Improving phosphoproteomic kinase activity inference by contextual refinement of Kinase Library relationships

Abstract

Kinase activity inference from phosphoproteomic data relies on kinase-specific substrate sets defined by kinase-substrate relationship (KSR) priors. Kinase Library (KL) expands candidate substrate coverage using experimentally derived kinase specificity profiles, but its predictions reflect sequence compatibility rather than established intracellular regulation. We benchmarked KL-expanded KSR priors using 222 perturbation phosphoproteomic samples covering 396 annotated activation or inhibition events across 78 kinases. Broad KL expansion increased the number of scoreable kinases from 108 to 391, but reduced perturbed-kinase discrimination, directional accuracy and ranking performance. Higher error rates occurred among kinases with limited curated substrate support, while broad expansion also increased redundancy among inferred activity profiles. To improve the reliability of motif-derived relationships, we constrained KL-derived KSRs using predicted protein co-localization and protein-association evidence. Protein-association filtering consistently improved inference performance, reduced activity redundancy and provided a more favorable accuracy-coverage balance than co-localization filtering. The selected protein-association-filtered prior retained activity estimates for 308 kinases with an area under the receiver operating characteristic curve (AUROC) above 0.75 and extended inference to 200 kinases not scoreable using curated KSRs, which were significantly enriched for dark kinases. These newly scoreable kinases showed structured activity patterns across perturbation contexts and interpretable signaling associations. Our results show that protein-association evidence can refine motif-derived KSR priors while preserving broad kinome coverage, providing a practical framework for more reliable phosphoproteomic kinase activity inference.