Open-source LearnM8 lets any scoring method, from docking to lab assays, drive active learning
- Preprint (not peer reviewed): an open-source Python package that treats the scoring oracle as a swappable function, so docking, free-energy calculations, stored scores or assay data can all feed one active-learning loop.
- It bundles many surrogate models, acquisition strategies and molecular featurizers, plus cycle scheduling, pruning and diversity tracking.
- Retrospective tests on precomputed docking scores for AmpC and D4 libraries of up to over a hundred million compounds show most top-scoring molecules are found after scoring only a small fraction, far beyond random picking.
- Smaller batches help recovery at higher cost; pruning cuts workload with a slight loss. No prospective validation is reported.
💡 Points Based on the abstract
Recovery on precomputed docking scores mostly shows the surrogate can learn the docking function, not that the hits are real binders, and AmpC and D4 are the familiar, well-behaved benchmarks. The test that matters for a team is a noisy oracle: run it on your own assay or FEP data, with a random-pick control and a scaffold-diversity check on what it recovers.
LearnM8: A Simple and Modular Molecular Active Learning Framework for any Oracle ↗