HuKSA: A Machine-Learning-Based Framework for Predicting Human Kinome Selectivity and Polypharmacology

Off-target polypharmacology across the conserved kinase ATP pocket can limit the safety and efficacy of kinase-directed compounds. HuKSA is a transparent ligand-based framework that predicts kinase targets and estimates measured-kinome selectivity from a SMILES string while exposing the reference activities behind each estimate. The method compares ECFP4 fingerprints by Tanimoto similarity and applies five-neighbour, similarity-weighted read-across to an atlas of 794 compounds with measurements spanning 464 kinases. Selectivity is summarized by the S-score, the fraction of measured kinases with pActivity above 6. In this sparse matrix, a Gini coefficient calculated after padding unmeasured kinases closely tracked assay coverage, whereas the measured-panel S-score showed weak assay-count dependence. Nevertheless, S-scores remain conditional on which kinases were assayed and on the mixture of IC50, Ki and Kd measurements. In 125 atlas-excluded KLIFS and PKIDB inhibitors, an annotated target ranked first for 19.2% of compounds and within the top 10 for 55.2%. Top-10 recovery was 83.3% among compounds in the high structural-support band, but these bands are empirical similarity descriptors, not calibrated probabilities. Internal leave-one-compound-out validation gave a pooled target-activity ROC-AUC of 0.83 and S-score prediction RMSE of 0.123. Scaffold and similarity-exclusion analyses defined the applicability domain, indicating strongest performance for queries supported by close measured analogues. By displaying measured neighbours, target-specific support counts, and prediction coverage, HuKSA enables users to inspect target and selectivity hypotheses before experimental testing and follow-up biochemical validation. The publicly accessible online Streamlit web server is available at https://huksalab-tool.hf.space.