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 with heterogeneous clinical presentations and treatment responses. With expanding therapeutic options for relapsing-remitting MS (RRMS), individualized treatment selection is increasingly needed. Genetic factors contribute to MS susceptibility and may influence disease severity and response to first-line therapies; however, little is known about genetic determinants of response to second-line drugs, and no prior studies have examined genetics and response to fingolimod (FTY). This study aimed to apply machine learning (ML) to identify a genetic signature, in combination with clinical and demographic features, to predict response to FTY in RRMS patients.

Large GWAS have established genetic predisposition to MS. Smaller studies suggested genetic influence on disease severity and on response to first-line therapies such as interferon-beta and glatiramer acetate, but data on second-line drugs are limited and genetics of FTY response had not been explored. ML methods are increasingly used in precision medicine and have been applied in MS for imaging tasks and, less frequently, to clinical data for predicting disease course. One prior study applied ML to genetics for predicting response to glatiramer acetate. The present work builds on this by evaluating genetic and clinical predictors of FTY response.

Design: Two cohorts of RRMS patients on FTY with available genetic and clinical data were collected: OSR (San Raffaele Hospital, Milan; initial n=364) and CHUT (Toulouse; initial n=108). After exclusions (NTZ in prior 9 months; genotyping QC; missing clinical data for NEDA assessment), 381 patients remained (OSR 342, CHUT 78 before NEDA exclusions; final combined n=381; 197 EDA, 184 NEDA). Patients were split into training (TR) 40% (n=152), validation (V) 40% (n=152), and independent test (TE) 20% (n=77), stratified by cohort and EDA/NEDA. Clinical data: Gender; age at onset; disease duration at start; ARR in 2 years pre-FTY; prior DMT categories; EDSS at start and 2 years; relapses; MRI activity (new/enlarging T2 and Gd+ lesions). Response endpoint: NEDA-3 at 2 years (no relapses, no MRI activity, no disability progression). Genotyping: Illumina HumanOmniExpress arrays. QC with PLINK v1.9: remove SNPs with MAF<0.01, call rate<0.97, HWE p<1e-4; remove samples with call rate<0.95, relatedness, sex mismatch, ancestry outliers. Datasets merged; LD pruning (r^2>0.2) to reduce redundancy; genotypes coded additively (minor allele dosages). Modeling: Random Forests (RFs) were selected after preliminary comparisons to SVM, KNN, decision trees, bagging, boosting. A robust feature selection was performed on TR/V: repeated 10-fold CV 100 times, storing the top-k SNPs per fold; SNPs retained if selected with relative frequency ≥f across repetitions. Approximately 2000 combinations were explored varying RF hyperparameters (ntrees 10–100; max nodes 1–100; node size 1–10) and feature selection parameters (k=50–1000; f=0.05–1). Vset used to choose SNP signatures and RF hyperparameters; TEset used only for final evaluation. Performance metrics: AUROC, AUPRC, F-score, accuracy. Implementation in R 3.6.3 with randomForest, caret, precrec. Clinical-only models: The same robust selection and RF training were applied to 17 clinical features (including prior DMT categories). Top four clinical signatures/models were selected based on Vset performance. Combined model: Multi-view RF ensemble combining genetic and clinical views: trees trained on bootstrap samples from either genetic or clinical feature sets; predictions aggregated for consensus. External evaluation: The best genetic model was applied to independent cohorts treated with first-line therapies (IFN n=304; GA n=273) to assess specificity for FTY response. Tertile analysis of predictions: On TEset, patients were grouped by tertiles of predicted non-response probability (from the selected combined model). Highest tertile defined predicted non-responders (PrNR), lowest defined predicted responders (PrR). Groups compared for MRI activity (number of active lesions; proportion with active MRI), relapses, clinical activity, and NEDA status using Mann–Whitney and chi-square tests.

- Cohort and splits: Final n=381 (197 EDA, 184 NEDA); TR n=152, V n=152, TE n=77. - Genetic model: A robust feature selection identified signatures; model g2 with 123 SNPs (k=500 top-ranked per fold; min frequency=0.1) yielded TE AUROC=0.6446, AUPRC=0.6630, F=0.7142, accuracy=0.5844. Other genetic models achieved TE AUROC around 0.58–0.65. - Pathway enrichment (73 annotated genes; 30 mapped in KEGG): Top uncorrected pathways included sphingolipid signaling (p=0.008), sphingolipid metabolism (p=0.011), CAMs (p=0.013), IBD (p=0.021); none survived multiple testing correction (FDR ~0.65–0.67). - Clinical-only models: Best AUROC on TE up to 0.6895 (AUPRC up to 0.7320; F up to 0.7379; accuracy up to 0.6494). Consistently selected clinical predictors included age at treatment start, pre-FTY ARR, and presence of new T2 and Gd+ lesions at baseline. - Combined clinical-genetic model: Multi-view RF (g2-c1) achieved TE AUROC=0.7095, AUPRC=0.7328, F=0.7328, accuracy=0.6623. - Tertile analysis on TE: Predicted non-responders (PrNR) vs predicted responders (PrR): EDA proportion 75% vs 27% (p=0.0019); mean number of active MRI lesions 2.1 vs 0.61 (p=0.034); MRI activity present in 50% vs 11.5% (p=0.005); mean relapses 0.25 vs 0.15 (p=0.40); clinical reactivation 20.8% vs 7.7% (p=0.24). - External cohorts (specificity): Applying the FTY-trained genetic model to IFN- and GA-treated cohorts produced AUROC=0.55 and 0.51, respectively, suggesting specificity to FTY response rather than general MS activity.

The study demonstrates that ML can extract predictive signals for FTY response from both genetic and clinical data in RRMS. A 123-SNP signature provides modest discrimination (TE AUROC ~0.65), and clinical variables alone yield slightly better performance (AUROC ~0.69), reflecting known prognostic factors (age at start, prior ARR, baseline MRI activity). Combining genetic and clinical features via multi-view RF improves performance to AUROC ~0.71 and AUPRC ~0.73, indicating complementary information and supporting a precision medicine approach to FTY selection. Biological plausibility is supported by enrichment of sphingolipid-related pathways, consistent with FTY’s mechanism of action, and cell adhesion pathways. The tertile-based analysis further shows meaningful stratification of patients by predicted non-response with significant differences in MRI and overall disease activity. External testing on IFN and GA cohorts yields near-random performance, suggesting the genetic signature is specific for FTY rather than a general disease activity marker. While not yet clinically actionable, the framework illustrates the potential of integrated ML to guide therapy decisions in MS.

The study identifies a 123-SNP genetic signature and shows that integrating genetic with clinical data via multi-view random forests improves prediction of response to fingolimod in RRMS (TE AUROC ~0.71). Although current accuracy is insufficient for clinical deployment, results support the promise of ML-driven precision medicine in MS. Future work should involve larger, independent cohorts, longer follow-up, and integration of additional modalities (omics and imaging) to enhance generalizability and predictive power, potentially enabling clinically applicable response prediction tools.

Key limitations include modest sample size relative to genetic feature dimensionality, an independent test set of only 77 patients, and relatively short follow-up (2 years). These factors may limit generalizability and increase risk of overfitting despite robust validation procedures. Pathway enrichment did not survive multiple testing correction. Clinical predictors may capture general prognostic factors rather than treatment-specific effects in some contexts.