Explainable ML force-field for evaluating protein–ligand binding energy using SO3LR
Year: 2026
Type: preprint
Abstract: Accurate estimation of protein-ligand binding affinities remains central to structure-based drug design. Classical force fields neglect important quantum-mechanical effects, while quantum mechanical and semi-empirical methods remain impractical for large biomolecular systems. Machine-learned force fields (MLFFs) offer an alternative, but their performance may be inflated by training on protein-ligand complexes. Here, we introduce SO3LR-SF, a physics-based scoring framework that can be instantiated with any of the four pretrained SO3LR MLFF models, combining an equivariant message-passing network for semi-local interactions with physically motivated terms for short-range repulsion, electrostatics, and long-range dispersion. On PLA15 (15 active-site models), SO3LR-SF attains a relative interaction error of 2.1-5.8% against DLPNO-CCSD(T) references depending on the underlying SO3LR model, with SO3LR-T-M reaching 2.1%, the lowest of all tested methods. On the FEP benchmark (8 targets, 264 ligands), all SO3LR-SF variants achieve an average Spearman correlation of around 0.50, on par with the best semi-empirical quantum-mechanical method tested (0.47, GFN-FF) and with MMGBSA (0.44), and surpassing Glide (0.27). On the Wang dataset (8 targets, 199 ligands), the SO3LR-SF models reach around 0.59, matching the 0.59-0.62 range of molecular dynamics-based free-energy methods at a fraction of their cost. Each complex is scored in 3 s to 47 s on 12 CPU cores. An 8 Å trimming protocol reduces runtime threefold without loss of accuracy, and a restrained geometry optimization recovers ranking accuracy for targets with strained input structures. A multi-level explainability framework adds energy decomposition, per-atom energy contributions reported either for ligand or as 2D protein-ligand interaction maps, and 3D binding site visualization. We found that the polar SASA ratio of the binding pocket is predictive of electrostatic scoring performance, enabling estimation of ligand-ranking accuracy before scoring (r = 0.74-0.81 across models). Together, SO3LR-SF presents a fast, interpretable, and competitive scoring function for drug discovery that requires no experimental binding-affinity data for training.
Source: ChemRxiv
Institutions: Saarland University, University of Luxembourg
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7211956891