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

Area of Science:

• Biochemistry and Molecular Biology
• Artificial Intelligence in Chemistry
• Drug Discovery and Development

Background:

• Protein-ligand binding (PLB) is crucial for biological processes and drug development.
• Current AI models often lack mechanistic interpretability, functioning as black boxes.
• Understanding non-covalent interactions is key to predicting binding affinity and specificity.

Purpose of the Study:

• To develop an AI framework, ExplainBind, for interpretable protein-ligand binding prediction.
• To learn residue-level interaction patterns grounded in physical non-covalent interactions.
• To enable mechanistic reasoning and functional modulation beyond binary binding prediction.

Main Methods:

• Developed ExplainBind, an interaction-aware, sequence-based framework.
• Constructed InteractBind, a curated database of protein-ligand complexes with residue-atom interaction maps.
• Aligned token-level cross-attention with interaction maps for physically grounded predictions.

Main Results:

• ExplainBind outperforms existing sequence-based baselines in in-distribution and out-of-distribution evaluations.
• Learned interaction maps accurately localize binding pockets and identify interaction motifs.
• Successfully ranked inhibitors for ACE and discovered modulators for L2HGDH.

Conclusions:

• ExplainBind provides a scalable and interpretable paradigm for protein-ligand binding prediction.
• The framework enhances mechanistic understanding for drug discovery and enzyme engineering.
• ExplainBind supports functional modulation applications in molecular design.