Combining Clinical and Genetic Data to Predict Response to Fingolimod Treatment in Relapsing Remitting Multiple Sclerosis Patients: A Precision Medicine Approach

Multiple sclerosis (MS) is a chronic inflammatory disease of the central nervous system characterized by clinical heterogeneity and varied treatment responses. The advent of numerous therapies for relapsing-remitting MS (RRMS) necessitates a more personalized approach to treatment. While genetic factors influencing MS susceptibility are well-established, their role in determining disease severity and treatment response remains less clear, particularly concerning second-line drugs like fingolimod (FTY). FTY is a highly effective treatment for RRMS; however, some patients show persistent disease activity, making early identification of non-responders crucial for switching to alternative therapies. Machine learning (ML) algorithms, capable of handling complex relationships between variables, offer a promising approach to precision medicine in MS. This study aimed to use ML methods to identify a genetic signature that, when combined with clinical and demographic characteristics, can predict response to FTY treatment in RRMS patients. The hypothesis was that a combination of genetic and clinical features could improve the prediction of fingolimod treatment response compared to using clinical features alone.

Extensive genome-wide association studies (GWAS) have identified genetic risk factors for MS. However, smaller studies have yielded inconsistent results regarding the influence of genetics on disease severity and response to first-line therapies like interferon and glatiramer acetate. Even fewer studies have explored genetic markers associated with response to second-line drugs. Notably, there is a lack of research on genetic factors associated with fingolimod response. Existing research primarily focuses on clinical predictors of response, and the potential for integrating genetic information for improved prediction is largely unexplored. Several studies have utilized machine learning for various applications in MS, such as MRI analysis and disease course prediction, but its application to genetic data to predict treatment response remains limited. This gap in the literature provides the strong rationale for the present study.

This study employed a two-cohort design, comprising 364 RRMS patients treated with FTY at the San Raffaele Hospital in Milan, Italy (OSR cohort) and 108 patients from the Centre Hospitalier Universitaire de Toulouse, France (CHUT cohort). Baseline and 2-year follow-up clinical data were collected, including demographics, disease history, relapse rate, EDSS scores, and MRI findings. Patients treated with natalizumab within 9 months prior to FTY initiation were excluded. Treatment response was assessed using the NEDA-3 criterion (No Evidence of Disease Activity). Genotyping was performed using the Illumina HumanOmniExpress Kit. Stringent quality control measures were applied to the genetic data using PLINK. Linkage disequilibrium (LD) pruning reduced redundancy among SNPs. The combined cohort was divided into training (TRset), validation (Vset), and test (TEset) subsets (40%, 40%, 20% respectively). Random Forests (RFs) were chosen as the ML algorithm due to their superior performance compared to other algorithms in preliminary analyses. A robust cross-validated feature selection method was used to identify SNPs associated with treatment response. The Vset was used for SNP signature selection to mitigate bias. The optimal RF parameters were selected using the validation set. A similar process was followed for the clinical model. The clinical and genetic models were then combined using multi-view random forests. Finally, the genetic model was applied to independent cohorts treated with first-line drugs (glatiramer acetate and interferon beta) to assess its specificity for FTY.

The study identified a genetic signature of 123 SNPs (g2 model) that predicted FTY response with an AUROC of 0.65 in the independent test set. This model’s performance was enhanced when combined with clinical data (g2-c1 model), resulting in an AUROC of 0.71. Key clinical predictors selected by the algorithm were age at treatment start, annualized relapse rate (ARR) in the 2 years prior to FTY initiation, and the presence of new T2 and Gd+ lesions at baseline MRI. Over-representation analysis of the genes associated with the 123 SNPs revealed enrichment in pathways related to sphingolipid metabolism, cell adhesion molecules, and inflammatory bowel disease, suggesting biological relevance to FTY's mechanism of action. Comparison of patients predicted as likely responders (PrR) versus likely non-responders (PrNR) to FTY showed significant differences in disease activity during the 2-year follow-up: PrNR patients exhibited higher rates of disease reactivation (75% vs. 27%, p=0.0019), increased MRI activity, and a non-significant trend toward more relapses. Testing the genetic model on independent cohorts treated with first-line drugs yielded AUROCs of 0.55 and 0.51, suggesting model specificity for FTY response.

This study demonstrates the potential of integrating genetic and clinical data through ML methods to improve prediction of FTY response in RRMS patients. The identification of a 123-SNP signature and clinically meaningful clinical features highlights the model's ability to capture relevant biological and clinical factors. The enrichment analysis of related pathways further supports the biological plausibility of the findings. While the predictive accuracy is not yet sufficient for routine clinical use, the AUROC of 0.71 for the combined model represents a significant improvement over relying solely on clinical factors. The model’s specificity for FTY, demonstrated by testing on independent cohorts treated with first-line therapies, strengthens the findings. The study’s methodological rigor, including a robust feature selection procedure and validation in an independent cohort, increases the reliability of the results. The methodology used provides a framework for future studies focusing on other complex diseases and treatment responses.

This study provides evidence that machine learning can be used to combine clinical and genetic data to improve the prediction of response to fingolimod in RRMS patients. Although the predictive accuracy requires improvement before clinical implementation, the findings suggest a promising direction for personalized medicine in MS. Future studies with larger cohorts and integration of other data types (e.g., imaging) are warranted to refine predictive models and enhance their clinical utility.

The study's limitations include the relatively small sample size, particularly in the test set (77 patients), which may limit the generalizability of the findings. The relatively short follow-up period (2 years) may also underestimate the long-term predictive power of the model. The identified genetic signature may not be universally applicable due to population-specific genetic variations. Further research with larger and more diverse cohorts is necessary to validate these findings and overcome these limitations.