An RNA Profile Identifies Two Subsets of Multiple Sclerosis Patients Differing in Disease Activity

Abstract

The multiple sclerosis (MS) patient population is highly heterogeneous in terms of disease course and treatment response. We used a transcriptional profile generated from peripheral blood mononuclear cells to define the structure of an MS patient population. Two subsets of MS subjects (MS_A and MS_B) were found among 141 untreated subjects. We replicated this structure in two additional groups of MS subjects treated with one of the two first-line disease-modifying treatments in MS: glatiramer acetate (GA) (n = 94) and interferon-β (IFN-β) (n = 128). One of the two subsets of subjects (MS_A) was distinguished by higher expression of molecules involved in lymphocyte signaling pathways. Further, subjects in this MS_A subset were more likely to have a new inflammatory event while on treatment with either GA or IFN-β (P = 0.0077). We thus report a transcriptional signature that differentiates subjects with MS into two classes with different levels of disease activity.

INTRODUCTION

Multiple sclerosis (MS) is thought to emerge when genetically susceptible individuals encounter environmental triggers that initiate an inflammatory reaction against self-antigens in the central nervous system. These events result in recurring episodes of inflammatory demyelination and, in many cases, a progressive neurodegenerative process (1). However, the prevalent syndromic definition of MS obscures extensive interindividual variation in terms of disease course, response to a given treatment, and distribution of deficits. Currently, although imaging may help to stratify risk of a second demyelinating event in the context of a clinically isolated demyelinating syndrome (CIS) (2), there are no validated molecular biomarkers that offer meaningful predictions for MS. As a result, in a clinical setting, one can only categorize subjects after prolonged observation of their disease course. Further, treatment decisions are based on study results from the highly selected subject populations used in clinical trials.

As illustrated by the success of natalizumab (an anti–VLA-4 antibody) and fingolimod (a sphingosine 1-phosphate receptor modulator), immune cells in the peripheral circulation play an important role in MS (3, 4). Sequestering these cells in the peripheral compartment reduces relapse rates and evidence of inflammatory demyelination on magnetic resonance imaging (MRI) in the brain. These and other studies suggest that sampling peripheral blood is likely to be informative in exploring the structure of the MS patient population. Here, we assess whether we can use such information to identify a priori subsets of subjects whose clinical course will differ over time.

Studies of the peripheral blood transcriptome in MS have been performed in many different ways; however, they have focused, for the most part, on identifying transcriptomic signatures of clinically defined classes of subjects, such as progressive versus relapsing subjects or subjects that respond to a given treatment. Here, similar to the work of Corvol and colleagues, which examined a population of subjects with CIS (5), we use an unsupervised clustering approach to explore whether there is more than one class of MS subjects that can be distinguished using transcriptomic data extracted from peripheral blood mononuclear cells (PBMCs). We define two subsets of MS that are present not only in our large collection of untreated subjects but also in two additional cohorts, one treated with glatiramer acetate (GA) and the other treated with interferon-β (IFN-β). Further, we report that the subset of subjects with higher expression of lymphocyte activation pathway genes is more likely to have another inflammatory event over the course of their disease.

RESULTS

Characteristics of the transcriptome in our three subject collections

In Table 1, we summarize the demographic and clinical characteristics of the subjects with demyelinating disease [relapsing-remitting (RR) MS or CIS] that were considered for this study. All subjects are participants in the Comprehensive Longitudinal Investigation of MS at the Brigham and Women’s Hospital (CLIMB), a prospective study of the early phase of demyelinating disease. Transcriptional profiles were generated from frozen PBMCs collected prospectively as part of CLIMB. There were some differences in the clinical characteristics among the three groups of subjects: (i) a slightly increased frequency of women in the untreated group; (ii) a higher number of CIS subjects in the untreated group; and (iii) a lower level of brain parenchymal fraction (BPF), a normalized measure of whole-brain volume, in the IFN-β–treated subjects, after correction for age at first symptom, sex, and disease duration (Table 1).

Table 1.

Demographic characteristics of subjects used in the study. Values are reported as means (SD). T₂LV: T₂ hyperintense lesion volume.

	GA (n = 94)	IFN-β (n = 128)	Untreated (n = 141)	Adjusted P value*
Female, n (%)	68 (73)	95 (74)	117 (85)	0.0344
Disease subtype (CIS/RR)	7/87	9/119	32/109	0.0001
Symptom duration at sampling	6.12 (6.36)	6.65 (6.36)	7.48 (8.06)	0.5644
Age at first symptom	34.23 (9.19)	34.77 (9.30)	33.35 (9.07)	0.4841
BPF	0.885 (0.034)	0.871 (0.036)	0.880 (0.040)	0.0178
EDSS	1.04 (1.01)	1.19 (1.04)	1.07 (1.09)	0.5539
T2LV**	3.19 (2.58)	3.65 (2.64)	3.93 (4.75)	0.1720

^*Adjusted for symptom duration, age at first symptom, and gender (unless outcome of interest). MRI measurements were captured in a time window of 6 months before or after sampling.

^**P value from model with natural log transformation of T₂ lesion volume value.

At the end of our quality control pipeline, 20,527 probe sets from 363 subjects passed our rigorous quality control measures. The first three principal components derived from these standardized data capture 41.4% of the variance in the data, and these three principal components are not associated with treatment category (PC1, P = 0.49; PC2, P = 0.16; PC3, P = 0.20). This is appreciated in Fig. 1A in which the distribution of PBMC RNA profiles from subjects of the three treatment categories (GA, IFN-β, and untreated) overlap. This figure shows that these major axes of transcriptome-wide variation are not significantly different among the three subject categories. When assessing individual probe sets for evidence of pairwise differential expression in the treatment groups, we find that there are no significant differences [false discovery rate (FDR) < 0.05] between the 94 GA-treated and 141 untreated subjects (Fig. 1B) (table S1). On the other hand, both of these groups reveal the expected altered expression of type 1 interferon response genes when compared to the 128 IFN-β–treated subjects: (i) for untreated versus IFN-β, 1649 probe sets are up-regulated and 1453 probe sets are down-regulated at an FDR < 0.05 in IFN-β subjects (table S2); (ii) for GA versus IFN-β, 568 probe sets are up-regulated and 840 probe sets are down-regulated at FDR < 0.05 in IFN-β–treated subjects (table S3). As illustrated in Fig. 1B, the differential expression patterns of the two comparisons (untreated and GA) with IFN-β overlap, and the direction of effect (up- or down-regulation) is the same in the two comparisons. The most up-regulated probe sets (targeting the IFI44L, IFIT1, IFI44, IFIT3, IFI6, OASL, MX1, IFIT2, OAS1, and OAS2 genes) quantify the RNA expression level of well-described interferon response genes (6–8) (table S4). Thus, whereas we find the expected transcriptional changes in PBMCs after treatment with IFN-β (9, 10), we see a lack of transcriptional changes in PBMCs after treatment with GA even though our large sample size gives us >90% power to identify an RNA expression difference between GA-treated and untreated subjects that is of the same magnitude as that seen in IFN-β–treated subjects relative to untreated subjects. This suggests that, at the patient population level, the effects of chronic GA treatment on transcription, which have not been explored deeply so far, (i) may be modest for any one gene, (ii) may be limited to a small subset of circulating cells, or (iii) may be limited to specific immunological compartments and are not reflected in peripheral blood. With this baseline assessment of the transcriptomic profiles of the three subject groups completed, we turned to analyses exploring the heterogeneity of the subject population.

Fig. 1.

Transcriptomic data from MS subjects. (A) Distribution of the study subjects using the first three principal components calculated from the transcriptomic data. Each dot represents one individual, color-coded according to the treatment group to which they belong [green, untreated (UNTR); red, GA-treated; blue, IFN-β–treated]. (B) Venn diagram of the number of the differentially expressed probe sets in each pairwise comparison among the three treatment groups. The number of probe sets meeting a Benjamini-Hochberg threshold of FDR < 0.05 was reported. Between GA and untreated subsets, there were no probe sets meeting that threshold. The inset graph plots each of the 1061 probe sets shared by the two comparisons to IFN-β using the magnitude and direction of the differential expression in each comparison (x axis, untreated and IFN-β; y axis, GA and IFN-β) to illustrate the consistency of gene expression changes in PBMCs relative to the IFN-β subject group.

Unsupervised clustering identifies two untreated MS subgroups

To empirically determine the number of subsets of subjects present within our untreated group, we used an unsupervised clustering method, non-negative matrix factorization (NMF) (11). NMF allowed exploration of the structure of our subject population without any assumptions as to which genes may be important. It identified metagenes, or groups of coregulated genes, which were used to cluster subjects into a range of subsets. We tested a range of models (from two to five clusters of subjects) and used the cophenetic coefficient, a measure of how well a model fits the data, to identify the optimal k (k is the number of clusters). As illustrated in Fig. 2, the solution of k = 2 subsets is the one that returns the highest cophenetic coefficient for the untreated subjects (Fig. 2, A and B, and table S5A). We called the smaller subset (n = 58) MS_A and the larger subset (n = 83) MS_B. This structure in our 141 untreated subject population was not driven by the subset of 33 CIS subjects: There was no significant difference in the proportion of CIS subjects in MS_A and MS_B (table S6). Moreover, when the unsupervised clustering approach was repeated after excluding the CIS subjects, the same structure was identified, and there was a strong correlation between subject class assignments in the two independent classifications (with and without CIS subjects; P < 0.0001, Fisher’s exact test) (fig. S1).

Fig. 2.

Two subsets of MS subjects in the untreated, GA, and IFN-β groups. (A) NMF consensus clustering results of mRNA expression array data from 141 untreated MS cases. Four models with two to five clusters of subjects (k = 2 to 5) were tested, and the results were summarized in the four consensus clustering plots that capture all pairwise comparisons for the 141 untreated subjects (one for each model). Each square is colored by the extent of correlation within the transcriptome of that subject pair: red, high correlation; yellow/green, moderate to low correlation; blue, no correlation. The cophenetic coefficient, a measure of how well a model fits the data, is reported below each plot, with 1.0 being maximal fit. (B) Plots summarizing the cophenetic coefficient for each of the four tested models (k = 2 to 5) in each of the three groups of subjects (untreated, GA, and IFN-β). In each case, the solution that maximizes the cophenetic coefficient is k = 2, meaning that the transcriptomic data are best explained by a model in which two subsets of subjects are present. The maximal cophenetic coefficient is reported below each plot.

Given that our three subject groups (GA, IFN-β, untreated) were not significantly different in the major axes of variation in the data, we attempted to replicate the observation of a two-subset population structure in the untreated subjects by deploying the NMF method in GA-treated and IFN-β–treated subjects. In both of the latter groups, the k = 2 subset solution is also the one that maximized the cophenetic coefficient (Fig. 2B) when testing models from k = 2 to k = 5. In untreated subjects, 205 probe sets were differentially expressed (FDR < 0.05) between the MS_A and MS_B subsets (fig. S2); thus, these 205 probe sets offer a transcriptional signature that can differentiate the two subsets of untreated subjects.

To further validate this signature and our hypothesis that the same two subsets of subjects with MS exist in the different treatment groups, we used a support vector machine (SVM) method (a form of supervised clustering) in which a predefined list of probe sets is used to separate subjects into two classes (12–14). With the SVM method and the 205 probe set list discriminating the untreated MS_A and MS_B subject subsets, we partitioned GA-treated subjects into two groups. As seen in Fig. 3A, we then compared the SVM-determined group assignments to the assignments determined by the unsupervised NMF method that uses the entire RNA data set. We observed that the structure of the population determined independently by each method was very similar (class assignment overlaps in 92.6% of subjects). Figure 3A more precisely displays the consistency of the two clustering strategies: Subjects that were classified differently in the two analyses were those subjects where the SVM method made a less confident assignment. Thus, taking the NMF-defined subject class (MS_A versus MS_B) in GA-treated subjects as our best estimate of the “correct” subject classification, we observed that the top differentially expressed genes between the untreated subgroups captured the structure of the GA-treated subject population. Also, the direction of the differential expression was the same: The same genes were over-expressed when comparing MS_A versus MS_B in either untreated or GA-treated subjects. Similarly, when we repeated this process with IFN-β–treated subjects, we were able to correctly predict the unsupervised, NMF-determined class of 90.6% of IFN-β–treated subjects with the SVM strategy and the expression signature determined in untreated subjects (Fig. 3B). As noted above, misclassified individuals were found among individuals for whom the SVM method returned a classification with lower confidence.

Fig. 3.

Top probe sets differentially expressed between MS_A and MS_B subsets in untreated subjects predict subject class in the GA- and IFN-β–treated subject groups. (A and B) The results of supervised SVM classification for GA (A) and IFN-β (B) using the untreated signature are shown. The confidence score for each SVM classification is plotted on the y axis. On the x axis, each subject is arrayed on the basis of the rank order of the likelihood of belonging to MS_A after SVM classification. Each point represents one subject. The upper portion of each graph contains those classified as MS_A by the SVM classifier; the lower portion contains MS_B subjects. The unsupervised NMF-derived classification, which is used as the reference classification, is reported in color: blue, MS_A; green, MS_B. Thus, the misclassified subjects are readily apparent. The confusion matrix that compares the NMF and SVM classifications is reported in each graph; the number and percentage (in parenthesis) of subjects found in each cell of the confusion matrix is shown. A total of 94 GA-treated subjects and 128 IFN-β–treated subjects were classified. (C) The table reports the success of the classifier at different confidence thresholds for the SVM classification and the proportion of subjects (in parenthesis) that meet each threshold. Most of the misclassified individuals were the result of low confidence classifications.

If we impose a confidence threshold >80% before we accept an SVM-determined classification, the accuracy becomes 98.8% for the 81 of 94 GA-treated subjects that have an SVM classification meeting this confidence threshold. Similarly, classification accuracy increases to 96.5% for the 114 of 128 IFN-β–treated subjects meeting the 80% confidence threshold (Fig. 3C). Thus, with a reasonable confidence threshold (>80% confidence), more than 86% of subjects treated with first-line disease-modifying agents (GA or IFN-β) could be classified to >96% accuracy with a signature derived from untreated subjects (Fig. 3C). These results validate the existence of the same two subsets of subjects (MS_A and MS_B) that can be differentiated by a single transcriptional signature in the three subject groups (GA, IFN-β, and untreated). In this signature, 98 probe sets were differentially expressed in the same direction in all three MS_A versus MS_B comparisons (table S7). Table S8 reports, by subject group, the probe sets differentially expressed for each MS_A versus MS_B comparison.

Functional annotation of the transcriptional signature

To search for patterns within the signature differentiating the MS_A and MS_B subsets identified by NMF, we used the Ingenuity Pathway tool (http://www.ingenuity.com) and its repertoire of known pathways. We focused this analysis on the pathways that are shared by the MS_A versus MS_B comparison in each of the three subject groups: Fig. S3 highlights those pathways that are enriched for genes differentiating the MS_A and MS_B subsets and meet a Benjamini-Hochberg P < 0.0005 (which is corrected for the testing of multiple comparisons) in each subject group. Seventeen pathways meet this criterion, but many more show robust enrichment below a Benjamini-Hochberg P < 0.05 threshold (table S9). Overall, many of the pathways have overlapping sets of genes and primarily involve adaptive immune functions. For example, the MS_A subjects displayed, on average, greater expression of genes in the predefined “T cell receptor” and “B cell receptor” pathways (fig. S3); genes found in the NFAT (nuclear factor of activated T cells), ILK (integrin-linked kinase), PI3K (phosphatidylinositol 3-kinase), and EGF (epidermal growth factor) signaling pathways were also expressed more highly in the MS_A subset. Thus, overall, it appears that MS_A subjects may have a greater proportion of lymphocytes or perhaps more activated lymphocytes in their peripheral blood.

An increased likelihood for a demyelinating event in MS_A subjects

Having defined two subsets of subjects with MS by NMF, we went on to assess whether this structure plays a role in disease course. First, we compared the demographic features of the two subsets of subjects in each treatment group to evaluate them for any differences that could confound our treatment response analysis. As outlined in table S5, A to C, none of the available clinical and paraclinical data appeared to be different between the two MS subsets after correcting for the testing of multiple hypotheses, except for a difference in disease duration at the time of sampling in GA-treated subjects. The following analysis therefore included a covariate for disease duration in addition to covariates for sex and age at symptom onset.

To maximize our power in our primary analysis of the role of the subject subsets in disease course, we included GA-and IFN-β–treated subjects that had information on new inflammatory events occurring after the time of blood sampling. The NMF-defined subject subsets were used in this analysis. Given their shared population structure, we merged GA- and IFN-β–treated subjects to assess the role of the subject subsets in disease course, independent of treatment. It is important to note that previous analysis of the extended CLIMB cohort from which our subjects were drawn revealed no significant difference in the likelihood of inflammatory events after the onset of GA or IFN-β treatment (15).

Our primary analysis therefore assessed the trajectories of disease activity in the merged MS_A and MS_B subsets of GA-and IFN-β–treated subjects after transcriptional profiling. Evidence of disease activity (the analysis end point) included (i) a clinical relapse, or (ii) new T₂ hyperintense or gadolinium-enhanced lesions, or (iii) an expanded disability status scale (EDSS) increase of 1, sustained more than 6 months. Among the 56% of GA-treated subjects who reached one of these end points, 39.2% had a clinical relapse, 51.0% had a new lesion on MRI, and 9.8% had an increase in EDSS. For IFN-β–treated subjects who reached an end point (57% of subjects), the distribution included 46.7% with relapses, 38.7% with an MRI event, and 14.7% with an increase in EDSS. There was no significant difference in the distribution of these events between the two treatment categories (P = 0.37). After merging subjects on first-line disease-modifying treatments for our disease course analysis, we observed that subjects classified as MS_A at the time of sampling were more likely to exhibit evidence of disease activity over time than MS_B subjects (Cox proportional hazard ratio = 0.6, P = 0.0077). Specifically, this hazard ratio suggests that MS_B subjects are 40% less likely to have a relapse than MS_A subjects (Fig. 4). Secondarily, we assessed whether this association was driven by one of the two treatment groups. Given that there was no statistically significant difference in the trajectories of the MS_A and MS_B subject subsets when comparing GA- and IFN-β–treated subjects (P = 0.37, interaction term for an effect of treatment on the disease course associated with MS_A), the role of the MS_A and MS_B subject subsets may be the same in the context of disease course while a subject is treated with either one of these two disease-modifying treatments. When the treatment groups were analyzed separately, we observed that, in GA-treated subjects, the MS_A subset had more events than the MS_B subset (P = 0.02) (fig. S4A). In IFN-β–treated subjects, the trend was the same, but the analysis returned a nonsignificant result (P = 0.17) (fig. S4B). Data on the trajectory of untreated subjects were unfortunately not available for comparison because most of these subjects were started on a disease-modifying treatment shortly after being sampled.

Fig. 4.

The MS_A and MS_B gene signatures are associated with disease outcome. The proportion of subjects in the MS_A subgroup (black line) and in the MS_B subgroup (red line) that were free of evidence of inflammatory demyelination was plotted. Subjects treated with either IFN-β or GA were pooled in this analysis. The results of the Cox proportional hazard (Cox-PH) models were adjusted for gender, disease duration from symptom onset, and treatment duration. The hazard ratio (HR) for the comparison of MS_A versus MS_B (with MS_A as the reference group) is shown, along with its 95% confidence interval.

DISCUSSION

We have gathered a large set of transcriptomic data generated from MS patients and leveraged an unsupervised clustering method to identify two subsets of MS subjects (MS_A and MS_B) that can be distinguished by a transcriptional signature. We validated this architecture of the MS patient population in subjects treated with either GA or IFN-β, suggesting that the underlying architecture of the disease may not be fundamentally altered by these two treatments.

The MS_A and MS_B architecture of the MS subject population requires rigorous further validation in prospective cohorts. If this validation is achieved, we speculate that it may be useful to recognize this population architecture when performing studies of MS patients, particularly as the likelihood for further inflammatory events differs in the two subsets of MS patients. As virtually all MS patients are treated, we unfortunately do not have longitudinal clinical and MRI data on the untreated subject group and do not know whether there may be a possible differential disease course in the two subject subsets among individuals who remain untreated. However, looking at our data generated from subjects on chronic treatment, the population of MS_A subjects had a more active disease course than the MS_B subjects on either GA or IFN-β treatment, suggesting that the difference in disease course was not treatment-specific. Thus, our transcriptional signature may have uncovered information relating to the pathophysiology of MS, and MS_A subjects may benefit from consideration of more aggressive treatments at an earlier stage of disease compared to MS_B subjects. Further, although the two subsets appear to behave similarly in the context of GA and IFN-β treatment, it is possible that other interventions may have a different effect in the two subsets.

The nature of this transcriptional signature requires more detailed study; our initial pathway analysis highlighted lymphocyte activation pathways as being overrepresented among the genes that were differentially expressed between MS_A and MS_B. This observation fits with our existing knowledge of MS pathophysiology in which both B and T cells play an important role in the early, inflammatory phase of MS, which was the focus of this study (1). Whether this signature is driven by a higher frequency of lymphocytes among PBMCs of MS_A subjects or by overexpression of specific molecules of activation pathways in lymphocytes remains to be determined. Investigation of specific cell populations is the next step. Such investigation will also need to sample subjects before starting therapy to integrate our results with expression signatures identified by other investigators, such as the higher expression of interferon response genes in pretreatment samples that appears to be associated with more disease activity after the initiation of IFN-β treatment (9, 10).

Our study has a number of limitations, including the lack of transcriptional profiles from longitudinal samples of our subjects. Thus, we cannot comment on the stability of our subject classification over time. Although there appear to be two MS states in our study (MS_A and MS_B), we do not know whether an individual subject fluctuates between these two states or remains in a single state over time. This important question needs to be explored in future longitudinal studies to determine how the molecular signature performs as a prognostic marker for predicting disease activity prospectively. The distinction is important because it would inform the frequency at which a subject’s state should be characterized; repeated sampling would be required if, for example, our transcriptional signature distinguishes active subjects with ongoing, acute episodes of inflammation from subjects who are quiescent at the time of sampling. None of our subjects had a clinical relapse at the time of sampling, but much of the disease activity in MS is asymptomatic and is probably imperfectly captured by MRI (16). Thus, whether our signature provides a surrogate marker for active, asymptomatic inflammation or captures an underlying difference in the pathophysiology of different subjects with MS, this information could be leveraged in the future to enhance patient care and drug development.

Another limitation of our study is the lack of data relating to disease course in our untreated subjects; as a result, we cannot assess whether GA and IFN-β treatments altered the natural history of disease activity in MS_A and MS_B subjects. It is possible that the difference in disease trajectories may be more pronounced in the untreated state. In addition, although the use of a single RNA detection platform and cellular substrate (PBMCs) enhanced our discovery effort, we have not addressed whether the signature is transportable to other settings that more closely approximate a clinical laboratory. Finally, our focus on the early, inflammatory phase of MS does not address the pressing need to understand the pathophysiology of progressive disease, that is, typically, the more disabling phase of MS. It is unclear, at this time, whether the architecture of the early MS patient population that we describe persists in the progressive stage and whether the MS_A and MS_B classification affects the timing or likelihood of entering the progressive phase of MS.

Overall, we report a transcriptional signature that distinguishes a subset of MS patients with more active disease. Stratifying MS subjects into meaningful subsets in this manner has potential for personalizing patient care and for enhancing our understanding of this disease.

MATERIALS AND METHODS

Patients and samples

Three hundred and sixty-three subjects with demyelinating disease and frozen PBMC samples were selected from the CLIMB (17) at the Partners Multiple Sclerosis Center in Boston. At the time of sampling, 315 subjects (86.77%) were diagnosed with RRMS per McDonald criteria (18) and 48 subjects (13.22%) were diagnosed with CIS. Only patients who were untreated or treated with one of the two first-line disease-modifying treatments, IFN-β and GA, at the time of sampling were part of the study. The untreated subjects (i) had never been treated with cytotoxic, monoclonal, or investigational agents and (ii) had been sampled either before their first treatment or at least 6 weeks after their last treatment. The disease-modifying treatment groups were limited to subjects that (i) had never been treated with cytotoxic, monoclonal, or investigational agents and (ii) had been treated for at least 3 months at the time of sampling. Finally, we excluded all samples collected within 4 weeks of a methylprednisolone pulse.

Computational methods

Unsupervised clustering

NMF consensus clustering was used to uncover the model that best fit each of the three data sets. It was performed with the algorithm proposed by Brunet et al. in 2004 (11). Briefly, NMF is an unsupervised learning algorithm (19, 20) that identifies molecular patterns in gene expression data. Rather than separating gene clusters on the basis of distance computation (Euclidean), NMF detects context-dependent patterns of gene expression in complex biological systems.

First, we reduce the dimensionality of the expression data from thousands of genes to a few metagenes by applying NMF consensus as implemented in GenePattern version 4 (14). We computed multiple factorizations of our expression matrix A. The number of classes k that we tested on the basis of the expected heterogeneity of our MS data set was between 2 and 5. In NMF, we repeated the clustering process many times to assess its stability and robustness and to define consensus membership. We have tested 20 and 40 initial clustering attempts per k to define the factorization of matrix A (20 is a standard default setting and 40 matrices require more computation time but may provide more robust results). Matrix A derives from the orthogonal combination of W × H, where W and H are the dimensionality-reduced matrices obtained by number of genes and number of experiments, respectively, per number of classes that are tested to assess heterogeneity of the data set (k = 2 to 5) (11). k = 2 was always the most stable model when testing either 20 or 40 initial different W × H matrices. In each of the three classes, only one sample was differently classified when we compare the classifications generated with n = 20 and n = 40 permutations of the matrix. We therefore selected the results based on the 40 matrices for each model (k = 2 to 5).

Supervised SVM clustering

To validate the structure of the MS patient population observed in untreated subjects, we used SVM (as implemented in GenePattern version 4, namely, with SVM R algorithm in e1071 package) (14) and the set of 205 probe sets identified by a linear regression model (Limma, R package) (21) as significantly differentially expressed between the untreated MS_A and MS_B subgroups to predict class membership of each individual in the GA- and IFN-β–treated subject groups. For each subject, the class defined by NMF (unsupervised) was used as the reference “true” class in our predictive analyses. We used the Prediction Results Viewer tool (version 4.5) written in Java and available in the GenePattern toolkit to visualize the results.