Active learning cuts docking cost when searching huge kinase fragment-recombination libraries
- Framework combines subpocket-aware fragment recombination from KinFragLib, docking, ML surrogate models, and genetic-algorithm-style fragment additions, deletions and substitutions around already-scored ligands.
- Tested retrospectively on protein kinase A in a very large recombination space; only a tiny fraction of ligands needed expensive docking, yet many top-scoring ones were found.
- A greedy acquisition strategy recovered a large share of the best ligands, clearly beating random sampling; a diversity-aware variant broadened fragment and hinge-binder coverage while staying competitive.
💡 Points Based on the abstract
PKA is about the friendliest kinase for this: a rigid, well-studied pocket where docking scores correlate decently with enrichment. Retrospective recovery of docking top-hits only shows the surrogate learns the docking function, not affinity. Before porting this to my target I'd want an orthogonal check, such as FEP or experimental data on a handful of picks, plus synthesizability of the recombined molecules.
Active Learning for Fragment-Based Kinase Inhibitor Discovery using Docking ↗