Developing a Machine-Learning Interatomic Potential for Non-Covalent Interactions in Proteins

Abstract

Machine learning interatomic potentials (MLIPs) enable efficient modeling of molecular interactions with quantum mechanical (QM) accuracy for complex systems. However, constructing robust and representative training datasets that capture subtle, system-specific interaction motifs remains challenging. Here, a PAirwise Non-covalent Interaction Potential (PANIP) is introduced, an ensemble MLIP model built upon the Neural Equivariant Interatomic Potentials (NequIP) framework and trained on non-covalent interactions (NCIs) between protein-derived fragments. PANIP is trained using an automated multi-fidelity active learning (MFAL) workflow that distills a diverse and information-rich subset from an otherwise prohibitively large pool of fragment dimers extracted from high-resolution protein structures in the Protein Data Bank (PDB). This strategy concentrates high-level QM calculations (ωB97X-D3BJ/def2-TZVPP) on the most informative structures while preserving comprehensive coverage of the data distribution. Applied to dimers constructed from 17 chemically distinct protein fragments (side chains, backbone motif, and water), this workflow yields the PDB Fragment Interaction Dataset (PDB-FRAGID), a condensed yet representative subset comprising only 8.7% of the original 36.3 million-dimer pool while maintaining structural and chemical diversity. Despite this drastic reduction, PANIP retains ωB97X D3BJ/def2 TZVPP-level accuracy on both equilibrium and non‑equilibrium fragment configurations and achieves mean absolute errors below 0.2 kcal/mol on out‑of‑distribution systems, demonstrating excellent transferability across diverse NCI motifs. Compared to the widely used ANI 2x potential, PANIP delivers substantially lower errors on protein‑derived fragments and exhibits superior generalization, particularly for charged and strongly interacting dimers. By combining a fragmentation-based energy decomposition scheme with PANIP, protein–ligand binding energies can be estimated at near force‑field computational cost while preserving QM‑level accuracy for pairwise NCIs, enabling its use as a fragment‑based scoring function that rivals specialized docking scoring functions despite being trained solely on fragment dimers. The PANIP models and associated benchmark sets are available at https://github.com/hnlab/PANIP, and the PDB‑FRAGID dataset is available at https://github.com/hnlab/PDB‑FRAGID.