Pose-ensemble MLPs rerank docking hits for the hard nuclear receptor target LRH-1, yielding four putative binders
- Poses from physics-based targeted docking and generative blind docking, across several receptor conformations, train two MLPs that predict binders or activators at the LRH-1 orthosteric pocket, avoiding dependence on one pose or one score.
- On a chemically distinct out-of-distribution set, the models found dissimilar compounds and responded to single-scaffold edits, suggesting use in lead optimization.
- A small prospective campaign produced four putative binders, and combining docking scores with MLP predictions improved hit enrichment across library sizes.
💡 Points Based on the abstract
Four putative binders is a docking-plus-ML result, not a validated one, so I'd want biophysical confirmation (DSF, SPR, or a cocrystal) before calling them hits. LRH-1 pockets are lipophilic and promiscuous, so I'd also check the MLPs against a property-matched decoy set, since pose-derived features can mostly encode lipophilicity and size. The consensus gain over docking alone is the useful, transferable bit.