Computational & AI chemRxiv
Affinity model derives JAK-family selectivity from predicted potency differences
- panJAK combines molecular representations with a ligand-conditioned encoding of aligned binding pockets to predict potency at four JAK paralogs, deriving selectivity from the differences without a dedicated selectivity objective.
- It was trained on a curated set of roughly ten thousand compounds and tested retrospectively on scaffold-disjoint splits.
- It beat baseline models at quantitative selectivity; correlated paralog errors partly cancelled, and calibrated intervals flagged a subset of comparisons with reliable direction.
Deriving selectivity from the difference of two noisy potency predictions is risky, so the cancellation of correlated errors is the interesting part. Scaffold-disjoint splits on mined ChEMBL-like data still leave assay heterogeneity and series bias, though. Open questions are how it handles a new chemotype's TYK2 versus JAK1 call and whether the intervals hold on in-house measured compounds.
Paper: chemRxiv ↗ PanJAK: predicting selectivity across the JAK family with deep learning