Advances in Machine Learning for Drug Repurposing: From Methodologies to Precision Medicine with Case Studies in COVID-19, Parkinson's Disease, and Cancer.

This review investigates recent advances in machine learning (ML) for drug repurposing, with applications spanning COVID-19, Parkinson's disease, and cancer. We provide a methodological taxonomy of ML techniques including supervised learning (e.g., random forests, SVMs), unsupervised clustering and dimensionality reduction, deep neural networks, reinforcement learning, and graph-based models such as MLGANN and DTD-GNN highlighting their roles in scalable drug-target interaction prediction and phenotype mapping. Case studies demonstrate how integrated models leveraging heterogeneous data (e.g., gene expression, chemical structure, EHRs) facilitated the identification of repurposable candidates like baricitinib, raloxifene, and rapamycin. Across the reviewed case studies, top-performing models achieved AUROC values ranging from 0.88 to 0.99 on their respective benchmarks, though significant performance degradation (AUROC drops of 0.05-0.15) was observed under cross dataset generalization, underscoring the translational gap between benchmark and real-world performance. We analyze computational challenges, including data sparsity, model overfitting, limited generalizability across datasets, and the interpretability gap in deep learning and GNNs. We discuss integration of explainable AI (e.g., SHAP, LIME), interpretable models (e.g., rule lists, causal graphs), network biology, and federated learning as avenues to improve robustness and translational relevance. Unlike prior reviews focused on single diseases or ML paradigms, this work provides a cross disease analysis exposing shared bottlenecks and identifying disease-agnostic best practices. We examine the role of foundation models and large language models (LLMs) in drug repurposing, including zero-shot molecular property prediction and literature-informed mining. This work outlines an ML-driven framework that balances computational scalability with biological validity, enabling actionable predictions in precision medicine.