MoleProLink-RL: geometric transport for domain-policy reinforcement learning in drug-target interaction prediction

Accurate drug-target interaction (DTI) prediction is vital to modern discovery workflows, yet models trained in one setting often underperform when chemotypes, protein families, or assay conditions shift. MoleProLink-RL addresses this challenge as a model-first solution that couples chemically faithful representations with geometry-aware alignment and stability-aware decision making. The system combines graph-transformer drug encoders and residual protein embeddings with an unsupervised distribution alignment layer that respects molecular similarity through a DTI-aware transport distance complemented by kernel mean embeddings, while a lightweight reinforcement-learning objective regulates ranking stability across latent sub-environments under a small trust region to a supervised reference. Unlike pipelines that bolt adaptation onto frozen features or tune policies post hoc, MoleProLink-RL trains representation, alignment, and decisions end-to-end so that what the model learns to represent directly improves how it prioritizes candidates under shift. We provide a Fisher-Wasserstein view that clarifies why joint updates are safe in parameter space and effective on distribution space, and we validate the approach on Human, C. elegans, and Davis benchmarks with realistic class imbalance, comprehensive ablations, and diagnostics for calibration and interpretability. The result is a practical, data-efficient DTI model that maintains top-of-list precision when conditions change, narrowing the gap between benchmark performance and laboratory impact.