Mechanistic homogeneity, not model sophistication, sets the ceiling for mutation-effect prediction: a five-family FGFR benchmark against deep mutational scanning

Which computational features can predict the functional consequences of kinase mutations, and where do they stop working? We benchmark five feature families spanning physics-based, docking and learned paradigms — FoldX folding stability, AutoDock Vina docking perturbation, FoldX interaction energy, AlphaMissense and ThermoMPNN-D — against deep mutational scanning of FGFR1–4 kinase domains under pemigatinib and futibatinib selection (7,522 unique variants; Tangermann et al. 2026), adding an erdafitinib-specific docking dimension the screen did not cover. The capability map is non-uniform in three ways. Classical features stay below 0.74 across the panorama but transmit strong signal inside annotated hotspots (0.65–0.88, region-level q ≤ 4.5 × 10⁻⁶). ThermoMPNN-D crosses the conventional 0.7 line where classical FoldX does too: FGFR3 activation reaches 0.877 (95% CI 0.798–0.938, q = 5.1 × 10⁻⁶) and resistance 0.702 — distributed across the domain, not hotspot-driven. Resistance to erdafitinib specifically never exceeds 0.63 on the experimental structure (0.663 on one modelled pose). We trace this staircase to the mechanistic homogeneity of each phenotype and document three distinct failure modes: drug-resistant variants are enriched among those AlphaMissense calls benign; docking features degrade whenever the ligand pose is modelled rather than experimental; classical FoldX has an isoform-specific blind spot on FGFR1. The map converts into 30 falsifiable predictions absent from the screen and from clinical databases. Mechanistic homogeneity, not model sophistication, sets the ceiling for mutation-effect prediction.