Answers are grounded in this paper and cite their sources.
Motivation: Drug-resistance mutations are a clinically central but poorly predicted phenotype: pathogenicity predictors trained on disease labels misrank them, and our recent benchmark of five static predictor families (AlphaMissense, FoldX stability, FoldX interaction energy, ThermoMPNN-D and AutoDock Vina) showed that biophysical scores predict FGFR resistance poorly. Whether supervised recalibration on functional deep mutational scanning (DMS) labels can repair this, and what feature class carries the signal, remains untested. Results: Using the DMS of the FGFR1–4 kinase domains (Tangermann et al. 2026; 7,522 unique variants) under leave-one-subtype-out validation, raw AlphaMissense correlates negatively with the resistance score in every subtype (ρ = −0.46 to −0.65), and on the binary task models built on biophysical scores reach AUROC 0.54–0.74 and never exceed the position-based model (scores 0.56–0.64 vs position 0.73–0.87 on the complete-case subset, n = 5,135). On the continuous score, a one-parameter sign flip remains the strongest recalibration. The signal is instead carried by an aligned-position prior: a parameter-free lookup of resistance rates at each aligned kinase-domain coordinate, transferred from the training subtypes. The lookup matches or exceeds a gradient-boosted model (AUROC 0.88, 0.89, 0.84, 0.70 for FGFR1–4), is strongly resistance-skewed (pure-resistance 0.86, 0.84, 0.84, 0.61; pure-activating 0.62, 0.60, 0.42, 0.46), peaks at single-residue resolution, sits 0.16–0.30 above the strongest clinical-hotspot baseline, and reaches AUPRC 0.31–0.48 with 37–63% precision in the top 50; a two-stage blend that resolves within-site ordering raises this to 36–64% with a tie-free ranking at no AUROC cost. The prior transfers across drugs (pemigatinib → futibatinib, 0.74–0.83) and across FGFR1–3, weakening for FGFR4. Conclusion: FGFR resistance phenotypes are recoverable from a homology-transferred positional label statistic — a lookup table, not a fitted model; the lookup alone recovers the discrimination, and a within-site score ranker refines the top of the list without adding discrimination. The statistic is specific to resistance, contains information beyond known clinical hotspots, and transfers across drugs within the FGFR inhibitor class.
No takes yet. Share an insight, caveat, or question.
Jia He (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: