DTBind: A Mechanism-Driven Deep Learning Framework for Accurate Prediction of Drug-Target Molecular Recognition

Accurate prediction of drug-target molecular recognition is essential for early-stage drug discovery, spanning binding occurrence, binding site localization, and binding affinity estimation. However, current approaches frequently treat these tasks independently, thereby overlooking the shared mechanistic principles that underlie them. We present DTBind, a unified and mechanism-driven framework that hierarchically adapts to sequence, structure, and complex-level inputs for predicting binding occurrence, site, and affinity, grounded in the shared mechanistic determinants of molecular recognition. Comprehensive benchmarking demonstrates that DTBind achieves superior performance over state-of-the-art methods in terms of accuracy and generalizability. Analysis of hierarchical protein encodings demonstrates that determinant-based protein representations gradually drive the accurate prediction of molecular recognition. Validation based on molecular dynamics simulations shows that DTBind reliably predicts drug-target binding regions for 3 famous proteins that lack experimental structures.