Geometry-Constrained Prediction of Catalytically Competent Kinase Domains Across the Human Kinome

Abstract

Understanding cellular signaling networks benefits from accurate structural modeling of protein kinases across their different functional states. However, most human kinases lack experimentally determined active-state structures, and existing computational methods often fail to reliably predict catalytically competent conformations. Here, we present a geometry-constrained modeling framework based on AFEXplorer (AFEX) that enables robust prediction of active, ATP- and magnesium-bound states for 436 catalytic kinase domains across the human kinome. By explicitly enforcing key biophysical criteria, AFEX overcomes the template dependence and conformational bias inherent in AlphaFold2 and AlphaFold3. Benchmarking against 156 kinase-substrate complexes shows that AFEX achieves sub-2{square}[A] backbone RMSD in 99% of cases and accurately recapitulates critical catalytic features with precise ATP coordination. Notably, AFEX correctly predicted active conformations for EGFR and CDK2, where AlphaFold3 failed. These improvements are consistent across diverse kinase families, including both well-characterized kinases and those underrepresented or absent from AlphaFolds training data. AFEX bridges a critical gap in kinase biology by providing a generalizable framework for studying active conformations and dynamics in kinases and other signaling proteins.