A pretrained ML force field scores protein–ligand binding at near-MM/GBSA cost and holds up against FEP benchmark sets
- SO3LR-SF combines pretrained SO3LR machine-learned force fields with physics-based terms for repulsion, electrostatics and dispersion to score active-site models of protein–ligand complexes.
- It was tested against high-level quantum references on small active-site models and on two ligand-ranking sets of the kind used for free-energy benchmarks.
- Ranking quality was comparable to the best semi-empirical method and MM/GBSA, above Glide, and in the range of MD-based free-energy methods on one set, at a few seconds to under a minute per complex on CPUs.
- The authors frame it as less exposed to inflated results from training on protein–ligand complexes.
💡 Points Based on the abstract
Matching MM/GBSA and edging Glide is useful, but the FEP sets here are retrospective congeneric series, which are the easy case for any scoring function. I would run it on our own series with known SAR cliffs and see whether it ranks the ones FEP+ gets wrong. If it holds up, 3 to 47 seconds per complex makes it a realistic rescoring layer after docking.
Explainable ML force-field for evaluating protein–ligand binding energy using SO3LR ↗