Deciphering the Role of the Gut Microbiota in Multiple Sclerosis

Description of methods and plan of operation

Our research plan consists of the following steps:

1. Cohort assembly. For each of the above aims the investigators will use the unique database of the Sheba Medical Center to identify the relevant individuals and invite them to take part in the study. Prof. Achiron has much experience in conducting many research projects that utilize the unique patient database available to the Center. For the first aim comparing MS patients to healthy individuals the investigators will select sex-, age- and diet-matched healthy individuals, ideally selecting spouses of MS patients as healthy controls as individuals living in the same environment have more similar microbiota. In our second aim comparing MS patients with similar time from diagnosis but different disease severity, the investigators will select MS patients that span the largest possible spectrum of disease severity as judged by the EDSS score employed by the Center. For the final aim individuals at high risk of relapse will be invited for profiling every 6 months and if relapse occurs, they will be profiled upon their visit to the Center as well as one month after the relapse event.
2. Cohort profiling. From each patient, the investigators will obtain a multi-dimensional data from the MS database consisting, as appropriate, of a subset of: (1) Clinical metadata, including: Consent form; Medications; annual relapse rate; (2) Blood tests, including a complete blood count, complete biochemistry, lipid profile, cholesterol profile; (3) Complete neurological examination for obtaining an EDSS score, cognitive assessment, gait assessment; MRI imaging data, evoked potentials, treatment response; (4) Blood samples will be processed for protein mRNA expression and peripheral blood mononuclear cells (PBMCs) will be separated on Ficoll-Hypaque gradient, total RNA purified, labeled, hybridized to Genechip array (U133A2), and scanned (GeneArray-TM scanner G2500A; Hewlett Packard) according to the manufacturer's protocol (Affymetrix, Santa Clara, CA). MAS5 software (Affymetrix) will be used to analyze the scanned arrays containing ~22,000 gene transcripts corresponding to 14,500 well-annotated human genes. (5) Gut microbiota profile obtained from stool samples will be processed for shotgun metagenomic sequencing and 16S rRNA profiling. Gut microbiota profiling will be done from stool samples that will be immediately flash-frozen in liquid nitrogen and preserved at a minimum of -80°C until further processing. Samples will then be processed by an automated robotic pipeline that was developed in the Segal lab at Weizmann. This pipeline works in 96-well format and can extract DNA from 96 stool samples within one day, prepare DNA Illumina libraries for shotgun metagenomic sequencing within another day, and carry out multiplexed polymerase chain reaction (PCR) amplification of the 16S rRNA gene in another day. Thus, every 96-stool sample group collected can be processed robotically for both 16S and metagenomic sequencing within 3 days under the supervision of one lab technician.
3. Data analysis and algorithmic development. (I) Microbiota: To comprehensively study the role of the microbiome in MS, the investigators will go much beyond the standard 16S rRNA analysis and into analysis of full shotgun metagenome samples. By sequencing the entire DNA content of stool samples, metagenome sequencing can potentially provide much more information as compared to 16S, as it allows to study genome structure, structural variants, and gene and metabolic pathway functions. After extracting these features from the microbiome (see below in Preliminary Results), the investigators will start by employing basic univariate and multivariate association tests, and continue with more complex machine learning models that attempt to distinguish individuals with MS from those without based on microbiome features (aim 1), to classify disease severity (aim 2), to predict relapse risk (aim 3), to differentiate between MS disease phenotypes i.e., radiologically isolated syndrome (RIS), clinically isolated syndrome (CIS), relapsing-remitting MS (RRMS), primary-progressive MS (PPMS), (aim 4), and to identify treatment responders (aim 5). (II) Blood: To analyze protein expression Partek Genomics Software (www.partek.com) will be used.
4. Univariate and multivariate analyses. The investigators will first compute the correlation (Pearson and Spearman) between all microbiome features extracted across all profiled individuals and the different patient measurements (EDSS score, time from relapse, etc.), and correct for the multiple hypotheses performed. Since the investigators will generate a vast number of microbiome features and many of them are highly correlated to each other, this analysis may suffer from lack of statistical power, especially given that the number of participants will be far smaller than the number of features. For this reason, the investigators will also perform multivariate analyses (e.g., singular value decomposition, principal component analysis) since the key components identified by these methods capture the main variation in the data in a way that takes into account the internal structure and relationships between the different input features. The investigators will then test whether projections of the data by any of the main principal components in this analysis provides a significant segregation of the participants by their measured metabolic parameters. As a different type of multivariate analysis, the investigators will also employ different unsupervised clustering methods (e.g., hierarchical clustering, naïve Bayes) to cluster the participants by their microbiome feature data, and then examine the clusters for enrichment in normal or abnormal metabolic parameters.

Machine learning algorithms. As a more global approach aimed at quantifying the overall contribution of the microbiome to MS and at unraveling the relative contribution of the different microbiome features, the investigators will classify the study participants into several groups in each aim (e.g., in aim 1 patients versus healthy individuals; in aim 2 individuals with high versus low EDSS score for the similar time from MS diagnosis), and develop different computational methods (e.g., boosted decision trees, Support Vector Machine algorithms (SVMs)) for this classification problem using only the microbiome features generated above. The investigators will use a cross validation scheme, whereby the model training is done on the data of a randomly chosen subset of participants and then tested on the data of the remaining held out participants. In addition, the investigators will leave aside a test set on which the investigators will evaluate the final model that is derived in cross validation, allowing a true estimate of the performance of our models. As the number of microbiome features and thus the number of dimensions is large, the investigators will employ various feature selection approaches as means of avoiding overfitting and reducing dimensionality. The Segal lab (Weizmann) has pioneered the development of several such methods in similar settings in the area of gene regulation. The investigators will also use a similar scheme to predict the continuous EDSS score representing MS severity. The problem setup is similar to classification, but the method development is quite different as the classification methods are replaced with regression type of methods (e.g., linear regression, probabilistic models, stochastic gradient descent).