Chemical features consistently associated with kinase inhibitor activity: a cross-target SHAP analysis under scaffold-split evaluation.

Abstract

Interpretable machine learning approaches to quantitative structure-activity relationship (QSAR) modelling are increasingly applied in drug discovery, yet most studies remain confined to single targets and report feature attributions without translating them into chemically meaningful insights. We introduce a cross-target SHAP entropy framework for quantifying shared versus target-specific structure-activity relationships across protein families, applied to 31 human kinase targets from BindingDB under scaffold-based train-test evaluation. Random Forest classifiers trained on Morgan ECFP4 fingerprints achieved a median AUROC of 0.994, AUPRC of 0.9998, and MCC of 0.674, confirming genuine SAR learning beyond class prevalence exploitation. Pairwise Spearman rank correlation of mean absolute SHAP profiles across targets yielded moderate cross-target consistency (mean r = 0.332; 465 pairs). Shannon entropy-based classification of the top 200 fingerprint bits identified 15 consensus features dominated by aromatic N-heterocycles, aliphatic rings, and hydrogen bond environments, and 50 divergent features showing 2.8-fold higher SHAP magnitude than consensus features. SAR validation confirmed genuine enrichment of two top consensus fragments in active compounds. These findings indicate that kinase QSAR models share a low-magnitude consensus descriptor signal across the kinase family, while target-specific features dominate predictive decision boundaries. All SHAP attributions describe model decision behavior and should not be interpreted as causal binding mechanisms. The entropy decomposition framework is generalisable to other protein families and provides a transferable workflow for converting SHAP outputs into chemically actionable insights. All code and data are publicly available.