Resolving Kinase Resistance Landscapes via Antisymmetric Siamese Graph Networks and Co-Folding Ensembles

Abstract

Acquired clinical resistance to small-molecule kinase inhibitors arising from single-point mutations within the ATP-binding cleft presents a critical challenge in precision oncology. Conventional computational modeling relies either on rigid docking—which overlooks induced structural alterations—or computationally demanding alchemical free energy perturbation (FEP) methods that cannot scale across wide mutational panels. In this work, we present an end-to-end computational methodology integrating multi-state conformational ensemble generation via the generative biomolecular co-folding engine Boltz with a continuous distance-aware Siamese Graph Neural Network (GNN). Crucially, we introduce an unconstrained bias-free difference head that strictly enforces thermodynamic antisymmetry (f (0) ≡ 0.0 kcal/mol), eliminating trivial baseline drift. Under rigorous 5-fold cross-validation on experimentally characterized kinase resistance pairs, the model achieves an out-of-fold root-mean-square error (RMSE) of 1.637 kcal/mol (MAE = 1.459 kcal/mol, Pearson R = 0.239, Spearman ρ = 0.260). Across a prospective panel of 52 clinical kinase-inhibitor complexes, the architecture structurally rationalizes key resistance patterns, resolving the steric disruption of the ABL1 T315I gatekeeper mutation circumvented by ponatinib and the covalent ablation of EGFR C797S against osimertinib.