.
When zooming into the results from Figure 7a, b further to observe the effect of the neighborhood size, we see that the results hold independent of the neighborhood size (Additional file 1: Figures S3–S7).

Real networks with unknown node mapping
The results that we observe for the synthetic networks in general hold for this network set as well. Namely, View MathML results in the worst topological alignment quality, while the other View MathML values are somewhat comparable, with a slight dominance of the larger values, as expected (Figure 7c). Interestingly, for this network set, the lowest value of View MathML results in the most of highest-scoring alignments with respect to biological alignment quality; yet, even the largest View MathMLs often lead to good alignments with respect to biological alignment quality (Figure 7d).

When zooming into the results further to observe the effect of the aligner, as with synthetic networks, the general results from Figure 7c, d hold independent of the aligner for real networks as well (Additional file 1: Figures S11–S16). However, unlike for synthetic networks, for real networks we now see result stability across all NCFs and all ASs, and not just for GHOST’s AS. Also, GHOST’s AS no longer suggests that View MathML should not be used.

When zooming into the results from Figure 7c, d further to observe the effect of the neighborhood size, just as with the synthetic networks, we again see that the results hold independent of the neighborhood size (Additional file 1: Figures S11–S16).

Summary
Overall, at least some amount of topological information should be included within NCF, as this results in good topological as well as biological alignment quality. While View MathML may (but does not always) result in biologically high-quality alignments, in every case it fails to produce topologically superior results. Thus, View MathML should not be used.

The size of nodes’ neighborhoods within NCF?
Intuitively, one would expect that the increase in the size of nodes’ network neighborhoods within NCF (i.e., in the amount of network topology) would result in higher-quality alignments. However, this assumption has not been tested to date. Instead, the existing methods blindly use the largest neighborhood size that is allowed by available computational resources (that is, MI-GRAAL uses all 2–5-node graphlets, whereas GHOST uses View MathML “Aligners resulting from combining existing NCFs and ASs, and their parameters”). Thus, within each aligner, we vary the neighborhood size from T1 to T4 (Table 2) to systematically evaluate the effect of this parameter.

Synthetic networks with known node mapping
Overall, the larger the neighborhood size, the better the alignment quality, even though all neighborhood sizes except T1 can in some cases result in higher-quality alignments than any other neighborhood size (Figure 9a, b). That is, for some values of network alignment parameters, smaller neighborhoods can produce higher-quality alignments than larger neighborhoods, which is a surprising though not alarming result. It is possible for larger neighborhood sizes to produce lower quality alignments due to nodes in one network having denser, more complex neighborhoods than nodes in the other network. For example, two nodes View MathML and View MathML from different networks can have similar neighborhoods at size e.g., T2 but different neighborhoods at larger size e.g., T3, if e.g., the 3-hop neighborhood of node View MathML in one network is empty while the 3-hop neighborhood of node View MathML in another network is not. Thus, although larger network neighborhoods include more of the network topological information, they could also “confuse” the network signal, depending on the topology of the aligned networks, in which case smaller neighborhoods may be preferred.

thumbnailFigure 9. The ranking of the four neighborhood sizes (T1–T4). The ranking is shown over all alignments for all aligners and all values of View MathML, with respect to: a all topological scores of alignments with known ground truth node mapping, b all biological scores of alignments with known node mapping, c all topological scores of alignments with unknown node mapping, and d all biological scores of alignments with unknown node mapping. Percentages represent the scores achieved by an alignment quality measure.
When zooming into the results further to observe the effect of the aligner, the general trends from Figure 9a, b still hold independent of the aligner, but some fluctuations in the results exist (Additional file 1: Figures S17–S21). Namely, M-M generally prefers T3 and T4 neighborhood sizes. G-M prefers T2 in addition to T3 and T4, where T3 or T4 are actually inferior to T2 in some cases, depending on the noise level. G-G performs well on of T1-T4, with a slight preference of T3 or T4, depending on the noise level. See Figure 10a for an illustration.

thumbnailFigure 10. Detailed illustration of the effect of the neighborhood size for a synthetic and b real network data. In particular, results are shown for all three aligners, for the yeast-yeast 5% alignment at View MathML in a and for the fly-worm alignment at View MathML in b. Percentages represent the scores achieved by an alignment quality measure. For other network pairs and other values of View MathML, see Additional file 1: Figures S17–S21 and S22–S27.
When zooming into the results further to observe the effect of the View MathML parameter, general trends from Figure 9a, b are overall the same for all values of View MathML (Additional file 1: Figures S17–S21). The only exception is View MathML, which should not be used in the first place (“Summary”).

Real networks with unknown node mapping
Unlike for the synthetic networks, the largest neighborhood size (T4) is now not overly dominant over the smaller network sizes. Specifically, for real network data set, it is T3 that is the most dominant, followed by T4 and T2, which are tied, and followed by T1, which is inferior (Figure 9c, d).

When zooming into the results further to observe the effect of the aligner, we see that each aligner has an interesting behavior (Additional file 1: Figures S22–S27). Namely, M-M’s and G-G’s preference on the neighborhood size is mainly dictated by the choice of species whose networks are aligned. For G-M, in general, the larger neighborhood sizes are preferred; in some cases, depending on the species, G-M prefers T3 more than other neighborhood sizes. See Figure 10b for an illustration.

When zooming into the results further to observe the effect of the View MathML parameter, just as for synthetic networks, the results from Figure 9c, d do not drastically change with the change of View MathML value (Additional file 1: Figures S22–S27).

Summary
In general, the larger the neighborhood size within NCF, the higher the alignment quality. However, it is not necessarily the case that the largest neighborhood size always produces the best alignments nor that it is always dominant to the smaller neighborhood sizes. This means that slightly smaller neighborhood sizes (and T3 in particular) might be desirable, as this could not only produce better alignments in some cases but also decrease the computational complexity of the given method.

Relationships between different alignment quality measures
We use a total of seven alignment quality measures: the ground truth NC measure that can only be measured in alignments of synthetic networks with known node mapping, four additional topological measures (EC, ICS, View MathML, and LCCS), and two biological measures (GO and EXP) (“Network alignment quality measures”). Here, we briefly comment on the relationship between the different measures.

NC significantly correlates with both topological and biological alignment quality measures (Figure 11a), which is encouraging. Further, for the synthetic network data set, it is also encouraging that all other measures significantly correlate well (Pearson correlation coefficient of at least 0.8), even though we see some clustering of the topological measures and also of the biological measures (Figure 11a). Interestingly, each of the two biological measures, GO and EXP, correlates better with some of the topological measures (e.g., EC) than with each other.

thumbnailFigure 11. Pairwise correlations between different alignment quality measures. The results are shown for: a synthetic networks with known ground truth node mapping and b real networks with unknown node mapping. Correlations were computed over alignments with the highest NC scores in a and over alignments with the highest EC scores in b (because we do not known NC scores for alignments of real networks). Note that color scales for the two panels are different.
Unlike for the synthetic network data, for the real network data, the topological measures now correlate poorly with the biological measures (Pearson correlation coefficient of at most 0.2; Figure 11b). Importantly, this implies that for the real network data set it might be hard to produce an alignment that is of excellent quality both topologically and biologically. Also, while we again see clustering of the topological measures, the two biological measures now correlate weakly (Figure 11b), indicating that the choice of GO annotation data obtained by experimental evidence code matters (“Network alignment quality measures”).

The result differences between the synthetic networks and the real networks could be due to differences in their properties (“Data sets”).

Note that when measuring the correlations between the different alignment quality measures, we have used the Pearson correlation coefficient. In case that the data is not necessarily normally distributed, using a non-parametric (i.e., distribution-free) measure of correlation would be appropriate. Hence, we repeat the above analysis with respect to such a measure, namely the Spearman correlation coefficient. Importantly, our results produced in this way are mostly consistent to the results produced when using the Pearson correlation coefficient (Additional file 1: Figure S28).

Conclusions
We have aimed to systematically answer three questions in the context of MI-GRAAL and GHOST network aligners: (1) what is the contribution of each method’s NCF and AS to the alignment quality, (2) how much sequence versus topology information should be used within NCF when generating an alignment, and (3) how large the size of the neighborhoods of the compared nodes from different networks should be. Our results show that: (1) MI-GRAAL’s NCF is superior to GHOST’s, while the performance of their ASs is data-dependent, (2) some amount of topological data should be used in the NCF, and (3) the larger the amount of topology, the better, although using the second largest neighborhood size can result in better results and lower computational complexity compared to using the largest neighborhood size. Our results represent a set of general recommendations for a fair evaluation of any GNA method (and especially if the method falls into the two-state NCF-AS category), not just MI-GRAAL and GHOST.

Genomic sequence alignment has revolutionized our biomedical understanding. Biological network alignment has already had similar impacts. And given the tremendous amounts of biological network data that continue to be produced, network alignment will only continue to gain importance. The hope is that it could lead to new discoveries about the principles of life, evolution, disease, and therapeutics. Network alignment has also strived in other domains as well, with applications such as semantically matching entities in different ontologies [8] or comparing online social networks with impacts on user privacy [9].

Abbreviations
GNA: global network aligner

LNA: local network aligner

NCF: node cost function

AS: alignment Strategy

PPI: protein–protein interaction

GDV: graphlet degree vector.

Additional files
Additional file 1:. Supplementary material containing additional results.
Format: PDF Size: 4.1MB Download file
This file can be viewed with: Adobe Acrobat ReaderOpen Data
Authors’ contributions
JC participated in the design of the study, performed most of the analyses presented in the paper, helped analyze the results, and helped write the paper. YS performed the rest of the analyses (and those concerning M-M aligner in particular) and helped write the paper. TM designed and supervised all aspects of the study, analyzed the results, and wrote the paper. All authors read and approved the final manuscript.

Acknowledgments
We thank Dr. R. Patro and Dr. C. Kingsford for their assistance with running GHOST. This work was supported by the National Science Foundation CAREER CCF-1452795, CCF-1319469 and EAGER CCF-1243295 Grants.

Compliance with ethical guidelines
Competing interests The authors declare that they have no competing interests.

References
Sharan R, Ideker T: Modeling cellular machinery through biological network comparison.
Nat Biotechnol 2006, 24(4):427-433. PubMed Abstract | Publisher Full Text OpenURL
Milenković T, Zhao H, Faisal FE (2013) Global network alignment in the context of aging. In: Proceedings of the international conference on bioinformatics, computational biology and biomedical informatics (BCB’13), pp 23–32
Faisal F, Zhao H, Milenkovic T: Global network alignment in the context of aging.
IEEE/ACM Trans Comput Biol Bioinform 2014, 12(1):40-52. Publisher Full Text OpenURL
Faisal F, Milenković T: Dynamic networks reveal key players in aging.
Bioinformatics 2014, 30(12):1721-1729. PubMed Abstract | Publisher Full Text OpenURL
Kuchaiev O, Milenković T, Memisević V, Hayes W, Pržulj N: Topological network alignment uncovers biological function and phylogeny.
J R Soc Interface 2010, 7(50):1341-1354. PubMed Abstract | Publisher Full Text OpenURL
Milenković T, Ng WL, Hayes W, Pržulj N: Optimal network alignment with graphlet degree vectors.
Cancer Inform 2010, 9:121-137. PubMed Abstract | Publisher Full Text OpenURL
Kuchaiev O, Pržulj N: Integrative network alignment reveals large regions of global network similarity in yeast and human.
Bioinformatics 2011, 27(10):1390-1396. PubMed Abstract | Publisher Full Text OpenURL
Li J, Tang J, Li Y, Luo Q: RiMOM: a dynamic multistrategy ontology alignment framework.
IEEE Trans Knowl Data Eng 2009, 21(8):1218-1232. Publisher Full Text OpenURL
Narayanan A, Shi E, Rubinstein B (2011) Link prediction by de-anonymization: How we won the Kaggle social network challenge. In: The 2011 international joint conference on neural networks (IJCNN). IEEE, pp 1825–1834
Kelley BP, Yuan V, Lewitter F, Sharan R, Stockwell BR, Ideker T: PathBLAST: a tool for alignment of protein interaction networks.
Nucleic Acids Res 2004, 32:83-88. Publisher Full Text OpenURL
Sharan R, Suthram S, Kelley RM, Kuhn T, McCuine S, Uetz P, et al.: Conserved patterns of protein interaction in multiple species.
Proc Natl Acad Sci 2005, 102(6):1974-1979. PubMed Abstract | Publisher Full Text OpenURL
Flannick J, Novak A, Balaji SS, Harley HM, Batzglou S: Graemlin general and robust alignment of multiple large interaction networks.
Genome Res 2006, 16(9):1169-1181. PubMed Abstract | Publisher Full Text OpenURL
Koyuturk M, Kim Y, Topkara U, Subramaniam S, Szpankowski W, Grama A: Pairwise alignment of protein interaction networks.
J Comput Biol 2006, 13(2):182-199. PubMed Abstract | Publisher Full Text OpenURL
Berg J, Lassig M: Local graph alignment and motif search in biological networks.
Proc Natl Acad Sci 2004, 101(41):14689-14694. PubMed Abstract | Publisher Full Text OpenURL
Liang Z, Xu M, Teng M, Niu L: NetAlign: a web-based tool for comparison of protein interaction networks.
Bioinformatics 2006, 22(17):2175-2177. PubMed Abstract | Publisher Full Text OpenURL
Berg J, Lassig M: Cross-species analysis of biological networks by Bayesian alignment.
Proc Natl Acad Sci 2006, 103(29):10967-10972. PubMed Abstract | Publisher Full Text OpenURL
Mina M, Hiram Guzzi P: Improving the robustness of local network alignment: design and extensive assessment of a markov clustering-based approach.
IEEE/ACM Trans Comput Biol Bioinform (TCBB) 2014, 11(3):561-572. Publisher Full Text OpenURL
Ciriello G, Mina M, Guzzi PH, Cannataro M, Guerra C: AlignNemo: a local network alignment method to integrate homology and topology.
PLoS One 2012, 7(6):e38107. PubMed Abstract | Publisher Full Text OpenURL
Hasan M, Kahveci T (2013) Color distribution can accelerate network alignment. In: Proceedings of the international conference on bioinformatics, computational biology and biomedical informatics, p 52
Singh R, Xu J, Berger B (2007) Pairwise global alignment of protein interaction networks by matching neighborhood topology. In: Research in computational molecular biology. Springer, Berlin, Heidelberg, pp 16–31
Flannick J, Novak A, Do CB, Srinivasan BS, Batzoglou S: Automatic parameter learning for multiple network alignment.
J Comput Biol 2009, 16(8):1001-1022. PubMed Abstract | Publisher Full Text OpenURL
Singh R, Xu J, Berger B: Global alignment of multiple protein interaction networks.
Proc Pac Symp Biocomput 2008, 13:303-314. OpenURL
Zaslavskiy M, Bach F, Vert JP: Global alignment of protein-protein interaction networks by graph matching methods.
Bioinformatics 2009, 25(12):259-267. Publisher Full Text OpenURL
Liao C, Lu K, Baym M, Singh R, Berger B: IsoRankN: spectral methods for global alignment of multiple protein networks.
Bioinformatics 2009, 25(12):253-258. Publisher Full Text OpenURL
Patro R, Kingsford C: Global network alignment using multiscale spectral signatures.
Bioinformatics 2012, 28(23):3105-3114. PubMed Abstract | Publisher Full Text OpenURL
Neyshabur B, Khadem A, Hashemifar S, Shahriar Arab S: NETAL: a new graph-based method for global alignment of protein-protein interaction networks.
Bioinformatics 2013, 29(13):1654-1662. PubMed Abstract | Publisher Full Text OpenURL
Guo X, Hartemink AJ: Domain-oriented edge-based alignment of protein interaction networks.
Bioinformatics 2009, 25(12):240-1246. Publisher Full Text OpenURL
Klau GW: A new graph-based method for pairwise global network alignment.
BMC Bioinform 2009, 10(Suppl 1):59. BioMed Central Full Text OpenURL
El-Kebir M, Heringa J, Klau GW (2011) Lagrangian relaxation applied to sparse global network alignment. In: Pattern recognition in bioinformatics. Springer, Berlin, Heidelberg, pp 225–236
Saraph V, Milenković T: MAGNA: maximizing accuracy in global network alignment.
Bioinformatics 2014, 30(20):2931-2940. PubMed Abstract | Publisher Full Text OpenURL
Clark C, Kalita J: A comparison of algorithms for the pairwise alignment of biological networks.
Bioinformatics 2014, 30(16):2351-2359. PubMed Abstract | Publisher Full Text OpenURL
Pache RA, Céol A, Aloy P: NetAligner-a network alignment server to compare complexes, pathways and whole interactomes.
Nucleic acids Res 2012, 40((Web Server issue)):157-161. Publisher Full Text OpenURL
West DB: Introduction to graph theory. 2nd edition. Prentice Hall, Upper Saddle River; 2001. OpenURL
Aladağ AE, Erten C: SPINAL: scalable protein interaction network alignment.
Bioinformatics 2013, 29(7):917-924. PubMed Abstract | Publisher Full Text OpenURL
Crawford J, Milenković T (2014) GREAT: GRaphlet Edge-based network AlignmenT. arXiv:1410.5103 [q-bio.MN]. Accessed 19 Oct 2014
Sun Y, Crawford J, Tang J, Milenković T (2014) Simultaneous optimization of both node and edge conservation in network alignment via wave. arXiv preprint arXiv:1410.3301
Milenković T, Pržulj N: Uncovering biological network function via graphlet degree signatures.
Cancer Inform 2008, 6:257-273. PubMed Abstract | Publisher Full Text OpenURL
Malod-Dognin N, Pržulj N: GR-Align: fast and flexible alignment of protein 3D structures using graphlet degree similarity.
Bioinformatics 2014, 30(9):1259-1265. PubMed Abstract | Publisher Full Text OpenURL
Milenković T, Memisević V, Ganesan AK, Pržulj N: Systems-level cancer gene identification from protein interaction network topology applied to melanogenesis-related interaction networks.
J R Soc Interface 2010, 7(44):423-437. PubMed Abstract | Publisher Full Text OpenURL
Solava RW, Michaels RP, Milenković T: Graphlet-based edge clustering reveals pathogen-interacting proteins.
Bioinformatics 2012, 18(28):480-486. Publisher Full Text OpenURL
Pache RA, Aloy P: A novel framework for the comparative analysis of biological networks.
PLoS One 2012, 7(2):31220. Publisher Full Text OpenURL
Collins SR, Kemmeren P, Zhao XC, Greenblatt JF, Spencer F, Holstege FCP, et al.: Toward a comprehensive atlas of the phyisical interactome of Saccharomyces cerevisiae.
Mol Cell Proteomics 2007, 6(3):439-450. PubMed Abstract | Publisher Full Text OpenURL
Breitkreutz BJ, Stark C, Reguly T, Boucher L, Breitkreutz A, Livstone M, et al.: The BioGRID interaction database: 2008 update.
Nucleic Acids Res 2008, 36:637-640. Publisher Full Text OpenURL
Altschul SF, Gish W, Miller W, Lipman DJ: Basic local alignment search tool.
J Mol Biol 1990, 215:403-410. PubMed Abstract | Publisher Full Text OpenURL
Milenković T, Lai J, Pržulj N (2008) GraphCrunch: a tool for large network analyses. BMC Bioinform 9(70)
Milenković T, Memisević V, Bonato A, Pržulj N: Dominating biological networks.
PLoS One 2011, 6(8):23016. Publisher Full Text OpenURL
Hulovatyy Y, Solava RW, Milenković T: Revealing missing parts of the interactome via link prediction.
PLoS One 2014, 9(3):90073. Publisher Full Text OpenURL
The Gene Ontology Consortium: Gene Ontology: tool for the unification of biology
Nat Genet 2000, 25:25-29. Publisher Full Text OpenURL
Pržulj N: Biological network comparison using graphlet degree distribution.
Bioinformatics 2007, 23:177-183. Publisher Full Text OpenURL

Oscillometric measurement of systolic and diastolic blood pressures validated in a physiologic mathematical model

Charles F Babbs

Correspondence: Charles F Babbs babbs@purdue.edu

Author Affiliations
Department of Basic Medical Sciences, Weldon School of Biomedical Engineering, Purdue University, 1426 Lynn Hall, West Lafayette, IN, 47907-1246, USA

BioMedical Engineering OnLine 2012, 11:56 doi:10.1186/1475-925X-11-56

The electronic version of this article is the complete one and can be found online at: http://www.biomedical-engineering-online.com/content/11/1/56

Received: 28 June 2012
Accepted: 3 August 2012
Published: 22 August 2012
© 2012 Babbs; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Formula display:
Abstract
Background
The oscillometric method of measuring blood pressure with an automated cuff yields valid estimates of mean pressure but questionable estimates of systolic and diastolic pressures. Existing algorithms are sensitive to differences in pulse pressure and artery stiffness. Some are closely guarded trade secrets. Accurate extraction of systolic and diastolic pressures from the envelope of cuff pressure oscillations remains an open problem in biomedical engineering.

Methods
A new analysis of relevant anatomy, physiology and physics reveals the mechanisms underlying the production of cuff pressure oscillations as well as a way to extract systolic and diastolic pressures from the envelope of oscillations in any individual subject. Stiffness characteristics of the compressed artery segment can be extracted from the envelope shape to create an individualized mathematical model. The model is tested with a matrix of possible systolic and diastolic pressure values, and the minimum least squares difference between observed and predicted envelope functions indicates the best fit choices of systolic and diastolic pressure within the test matrix.

Results
The model reproduces realistic cuff pressure oscillations. The regression procedure extracts systolic and diastolic pressures accurately in the face of varying pulse pressure and arterial stiffness. The root mean squared error in extracted systolic and diastolic pressures over a range of challenging test scenarios is 0.3 mmHg.

Conclusions
A new algorithm based on physics and physiology allows accurate extraction of systolic and diastolic pressures from cuff pressure oscillations in a way that can be validated, criticized, and updated in the public domain.

Background
The ejection of blood from the left ventricle of the heart into the aorta produces pulsatile blood pressure in arteries. Systolic blood pressure is the maximum pulsatile pressure and diastolic pressure is the minimum pulsatile pressure in the arteries, the minimum occurring just before the next ventricular contraction. Normal systolic/diastolic values are near 120/80 mmHg. Normal mean arterial pressure is about 95 mmHg [1].

Blood pressure is measured noninvasively by occluding a major artery (typically the brachial artery in the arm) with an external pneumatic cuff. When the pressure in the cuff is higher than the blood pressure inside the artery, the artery collapses. As the pressure in the external cuff is slowly decreased by venting through a bleed valve, cuff pressure drops below systolic blood pressure, and blood will begin to spurt through the artery. These spurts cause the artery in the cuffed region to expand with each pulse and also cause the famous characteristic sounds called Korotkoff sounds. The pressure in the cuff when blood first passes through the cuffed region of the artery is an estimate of systolic pressure. The pressure in the cuff when blood first starts to flow continuously is an estimate of diastolic pressure. There are several ways to detect pulsatile blood flow as the cuff is deflated: palpation, auscultation over the artery with a stethoscope to hear the Korotkoff sounds, and recording cuff pressure oscillations. These correspond to the three main techniques for measuring blood pressure using a cuff [2].

In the palpatory method the appearance of a distal pulse indicates that cuff pressure has just fallen below systolic arterial pressure. In the auscultatory method the appearance of the Korotkoff sounds similarly denotes systolic pressure, and disappearance or muffling of the sounds denotes diastolic pressure. In the oscillometric method the cuff pressure is high pass filtered to extract the small oscillations at the cardiac frequency and the envelope of these oscillations is computed, for example as the area obtained by integrating each pulse [3]. These oscillations in cuff pressure increase in amplitude as cuff pressure falls between systolic and mean arterial pressure. The oscillations then decrease in amplitude as cuff pressure falls below mean arterial pressure. The corresponding oscillation envelope function is interpreted by computer aided analysis to extract estimates of blood pressure.

The point of maximal oscillations corresponds closely to mean arterial pressure [4-6]. Points on the envelope corresponding to systolic and diastolic pressure, however, are less well established. Frequently a version of the maximum amplitude algorithm [7] is used to estimate systolic and diastolic pressure values. The point of maximal oscillations is used to divide the envelope into rising and falling phases. Then characteristic ratios or fractions of the peak amplitude are used to find points corresponding to systolic pressure on the rising phase of the envelope and to diastolic pressure on the falling phase of the envelope.

The characteristic ratios (also known as oscillation ratios or systolic and diastolic detection ratios [8]) have been obtained experimentally by measuring cuff oscillation amplitudes at independently determined systolic or diastolic points, divided by the maximum cuff oscillation amplitude. The systolic point is found at about 50% of the peak height on the rising phase of the envelope. The diastolic point is found at about 70 percent of the peak height on the falling phase of the envelope [7]. These empirical ratios are sensitive however to changes in physiological conditions, including most importantly the pulse pressure (systolic minus diastolic blood pressure) and the degree of arterial stiffness [9,10]. Moreover, a rational physical explanation for any particular ratio has been lacking. Since cuff pressure oscillations continue when cuff pressure falls beneath diastolic blood pressure, the endpoint for diastolic pressure is indistinct. Most practical algorithms used in commercially available devices are closely guarded trade secrets that are not subject to independent critique and validation. Hence the best way to determine systolic and diastolic arterial pressures from cuff pressure oscillations remains an open scientific problem.

The present study addresses this problem with a new approach based upon the underlying physics, anatomy, and physiology. This task requires modeling the cuff and arm and the dynamics of a partially occluded artery within the arm during cuff deflation. A second phase of the problem is the development of a regression procedure for analysis of recorded cuff pressure oscillations to extract model parameters and predict the unique systolic and diastolic pressure levels that would produce the observed cuff pressure oscillations.

Methods Part 1: Modeling cuff pressure oscillations
Model of the cuff and arm
As shown in Figure 1, one can regard the cuff as an air filled balloon of dimensions on the order of 30 cm x 10 cm x 1 cm, which is wrapped in a non-expanding fabric around the arm. After inflation the outer wall of the cuff becomes rigid and the compliance of the cuff is entirely due to the air it contains. During an oscillometric run the cuff is inflated to a pressure well above systolic, say 150 to 200 mmHg, and then vented gradually at a bleed rate of r = 3 mmHg / second [11]. Small oscillations in cuff pressure happen when the artery fills and empties with blood as cuff pressure passes between systolic and diastolic pressure in the artery.

thumbnailFigure 1. Arrangement of cuff, skin, muscle, bone, and artery for a simple model of the arm during oscillometric blood pressure recording.
Let P0 be the maximal inflation pressure of the cuff at the beginning of a run. The pressure is bled down slowly at rate, r mmHg/sec (about 3 mmHg/sec [11]). During the brief period of one heartbeat the amount of air inside the cuff is roughly constant. In addition to smooth cuff deflation, small cuff pressure oscillations are caused by pulsatile expansion of the artery and the corresponding compression of the air in the cuff. One can model the cuff as a pressure vessel having nearly fixed volume, V0 − ΔVa, where V0 is cuff volume between heartbeats and ΔVa is the small incremental volume of blood in the artery beneath the cuff as it expands with the arterial pulse.

To compute cuff pressure oscillations from the volume changes, ΔVa, in the occluded artery segment it is necessary to know the compliance of the cuff, C = ΔV/ΔP, which is obtainable from Boyle’s law as follows. Boyle’s law is PV = nRT, where P is the absolute pressure (760 mmHg plus cuff pressure with respect to atmospheric), V is the volume of air within the cuff, n is the number of moles of gas, R is the universal gas constant, and T is the absolute temperature. During the time of one heartbeat, n, R, and T are constants and n is roughly constant owing to the slow rate of cuff deflation. Hence to relate the change in cuff pressure, ΔP to the small change in cuff volume, ΔV, from artery expansion we may write PV≈(P+ΔP)⋅(V+ΔV)≈PV+PΔV+VΔP, for absolute cuff pressure P. So

0≈PΔV+VΔPandCcuff=−ΔVΔP=ΔVaΔP≈V0P. (1)
The negative change in cuff volume represents indentation by the expanding arm when the artery inside fills with blood. The effective cuff compliance, Ccuff , or more precisely the time-varying and pressure-varying dynamic compliance of the sealed air inside the cuff, is

Ccuff=dVcuffdP=V0P+760mmHg, (2)
with cuff pressure, P, expressed in normal clinical units of mmHg relative to atmospheric pressure. In turn, the time rate of change in cuff pressure is

dPdt=−r+1CcuffdVa(t)dt≅−r+(P0+760−rtV0)dVa(t)dt. (3)
In this problem as cuff pressure is slowly released, even as cuff volume remains nearly constant, the dynamic compliance of the cuff increases significantly and its stiffness decreases. Hence a suitably exact statement of the physics requires a differential equation (1a), rather than the constant compliance approximation P=P0−rt+Va(t)/Ccuff. However, equation (1a) may be integrated numerically to obtain a sufficiently exact representation of cuff pressure changes with superimposed cardiogenic oscillations.

Model of the artery segment
Next to characterize the time rate of volume expansion of the artery, dVa/dt, one can regard the artery as an elastic tube with a dynamic compliance, Ca, which varies with volume and with internal minus external pressure. The dynamic compliance Ca = dVa/d(Pa – Po), where Pt = Pa – Po is the transmural pressure or the difference between pressure inside the artery and outside the artery. Then

dVadt=dVad(Pa−Po)⋅d(Pa−Po)dt≅Ca(dPadt+r), (4)
where the artery “feels” the prevailing difference between internal blood pressure and external cuff pressure, neglecting the small cuff pressure oscillations. The time derivative of arterial pressure can be determined from a characteristic blood pressure waveform and the known rate, r, of cuff deflation. Hence, the crucial variable to be specified next is the dynamic arterial compliance, Ca.

Specifying the compliance of the artery is more difficult than specifying the cuff compliance, because the pressure across the artery wall during an oscillometric measurement varies over a wide range from negative to positive. Most research studies, such as the classical ones of Geddes and Posey [12], explore only positive distending pressures. A few sources however [9,13] describe pressure-volume functions like the one sketched in Figure 2 for arteries subjected to both positive and negative distending pressure.

thumbnailFigure 2. Hypothetical pressure-volume relationship for an artery including negative transmural pressures and collapse. Pc is collapse pressure, and Pmid is normal mid-level arterial pressur.
For classical biomaterials one can use two exponential functions to model the nonlinear volume vs. pressure relationship over a wide range of distending pressures. Here we shall use two exponential functions: one for negative pressure range and another for the positive pressure range in a manner similar to that described by Jeon et al. [13]. The first exponential function for negative transmural pressure, Pt<0, is easy to imagine. For the negative transmural pressure domain artery volume Va=Va0eaPt for positive constant, a, and zero pressure volume, Va0, of the artery. Here the dynamic compliance is clearly

dVadPt=aVa0eaPtfor Pt<0. (5)
For the positive transmural pressure domain one can use a similar, but decelerating exponential function [12]. However, there should be no discontinuity at the zero-transmural pressure point (0, V0). This means that for positive constant, b, and typically b < a,

dVadPt=aVa0e−bPtfor Pt≥0. (6)
These two exponential functions can be used to characterize the dynamic compliance of the artery model in terms of easily obtained data, including the collapse pressure, Pc < 0, defined as the pressure when the artery volume is reduced to 0.1Va0 (for example, Pc = −20 mmHg) and the normal pressure arterial compliance, Cn, measured at normal mid-level arterial pressure, Pmid , halfway between systolic and diastolic pressure.

Solving for constant, a, we have

a=ln(0.1)Pc. (7)
Solving for constant, b, we have

b=−ln(CnaVa0)Pmid. (8)
The zero pressure volume, Va0, can be known from anatomy if necessary, but as shown later is not needed if one is interested only in the relative amplitude of cuff pressure oscillations.

One can integrate the expressions (2a) and (2b) to obtain analytical volume versus pressure functions similar to Figure 2. Thus for Pt < 0

Va=Va0+aVa0∫Pt0eaPtdPt=Va0eaPt, (9)
and for Pt ≥ 0

Va=Va0+aVa0∫Pt0e−bPtdPt=Va0−abVa0(e−bPt−1)=Va0(1+ab(1−e−bPt)). (10)
Figure 3 shows a plot of the resulting pressure-volume curve for a normal 10-cm long artery segment and constants a and b as described for initial conditions below. The form of the function is quite reasonable and consistent with prior work [9,13]. When bi-exponential constants a and b are varied, a wide variety of shapes for the pressure-volume curve can be represented. When volume changes more rapidly with pressure, the artery is more compliant. When volume changes less rapidly with pressure, the artery is stiffer. Increasing a and b in proportion allows greater volume change for a given pressure change and represents a more compliant artery. Decreasing a and b in proportion reduces the volume change for a given pressure change and represents a stiffer artery. Increasing the ratio a/b represents a greater maximal distension. Decreasing the ratio a/b represents a smaller maximal distension.

thumbnailFigure 3. Representative volume vs. pressure curves for an artery segment over a wide range of positive and negative transmural pressure. Standard normal variables a = 0.11 mmHg-1 , b = 0.03/mmHg-1, Va0 = 0.3 ml. Variations in shape occur with combinations of increased (2x normal) and decreased (1/2 normal) values of parameters a and b.
Forcing function—the time domain blood pressure waveform
For proof of concept and validity testing one can use a Fourier series to represent blood pressure waveforms in these models [2]. A suitable and simple one for initial testing here is

Pa=DBP+0.5PP+0.36PP[sin(ωt)+12sin(2ωt)+14sin(3ωt)] (11)
for arterial pressure, Pa, as a function of time, t, with ω being the angular frequency of the heartbeat, that is ω = 2πf for cardiac frequency, f, in Hz. Here SBP is systolic blood pressure, DBP is diastolic blood pressure, and PP is pulse pressure (SBP − DBP). In turn, the derivative of the arterial pressure waveform is

dPadt=0.36ω[cos(ωt)+cos(2ωt)+34cos(3ωt)]. (12)
Combining the cuff compliance, pressure-volume functions for the artery, and the arterial pressure waveform, one can write a set of equations for the rate of change in cuff pressure during an oscillometric pressure measurement in terms of P0, r, Va, Ccuff, and time. We must work with the time derivative of cuff pressure, rather than absolute cuff pressure, because the compliance of the cuff and also the form of the artery volume vs. pressure function vary with time and pressure during a run. Cuff pressure can then be computed numerically by integrating equation (1a),

dPdt≅−r+(P0+760−rtV0)dVa(t)dt. (13)
Using the chain rule of calculus, and taking transmural pressure as arterial blood pressure minus cuff pressure,

dVadt=aVa0ea(Pa−P0+rt)⋅(dPadt+r)for Pa–P0+rt<0 (14)
dVadt=aVa0e−b(Pa−P0+rt)⋅(dPadt+r)for Pa–P0+rt≥0, (15)
with artery pressure, Pa, and its time derivative given by equations (5). Combining equations (1a) and (6) gives a precise model for cuff pressure oscillations.

Initial conditions
Artery dimensions: As a standard normal model consider a brachial artery with internal radius of 0.1 cm under zero distending pressure. The resting artery volume is Va0 = πr2L or

Va0=3.14⋅(0.1cm)2⋅10cm=0.3cm3. (16)
Stiffness constant a: For collapse to 10 percent at −20 mmHg transmural pressure we have

a=ln(0.1)Pc=−2.3−20mmHg=0.11mmHg−1. (17)
Stiffness constant b: It is easy to estimate the normal pressure compliance of the brachial artery in humans, Cn , from experiments using ultrasound. For example, using the data of Mai and Insana [14], the brachial artery strain (Δr/r) during a normal pulse is 4 percent for a blood pressure of 130/70 mmHg with pulse pressure 60 mmHg. In turn the volume of expansion during a pulse is 2πrΔrL, where r is the radius and L is the length of the compressed artery segment. Hence for a normal pressure radius of 0.2 cm the change in volume would be

ΔVa=6.28⋅0.2cm⋅0.04⋅0.2cm⋅10cm=0.10cm3. (18)
The normal pressure compliance for the artery segment is the volume change divided by pulse pressure or

Cn = 0.10 ml / 60 mmHg = 0.0016 ml/mmHg.

For normal artery the pressure halfway between systolic and diastolic pressure, Pmid , would be 100 mmHg, so

b=−ln(CnaVa0)Pmid=−ln⎛⎝⎜⎜⎜0.0016cm3mmHg0.11mmHg⋅0.3cm3⎞⎠⎟⎟⎟100mmHg=0.03mmHg−1. (19)
Jeon et al. [13] working with a similar model used a = 0.09 mmHg-1., b = 0.03 mmHg.

Numerical methods
In this model equations (1), (5), and (6) govern the evolution of cuff pressure as a function of time during cuff deflation. Equation (1) can be integrated numerically using techniques such as the simple Euler method coded in Microsoft Visual Basic, Matlab, or “C”. In the results that follow cuff deflation is started from a maximal level of 150 mmHg and continues over a period of 40 sec. Pressures are plotted every 1/20th second. To extract the small oscillations from the larger cuff pressure signal, as would be done in an automatic instrument by an analog high pass filter, cuff pressure at time, t, is subtracted from the average of pressures recorded between times t − Δt/2 and t + Δt/2 , where Δt is the period of the pulse. For simplicity, filtered oscillations are not computed for time points that are Δt/2 seconds from the beginning or from the end of the time domain sample.

Methods Part 2: Interpreting cuff pressure oscillations
Given this model and the associated insight into the physics of cuff pressure oscillations, one can also devise a scheme for estimating true systolic and diastolic blood pressures from an observed time domain record of cuff pressure and filtered cuff pressure oscillations. The method is based upon the ability, just described, to predict the amplitude of pulse pressure oscillations for a given diastolic pressure and pulse pressure and the ability to deduce exponential constants, a and b, from the rising and falling regions of the oscillation amplitude envelope. Details are as follows.

Artery motion during cuff deflation
The shape of the volume vs. pressure curve for arteries determines the driving signal for cuff pressure oscillations during an oscillometric measurement, as shown in Figure 4.

thumbnailFigure 4. Pressure-volume relationship for an artery (solid curve) including positive and negative transmural pressures. Dashed triangles have equal bases indicating the range of transmural pressure (internal artery blood pressure minus cuff pressure) that determines the change in volume with each pulse. (a) Cuff pressure well above systolic with net distending pressure always negative. (b) Cuff pressure close to systolic. (c) Cuff pressure near mean arterial pressure with maximal volume changes. (d) Cuff pressure just below diastolic. (e) Cuff pressure well below diastolic.
The pulsatile component of transmural pressure causes the artery to change in volume with each heartbeat. The magnitude of the change in transmural pressure is always equal to the pulse pressure (say, 40 mmHg) which is assumed to be constant during cuff deflation. As cuff pressure gradually decreases from well above systolic to well below diastolic pressure, the range of transmural pressure, Pt, experienced by the artery changes. At (a) cuff pressure is well above systolic and net distending pressure is always negative. There is a small change in arterial volume because the artery becomes less collapsed as each arterial pulse makes the transmural pressure less negative. As cuff pressure approaches systolic the relative unloading of negative pressure becomes more profound. Because of the exponential shape of the arterial pressure-volume curve, the amount of volume change accelerates. At (b) cuff pressure is close to systolic. After this point the volume change continues to increase but at a decelerating rate, because of the shape of the pressure-volume curve. Hence (b) is the inflection point for systolic pressure. At (c) cuff pressure is near mean arterial pressure and the volume change is maximal. At (d) cuff pressure is just below diastolic. After this point, as shown in (e), the volume change becomes less and less with each pulse as the increasingly distended artery becomes stiffer. Hence (d) is the inflection point for diastolic pressure. Thus the nonlinear compliance of arteries and the shape of the arterial pressure-volume curve govern the amplitude of cuff pressure oscillations.

The particular volume change of the artery from the nadir of diastolic pressure to the subsequent peak of systolic pressure can be specified analytically from Equations (4a) and (4b) as follows. Consider Pt as the transmural pressure at the diastolic nadir of the arterial blood pressure wave and let PP be the pulse pressure. One can imagine three domains of transmural pressure. In Domain (1) Pt + PP < 0. In Domain (2) Pt < 0 and Pt + PP ≥ 0. In Domain (3) Pt > 0. The largest artery volume oscillations occur in Domain (2) when transmural pressure oscillates between positive and negative values. Doman (1) represents the head of the oscillation envelope in time, and Domain (3) represents the tail.

Using equations (4), the artery volume changes during the rising phase of the arterial pulse in each of the three domains are

Domain (1):

ΔVa=Va0[ea(Pt+PP)−eaPt] (20)
Domain (2):

ΔVa=Va0[1+ab(1−e−b(Pt+PP))−eaPt] (21)
Domain (3):

ΔVa=Va0[ab(1−e−b(Pt+PP))−ab(1−e−bPt)], (22)
where for cuff pressure, P, systolic blood pressure SBP, and diastolic blood pressure DBP, the transmural pressure Pt = DBP − P, and the pulse pressure PP = SBP − DBP.

It is easy to show by differentiating expressions (7) for Domains (1), (2), and (3) that the systolic and diastolic pressure points correspond exactly to the maximal and minimal slopes d(ΔVa)/dPt. Therefore a simple analysis for finding systolic and diastolic pressures points would involve taking local slopes of the oscillation envelope vs. pressure function. Slope taking, however, is vulnerable to noise in practical applications. An alternative approach that does not involve slope taking creates a model of each individual subject’s arm in terms of exponential constants a and b and then numerically finds the unique combination of systolic and diastolic arterial pressures that best reproduces the observed oscillation envelope.

Regression analysis for exponential constants
To obtain exponential constant, a, note that in the leading edge of the amplitude envelope at pressures near systolic blood pressure in Domain (1) the pulsatile change in cuff pressure is

ΔP=ΔVaCcuff=Va0Ccuff[ea(Pt+PP)−eaPt]=Va0Ccuff[ea(PP)−1]eaPt=Va0Ccuff[ea(PP)−1]ea(DBP−P)=Va0Ccuff[ea(PP)−1]ea(DBP)e−aP=k1e−aP (23)
for constant, k1, during a cuff deflation scan in which cuff pressure, P, varies and the other variables are constant. (Note that here Ccuff is very nearly constant because the rising phase of the pulse happens in a very short time, roughly 0.1 sec.) Hence, ln(ΔP)=ln(k1)−aP, and a regression plot of the natural logarithm of the amplitude of pulse oscillations in the leading region of the envelope versus the instantaneous cuff pressure, P, yields a plot with slope − a. Thus we can obtain by linear regression an estimate of stiffness constant, a, as aˆ=slope1. The range of the rising phase of the oscillation envelope from the beginning of the envelope to the first inflection point (maximal slope) can be used for the first semi-log regression. More simply, the range of the rising phase of the oscillation envelope from its beginning to one third of the peak height provides reasonable estimates of slope1.

Similarly in Domain (3) during the tail region of the amplitude envelope at cuff pressures less than the maximal negative slope of the falling phase

ΔP=ΔVaCcuff=Va0Ccuff[ab(1−e−b(Pt+PP))−ab(1−e−bPt)]=Va0Ccuffab[1+e−b(PP)]e−bPt=Va0Ccuffab[1+e−b(PP)]e−b(DBP−P)=k3ebP (24)
hence, ln(ΔP)=ln(k3)+bP, and a regression plot of the natural logarithm of the amplitude of pulse oscillations in the envelope tail versus cuff pressure at the time of each pulse yields a plot with slope b. In turn, we can obtain by linear regression an estimate of stiffness constant, b, as bˆ=slope3. The range of the falling phase of the oscillation envelope from the second inflection point (maximal negative slope) of the oscillation envelope to the end of the envelope can be used to define the range of the second semi-log regression. More simply, the range of the falling phase of the oscillation envelope from two thirds of the peak height to the end of the envelope provides reasonable estimates of slope3. The slope estimates from the head and tail regions of the amplitude envelope include multiple points and so are relatively noise resistant. Other variables involved in the lumped constants, k1 and k3, are not relevant to the estimation of exponential constants a and b.

Least squares analysis
Having estimated elastic constants a and b for a particular envelope of oscillations from a particular patient at a particular time, it is straightforward in a computer program to find SBP and DBP values that reproduce the observed envelope function most faithfully. Let y(P) be the observed envelope amplitude as a function of cuff pressure, P, and let ymax(Pmax) be the observed peak amplitude of oscillations at cuff pressure Pmax. Let yˆ(P, SBP, DBP) be the simulated envelope amplitude as a function of cuff pressure, P, for a particular pulse and a particular test set of systolic and diastolic pressure levels. The values of yˆ are obtained from equations (7) and the prevailing cuff compliance as follows

Domain (1):

yˆ=ΔVaCcuff=Va0[ea(SBP−P)−ea(DBP−P)]⋅P+760V0 (25)
Domain (2):

yˆ=ΔVaCcuff=Va0[1+ab(1−e−b(SBP−P))−ea(DBP−P)t]⋅P+760V0 (26)
Domain (3):

yˆ=ΔVaCcuff=Va0[ab(1−e−b(SBP−P))−ab(1−e−b(DBP−P))]⋅P+760V0. (27)
Let yˆmax(Pmax, SBP, DBP) be the predicted peak of the oscillation envelope at cuff pressure Pmax . A figure of merit for goodness of fit between modeled and observed oscillations for particular test values of SBP and DBP is the sum of squares over all measured pulses

SS(SBP,DBP)=∑allpulses(yymax−yˆyˆmax)2. (28)
The values of SBP and DBP that minimize this sum of squares are the taken as the best estimates of systolic and diastolic pressure by the oscillometric method.

Here cuff pressure, P, is the cuff pressure at the time of each oscillation. Use of the amplitude normalized ratios y/ymax and yˆ/ yˆmax, means that it is not necessary to know the zero pressure volume of the artery, Va0 , or cuff volume V0, which depend on anatomy and geometry of a particular arm and cuff and are constants. It is the shape of the amplitude envelope in the pressure domain that contains the relevant information. The least squares function, SS, includes information from all of the measured oscillations and so is relatively noise resistant.

A variety of numerical methods may be used to find the unique values of SBP and DBP corresponding to the minimum sum of squares. Here, to demonstrate proof of concept, we evaluate the sum of squares, SS, over a two-dimensional matrix of candidate systolic and diastolic pressures at 1 mmHg intervals and identify the minimum sum of squares by plotting. The values of SBP and DBP corresponding to this minimum sum of squares are the best fit estimates for a particular oscillometric pressure run. The best fit model takes into account the prevailing artery stiffness and also the prevailing pulse pressure.

Results and discussion
Normal model
Particular parameter values for the standard normal model are as shown in Table 1.

Table 1. Standard parameters for the oscillometric blood pressure model
Figures 5 (a) and (b) show plots of cuff pressure and arterial pressure vs. time and high pass filtered cuff pressure oscillations vs. time. Figure 6 shows cuff pressure oscillations vs. cuff pressure and the amplitude envelope of cuff pressure for the standard normal model. Cuff pressure oscillations were obtained by subtracting each particular value from the moving average value over a period of one heartbeat.

thumbnailFigure 5. Simulated oscillometric blood pressure determination in a normal patient. (a) Blood pressure and cuff pressure vs. time. (b) High pass filtered cuff pressure oscillations.
thumbnailFigure 6. Simulated oscillometric blood pressure determination in a normal patient. (a) Cuff pressure oscillations vs. pressure. (b) Amplitude envelope obtained from maximum minus minimum cuff pressure over each heartbeat.
Varying arterial compliance
Prior studies have suggested that variations in arterial wall stiffness and arterial pulse pressure cause errors in systolic and diastolic blood pressure estimates using the oscillometric method [4,14,15]. Hence, these variables were studied explicitly. Figure 7 shows effects of varying arterial stiffness, represented by the constants a and b in the bi-exponential artery model. Actual blood pressure was 120/80 mmHg. The cuff oscillation ratios for systolic pressure are similar with varying stiffness. However, the cuff oscillation ratios for diastolic pressure differ greatly among more compliant, normal, and stiffer arteries, indicating that the same oscillation ratios cannot be used to determine diastolic pressures from the amplitude envelope when artery stiffness varies. The diastolic oscillation ratios decrease from about 94% to 88% to 75% as stiffness decreases from high to normal to low. Oscillation amplitude ratios for diastolic pressure in particular are highly dependent upon the stiffness of arteries. Since artery stiffness varies with age, this phenomenon may be a problem clinically. Note, however, that the maximum and minimum slopes of the envelope in the pressure domain still correlate well with true systolic and diastolic pressures.

thumbnailFigure 7. Amplitude envelopes for varying arterial stiffness. Stiffness is represented as inverse compliance. Exponential constants a and b for 1/2 normal stiffness are multiplied by ln(2) = 1.44. Exponential constants a and b for 2x normal stiffness are divided by 1.44. In all cases actual blood pressure was 120/80 mmHg.
Varying pulse pressure
Figures 8 and 9 show raw data and amplitude oscillation envelopes for cases of high and low pulse pressure. The amplitude of cuff pressure oscillations is greater for widened pulse pressure than for narrowed pulse pressure. The shape of the amplitude envelope is distorted for widened pulse pressure; however the maximum and minimum slopes of the envelope in the pressure domain still correlate well with true systolic and diastolic pressures. Characteristic ratios for systolic and diastolic pressures vary with pulse pressure. The characteristic ratio for systolic pressure is substantially smaller for widened pulse pressure and significantly larger for narrowed pulse pressure. The characteristic ratio for diastolic pressure is substantially larger for widened pulse pressure than for narrowed pulse pressure.

thumbnailFigure 8. Simulations of varying arterial pulse pressure. (a) and (b) blood pressure and cuff pressure vs. time, 140/60 mmHg vs. 110/90 mmHg.
thumbnailFigure 9. Simulations of varying arterial pulse pressure. (a) and (b) amplitude envelopes for 140/60 mmHg vs. 110/90 mmHg.
Regression analysis for systolic and diastolic pressures
Figure 10 shows semi-log plots for the envelope functions shown in Figure 7 representing arteries of varying stiffness. The linear portions of the plots in the head and tail regions of log envelope amplitude vs. cuff pressure curves are evident. The artery stiffness constants a and b obtained from linear regression slopes for these head and tail regions are close to the nominal input values (data in Table 2).

thumbnailFigure 10. Semi-log plots for determining model constants from amplitude envelope data. Note straight line regions in rising and falling phases of the curves.
Table 2. Validation of algorithm for estimation of systolic and diastolic pressures
Figure 11 shows a contour map of the sum of squares function in equation (11) for different test values of systolic and diastolic blood pressure using the previously determined regression values for stiffness constants a and b. The semi-log regression slopes give values for constants a and b of 0.1074 and 0.0303, respectively, versus the actual values of 0.110 and 0.030 used in the model to create the analyzed cuff pressure oscillations. The minimum sum of squares indicates the best fit between the oscillation envelope predicted by the mathematical model and the observed oscillation envelope. The minimum sum of squares is shown in Figure 11 as the center of the target-like pattern of colored, equal value contours. This point indicates the least squares solutions both for systolic pressure on the vertical scale and for diastolic pressure on the horizontal scale. Larger diameter ring-shaped contours indicate progressively greater sums of squares and therefore progressively greater disagreement between observed and predicted oscillation envelopes. The contour interval is 0.1 dimensionless units. The flat background indicates exceedingly large, off-scale sums of squares > 1.5 units. The minimum sum of squares occurs for test values SBP/DBP of 119/80 mmHg. The actual pressure was 120/80 mmHg.

thumbnailFigure 11. Contour plot of sum of squares goodness of fit measure showing a minimum value and best agreement at an estimated blood pressure of 119/80 mmHg, evaluated for input data computed with known pressure of 120/80 mmHg. Flat background indicates exceedingly large, off-scale sums of squares.
Figure 12 illustrates the sensitivity of the reconstruction algorithm to differences between various test levels and the actual values of systolic and diastolic blood pressure, in this case 120/80 mmHg. A low value of test pressure (110/70) creates a reconstructed envelope (dashed curve to left) that is clearly discordant with the observed normalized envelope values, E/Emax, shown as filled circles. A high value of test pressure (130/90) leads to equally discordant reconstructions in the opposite direction (heavy dashed curve to right). For both low and high test values the sum of squared differences is obviously large. The reconstructed model for the actual pressure (120/80) is shown as the solid curve. This illustration demonstrates the sensitivity of the least squares approach.

thumbnailFigure 12. Agreement of model (curves) and input (filled circles) amplitude functions in the normal pressure case.
Validation of the regression procedure
The cuff-arm-artery model of an oscillometric pressure measurement described in Methods Part One can be used to validate the regression and analysis procedure of Methods Part Two. An unlimited number and wide variety of test scenarios can be simulated in the model as unknowns for testing by the regression scheme, including a wide range of arterial stiffnesses and a wide range of pulse pressures, heart rates, blood pressure waveforms, cuff sizes, arm sizes, cuff lengths, artery diameters, etc. Importantly, the regression analysis assumes no prior knowledge of these model parameters or of the blood pressure used to generate the simulated oscillations. Cuff pressure oscillations and absolute cuff pressure are the only inputs to the algorithm for obtaining systolic and diastolic pressures.

The data summarized in Table 2 show the effectiveness of the proposed regression procedure in small sample of various possible test scenarios, including varying artery stiffness and varying pulse pressure. This small, systematic sample includes challenging cases for the algorithm. The accuracy is quite satisfactory, with reconstructed pressures within 0, 1, or 2 mmHg of the actual pressures in the face of varying artery stiffness and varying pulse pressure. The root mean squared error is 8√/10 = 0.28 mmHg.

Discussion
The challenge of creating a satisfactory theoretical treatment of the genesis and interpretation of cuff pressure oscillations has attracted a diverse community of thinkers [4,5,7-10,16]. Nevertheless, specifying a valid method for extracting systolic and diastolic pressures from the envelope of cuff pressure oscillations remains an open problem. Here is presented a mathematical model incorporating anatomy, physiology, and biomechanics of arteries that predicts cuff pressure oscillations produced during noninvasive measurements of blood pressure using the oscillometric method. Understanding of the underlying mechanisms leads to a model-based algorithm for deducing systolic and diastolic pressures accurately from cuff pressure oscillations in the presence of varying arterial stiffness or varying pulse pressure.

The shape of the oscillation amplitude envelope dictates the stiffness parameters for the artery during both compression and distension. Semilog regression procedures give good estimates of the artery stiffness parameters that characterize each individual cuff deflation sequence. Using these parameters one can create and exercise an individualized cuff-arm-artery model for a wide range of possible systolic and diastolic pressures. The pair of systolic and diastolic pressures that best reproduces the observed oscillation envelope according to a least squares criterion constitutes the output of the algorithm.

When applied to amplitude normalized oscillation data the algorithm is insensitive to variations among subjects in zero pressure artery volume, Va0 , or initial cuff volume V0, since these terms are constants that are eliminated by the normalization procedure. Compression of the entire length of artery underlying the cuff is not necessary. Incomplete coupling of cuff pressure to the artery near the ends of the cuff merely decreases the ratio Va0 / V0 without effecting the extracted systolic and diastolic pressures.

The cuff-arm-artery model can be used as well to test the validity of the algorithm for over a wide range of possible conditions by generating trial cuff pressure data for known arterial pressure waveforms. A stress test for the algorithm can be done by comparing systolic and diastolic pressure levels extracted from synthesized cuff pressure oscillations with the arterial pressure that generated the synthesized oscillations over a wide range of test conditions. These conditions may include extreme cases that are hard to reproduce experimentally, contamination with excessive noise, any conceivable blood pressure waveforms, cardiac arrhythmias such as atrial fibrillation, etc. Such computational experiments, in addition to future animal and clinical studies, can boost confidence in the reliability of the oscillometric method and can suggest further refinements.

Here for convenience we have used the bi-exponential model to generate cuff pressure oscillations for algorithm testing. However, the regression algorithm does not “know” where the sample data came from. It tries to extract constants a and b from the head and tail portions of the semi-log plot of oscillation amplitude versus cuff pressure. The resulting best fit values of a and b will still work for non-ideal or noise contaminated data to produce a model envelope that can be matched to the actual data. An extremely stiff artery with a linear pressure volume curve is easily accommodated by this process, since ex ≈ 1 + x for small values of x. In this limiting case the exponential pressure-volume curve becomes linear. An exceptionally flabby artery, rather like dialysis tubing, is well described by larger values of a and b and a larger ratio a/b. Thus the family of bi-exponential models is very inclusive of a wide range of arterial mechanical properties, as suggested in Figure 3.

Classically the oscillometric method has been relatively well validated as a measure of mean arterial pressure, which is indicated by the peak of the oscillation amplitude envelope [4]. Automated oscillometric pressure monitors have found use in hospitals for critical care monitoring in which the goal is to detect any worrisome trend in blood pressure more so than the exact absolute value. Out of hospital use of the oscillometric method in screening for high blood pressure is more problematic, because heretofore the accuracy of systolic and diastolic end points has been questioned and doubted. For example Stork and Jilek [17] studied two published algorithms, differing in detail and based on cuff oscillation ratios of either 50% for systolic and 80% for diastolic or 40% for systolic and 55% for diastolic. Compared to a reference pressure of 122/78 mmHg the algorithmic methods applied to oscillometric data gave pressures of 135/88 and 144/81 mmHg, respectively. An advisory statement from the Council for High Blood Pressure Research, American Heart Association [18] stressed the need for caution in the selection of all instruments used for blood pressure determination and the need for continuing studies to validate their the safety and reliability.

Accurate measurements of blood pressure in routine clinic and office settings are important because systemic arterial hypertension is a major cause of serious complications, including accelerated atherosclerosis, heart attacks, strokes, kidney disease, and death. These serious complications increase smoothly with every point above the nominal 120/80 mmHg, hence even small increases in blood pressure are important to detect. In screening for hypertension systematic bias or inaccuracy in blood pressure readings of a few mmHg can be significant, since the difference between high normal (85 diastolic) and abnormal (90 diastolic) is only a few mmHg. A recent 1 million-patient meta-analysis suggests that a 3–4 mmHg increase in systolic blood pressure would translate into 20% higher stroke mortality and a 12% higher mortality from ischemic heart disease [19].

False negative readings would be problematic because untreated high blood pressure can lead to strokes, blindness, kidney failure, and lethal heart attacks. False positive readings would be undesirable because the usual drugs for hypertension must be taken every day for life and can be expensive. They also have side effects. Hence accurate readings are essential. Given a reliable algorithm for extracting systolic and diastolic pressures, an automatic oscillometric device could provide screening for high blood pressure that is performed in the same way each time without inter-observer variation. The present research could lead to a wider role for oscillometric blood pressure monitors in physicians’ offices and clinics.

Conclusions
The analytical approach and algorithm presented here represent a solution to an open problem in biomedical engineering: how to determine systolic and diastolic blood pressures using the oscillometric method. Current algorithms for oscillometric blood pressure implemented in commercial devices may be quite valid but are closely held trade secrets and cannot be independently validated. The present paper provides a physically and physiologically reasonable approach in the public domain that can be independently criticized, tested, and refined. Future demonstration of real-world accuracy will require data comparing oscillometric and intra-arterial pressures in human beings over a range of test conditions including variable cuff size, arm diameter, cuff tightness, cuff deflation rate, etc. Further development and incorporation of this algorithm into commercial devices may lead to greater confidence in the accuracy of systolic and diastolic pressure readings obtained by the oscillometric method and, in turn, an expanded role for these devices.

Competing interests
The author declares that he has no competing interests.

Author’s contributions
CB is the only author and is responsible for all aspects of the research and the intellectual and technical content of the manuscript.

References
Boron WF, Boulpaep EL: Medical Physiology. 2nd edition. Philadelphia: Elsevier; 2005. OpenURL
Geddes LA, Baker LE:
Prinicples of Applied Biomedical Instrumentation. Second edition. 1975. OpenURL
Soueidan K, Chen S, Dajani HR, Bolic M, Groza V: Augmented blood pressure measurement through the noninvasive estimation of physiological arterial pressure variability.
Physiol Meas 2012, 33:881-899. PubMed Abstract | Publisher Full Text OpenURL
Mauck GW, Smith CR, Geddes LA, Bourland JD: The meaning of the point of maximum oscillations in cuff pressure in the indirect measurement of blood pressure–part ii.
J Biomech Eng 1980, 102:28-33. PubMed Abstract | Publisher Full Text OpenURL
Geddes LA, Voelz M, Combs C, Reiner D, Babbs CF: Characterization of the oscillometric method for measuring indirect blood pressure.
Ann Biomed Eng 1982, 10:271-280. PubMed Abstract | Publisher Full Text OpenURL
Ursino M, Cristalli C: A mathematical study of some biomechanical factors affecting the oscillometric blood pressure measurement.
IEEE Trans Biomed Eng 1996, 43:761-778. PubMed Abstract | Publisher Full Text OpenURL
Baker PD, Westenskow DR, Kuck K: Theoretical analysis of non-invasive oscillometric maximum amplitude algorithm for estimating mean blood pressure.
Med Biol Eng Comput 1997, 35:271-278. PubMed Abstract | Publisher Full Text OpenURL
Drzewiecki G, Hood R, Apple H: Theory of the oscillometric maximum and the systolic and diastolic detection ratios.
Ann Biomed Eng 1994, 22:88-96. PubMed Abstract | Publisher Full Text OpenURL
Forster FK, Turney D: Oscillometric determination of diastolic, mean and systolic blood pressure–a numerical model.
J Biomech Eng 1986, 108:359-364. PubMed Abstract | Publisher Full Text OpenURL
van Popele N, Bos W, de Beer N, van der Kuip D, Hoffman A, Diederick E, CM W: Arterial stiffness as underlying mechanism of disagreement between oscillometric blood pressure monitor and sphygmomanometer.
Hypertension 2000, 36:484-488. PubMed Abstract | Publisher Full Text OpenURL
Pickering TG, Hall JE, Appel LJ, Falkner BE, Graves J, Hill MN, Jones DW, Kurtz T, Sheps SG, Roccella EJ: Recommendations for blood pressure measurement in humans and experimental animals: part 1: blood pressure measurement in humans: a statement for professionals from the Subcommittee of Professional and Public Education of the American Heart Association Council on High Blood Pressure Research.
Circulation 2005, 111:697-716. PubMed Abstract | Publisher Full Text OpenURL
Posey J, Geddes L: Measurement of the modulus of elasticity of the arterial wall.
Cardiovasc Res Cent Bull 1973, 11:83-103. OpenURL
Jeon G, Jung J, Kim I, Jeon A, Yoon S, Son J, Kim J, Ye S, Ro J, Kim D, Kim C: A simulation for estimation of the blood pressure using arterial pressure-volume model.
World Acad Sci Eng Technol 2007, 30:366-371. OpenURL
Mai JJ, Insana MF: Strain imaging of internal deformation.
Ultrasound Med Biol 2002, 28:1475-1484. PubMed Abstract | Publisher Full Text OpenURL
Ursino M, Cristalli C: Mathematical modeling of noninvasive blood pressure estimation techniques–Part I: pressure transmission across the arm tissue.
J Biomech Eng 1995, 117:107-116. PubMed Abstract | Publisher Full Text OpenURL
Ursino M, Cristalli C: Mathematical modeling of noninvasive blood pressure estimation techniques–Part II: brachial hemodynamics.
J Biomech Eng 1995, 117:117-126. PubMed Abstract | Publisher Full Text OpenURL
Stork M, Jilek J: Cuff pressure waveforms: thier current and prospective application in biomedical instrumentation. In Biomedical Engineering Trends in Electronics, Communications and Software. Edited by Laskovski AN. InTech; 2011:193-210. OpenURL
Jones DW, Frohlich ED, Grim CM, Grim CE, Taubert KA: Mercury sphygmomanometers should not be abandoned: An advisory statement from the Council for High Blood Pressure Research, American Heart Association.
Hypertension 2001, 37:185-186. PubMed Abstract | Publisher Full Text OpenURL
Lewington S, Clarke R, Qizilbash N, Peto R, Collins R: Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies.
Lancet 2002, 360:1903-1913. PubMed Abstract | Publisher Full Text OpenURL

(N)eural , (E)soterical , (B)enevolent , NEB Program

Neural Program is Stephen Fry proof thanks to caching by WP Super Cache