A Pathway-Driven Computational Drug Discovery Framework Integrating Machine Learning and Virtual Screening for Identification of Anti-Rheumatoid Arthritis Natural Small Molecules

Objective: This study aimed to establish a pathway-driven computational drug discovery framework integrating machine learning, target prioritization, virtual screening, molecular dynamics (MD) simulations, and in vitro evaluation to identify natural small molecules with potential anti-rheumatoid arthritis (RA) activity. Methods: Sixteen RA-related Gene Expression Omnibus (GEO) datasets were included, with 15 cohorts used for model development and one reserved for independent external validation. Pathway features quantified by gene set variation analysis (GSVA) were subjected to cohort-aware feature selection using a leave-one-dataset-out framework across seven machine-learning algorithms. Predictive modeling and SHapley Additive exPlanations (SHAP) analysis were subsequently performed, followed by random-effects meta-analysis (REMA) and network analysis for candidate-target prioritization. Natural compounds were screened by molecular docking and further evaluated using ligand efficiency, multiple scoring approaches, explicit-membrane MD simulations, and a TNF-α-induced MH7A inflammatory model. Results: Ten stable RA-associated pathways were identified. The final CAT+ET model achieved ROC-AUCs of 0.783 (95% CI, 0.731–0.836) and 0.610 (95% CI, 0.496–0.720) in the internal and independent external validation sets, respectively, indicating limited cross-cohort generalizability. SHAP and cross-cohort analyses supported Complement and coagulation cascades for downstream investigation, and integrated expression and network evidence prioritized C3AR1 as a candidate target. Virtual screening identified four candidate natural compounds (SM1–SM4). Ligand-size analysis, ligand efficiency, multiple scoring approaches, and explicit-membrane MD simulations provided complementary structural and energetic assessments of these candidates. Among them, SM1 showed comparatively favorable computational characteristics and a relatively consistent reduction in the inflammatory markers IL6 and MMP3 in TNF-α-stimulated MH7A cells, accompanied by changes in C3AR1 expression. These convergent computational and experimental findings supported the prioritization of SM1 for further investigation. Conclusions: This study established a pathway-driven framework linking cross-cohort pathway identification with candidate-target prioritization and natural-product screening in RA. Complement and coagulation cascades and C3AR1 were prioritized for downstream investigation, while integrated computational and experimental evidence identified SM1 as a candidate with promising anti-inflammatory potential. Although direct compound–C3AR1 engagement and its functional involvement require further validation, the present study provides a systematic and extensible strategy for translating robust RA-associated pathway signals into candidate targets and natural compounds, offering a valuable foundation for subsequent natural-product drug discovery in RA.