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

Machine learning interatomic potentials (MLIPs) enable efficient modeling of molecular interactions with quantum mechanical (QM) accuracy. However, constructing robust and representative training datasets that capture subtle, system-specific interaction motifs remains challenging. We introduce PANIP (PAirwise Non-covalent Interaction Potential), an ensemble MLIP model built upon the 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, in which a representative training subset, termed PDB-FRAGID (PDB Fragment Interaction Dataset), was distilled from an otherwise prohibitively large pool of fragment dimers extracted from the Protein Data Bank (PDB). PANIP retains $ω$B97X-D3BJ/def2-TZVPP-level accuracy 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, particularly for charged and strongly interacting dimers. Coupled with a fragmentation-based energy decomposition scheme, PANIP estimates protein-ligand binding energies at near force-field computational cost yet QM-level accuracy, enabling its use as a fragment-based scoring function that rivals specialized docking scoring functions.

💡 Research Summary

The authors present PANIP (PAirwise Non‑covalent Interaction Potential), a machine‑learning interatomic potential specifically designed to model the diverse non‑covalent interactions (NCIs) that dominate protein structure and protein‑ligand binding. Building on the equivariant NequIP architecture, PANIP is trained on a curated set of fragment dimers extracted from the Protein Data Bank (PDB). Because labeling all ~36 million possible dimers at a high‑level quantum‑mechanical (QM) method would be prohibitive, the authors devised a multi‑fidelity active learning (MFAL) workflow. First, the low‑cost r²SCAN‑3c functional screened the entire pool, providing rough interaction energies. A surrogate NequIP model was then iteratively refined, flagging high‑error dimers for high‑level labeling with ωB97X‑D3BJ/def2‑TZVPP. This process distilled the dataset to 3.15 million representative dimers (≈8.7 % of the original) while preserving coverage of 17 fragment types and 153 dimer combinations, with special emphasis on charged, polar, and otherwise chemically challenging pairs.

Training on this “PDB‑FRAGID” set, PANIP achieves near‑quantum accuracy: mean absolute errors (MAE) range from 0.09 kcal mol⁻¹ for low‑energy equilibrium dimers to 0.45 kcal mol⁻¹ for a broad set of non‑equilibrium conformations, and R² values consistently exceed 0.99. The model remains robust when applied to external benchmarks, including the Cambridge Structural Database (CSD) and randomly sampled high‑energy conformations, demonstrating excellent transferability beyond the training domain. In direct comparison, the widely used ANI‑2x potential performs poorly on the same CSD set (MAE ≈ 9 kcal mol⁻¹) and especially fails for charged or strongly interacting dimers, highlighting the advantage of a protein‑specific, high‑fidelity training set.

Computationally, PANIP delivers two orders of magnitude speed‑up over hybrid DFT: evaluating 15 300 random dimers on a single CPU core takes ~6 h, whereas the reference QM calculations would require over 463 days. This efficiency enables large‑scale applications. The authors showcase two such uses: (1) systematic mapping of NCI patterns across the full 36 million PDB‑derived dimers, revealing expected cation‑π geometries and uncovering previously under‑explored dimethyl‑sulfide‑aromatic interactions with distinct spatial preferences; (2) a fragment‑based scoring function for protein‑ligand binding, which reproduces QM‑level binding energies at near‑force‑field cost and competes with specialized docking scores.

Overall, PANIP demonstrates that a carefully constructed, multi‑fidelity training pipeline can produce an MLIP that combines quantum accuracy, broad chemical coverage, and practical speed for biomolecular simulations. The work sets a new benchmark for protein‑focused ML potentials and opens avenues for rapid, accurate drug‑discovery pipelines, protein engineering, and large‑scale NCI analyses.