An AI-enabled structural atlas decodes kinase specificity across the human proteome

Abstract

Of the 1.8 million serine/threonine/tyrosine residues in the human proteome, only 6% bear experimental validation of phosphorylation, and only 5% of these have been mapped to a kinase. Here we present KinoPlex, a computational framework that integrates predicted protein structures and kinase recognition motifs to assign phosphorylation potential and kinase specificity to all serine/threonine/tyrosine residues. Using ~20,000 AlphaFold models and positive-unlabeled transfer learning, we identified ~567,000 residues as structurally phospho-competent. We intersected these with kinase position-specific scoring matrices to quantify motif specificity, yielding ~250,000 high-confidence candidates with sequence recognition potential and optimal structural presentation. The structural atlas uncovered fundamental organizing principles guiding kinase substrate recognition and dynamics of phosphorylation, including a phenomenon we call sequence–structure selective coupling, whereby kinases achieve specificity through structural scarcity of their preferred motif (negative-selecting kinases) or promiscuity through its structural accessibility (positive-selecting kinases), rather than by motif discrimination alone. Deep phosphoproteomics in K562 cells validates KinoPlex predictions and kinase enrichment capacities.