Decoding FGFR inhibitor sensitivity in cholangiocarcinoma with interpretable machine learning and cross-platform pharmacogenomic validation

Background Fibroblast growth factor receptor (FGFR) inhibitors (FGFRis) benefit selected cholangiocarcinoma (CHOL) patients, yet responses remain heterogeneous and are not fully explained by canonical FGFR alterations. Clinically useful biomarkers are hindered by variable single-compound drug sensitivity readouts with limited concordance across FGFR-targeting agents and by optimistic bias arising from information leakage in pharmacogenomic modeling. Methods We integrated large-scale pharmacogenomic screening with baseline transcriptomes to derive a pathway-level FGFRi sensitivity phenotype. Drug responses were z-normalized per compound and aggregated across an FGFR-targeting panel to generate a composite FGFRi score, with reliability assessed by drug–drug concordance and split-half reproducibility. Transcriptome-based predictors were trained using strict group-aware cross-validation with out-of-fold (OOF) evaluation; performance was quantified by Spearman correlation and tested by permutation. An interpretable transcriptional program was extracted from linear model coefficients and projected into CHOL cohorts. Portability was assessed via cross-platform concordance in an orthogonal resource (GDSC) using matched cell lines and via biological coherence across multiple CHOL patient datasets, including rank-based scoring and a reduced feature panel. Results The composite FGFRi score was more stable than single-drug readouts and showed strong split-half reliability (median ρ = 0.63) without confounding by drug coverage. Baseline transcriptomes predicted FGFRi sensitivity under leakage-safe evaluation and yielded a compact, interpretable signature. In TCGA-CHOL, the signature mapped to structured tumor states and associated with FGFR-axis components, showing inverse correlations with FGFR1/2/3 and a positive correlation with KLB. In matched cell lines, PRISM-derived scores/signatures aligned with GDSC sensitivity for representative FGFR inhibitors. In external CHOL cohorts, rank-based scoring and leakage-controlled proxy-label models showed consistent performance across datasets, and a reduced 15-feature panel preserved concordance with the full signature. Conclusion A multi-compound, pathway-level FGFRi phenotype coupled with leakage-safe transcriptomic modeling identifies a transferable, interpretable FGFRi-associated program. This framework improves reliability relative to single-drug biomarker discovery and supports practical, portable scoring for CHOL stratification. Prospective validation in FGFRi-treated CHOL cohorts is warranted.