MolFuSE builds a UMAP space from same-function ligands to enrich potent inhibitors for targets with almost no SAR data
- For orphan or understudied targets, MolFuSE embeds compounds with UMAP in a similarity space built from ligands known to bind proteins with the same molecular function, needing no target-specific training data.
- Across nine targets, UMAP embeddings of Mordred 2D descriptors gave better early enrichment of high-potency inhibitors than ECFP4 fingerprints, although ECFP4 ranked actives better overall.
- Performance peaked at a reference set of roughly ten thousand compounds and fell with larger sets, as structurally diverse weak binders dilute the descriptor space.
💡 Points Based on the abstract
The reference-set sweet spot is the honest finding, and it means curation is the method: somebody has to choose that ten thousand, and that is where bias creeps in. Early enrichment versus global ranking is a real split, so on an orphan target I'd run this next to a plain ECFP similarity search and take the union. First check that the 'potency filter' survives when the reference set shares few chemotypes with the true actives.
Molecular function-guided unsupervised similarity-based enrichment for low-data virtual screening ↗