PanJAK: predicting selectivity across the JAK family with deep learning
Predicting selectivity between closely related protein targets remains a major challenge in drug discovery. While drug-target affinity (DTA) models typically focus on predicting activity against individual targets, selectivity depends on resolving differences in activity between them. Here, we introduce panJAK, a DTA framework to predict selectivity across the four Janus kinase (JAK) paralogs. The model combines molecular representations with a ligand-conditioned encoding of structurally aligned binding pockets and predicts cross-paralog selectivity from differences in potency without requiring a selectivity-specific training objective. Using a curated dataset of approximately 10,000 compounds, we demonstrate that panJAK improves the quantitative prediction of selectivitywindows compared to established baseline models under scaffold-disjoint evaluation. Errors in individual potency predictions were positively correlated across paralogs and partly cancelled upon subtraction, resulting in more accurate selectivity estimates than expected under independent errors. Although prediction errors increased with chemical novelty, calibrated prediction intervals allowed for the identification of a subset of comparisons with high retrospective directional accuracy. These findings highlight the potential of adapting affinity-prediction models to address quantitative selectivity across homologous targets, supporting compound prioritization based on predicted selectivity profiles and associated reliability of those predictions.
Authors
- Francesca Grisoni (ORCID: https://orcid.org/0000-0001-8552-6615)
- Sébastien Sueron (ORCID: https://orcid.org/0009-0005-6679-694X)
Institutions
- Eindhoven University of Technology (NL)
Publication Details
- Journal
- ChemRxiv
- Published
- 2026-11-17
- DOI
- https://doi.org/10.26434/chemrxiv.15012326/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint
Funders
- European Commission