Active learning lets docking search a 210,000-ligand kinase fragment space by scoring only a sliver of it
- Couples subpocket-aware fragment recombination from KinFragLib with docking, a machine-learning surrogate and genetic-style add/delete/swap moves around already-scored ligands.
- Retrospective case study on PKA with a recombination space of roughly 210,000 ligands.
- Needed docking for only a small fraction of the space, and a greedy strategy recovered about half of the best ligands, clearly beating random picks.
- A diversity-aware variant widened fragment and hinge-binder coverage with little loss in performance.
💡 Points Based on the abstract
"About half of the best ligands" is only meaningful against a docking score, and that score is the ground truth here, so this measures how well a surrogate reproduces Vina-type scoring, not affinity. Also one kinase, retrospective, with the whole space presumably pre-docked. On my project I'd ask for a second target and a synthesizability or experimental check before trusting the recombined hits.
Active Learning for Fragment-Based Kinase Inhibitor Discovery using Docking ↗