Prediction of acute multiple sclerosis relapses by transcription levels of peripheral blood cells
- Michael Gurevich†1,
- Tamir Tuller†1Email author,
- Udi Rubinstein2,
- Rotem Or-Bach1 and
- Anat Achiron1Email author
© Gurevich et al; licensee BioMed Central Ltd. 2009
Received: 10 December 2008
Accepted: 22 July 2009
Published: 22 July 2009
The ability to predict the spatial frequency of relapses in multiple sclerosis (MS) would enable physicians to decide when to intervene more aggressively and to plan clinical trials more accurately.
In the current study our objective was to determine if subsets of genes can predict the time to the next acute relapse in patients with MS. Data-mining and predictive modeling tools were utilized to analyze a gene-expression dataset of 94 non-treated patients; 62 patients with definite MS and 32 patients with clinically isolated syndrome (CIS). The dataset included the expression levels of 10,594 genes and annotated sequences corresponding to 22,215 gene-transcripts that appear in the microarray.
We designed a two stage predictor. The first stage predictor was based on the expression level of 10 genes, and predicted the time to next relapse with a resolution of 500 days (error rate 0.079, p < 0.001). If the predicted relapse was to occur in less than 500 days, a second stage predictor based on an additional different set of 9 genes was used to give a more accurate estimation of the time till the next relapse (in resolution of 50 days). The error rate of the second stage predictor was 2.3 fold lower than the error rate of random predictions (error rate = 0.35, p < 0.001). The predictors were further evaluated and found effective both for untreated MS patients and for MS patients that subsequently received immunomodulatory treatments after the initial testing (the error rate of the first level predictor was < 0.18 with p < 0.001 for all the patient groups).
We conclude that gene expression analysis is a valuable tool that can be used in clinical practice to predict future MS disease activity. Similar approach can be also useful for dealing with other autoimmune diseases that characterized by relapsing-remitting nature.
Multiple sclerosis (MS) is an autoimmune demyelinating central nervous system (CNS) disease characterized by an unpredictable relapsing-remitting course. In MS and other autoimmune diseases, a relapse is defined as the new onset or worsening of clinical neurological symptoms, and is followed by periods of remissions with no disease activity. Relapses are the basic feature of MS and other autoimmune diseases such as myasthenia gravis , systemic lupus erythemathosus , rheumatoid arthritis , and Crohn's disease . In MS, relapses are the consequence of complex immunological and neurodegenerative processes. Relapses in MS are associated with myelin and axonal loss; they may cause new acute inflammatory lesions or can activate old lesions within the CNS [5–7]. Accordingly, relapses are the visible clinical expression of the complicated immunopathological mechanisms operating in the CNS and peripheral blood. The ability to predict the occurrence of a subsequent relapse (yes/no) and to estimate the time when that process will occur has important clinical and practical implications. This knowledge can help in decisions related to treatment – e.g. either treat patients with more aggressive disease or avoid over-treatment of patients with a more favorable disease course. Prediction of the time to next relapse can also be useful in the design of clinical trials as an additional criterion for selecting active patients. For patients with clinically isolated syndrome (CIS), who have just experienced the first relapse, such a tool can be used for estimating the probability to convert to definite MS by predicting the time until the second relapse.
Biologically, analysis of genes and pathways that are related to predicting relapses may help to better understand the mechanisms underlying the progression of the disease, and more specifically the processes that trigger and operate in acute MS relapses.
Various demographic and disease-related variables have been utilized for predicting clinical outcome. Late age at disease onset, poly-symptomatic symptomatology at onset, higher annual relapse rate and short time-interval between attacks are correlated with poor outcome, while onset with the presentation of optic neuritis or sensory symptomatology have been associated with a good outcome [8–15]. Disease-related variables, measurements of autoantibodies, and gene expression were found to be useful for diagnosis and prognosis in MS and in other autoimmune diseases [16–18]. For example, in the case of MS, Martinez-Yelamos et al [19, 20] showed that CSF-TAU and 14-3 3 proteins are independent predictive factors for short time conversion to clinical definite MS. On the other hand, the correlation between anti-myelin antibodies and time to next relapse in CIS patients has produced controversial results [21–23].
Imaging is considered as a more sensitive tool for predicting MS progression; it was reported that various parameters like brain MRI lesion load including the number, volume and location of lesions, as well as the presence of enhancing lesions and brain atrophy can predict disease outcome . In CIS patients, T2 lesion volume at onset correlates with disability over the next 10 years, and with the time to progress to definite MS .
The possibility to use peripheral blood gene expression analysis for prediction of clinical outcome in MS patients was demonstrated in our previous work  where we showed that peripheral blood mononuclear cells (PBMC) gene expression based classifier correctly predicted disease progression for two years. Another relevant work is the work of Baranzini et al. , they showed that PBMC gene expression can be used to predict the response of MS patients to recombinant human interferon beta (rINFβ).
The aim of current study was to evaluate whether it is possible to use peripheral blood gene expression to predict the time to next acute relapse in CIS and relapsing-remitting MS patients.
Most of the new radiological MRI lesions are clinically silent. The frequency of new radiological lesions is ten times higher than the frequency of clinical relapses; i.e., on average, a cumulative effect of about 10 new radiological MRI lesions is equivalent to one clinical relapse . Since most clinical relapses are associated with new MRI lesions, and since MRI measurements was available only for small fraction of the patients, in the current study we focused only on clinically definite MS relapses.
We designed a comprehensive feature selection procedure that was implemented on different sets of feature including: 1) all genes represented in the microarray; 2) set of genes significantly discriminated between groups of different classes of time to next relapse; 3) genes significantly correlated with time to next relapse, and 4) clinical and demographical confounders.
This approach enabled us to identify a PBMC gene expression based predictors that envisioned the time until the next relapse with high accuracy.
The study was approved by the Sheba Medical Center Institutional Review Board, and all patients gave written informed consent for participation.
It was a prospective collection of data. The data set includes 94 patients, 62 patients diagnosed with definite MS according to McDonald criteria  and 32 patients with CIS. All patients were free of steroids and immunomodulatory treatments for at least 30 days before blood withdrawal, and were at least one year after treatment with cyclophosphamide (see Additional file 1). As can be seen, 62 patients had not experienced previous treatment; the rest of the patients had average distance of 300.4 ± 68.3 days from previous immunomodulatory treatment. The time to next relapse was not a selection criterion for inclusion patients in study; we randomly sampled the blood of 100 patients (94 microarrays passed quality control criteria).
We excluded from the study patients with Neuromylitis Optica (NMO) according to the criteria of Wingerchuk et al. .
Clinical and demographical characteristic of the analyzed patients.
Disease Duration (Years)
Annual Relapse Rate
Yes/No Future IMD Treatment
32.1 ± 1.5
0.20 ± 0.02
0.9 ± 0.2
38.5 ± 1.4
8.50 ± 1.09
0.92 ± 0.1
2.4 ± 0.2
The immunomodulatory treatments the analyzed patients underwent after blood sampling.
Non treated patients
Interferon β-1a Avonex
Interferon β-1b Betaferon
Interferon β-1a Rebif
Glatiramer acetate Copaxone
Intravenous Immunoglobulins Iv-Ig
No of Patients
As the aim of this study is the prediction of the time till the next relapse we gathered patients with large range of this parameter; our dataset included patients with long period between relapses. Some of these patients have benign MS (if they have EDSS < 3.0 after 15 years with the disease). Patients with benign MS are not necessarily treated in our country. Additionally, CIS patients (32 of the patients) are not treated according medical regulations in our country. Thus, in summery, our study included relatively high number untreated patients
Definition of Relapses
MS relapse was defined as the onset of new objective neurological symptoms/signs or worsening of existing neurological disability, not accompanied by metabolic changes, fever or other signs of infection, and lasting for a period of at least 48 hours accompanied by objective change of at least 0.5 in the EDSS score. Confirmed relapses and EDSS scores were recorded consecutively. Time from baseline gene expression analysis to next acute relapse was recorded and used as a variable for clinical outcome prediction.
RNA isolation and microarray expression profiling
The blood samples were collected for this analysis. After blood withdrawal, PBMC was immediately purified and frozen in liquid nitrogen for the future microarray analysis. Microarray analysis was performed each time that a large enough set of samples was collected (between 10–12 samples for a working set).
PBMC were separated on Ficol-Hypaque gradient, total RNA was purified, labelled, hybridized to a Genechip array
(either HU-133A or HU133A-2) and scanned (Hewlett Packard, GeneArray-TM scanner G2500A) according to the manufacturer's protocol (Affymetrix Inc, Santa Clara, CA, USA). All microarrays used in analysis passed all the stringent quality control criteria. The gene expression measurements used in this study are available and can be downloaded from Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/; accession number GSE15245).
Data analysis was performed by Partek Genomics Solution software http://www.partek.com. Expression values were computed from raw CEL files by applying the Robust Multi-Chip Average (RMA) background correction algorithm. The RMA correction included: 1) values background correction; 2) quantile normalization; 3) log2 transformation; 4) median polish summarization. In order to avoid the noise caused by variable set effects we normalized each set to pre-saved distribution pattern of a well balanced set used as a reference distribution.
To reduce batch effect, Analysis Of Variance (ANOVA) multiple model analysis was applied. Source of variation was analyzed; non-relevant batches effects such as array type, working batch, patient age and gender were eliminated. Most Informative Genes (MIGs) were defined as genes that distinguished between the different classes of time to next acute relapse with p < 0.01 by ANOVA test. Most Correlated Genes (MCGs) were defined as genes that were correlated with time to next acute relapse by Spearman, Pearson or Kendal method with p < 0.05.
Dividing the patients to classes
We divided the patients to three classes comprising at least 20 patients. Each class corresponded to relatively similar time ranges until their next relapse: a) 31 patients that had not experienced relapse in the 1264 days of the follow up period; b) 40 patients that experienced relapse in less than 500 days and c) 23 patients that experienced relapse in 500 to 1264 days. The boundaries of the classes were chosen according to various constrains: 1) the upper bound (1264) reflects that the patients' follow up period after blood withdrawal. 2) We wanted to divide the time rang 0–1264 to relatively similar ranges such that the number of patients in each group will be similar (at least 20 patients). Thus, we decided on the second boundary (500), a larger boundary would decrease the number of patients in the second group while a smaller one would result in a too small time range of the first group.
The distribution of subjects with CIS and Definite MS across these three categories is [30:16:15] for definite MS and [9: 5:14] for CIS for the three classes (less than 500 days: between 500 and 1264 days: more than 1264 days) respectively. As can be seen, in both groups there are subjects in each of the three categories.
In the case of the FTP, we used resolution of 50 days since we wanted to divide the range related to the first group to equally spaced sub-ranges (thus, for example, 30 or 90 days were inappropriate). It is important to note that the concept used here could be used with different resolutions and still give qualitatively similar results (for example lower resolution of 100 days gave error-rate of less than 0.2, p < 0.0001).
Implementations of the predictors
The predictor has two major parts: 1) a Support Vector Machine (SVM) classifier, that we named First Level Predictor (FLP), and 2) a multivariate linear regressor that we named Fine Tuning Predictor (FTP). We used the Spider http://www.kyb.tuebingen.mpg.de/bs/people/spider/ implementation for the SVM multi-class classifier. We used a multivariate linear regressor for the FTP (for patients with time till next relapse < 500). This multivariate linear predictor gave better results than SVM regressor (the Partek implementation of SVM regressor gave error rate that was much higher, around 0.8).
In the case of the FTP, the gene expression of each predictive gene is multiplied by a weight (positive or negative real number) and the results are summed. This sum is used as a prediction for the number of days till the next relapse. Thus, the expression level of each gene may have a positive or a negative affect on the prediction. In the case of the FLP, which is based on SVM classifier , the general idea is similar but more complicated – each gene may have a positive effect on the prediction in some cases and negative effect in other cases.
The feature selection procedure for finding the most predictive genes
We used different sets of genes as input to our feature selection procedure: a) genes that were differentially expressed between the three different classes of time till the next relapse, b) genes that were expressed with correlation (p < 0.05) with the time till the next relapse, c) all unique genes presented on the microarray.
Using a very large set of genes for constructing the predictors elevates the risk of overfitting (e.g. see ). The solution for overfitting problem is the Leave 20% Out Cross validation (L20OCV) procedure (a version of leave one out cross validation [33, 34]) that is described in this subsection.
We started with a set of 22,215 gene-transcripts present on each microarray. The expression levels of multiple probes related to the same gene were averaged, resulting in a set of 10,594 potential features (genes and annotated sequences). At the initial stage, we computed a vector of errors in 100 (80% training set, 20% test set) leave 20% out iterations for each gene. We chose the gene with the lowest mean error rate and all the genes whose mean error rate that is not significantly higher than the error rate of the first gene (p-value < 0.05 by t-test).
In the following stages, we expanded the initial sets while getting significantly better predictors which were based on 2, 3 or more genes. In each step, we tried to expand each predictor in the current set of best predictors by adding more genes to the predictors, and while keeping all the predictors whose error rates were not significantly worse than the error rate of the best predictor. This was done by an iterative cross validation procedure in which 20% of the initial data set was left apart and the remaining 80% were used as temporary training set in each of the iterations. We performed 100 iterations for the cross validation procedure at each stage for selecting sets of potential predictors. The output of the initial stage was a set of predictors with similar performances [i.e. according to the 100 leave 20% outs, none of them was significantly better than the best predictor; namely, all the p-values were > 0.05].
At the final stage, to discriminate between the predictors that have similar performances as the best predictor, we performed 10,000 cross validation iterations and selected the predictor(s) that was/were significantly better than the others.
Note that qualitatively similar results were gained when we learned the predictors based on 85 of the patients and afterward test them on a different group of 9 patients (the rest of the patients). In this case, the mean error rate of the FLPs was less than 0.11 on the training set and 0.13 on the test set (all p-values < 0.001). In the case of the FTPs, the mean error rates on the training set and test set were 0.44 and 0.66 respectively (all p-values < 0.001).
Evaluating the Performance of the Predictors on Subgroups of Patients
For evaluating the performances of the predictors on subgroups of the patients (e.g. MS vs. CIS, or non-treated vs. patients under the various treatments), we performed 1000 Leave 20% Out Cross Validation (L20OCV) procedures where in each L20OCV step we randomly chose subsets of 80% of each of the patient subgroups for training the predictors (all these subsets were unified to one subset), and tested the predictors on the rest of the samples (in the cases where the dataset was very small, at least one patient was chosen for the training set and for the test set). The predictors were based on the sets of genes that were found by the procedure that was described in the previous subsection, but in each iteration a different training set for inferring the weights of the different genes in the predictor was used; and in each iteration the predictors were implemented on different test set. The final error rate of each subgroup is the average error rate across all the 1000 L20OCV procedures for patients from the subgroup.
Another Set of Patients for Additional Validation
Clinical and demographical characteristic of the additional validation set of patients (4 CIS and 6 Definite MS).
Disease Duration (Years)
Annua Relapse Rate
Yes/No Future IMD Treatment
24 ± 5.01
0.34 ± 0.09
6.1 ± 2.05
2.58 ± 0.15
36 ± 7.61
5.3 ± 2.39
1 ± 0.51
5.3 ± 2.39
The Role of Clinical and Demographical Variables in the Predictors' Performance
We examined if clinical parameters are helpful for predicting the time to next relapse by 1) evaluating the performance of a FLP and FTP predictors that are based only on these parameters, and 2) examining if these variables can improve the performances of our predictors. To this end, we checked the following clinical parameters: age, MS stage (CIS or Definite), gender, annual relapse rate, EDSS at time of blood sampling, disease duration, age at onset, EDSS change in the last relapse.
Empirical p-values for the predictors
We computed empirically p-values for the best FLP and FTP by performing random permutation of the labels, learning best predictors for each such permutation, and computing the fraction of cases (out of 1000 permutations) that a predictor for permutated labels gave better error rate than the original predictor. All the 1000 random predictors were much worse than the original one (p-value < 0.001). For example, the average error rates of the random FLP were 0.67 and the average error rates of the random FTP were 0.8.
The Role of Relapse Severity in the Predictors' Performance
We examined if the performances of the predictors depend on the severity of the relapse (measured as the change in EDSS in the last relapse). To this end, we divided the patients to 8 groups according to their change in EDSS in the last relapse (the range was 0.5 – 6.5). We computed the error rate of the predictors for each of these groups (as was performed for the treatment groups).
Biological functional analysis
Gene functional annotation was performed using functional classification tools such as David Bioinformatics Resources http://david.abcc.ncifcrf.gov/home.jsp and Ingenuity Pathways Analysis web-software http://www.ingenuity.com. Information about the most predictive genes was extracted from NCBI http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene.
To learn about the three groups of patients, we performed Principal Component Analysis (PCA) and clustering analysis of the patients based on 1359 MIGs (Additional file 2). The patients with time until next relapse < 500 days exhibit a relatively coherent clustering where 29/40 of the patients appear in the same cluster. The clustering results of the other groups were much worse as they were partitioned among many clusters with up to 5 patients in the same cluster. This result demonstrates that it is much more complicated to cluster sub-groups of patients that will experience their next relapse in more than 500 days. Thus, an additional, finer predictor only for the first group (time until next relapse < 500) was justified.
We named the three groups classifier First Level Predictor (FLP). The FLP classified the time till next relapse of a patient to one of the following groups: <500 days, between 500 and 1264 days, and >1264 days. The finer predictor, for the patients with time till next relapse < 500 days was named Fine Tuning Predictor (FTP).
The performances (error rate) of the predictors in different subgroups of patients.
Interferon beta 1a Avonex
Interferon beta 1b Betaferon
Interferon beta 1a Rebif
Glatiramer acetate Copaxone
Intravenous Immunoglobulines Iv-Ig
Average error rate
The improvement in the error rate of the FLP as a function of the number of genes that were used by the predictor is demonstrated in Figure 2A. Genes were added iteratively to the FLP until there was no significant improvement in the performances of the predictor (a flow diagram appears in Figure 1).
The genes that were selected for the best FLPs (FLP1 – FLP3) and best FTPs (FTP1 – FTP4).
The performances of the predictor were significant (p < 0.001; details about the p-value calculation appear in the Methods section).
When we implemented the feature selection procedure using only MIG genes the performances of the result FLP were not improved (see Supplementary Note 1 in Additional file 5).
Next, we designed a more accurate predictor that was named Fine Tuning Predictor (FTP). It predicts the time until the next relapse only for patients that experience acute relapse during a period of 500 days. As a FTP we used a multivariate regressor (see Methods) that can predict the time until the next relapse with a resolution of a few days. In the case of the FTP, we defined a prediction error as a prediction that is more/less than 50 days (± 50) from the real date of relapse onset. We found 240 gene sets that gave error rate < 0.36. Our feature selection procedure combined with 10000 permutations of Leave One 20% Cross Validation (L20OCV) procedure found four FTP s; each FTP was based on 9 genes. The minimal error-rate of each FTP was 0.35 (p-value < 0.001); and was significantly better than the other gene sets. The error rate of the FTP after random permutations of the labels was 0.8; this is 2.3 folds higher than the error rate of the inferred FTP (see Methods for description about the p-value). The error rates of best 9-genes-FTPs are demonstrated in Table 4.
As in the case of the FLP, we performed a similar analysis of the improvement in the error rate of the best FTP (see Table 4) as function of the number of predictive genes (from 1 to 9; Figure 2C). Every time a gene was added to the FTP, the performances of the FTP were significantly improved (see Methods).
When we implemented the feature selection procedure only on genes whose expression is correlative with time until next relapse the FTP results were not improved (see Supplementary Note 2 in Additional file 5).
To further evaluate the performances of the predictors, we divided the patients in two ways. First, the dataset was divided to CIS and definite MS patients (32 and 62 patients respectively); second, the dataset was divided into 6 groups according to their future treatment after their blood was withdrawn (Non-treated, Interferon β-1a (Avonex), Interferon β-1b (Betaferon), Interferon β-1a (Rebif), Glatiramer Acetate (Copaxone) and Intravenous Immunoglobulins (Iv-Ig) with 61, 5, 2, 10, 10, and 6 patients respectively (see Table 2). The performances of the best predictors in different subgroups of patients were evaluated as was described in the Methods section.
Table 4 depicts the performances of each of the best predictors (FLPs and FTPs) on each of these groups. The error rates remain significant for each of these groups. The FLP error rates were 0.1 and 0.08 (p-value < 0.001) for the CIS and MS groups respectively; the FTP error rates were 0.21 and 0.355 (p value < 0.001), for the CIS and MS groups respectively.
In the analysis of the different treatments, the FLP error rate ranged between 0.087 (non-treated) to 0.18 (Copaxone); the FTP error rate ranged between 0.235 (Rebif) and 0.53 (Copaxone), see Table 4. The probabilities of the different types of classification errors for each treatment group appear in Additional file 7. In all sub-groups the best predictor exhibited significant performances (p-value < 0.001); this fact suggests that the signal of the next relapse is usually strong enough to be detected by our predictor even when the patients undergo various Immunomodulatory Drugs (IMD) treatments after blood withdrawal.
For additional validation of the FLP, we collected an additional independent dataset of 10 untreated RRMS patients (see more details about this dataset in the Methods section and Table 3). The total error rate of the FLP on these patients was 0.25. This result further supports the viability of our approach.
Clinical and demographical confounders did not improve the performance of the best gene expression based FLP and FTP (see Methods for the exact list of Clinical and demographical confounders we checked). FLP and FTP predictors that were based on combination of these confounders and that were found by our approach had much higher error rates than the predictors that were based only on gene expression (FLP error rate > 0.63, FTP error rate > 0.74; for all the predictors that were based on the above mentioned confounders). We also did not get a significant correlation between the error rate of a predictor and the severity of relapse (as was measured by the change in EDSS), Spearman correlation -0.29, p-value = 0.5, for FTP; Spearman correlation -0.6, p-value = 0.12, for FLP, see Methods for more details).
In this work, we demonstrated that gene expression in PBMC can be used for predicting the time of the upcoming relapse in MS. Using different prediction strategies to determine an appropriated gene set for accurate relapse prediction we found that the classifier that was based on all microarray genes had the best prediction. We describe a FLP, which is based on the expression levels of ten genes, which can predict the time till next relapse in a resolution of 500 days during 3.5 years of disease progression. An additional FTP, which is based on different set of nine genes, can be used for a prediction of a higher resolution (e.g. a resolution of 50 days).
At first glance, the error rate (about 0.35) of the FTP seems relatively high. However it is important to remember that the definition of error in this case was very tight (more than 50 days from the real value). As mentioned, this error rate was very significant (p-value < 0.001) and surprisingly good. For comparison, our simulation showed that a random FTP (i.e. the best FTP after a random shuffling the input labels) gave an error rate that was close to 0.8 – 2.3 fold higher than our error rate.
This study includes 94 patients for evaluating the predictors. It is clear that a larger dataset will give better performances. In order to estimate the potential improvement in the error rate when using larger datasets we performed the following analysis: we computed the error rate of the FLP and the FTP as function of the dataset size (% of the original dataset; see Additional file 8). The figure shows that the error rates decrease for larger datasets. This fact suggests that with the accumulation of more gene expression measurements we can design better predictors. Specifically, enlargement of the dataset to 200 patients (instead of 94) will give a classification error of about 0.05 and a regression error of about 0.2 (Additional file 8).
Another interesting conclusion from this work is that there are multiple predictors (FLPs and FTPs) that have similar performances. The predictors that were described in this work were significantly better than the other predictors; however, there were a few dozen predictors that gave similar results. For example, Additional file 3 includes 36 FLP with error rate < 0.1, and 240 FTP with error rate < 0.36. This means that the best predictors appear in this work can be replaced by other predictors with a relatively small influence on the error rate.
Finding a good predictor for the time to next relapse and finding a molecular explanation for relapses are different tasks with a possible overlap. There are a few explanations why the connection between the predictive genes and relapse associated mechanism is not necessarily immediate: First, the predictors were designed to include relatively small number of genes while the actual mechanisms may include dozens of signaling pathways. Second, our study was based on changes in gene transcription levels; it is possible that major parts of the relapse associated regulatory mechanisms are post transcriptional. In such cases, the most relevant genes are useless in terms of improving the predictions and the feature selection procedure finds genes that are less relevant but that exhibit significant change in their mRNA levels (e.g. genes that are regulated/regulate the genes that are directly related to relapse).
However, many of the FLP genes are linked to MS. This is an additional support of their predictive ability. For example, the gene TGFB2 is a one of the master genes in MS; it is closely related to a rapid recovery from relapses that is mediated by Th2/Th3 lymphocytes. Th2/Th3 lymphocytes produce anti-inflammatory cytokines (like IL10) . TGFB inhibits IL12 mediated inflammatory response, and it virtually decreases T cells proliferation and IFNgamma production . TGFB prevents induction of pro-inflammatory gene-program by inhibiting the expression of 25% of the TNFalpha/IFNgamma induced genes . The target genes that TGFB inhibits are various genes that are involved in MS pathogenesis processes (e.g. chemotaxis, adhesion and cell migration).
The gene for Familial Mediterranean Fever (MEFV) is expressed in early leukocyte development and is regulated in response to inflammatory mediators. Stimulation of cells with the proinflammatory agents interferon (IFN) gamma, tumor necrosis factor, and lipopolysaccharide induced MEFV expression, whereas the anti inflammatory cytokines (IL4, IL-10) and especially TFGB inhibited such expression .
The CA2 gene (carbonic anhydrase II) supports the transport of bicarbonate ions, sodium ions, and water from blood to the CSF; and in the myelin sheath CA2 supports compaction of myelin by stimulating co-transport of ions between the myelin membranes. The double mutant mice deficient by CA2 and myelin displayed tremors and seizures . Interestingly, the onset of seizures was delayed significantly in the double mutants, and the lifespan increased by several months, this fact corroborates with CA2's activity as predictor of acute onset in MS.
Another important group of genes that are part of the best FLP is related to the interferon regulation mechanism. This group includes RNA binding and signal transduction SLM1 gene, that is associated with Interferon Receptor 1 Binding Protein 4 (IR1B4)  and TRIM22, an important member of interferon related genes, that is involved in transduction of IFN activity .
The best FTP includes 2 inflammatory related genes (FPRL1 and PPFIA1). FPRL1 functions as a receptor component of inflammatory response , activation of FPRL1 results in leukocytes activation. FPRL1 involved in direct monocytes/microglia migration as was demonstrated in activated brain prion plaques and brain lesions in Alzheimer disease . PPFIA1 receptor gene is involved in cell motility, cells spreading, migration and adhesion. Up regulation of human PPFIA1 (LIPRIN) gene in peripheral blood is associated with psoriatic arthritis . Another interesting gene is G3BP that encodes a downstream effector protein of the Ras signaling pathway . Interestingly, as in the case of best FLP, the genes FPRL1 and G3BP are regulated by TGFB [49, 50]. This result suggests that a single pathway can explain both the FLP and the FTP genes.
To summarize, most of the predictive genes seem very relevant to the pathophysiology of MS. We have constructed a schematic regulatory network which unifies many of these genes to a single regulatory network (Figure 6). One important goal for further research is to better understand how these genes are involved in the biological mechanisms that lead to a clinical relapse.
We also found a few predictive genes whose their biological roles are unknown (see Additional file 4). We thus think that the potential connection of these genes to MS is a natural target for a further study.
As relapses have different level of severity [e.g. it can be measured by the increase in EDSS] one may think that in the cases of a more severe next relapse the performances of the predictors will be better. Analysis of the predictor error probability for the FLP and FTP as a function of the change in EDSS levels showed a negative relation (as expected – i.e. larger changes in EDSS are easier to detect). However, this correlation was not significant. Thus, a final answer to this question should be deferred till a larger gene expression dataset will be accumulated.
We used a heterogeneous dataset for inferring the predictors. The dataset included both CIS and Definite MS patients, and patients that underwent different immunomodulatory treatments after their blood was sampled. A survival analysis showed that the disease stage (CIS or Definite MS) had a statistical significant influence on the probability to experience next acute relapse (Additional file 9, Supplementary Note 3 in Additional file 5). Further, statistical analysis of the gene expression of these two groups of patients showed that there are dozens of genes that are differentially expressed in these groups (data not shown).
However, our predictor was insensitive to the disease stage and successfully dealt with this issue. The error rates for the two groups were significantly low (less than 0.1 for the FLP and less than 0.37 for the FTP). This fact may suggest that, in the case of the small sets of predictive genes, the changes in PBMC gene expression before the second relapse (CIS patients) or before any other relapse (Definite MS patients) are similar.
Our dataset includes patients that underwent various treatments after we sampled their gene expression. We believe that these patients are the major source of error for our predictors. On the other hand, we decided to include them in the analysis since they improved its statistical significance. We demonstrated that our predictor gave significantly good results, also when considering each of these datasets separately. This was unexpected since it is known that in general drugs change the relapse frequency. The explanation of this result is simple: Most of the treatments delay the next relapse by about 30% [51–55], and this fact increases the prediction error primarily for patients whose real time to next relapse is close to the boundaries of their classification group (e.g. close to 500 or close to 1264). Since the number of such patients is relatively low the error rate remains significantly low (see details in Supplementary Note 4 in Additional file 5).
We conclude that gene expression in PBMC can be used to accurately predict the time until the next acute relapse. In this work, we described a few sets of predictive genes that can be used for this purpose and demonstrated that other combinations may also yield significant results. It is possible that different technology for measuring the gene expression will yield different sets of the most predictive genes. Thus, our next goal is to find sets of predictive genes that give significant results when their gene expression is measured by cheaper, small-scale, technologies such as kinetic RT-PCR.
In this work, as information about clinically silent lesions was not available for most of the patients, we focused only on clinically definite MS relapses. In the future, when such information will be available, it can be used for improving the performances of our predictor. In addition, based on such data it will be feasible to study the possibility to predict radiological MRI lesions (that are possibly clinically silent) from gene expression in PBMC.
Finally, it is possible that the techniques described here will be valuable not only in future MS research but also in other autoimmune disease with relapsing-remitting nature.
Clinically Isolated Syndrome
Peripheral Blood Mononuclear Cells
Most Informative Genes
Most Correlated Genes
Support Vector Machine
Leave 20% Out Cross Validation
First Level Predictor
Fine Tuning Predictor
Relapsing Remitting MS
Expanded Disability Status Scale
Robust Multi-chip Analysis
Analysis of Variance
T.T. was supported by the Edmond J. Safra Bioinformatics program at Tel Aviv University and the Yeshaya Horowitz Association through the Center for Complexity Science.
- Conti-Fine BM, Milani M, Kaminski HJ: Myasthenia gravis: past, present, and future. J Clin Invest. 2006, 116 (11): 2843-2854.View ArticlePubMedPubMed CentralGoogle Scholar
- Hutchinson M, Bresnihan B: Neurological lupus erythematosus with tonic seizures simulating multiple sclerosis. J Neurol Neurosurg Psychiatry. 1983, 46 (6): 583-585.View ArticlePubMedPubMed CentralGoogle Scholar
- Majithia V, Geraci SA: Rheumatoid arthritis: diagnosis and management. Am J Med. 2007, 120 (11): 936-939.View ArticlePubMedGoogle Scholar
- Hanauer SB: Inflammatory bowel disease. N Engl J Med. 1996, 334 (13): 841-848.View ArticlePubMedGoogle Scholar
- McDonald E: Multiple sclerosis: common management issues. Aust Fam Physician. 1992, 21 (10): 1421-1424.PubMedGoogle Scholar
- McDonald WI, Ron MA: Multiple sclerosis: the disease and its manifestations. Philos Trans R Soc Lond B Biol Sci. 1999, 354 (1390): 1615-1622.View ArticlePubMedPubMed CentralGoogle Scholar
- McDonald WI: Relapse, remission, and progression in multiple sclerosis. N Engl J Med. 2000, 343 (20): 1486-1487.View ArticlePubMedGoogle Scholar
- Scott TF, Schramke CJ, Novero J, Chieffe C: Short-term prognosis in early relapsing-remitting multiple sclerosis. Neurology. 2000, 55 (5): 689-693.View ArticlePubMedGoogle Scholar
- Held U, Heigenhauser L, Shang C, Kappos L, Polman C: Predictors of relapse rate in MS clinical trials. Neurology. 2005, 65 (11): 1769-1773.View ArticlePubMedGoogle Scholar
- Bergamaschi R, Berzuini C, Romani A, Cosi V: Predicting secondary progression in relapsing-remitting multiple sclerosis: a Bayesian analysis. J Neurol Sci. 2001, 189 (1–2): 13-21.View ArticlePubMedGoogle Scholar
- Bergamaschi R: Prognosis of multiple sclerosis: clinical factors predicting the late evolution for an early treatment decision. Expert Rev Neurother. 2006, 6 (3): 357-364.View ArticlePubMedGoogle Scholar
- Bergamaschi R, Quaglini S, Trojano M, Amato MP, Tavazzi E, Paolicelli D, Zipoli V, Romani A, Fuiani A, Portaccio E, et al: Early prediction of the long term evolution of multiple sclerosis: the Bayesian Risk Estimate for Multiple Sclerosis (BREMS) score. J Neurol Neurosurg Psychiatry. 2007, 78 (7): 757-759.View ArticlePubMedPubMed CentralGoogle Scholar
- Weinshenker BG, Ebers GC: The natural history of multiple sclerosis. Can J Neurol Sci. 1987, 14 (3): 255-261.PubMedGoogle Scholar
- Phadke JG: Clinical aspects of multiple sclerosis in north-east Scotland with particular reference to its course and prognosis. Brain. 1990, 113 (Pt 6): 1597-1628.View ArticlePubMedGoogle Scholar
- Runmarker B, Andersson C, Oden A, Andersen O: Prediction of outcome in multiple sclerosis based on multivariate models. J Neurol. 1994, 241 (10): 597-604.View ArticlePubMedGoogle Scholar
- Allantaz F, Chaussabel D, Stichweh D, Bennett L, Allman W, Mejias A, Ardura M, Chung W, Wise C, Palucka K, et al: Blood leukocyte microarrays to diagnose systemic onset juvenile idiopathic arthritis and follow the response to IL-1 blockade. J Exp Med. 2007, 204 (9): 2131-2144.View ArticlePubMedPubMed CentralGoogle Scholar
- Oken O, Batur G, Gunduz R, Yorgancioglu RZ: Factors associated with functional disability in patients with rheumatoid arthritis. Rheumatol Int. 2008, 29 (2): 163-166.View ArticlePubMedGoogle Scholar
- Vojdani A: Antibodies as predictors of complex autoimmune diseases. Int J Immunopathol Pharmacol. 2008, 21 (2): 267-278.PubMedGoogle Scholar
- Martinez-Yelamos A, Rovira A, Sanchez-Valle R, Martinez-Yelamos S, Tintore M, Blanco Y, Graus F, Montalban X, Arbizu T, Saiz A: CSF 14-3-3 protein assay and MRI as prognostic markers in patients with a clinically isolated syndrome suggestive of MS. J Neurol. 2004, 251 (10): 1278-1279.View ArticlePubMedGoogle Scholar
- Martinez-Yelamos A, Saiz A, Bas J, Hernandez JJ, Graus F, Arbizu T: Tau protein in cerebrospinal fluid: a possible marker of poor outcome in patients with early relapsing-remitting multiple sclerosis. Neurosci Lett. 2004, 363 (1): 14-17.View ArticlePubMedGoogle Scholar
- Berger T, Rubner P, Schautzer F, Egg R, Ulmer H, Mayringer I, Dilitz E, Deisenhammer F, Reindl M: Antimyelin antibodies as a predictor of clinically definite multiple sclerosis after a first demyelinating event. N Engl J Med. 2003, 349 (2): 139-145.View ArticlePubMedGoogle Scholar
- Berger T, Reindl M: Lack of association between antimyelin antibodies and progression to multiple sclerosis. N Engl J Med. 2007, 356 (18): 1888-1889.View ArticlePubMedGoogle Scholar
- Kuhle J, Lindberg RL, Regeniter A, Mehling M, Hoffmann F, Reindl M, Berger T, Radue EW, Leppert D, Kappos L: Antimyelin antibodies in clinically isolated syndromes correlate with inflammation in MRI and CSF. J Neurol. 2007, 254 (2): 160-168.View ArticlePubMedGoogle Scholar
- Bergamaschi R: Prognostic factors in multiple sclerosis. Int Rev Neurobiol. 2007, 79: 423-447.View ArticlePubMedGoogle Scholar
- Sailer M, O'Riordan JI, Thompson AJ, Kingsley DP, MacManus DG, McDonald WI, Miller DH: Quantitative MRI in patients with clinically isolated syndromes suggestive of demyelination. Neurology. 1999, 52 (3): 599-606.View ArticlePubMedGoogle Scholar
- Achiron A, Gurevich M, Snir Y, Segal E, Mandel M: Zinc-ion binding and cytokine activity regulation pathways predicts outcome in relapsing-remitting multiple sclerosis. Clin Exp Immunol. 2007, 149 (2): 235-242.View ArticlePubMedPubMed CentralGoogle Scholar
- Baranzini SE, Mousavi P, Rio J, Caillier SJ, Stillman A, Villoslada P, Wyatt MM, Comabella M, Greller LD, Somogyi R, et al: Transcription-based prediction of response to IFNbeta using supervised computational methods. PLoS Biol. 2005, 3 (1): e2.View ArticlePubMedGoogle Scholar
- Filippi M, De Stefano N, Dousset V, McGowan JCE: MR imaging in white matter diseases of the brain and spinal cord. 2005, Berlin: Springer, 219.View ArticleGoogle Scholar
- McDonald WI, Compston A, Edan G, Goodkin D, Hartung HP, Lublin FD, McFarland HF, Paty DW, Polman CH, Reingold SC, et al: Recommended diagnostic criteria for multiple sclerosis: guidelines from the International Panel on the diagnosis of multiple sclerosis. Ann Neurol. 2001, 50 (1): 121-127.View ArticlePubMedGoogle Scholar
- Wingerchuk DM, Lennon VA, Pittock SJ, Lucchinetti CF, Weinshenker BG: Revised diagnostic criteria for neuromyelitis optica. Neurology. 2006, 66 (10): 1485-1489.View ArticlePubMedGoogle Scholar
- Burges CJC: A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery. 1998, 2: 121-167.View ArticleGoogle Scholar
- Dietterich T: Overfitting and undercomputing in machine learning. ACM Computing Surveys. 1995, 27 (3): 326-327.View ArticleGoogle Scholar
- Efron B: Estimating the error rate of a prediction rule improvement on cross-validation. Journal of the American Statistical Association. 1983, 78 (382): 316-330.View ArticleGoogle Scholar
- Geisser S: Predictive Inference: An Introduction. 1993, New York: CRC PressView ArticleGoogle Scholar
- Paglinawan R, Malipiero U, Schlapbach R, Frei K, Reith W, Fontana A: TGFbeta directs gene expression of activated microglia to an anti-inflammatory phenotype strongly focusing on chemokine genes and cell migratory genes. Glia. 2003, 44 (3): 219-231.View ArticlePubMedGoogle Scholar
- Bright JJ, Sriram S: TGF-beta inhibits IL-12-induced activation of Jak-STAT pathway in T lymphocytes. J Immunol. 1998, 161 (4): 1772-1777.PubMedGoogle Scholar
- Centola M, Wood G, Frucht DM, Galon J, Aringer M, Farrell C, Kingma DW, Horwitz ME, Mansfield E, Holland SM, et al: The gene for familial Mediterranean fever, MEFV, is expressed in early leukocyte development and is regulated in response to inflammatory mediators. Blood. 2000, 95 (10): 3223-3231.PubMedGoogle Scholar
- Cammer W, Zhang H, Tansey FA: Effects of carbonic anhydrase II (CAII) deficiency on CNS structure and function in the myelin-deficient CAII-deficient double mutant mouse. J Neurosci Res. 1995, 40 (4): 451-457.View ArticlePubMedGoogle Scholar
- Cote J, Boisvert FM, Boulanger MC, Bedford MT, Richard S: Sam68 RNA binding protein is an in vivo substrate for protein arginine N-methyltransferase 1. Mol Biol Cell. 2003, 14 (1): 274-287.View ArticlePubMedPubMed CentralGoogle Scholar
- Tissot C, Mechti N: Molecular cloning of a new interferon-induced factor that represses human immunodeficiency virus type 1 long terminal repeat expression. J Biol Chem. 1995, 270 (25): 14891-14898.View ArticlePubMedGoogle Scholar
- Andrianifahanana M, Agrawal A, Singh AP, Moniaux N, van Seuningen I, Aubert JP, Meza J, Batra SK: Synergistic induction of the MUC4 mucin gene by interferon-gamma and retinoic acid in human pancreatic tumour cells involves a reprogramming of signalling pathways. Oncogene. 2005, 24 (40): 6143-6154.View ArticlePubMedGoogle Scholar
- Baron BW, Anastasi J, Thirman MJ, Furukawa Y, Fears S, Kim DC, Simone F, Birkenbach M, Montag A, Sadhu A, et al: The human programmed cell death-2 (PDCD2) gene is a target of BCL6 repression: implications for a role of BCL6 in the down-regulation of apoptosis. Proc Natl Acad Sci USA. 2002, 99 (5): 2860-2865.View ArticlePubMedPubMed CentralGoogle Scholar
- Bouwmeester T, Bauch A, Ruffner H, Angrand PO, Bergamini G, Croughton K, Cruciat C, Eberhard D, Gagneur J, Ghidelli S, et al: A physical and functional map of the human TNF-alpha/NF-kappa B signal transduction pathway. Nat Cell Biol. 2004, 6 (2): 97-105.View ArticlePubMedGoogle Scholar
- Deb A, Haque SJ, Mogensen T, Silverman RH, Williams BR: RNA-dependent protein kinase PKR is required for activation of NF-kappa B by IFN-gamma in a STAT1-independent pathway. J Immunol. 2001, 166 (10): 6170-6180.View ArticlePubMedGoogle Scholar
- Klein C, Paul JI, Sauve K, Schmidt MM, Arcangeli L, Ransom J, Trueheart J, Manfredi JP, Broach JR, Murphy AJ: Identification of surrogate agonists for the human FPRL-1 receptor by autocrine selection in yeast. Nature biotechnology. 1998, 16 (13): 1334-1337.View ArticlePubMedGoogle Scholar
- Le Y, Yazawa H, Gong W, Yu Z, Ferrans VJ, Murphy PM, Wang JM: The neurotoxic prion peptide fragment PrP(106–126) is a chemotactic agonist for the G protein-coupled receptor formyl peptide receptor-like 1. J Immunol. 2001, 166 (3): 1448-1451.View ArticlePubMedGoogle Scholar
- Batliwalla FM, Li W, Ritchlin CT, Xiao X, Brenner M, Laragione T, Shao T, Durham R, Kemshetti S, Schwarz E, et al: Microarray analyses of peripheral blood cells identifies unique gene expression signature in psoriatic arthritis. Molecular medicine (Cambridge, Mass). 2005, 11 (1–12): 21-29.Google Scholar
- Barnes CJ, Li F, Mandal M, Yang Z, Sahin AA, Kumar R: Heregulin induces expression, ATPase activity, and nuclear localization of G3BP, a Ras signaling component, in human breast tumors. Cancer research. 2002, 62 (5): 1251-1255.PubMedGoogle Scholar
- Le Y, Iribarren P, Gong W, Cui Y, Zhang X, Wang JM: TGF-beta1 disrupts endotoxin signaling in microglial cells through Smad3 and MAPK pathways. J Immunol. 2004, 173 (2): 962-968.View ArticlePubMedGoogle Scholar
- Stasyk T, Dubrovska A, Lomnytska M, Yakymovych I, Wernstedt C, Heldin CH, Hellman U, Souchelnytskyi S: Phosphoproteome profiling of transforming growth factor (TGF)-beta signaling: abrogation of TGFbeta1-dependent phosphorylation of transcription factor-II-I (TFII-I) enhances cooperation of TFII-I and Smad3 in transcription. Mol Biol Cell. 2005, 16 (10): 4765-4780.View ArticlePubMedPubMed CentralGoogle Scholar
- IFNB Multiple Sclerosis Study Group E-U: Interferon beta-1b is effective in relapsing-remitting multiple sclerosis. I. Clinical results of a multicenter, randomized, double-blind, placebo-controlled trial. The IFNB Multiple Sclerosis Study Group. Neurology. 1993, 43 (4): 655-661.View ArticleGoogle Scholar
- (Commanditaire) IMSG, University of British Columbia MS/MRI Analysis Group E-UC: Interferon beta-1b in the treatment of multiple sclerosis: final outcome of the randomized controlled trial. The IFNB Multiple Sclerosis Study Group and The University of British Columbia MS/MRI Analysis Group. Neurology. 1995, 45 (7): 1277-1285.View ArticleGoogle Scholar
- Group PPoRaDbIb-aSiMSS: Randomised double-blind placebo-controlled study of interferon beta-1a in relapsing/remitting multiple sclerosis. Lancet. 1998, 352 (9139): 1498-1504.View ArticleGoogle Scholar
- Group UMMS, IFNB Multiple Sclerosis Study Group E-U: PRISMS-4: Long-term efficacy of interferon-beta-1a in relapsing MS. Neurology. 2001, 56 (12): 1628-1636.View ArticleGoogle Scholar
- Jacobs LD, Cookfair DL, Rudick RA, Herndon RM, Richert JR, Salazar AM, Fischer JS, Goodkin DE, Granger CV, Simon JH, et al: Intramuscular interferon beta-1a for disease progression in relapsing multiple sclerosis. The Multiple Sclerosis Collaborative Research Group (MSCRG). Ann Neurol. 1996, 39 (3): 285-294.View ArticlePubMedGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1755-8794/2/46/prepub