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

, a curated database of protein-ligand complexes from which residue-atom interaction maps are systematically extracted. ExplainBind aligns token-level cross-attention with these interaction maps, enabling physically grounded and mechanistically interpretable predictions without requiring explicit three-dimensional inputs at inference time. Across extensive in-distribution and out-of-distribution evaluations, ExplainBind consistently outperforms existing sequence-based baselines and demonstrates improved robustness to protein and ligand distribution shifts. Quantitative analyses and structural case studies show that the learned interaction maps accurately localize binding pockets and recover known interaction motifs. We further validate the framework by effectively ranking highly potent inhibitors for angiotensin-converting enzyme (ACE) and discovering both inhibitors and activators of the metabolic enzyme L-2-hydroxyglutarate dehydrogenase (L2HGDH), illustrating how ExplainBind supports functional modulation beyond binary binding prediction. Together, these results establish ExplainBind as a scalable and interpretable paradigm for protein-ligand binding prediction across drug discovery, enzyme engineering, and broader molecular design applications.