Explainable Physicochemical Determinants of Protein–Ligand Binding via Non-Covalent Interactions

Abstract Protein–ligand binding governs enzymatic catalysis, metabolic regulation, and therapeutic modulation, and its prediction underpins drug discovery. However, existing AI approaches broadly function as black-box predictors that cannot resolve which residues mediate binding or which non-covalent forces drive molecular recognition. We present ExplainBind, an interaction-aware framework that predicts binding likelihood, localizes specific binding residues rather than coarse pocket-level regions, and decodes the underlying non-covalent interaction patterns, all without requiring three-dimensional structural inputs. ExplainBind consistently outperforms representative baselines across proteins and ligands ranging from closely related to highly novel sequences and scaffolds. Applied to two unseen targets, ExplainBind successfully ranks potent angiotensin-converting enzyme (ACE) inhibitors and explains potency differences via affinity-stratified interaction landscapes, and prospectively discovers both inhibitors and activators of L-2-hydroxyglutarate dehydrogenase (L2HGDH), with mechanistically distinct interaction profiles rationalizing their divergent functional outcomes, establishing it as a broadly applicable tool for mechanistically informed drug discovery. Competing Interest Statement The authors have declared no competing interest. Footnotes The revised manuscript expands the evaluation of ExplainBind beyond binary binding prediction to include binding-site localization and non-covalent interaction prediction. Compared with the previous version, it now evaluates representative baselines across binding prediction and residue-level localization, revealing that high binding classification performance does not necessarily translate into accurate binding-site identification. The revision also introduces an interaction-type-specific evaluation using Interaction Hit Rate (IHR) to quantify recovery of individual non-covalent contacts at the residue-atom level. In addition, the abstract, introduction, results narrative, figures, and mechanistic analyses have been substantially revised, with expanded discussion of interaction-aware supervision, comparison with structure-based models, and updated ACE and L2HGDH validation analyses.