Integrating genetic, epigenetic, and clinical signatures via machine learning for robust prediction of leflunomide response in rheumatoid arthritis: a multi-center validation study.

OBJECTIVE

To develop and validate a machine learning(ML)-based integrated predictive model combining genetic, epigenetic, and clinical factors for predicting leflunomide (LEF) treatment response in rheumatoid arthritis (RA) patients.

METHODS

A total of 357 RA patients (231 in the model development cohort [MDC], 126 in the external validation cohort [EVC]) were recruited from multiple centers in China. Whole-exome sequencing(WES), genome-wide DNA methylation profiling, and comprehensive clinical data were integrated for model development. Feature selection was performed via univariate analysis, Least Absolute Shrinkage and Selection Operator(LASSO) regression, and clinical feasibility filtering. Ten ML algorithms were tested, with SHapley Additive exPlanations (SHAP) for interpretability, and external validation to assess generalizability.

RESULTS

The final integrated model included 3 single nucleotide polymorphisms (SNPs: ESR1-rs2813563, ABCC2-rs4148396, LMO4-rs983332), 7 differentially methylated positions (DMPs: cg13568171-MECR, cg07694252-ANGPT1, cg13401893-RNF39, cg19814518-UHMK1, cg26370237-HSF5, cg11136343-intergenic, cg15961042-intergenic), and 3 clinical variables (IgG, course of disease, baseline Disease Activity Score 28(DAS28)). The Random Forest (RF) algorithm achieved the highest performance among the ten ML algorithms, with an area under the curve (AUC) of 0.84 (95% CI: 0.73-0.94) in the MDC and 0.70 (95% CI: 0.60-0.80) in the EVC-outperforming SNPs-only (AUC: 0.77/0.70) and DMPs-only (AUC:0.80/0.70) models. SHAP analysis identified cg13568171-MECR as the most important predictive feature, and its hypermethylation was associated with an improved response to LEF. Functional enrichment revealed three interconnected biological modules (lipid metabolism, drug transport, endocrine signaling) regulating LEF efficacy. A covariate-adjusted interaction between cg07694252-ANGPT1 and ESR1-rs2813563 was observed in the MDC (P = 0.02).

CONCLUSION

The integrated clinical-genetic/epigenetic RF model enables reliable prediction of LEF response in RA. Multi-omics integration showed superior performance in the MDC, while maintaining robust and non-inferior performance in EVC. The methylation-dependent interaction between cg07694252-ANGPT1 and ESR1-rs2813563 implies a novel context-dependent transcriptional crosstalk mechanism, highlighting the value of multi-omics integration in advancing precision rheumatology.