Integrative multi-omics and single-cell analysis identifies EGFR pathway activation and metabolic reprogramming as potential synthetic lethal vulnerabilities in resistance to the FGFR inhibitor AZD4547 Authors: Linghui Tan, Tianlun Hou, Pingting Ying, Xian Wang, Hongchuan Jin, Jingfeng Luo Published: 2026 ABSTRACT BackgroundAlthough fibroblast growth factor receptor (FGFR) inhibitors (FGFRi) have demonstrated clinical promise, the inevitable emergence of acquired resistance remains a critical bottleneck, severely compromising their long-term clinical efficacy. The pan-cancer molecular landscape and heterogeneous mechanisms driving this resistance, ranging from genetic alterations to dynamic network rewiring, remain poorly understood.MethodsWe integrated large-scale pharmacogenomic profiling of the FGFR inhibitor AZD4547 from the GDSC2 and PRISM databases with single-cell RNA sequencing to dissect the multi-omics landscape of FGFRi resistance across 312 cell lines from 8 cancer types. This multi-omics framework was further extended by machine learning modeling and systematic synthetic lethality screening to uncover actionable therapeutic targets. In vitro viability assays and western blot analysis were subsequently conducted to experimentally evaluate the predicted FGFR-EGFR synthetic lethality.ResultsOur dual-database analysis unveiled a multi-dimensional atlas of FGFRi resistance. We identified cancer-specific genomic drivers, such as ELF4 amplification in glioblastoma, alongside key transcriptomic markers including UCP2 and FSCN1, highlighting a shift towards metabolic reprogramming and epithelial-mesenchymal transition (EMT). Single-cell analysis unveiled that resistance is linked to the heterogeneous enrichment of baseline subpopulations characterized by distinct metaprograms, including cell-cycle dysregulation. Furthermore, a random forest model built on a LASSO-derived transcriptomic signature was constructed, demonstrating promising predictive capability for AZD4547 sensitivity (mean test-set AUC = 0.73, 95% CI [0.63, 0.80]); the signature generalized well to erdafitinib but showed limited transferability to some other FGFR inhibitors (e.g. pemigatinib, BGJ398). Most notably, our synthetic lethal screening revealed a convergent reliance on compensatory RTK signaling (specifically EGFR pathway enrichment) and downstream MAPK/PI3K cascades in resistant phenotypes, providing converging computational evidence for EGFR pathway activation as an adaptive bypass mechanism. This predicted synthetic lethality was experimentally supported in two FGFR-dependent cell line models (RT112 and CCLP1), in which combined FGFR-EGFR inhibition produced marked synergistic antiproliferative effects.ConclusionsThis study establishes a comprehensive multi-omics atlas of resistance to the FGFR inhibitor AZD4547, delineating convergent mechanisms of metabolic reprogramming and EGFR-mediated bypass signaling. Our findings characterize the resistance as a dynamic network rewiring and nominate rational combination strategies to overcome this therapeutic bottleneck. While FGFR-EGFR co-inhibition is experimentally supported, metabolic co-targeting remains a computationally derived, hypothesis-generating strategy.Supplementary informationThe online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08534-y. Background The Fibroblast Growth Factor Receptor (FGFR) family (FGFR1-4), a subset of the receptor tyrosine kinase (RTK) superfamily, governs critical physiological processes ranging from embryonic development to tissue repair [1, 2]. Upon binding to Fibroblast Growth Factor (FGF) ligands, FGFR initiates cascading reactions of downstream signaling pathways, including PI3K-AKT and RAS-MAPK, thereby regulating cell proliferation, differentiation, migration, and survival [1, 3]. Aberrant activation of FGFR signaling—driven by gene fusion, mutation, or amplification—has been established as a pivotal oncogenic driver across a diverse spectrum of malignancies [2, 4]. Consequently, targeting this axis with small-molecule tyrosine kinase inhibitors (TKIs), such as the selective inhibitor AZD4547, has emerged as a cornerstone of precision oncology [5–9]. Despite the transformative potential of these therapies, the clinical reality remains sobering. While initial responses in patients with FGFR alterations—particularly those with urothelial carcinoma or cholangiocarcinoma—are encouraging, they are often transient [10, 11]. The inevitable emergence of acquired resistance severely limits the durability of response, typically resulting in disease progression within months. Current paradigms of FGFR inhibitor (FGFRi) resistance are predominantly categorized into four mechanisms: (1) Secondary activating mutations in the FGFR kinase domain, leading to impaired inhibitor binding or sustained enhancement of kinase activity [12–14]. (2) Compensatory activation of alternative RTKs (e.g., EGFR or MET), bypassing signal transduction mediated by the FGFR pathway [15, 16]. (3) Reactivation of downstream pathways such as PI3K-AKT and RAS-MAPK, offsetting the growth inhibitory effect after FGFR inhibition [14, 17]. (4) Mutations in tumor suppressor genes such as TP53, regulating cell cycle checkpoints or DNA damage repair capacity and reducing drug sensitivity [18–20]. Among these, RTK pathway remodeling has emerged as a prominent yet complex escape strategy. Notably, Wu et al. demonstrated that compensatory EGFR activation drives resistance specifically in FGFR2-fusion positive cholangiocarcinoma, providing a rationale for dual blockade [15]. However, our understanding of this phenomenon remains fragmented and confined to isolated cancer types. Critical systemic questions remain unanswered: Is this EGFR-mediated bypass a universal vulnerability across the pan-cancer landscape, or is it a context-dependent event? More importantly, does this signaling rewiring operate in isolation, or is it merely one facet of a multimodal adaptive strategy that also involves metabolic reprogramming? Crucially, existing studies predominantly focused on distinct molecular layers in isolation, thereby failing to capture the concurrent evolution of signaling networks and metabolic states. To bridge this gap, we designed a comprehensive multi-omics framework leveraging large-scale pharmacogenomic data from 312 cell lines across eight major cancer types: Bladder Urothelial Carcinoma (BLCA), Breast Invasive Carcinoma (BRCA), Colon Adenocarcinoma (COAD), Esophageal Squamous Cell Carcinoma (ESCC), Glioblastoma (GBM), Lung Adenocarcinoma (LUAD), Melanoma (MEL), and Pancreatic Adenocarcinoma (PAAD). To ensure reliability, we implemented a cross-validation framework utilizing two independent datasets, GDSC2 and PRISM. Beyond traditional bulk-level profiling, we integrated single-cell RNA sequencing (scRNA-seq) to characterize the cellular heterogeneity of resistance at cellular resolution. Finally, we translated these biological insights into actionable strategies by constructing a machine learning-based predictive model and conducting systematic synthetic lethality screening. In this study, we present a multi-dimensional atlas of FGFRi resistance. We find that resistance is associated not only with genetic alterations but also with multilayered rewiring involving metabolic reprogramming (e.g., UCP2 upregulation) and RTK pathway remodeling (notably towards EGFR signaling). By integrating these mechanistic insights with synthetic lethal vulnerability mapping, we propose a rational framework for tiered precision combination therapies, offering new avenues to address the clinical bottleneck of FGFRi resistance. Methods Download of cell line multi-omics data and AZD4547 sensitivity data Multi-dimensional omics data of pan-cancer single cell lines, including gene mutation data, copy number variation (CNV) data, DNA methylation data, transcriptome sequencing data, and proteome quantification data, were systematically downloaded from the DepMap database (https://depmap.org/portal/) and Cell Model Passports (CMP) database (https://cellmodelpassports.sanger.ac.uk/downloads). Meanwhile, drug response data from the Genomics of Drug Sensitivity in Cancer (GDSC2) database (https://www.cancerrxgene.org/) of the Sanger Institute and the PRISM Repurposing Public 24Q2 database (https://depmap.org/portal/prism/) of the Broad Institute were obtained. Proteomic data processing The proteomic data underwent comprehensive preprocessing to ensure comparability across samples. Missing values were processed separately for each cancer type using the R package impute. Specifically, proteins with missing values exceeding 50% in any cohort were removed prior to downstream analysis. The remaining missing entries in the filtered proteomic matrix were imputed via the k-nearest neighbor (KNN) algorithm to enhance data robustness and accuracy. Finally, log2 transformation was applied to stabilize variance and approximate a normal distribution for downstream statistical analyses. FGFR-activated cell line screening criteria FGFR signaling drives tumorigenesis primarily through gain-of-function alterations (amplifications, activating mutations, fusions), which are the canonical indications for FGFR inhibitors. We selected cell lines with these FGFR-activating events for analysis, consistent with clinical trial eligibility criteria for AZD4547 [5, 6, 21]. This design ensures our results are clinically relevant to patient populations enrolled in AZD4547 clinical trials. The screening criteria included: point mutations, covering functional impact mutations such as frameshift_variant, splice_acceptor_variant, splice_donor_variant, start_lost, stop_gained, stop_lost, protein_altering_variant, and missense mutation; copy number variations, gene amplification with copy number ≥ 4; gene fusions, cell lines with FGFR family gene fusion events. Across multiple pan-cancer cohorts, FGFR1 amplification was significantly associated with elevated mRNA expression in both clinical tumor tissues and cell line databases, and FGFR2-3 amplification was similarly correlated with higher expression in TCGA (Figure S1A-D). In the PRISM database, drug sensitivity was represented by the log-fold change (LFC), calculated as: LFC = log2 (viable cell count in the drug-treated group / viable cell count in the untreated control group), such that LFC = 0 represents the natural null point at which drug-treated cells proliferate equivalently to vehicle controls. An LFC < 0 indicates that the drug has an inhibitory effect on the cell line, and a smaller LFC value indicates a stronger inhibitory effect. Based on this metric, cell lines with LFC < 0 were classified as drug-sensitive, and those with LFC > 0 as drug-resistant. To ensure statistical test power and reliable feature extraction, only cancer types with ≥3 cell lines in both the sensitive and resistant groups were retained for subsequent analysis. To mitigate the potential bias from this binary classification, all candidate markers derived from this grouping were further cross-validated using continuous drug response metrics (AUC) from the independent GDSC2 database to ensure the robustness of the findings. After the above screening process, multi-omics data of 312 eligible tumor cell lines were finally obtained, covering 8 cancer types (BLCA, BRCA, COAD, ESCC, GBM, LUAD, MEL, and PAAD). Detailed information of the screened cell lines was provided in Tables S1 and S2. To assess the robustness of findings with respect to the binary classification threshold, three alternative grouping strategies were additionally applied in PRISM: (1) top/bottom 25% quantile thresholds, (2) cancer-type-specific median LFC thresholds, and (3) continuous LFC values (unbinned). For each strategy, differentially expressed genes were re-identified, intersected with GDSC2-derived candidates, and subjected to GSEA enrichment analysis. Drug response analysis in the GDSC2 database For each cancer cohort, Student’s t-test was used to compare the difference in the Area Under the Curve (AUC) of the AZD4547 drug response curve between FGFR-abnormal mutant and wild-type (WT) cell lines (a lower AUC value indicates higher drug sensitivity). p-values were corrected for multiple testing using the Benjamini-Hochberg method. Screening criteria: genes with adjusted p < 0.05 and Fold Change (FC) ≥1.2 were defined as FGFRi resistance-related mutant genes; genes with adjusted p < 0.05 and FC ≤ 0.8 were defined as FGFRi sensitivity-related mutant genes. For the correlation between CNV, mRNA, and protein levels and drug response, Spearman rank correlation analysis was used to detect their correlation with the AZD4547 AUC value, and p-values were corrected using the Benjamini-Hochberg method. Judgment criteria: molecules with adjusted p < 0.05 and Spearman correlation coefficient (r) ≥0.1 were defined as FGFRi resistance-related molecules; molecules with adjusted p < 0.05 and r ≤ −0.1 were defined as FGFRi sensitivity-related molecules. Drug response analysis in the PRISM database Fisher’s exact test was used to compare the distribution difference of gene variations (mutant/wild-type) between the FGFRi resistant and sensitive groups in each cancer cohort, and p-values were corrected using the Benjamini-Hochberg method; genes with adjusted p < 0.05 were considered drug response-related mutant genes, among which genes with significantly enriched mutants in the resistant group were FGFRi resistance-related mutant genes, and the rest were FGFRi sensitivity-related mutant genes. Unpaired Student’s t-test was used to compare the differences in CNV levels, mRNA expression levels, and protein abundances between the resistant and sensitive groups, and p-values were corrected using the Benjamini-Hochberg method. Screening criteria: molecules with adjusted p < 0.05 and FC ≥ 1.2 were defined as FGFRi resistance-related molecules; molecules with adjusted p < 0.05 and FC ≤ 0.8 were defined as FGFRi sensitivity-related molecules. Cross-database screening of high-confidence molecules To ensure the robustness of our findings, we employed a strict cross-validation strategy using GDSC2 and PRISM as independent discovery cohorts. Statistical analyses, including differential expression profiling and drug response correlation, were performed independently within each dataset. Overlapping cell lines between the two datasets were retained to assess technical robustness against experimental variations. High-confidence FGFRi resistance/sensitivity-related molecules were defined based on the following criteria:Consistency in Significance: The feature (mutation, CNV, mRNA, or protein) must be statistically significant (p < 0.05) in both datasets.Concordance in Direction: The feature must exhibit the same direction of association (e.g., consistently upregulated or positively correlated with drug sensitivity) in both cohorts. Only molecular features satisfying these intersection criteria were retained for subsequent functional enrichment and mechanistic analyses. Consistency in Significance: The feature (mutation, CNV, mRNA, or protein) must be statistically significant (p < 0.05) in both datasets. Consistency in Significance: The feature (mutation, CNV, mRNA, or protein) must be statistically significant (p < 0.05) in both datasets. Concordance in Direction: The feature must exhibit the same direction of association (e.g., consistently upregulated or positively correlated with drug sensitivity) in both cohorts. Only molecular features satisfying these intersection criteria were retained for subsequent functional enrichment and mechanistic analyses. Concordance in Direction: The feature must exhibit the same direction of association (e.g., consistently upregulated or positively correlated with drug sensitivity) in both cohorts. Only molecular features satisfying these intersection criteria were retained for subsequent functional enrichment and mechanistic analyses. GO and Reactome pathway enrichment analysis Gene Ontology (GO) and Reactome pathway enrichment analyses were performed on high-confidence FGFRi resistance/sensitivity-related mRNAs and proteins using the R package “BioEnricher” (version 0.1.0). An adjusted p < 0.05 was used as the statistical significance criterion to screen differentially enriched biological pathways. Gene set enrichment analysis (GSEA) GSEA was performed using the R package clusterProfiler (version 4.10.1), with the background gene set selected from 50 hallmark gene sets in the MSIGDB database. A gene ranking list was constructed for each cancer cohort: in the GDSC2 database, genome-wide mRNAs and proteins were included and ranked by their Spearman correlation coefficient with the AZD4547 AUC value; in the PRISM database, genome-wide mRNAs and proteins were included and ranked by the FC value between the resistant and sensitive groups. The analysis parameters were set to 1000 permutations, with a False Discovery Rate (FDR) <0.05 as the significant enrichment threshold; pathways with a Normalized Enrichment Score (NES) >1 were defined as FGFRi resistance-related candidate pathways, and pathways with NES < −1 were defined as FGFRi sensitivity-related candidate pathways. Criteria for defining significant pathways: each pathway was required to obtain four types of enrichment results (GDSC2-mRNA, GDSC2-protein, PRISM-mRNA, PRISM-protein), among which at least 2 types met the significant enrichment criteria, and the functional directions (resistance/sensitivity) of all significant results were consistent to be identified as a significantly enriched pathway related to FGFRi response in that cancer type. The intersection of molecules in significantly enriched pathways with high-confidence FGFRi resistance/sensitivity-related mRNAs and proteins was taken to determine the core regulatory molecules of the FGFRi response phenotype. Regulatory effect of gene mutations on mRNA/protein expression Among high-confidence FGFRi resistance/sensitivity-related mutant genes, the effect of their mutation status on the expression of core regulatory mRNAs and proteins was explored. By integrating SNV, mRNA, and protein data of cell lines from the GDSC2 and PRISM databases, Student’s t-test was used to compare the differences in the expression levels of core regulatory mRNAs and protein abundances between mutant and wild-type cell lines, and p-values were corrected using the Benjamini-Hochberg method; mutation-mRNA/protein combinations with adjusted p < 0.05 were defined as significant regulatory relationships. Correlation analysis of CNV, methylation, transcription factors with mRNA/protein expression To identify potential transcription factors (TFs) driving FGFRi resistance, we utilized the TRRUST v2 database, which provides high-quality, manually curated TF-target regulatory relationships for the human genome. Specifically, we intersected the high-confidence FGFRi resistance/sensitivity-related mRNAs (overlapping hits from GDSC2 and PRISM) with the comprehensive list of human TFs documented in TRRUST. This yielded a prioritized set of resistance-associated TFs for subsequent multi-omics integration. CNV, methylation, and resistance-associated TFs expression data from the GDSC2 and PRISM databases were integrated, and Spearman rank correlation analysis was used to detect the correlation between each CNV, methylation site, TFs expression level and the expression of core regulatory mRNAs/proteins, respectively. After p-value correction by the Benjamini-Hochberg method, significant CNV-mRNA/protein, methylation-mRNA/protein, and transcription factor-mRNA/protein regulatory associations were determined with adjusted p < 0.05 as the criterion. Download of bulk multi-omics data from public cancer patients Multi-omics datasets and matched clinicopathological metadata from The Cancer Genome Atlas (TCGA) were downloaded from the official TCGA data portal (https://gdc.cancer.gov). In contrast, data corresponding to the Clinical Proteomic Tumor Analysis Consortium (CPTAC) cohort were derived from a prior publication by Li et al. [22]. Download and processing of single-cell expression data of pan-cancer cell lines Single-cell expression profiles and clinical annotation data of cancer cell lines were downloaded from the GEO database (GSE157220), which contains single-cell expression data of 205 cancer cell lines (280 cells per cell line). Single-cell expression data were merged with drug sensitivity data from the PRISM database, and quality control and preprocessing were performed using the Seurat software package (version 4.4.0) [23]. The filtering criteria were as follows: cells with the number of detected genes (nFeature_RNA) >200 were retained; cells with nFeature_RNA > 5000 (high-complexity abnormal cells) were removed; cells with mitochondrial gene ratio < 15% were screened; cells with gene expression count (nCount_RNA) >400 were retained; genes expressed in at least 5 cells were retained. After the above filtering, 9997 cells were finally retained for subsequent analysis. The SCTransform function of the Seurat package was used to normalize the single-cell expression data, principal component analysis (PCA) was performed using the RunPCA function, and the top 20 principal components were selected for subsequent cell clustering and dimensionality reduction analysis. Non-linear dimensionality reduction and visualization were performed using the RunTSNE function. Pan-cancer gene metaprograms analysis To mitigate the potential confounding effect of lineage heterogeneity on unsupervised non-negative matrix factorization (NMF), lineage-enriched genes were first curated by analyzing human organ-specific bulk RNA-seq data from the GTEx project [24]. Following the standard criteria from the Human Protein Atlas, genes were defined as “lineage-enriched” if their expression level in a given organ was at least fourfold higher than that in all other organs [25]. Of note, bladder tissue transcriptomic profiles were not available in the GTEx database; given the epithelial homology of the urinary system, kidney tissue was selected as a surrogate to define lineage-enriched genes for bladder cancer (Table S3). Samples were then stratified by cancer type, and the corresponding lineage-enriched genes were removed from the single-cell expression matrix prior to downstream analysis. Gene metaprograms identification was subsequently performed using the R package geneNMF (version 0.9.2). First, the Seurat object was split by sample ID, and multi-k value (k = 4 ~ 9) NMF analysis was performed on each sample using the multiNMF function, and 800 highly variable genes were selected to identify gene programs of different dimensions; the similarity between gene programs was evaluated based on cosine similarity/Jaccard index, and highly similar gene programs were aggregated into consensus metaprograms (MPs) using the getMetaPrograms function, with a weight explanation threshold (weight.explained = 0.6) and a maximum number of genes per metaprogram (max.genes = 300) set to ensure the biological validity of metaprograms; for each candidate number of metaprograms (k = 4 ~ 9), quantitative indicators such as sample coverage, silhouette coefficient, average similarity, number of core genes, and number of original gene program integrations were calculated, and combined with heatmap visualization of the similarity clustering characteristics between metaprograms to determine the optimal number of metaprograms; GO functional enrichment analysis was performed on the core gene sets of each metaprogram, and significantly enriched biological pathways were screened with adjusted p < 0.05 as the threshold; metaprogram scores of individual cells were calculated using two algorithms, the AddModuleScore function embedded in the Seurat package (version 4.4.0) and the AUCell package (version 1.24.0), respectively. Construction of FGFRi sensitivity prediction model based on machine learning To develop a predictive model for FGFRi sensitivity, transcriptomic profiles were integrated with AZD4547 response annotations from the PRISM dataset, stratifying cell lines into drug-sensitive (n = 82) and drug-resistant (n = 54) subsets. The cancer-specific distributions were as follows: BLCA (9 sensitive, 6 resistant), BRCA (5 sensitive, 10 resistant), COAD (10 sensitive, 4 resistant), ESCC (9 sensitive, 5 resistant), GBM (10 sensitive, 7 resistant), LUAD (17 sensitive, 7 resistant), MEL (13 sensitive, 6 resistant), PAAD (9 sensitive, 9 resistant). A 10-times repeated nested holdout validation framework was implemented to ensure rigorous model training and unbiased performance evaluation. For the outer loop, the createDataPartition function from the caret package (version 7.0.1) was utilized to randomly partition the dataset into training (70%), validation (20%), and test (10%) sets. This stratified sampling approach maintained identical proportions of sensitive and resistant samples across all subsets. Within each training set, high-dimensional feature reduction was first implemented utilizing LASSO regression (alpha = 1), where the optimal penalty parameter (lambda.min) was determined via an embedded 10-fold cross-validation. Genes with non-zero coefficients were retained as candidate predictive features. Subsequently, based on the selected candidate features, seven mainstream machine learning algorithms were evaluated for binary classification within the R-based mlr3verse ecosystem [comprising mlr3 (version 1.5.0), mlr3fselect (version 1.5.0), and mlr3tuning (version 1.3.0)]. The candidate algorithms include Random Forest (RF; classif.ranger), Logistic Regression (LogReg; classif.log_reg), Decision Tree (DT; classif.rpart), Extreme Gradient Boosting (XGBoost; classif.xgboost), K-Nearest Neighbor (KNN; classif.kknn), Naive Bayes (classif.naive_bayes), and Multilayer Perceptron (MLP; classif.nnet). Hyperparameter tuning for each classifier was executed via an inner 3-fold cross-validation on the training set to optimize classification accuracy (classif.acc). Optimized models were then retrained on the full 70% training dataset. Crucially, the 20% validation cohort was leveraged for intermediate model selection (benchmarking and choosing the best-performing algorithm among the seven), whereas the holdout 10% test cohort was reserved strictly for the final, unbiased performance evaluation. To eliminate potential predictive bias introduced by the randomness of data partitioning, this entire nested pipeline was repeated independently ten times. Multidimensional evaluation metrics were averaged across the ten replicates to quantify final model efficacy, including the primary metric, area under the receiver operating characteristic curve (AUC; classif.auc), alongside accuracy (ACC), precision, recall, and negative predictive value (NPV). For external validation using the GDSC2 dataset, overlapping cell lines with the PRISM dataset were strictly excluded to guarantee cross-dataset independence. Model calibration curves were assessed via the Score function in the riskRegression package (version 2023.12.21) and visualized using plotCalibration, while Precision-Recall (PR) curves were generated using the ROCR package (version 1.0.11). Screening of synthetic lethal vulnerabilities To identify high-confidence synthetic lethal (SL) partners for the FGF/FGFR family, we developed a systematic computational framework integrating pharmacogenomic profiling, functional genomic screens, and biological relevance filtering. The screening pipeline was conducted across 8 cancer types based on the following steps: Genotype Stratification Genotype Stratification Genotype Stratification For each query gene (designated as Gene A), cell lines within a specific cancer type were stratified into “Gene A-inactive” and “Gene A-active” cohorts. A cell line was classified as “Gene A-inactive” if it met at least one of the following characteristics:i)Deleterious Mutation: Harboring a somatic mutation classified as damaging, including frameshift variants, splice acceptor/donor variants, start lost, stop gained, and stop lost. Missense variants were included only if predicted as “deleterious” by both SIFT and PolyPhen-2 algorithms.ii)Copy Number Loss: Gene copy number falling within the bottom 25th percentile of all cell lines in the corresponding cancer type.iii)Low Expression: mRNA expression levels (TPM) falling within the bottom 25th percentile of all cell lines in the corresponding cancer type. i)Deleterious Mutation: Harboring a somatic mutation classified as damaging, including frameshift variants, splice acceptor/donor variants, start lost, stop gained, and stop lost. Missense variants were included only if predicted as “deleterious” by both SIFT and PolyPhen-2 algorithms. Deleterious Mutation: Harboring a somatic mutation classified as damaging, including frameshift variants, splice acceptor/donor variants, start lost, stop gained, and stop lost. Missense variants were included only if predicted as “deleterious” by both SIFT and PolyPhen-2 algorithms. ii)Copy Number Loss: Gene copy number falling within the bottom 25th percentile of all cell lines in the corresponding cancer type. Copy Number Loss: Gene copy number falling within the bottom 25th percentile of all cell lines in the corresponding cancer type. iii)Low Expression: mRNA expression levels (TPM) falling within the bottom 25th percentile of all cell lines in the corresponding cancer type. Low Expression: mRNA expression levels (TPM) falling within the bottom 25th percentile of all cell lines in the corresponding cancer type. Cell lines failing to meet any of these criteria were classified as “Gene A-active”.2.Pharmacogenomic Interaction Profiling 2.Pharmacogenomic Interaction Profiling Pharmacogenomic Interaction Profiling We utilized drug sensitivity data from the PRISM database to identify compounds exhibiting selective lethality in the Gene A-inactive cells. For each compound, we compared the log-fold change (LFC) values between Gene A-inactive and Gene A-active groups using an unpaired Student’s t-test. Drugs showing significantly lower viability in the inactive group were considered significant hits (Benjamini-Hochberg adjusted p < 0.05). The annotated molecular targets of these compounds were extracted as potential synthetic lethal partners (Candidate Gene B).3.CRISPR-Cas9 Functional Validation 3.CRISPR-Cas9 Functional Validation CRISPR-Cas9 Functional Validation To corroborate pharmacological findings with genetic evidence, we analyzed genome-wide CRISPR-Cas9 gene dependency scores from the DepMap portal. We compared the dependency scores of the Candidate Gene B list between Gene A-inactive and Gene A-active groups. Genes exhibiting a significantly lower dependency score (indicating higher essentiality) in the Gene A-inactive group were prioritized.4.Biological Functional Similarity Filtering 4.Biological Functional Similarity Filtering Biological Functional Similarity Filtering To ensure biological plausibility and reduce false positives, we evaluated the functional relatedness between the query gene (Gene A) and candidate partners (Gene B). Functional similarity scores were calculated based on Gene Ontology (GO) terms using the R package GOSemSim. Only gene pairs with score > 0.3 were regarded as potential SL pairs. Finally, to define a high-confidence set of SL pairs, we applied an intersection approach: only genes that demonstrated consistent lethality in both pharmacogenomic and CRISPR screens, and shared significant functional similarity (score > 0.3) with Gene A, were retained as final SL targets. Cell culture RT112 human bladder cancer cells were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin. CCLP1 human cholangiocarcinoma cells were cultured in DMEM supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin. Both cell lines were maintained at 37 °C in a humidified atmosphere containing 5% CO₂ and were passaged using trypsin-EDTA at 70–80% confluence. RT112 was selected as a representative FGFR3-dependent bladder cancer model harboring the FGFR3–TACC3 fusion, a well-established context for studying bypass-mediated resistance to FGFR inhibition [26]. Concurrently, the intrahepatic cholangiocarcinoma line CCLP1 was selected to evaluate FGFR-EGFR co-targeting, capitalizing on the established paradigm that FGFR inhibition triggers compensatory feedback EGFR activation in biliary malignancies [15]. Of note, both cell lines are FGFR-active parental models that recapitulate the early adaptive or feedback EGFR-bypass response to FGFR inhibition; they are not chronically drug-selected, acquired FGFRi-resistant derivatives. They were therefore used to test the predicted FGFR-EGFR co-dependency and bypass mechanism, rather than to model the reversal of established acquired resistance. Cell viability and drug combination assay RT112 and CCLP1 cells were seeded in 96-well plates in triplicate and allowed to attach overnight. Cells were then treated with AZD4547 (0, 0.25, 0.5, 1, and 2 μM) and osimertinib (0, 0.5, 1, 2, and 4 μM), either alone or in a 5 × 5 dose-response combination matrix, for 48 h. Cell viability was measured using the CCK-8 assay by adding 10 μL of CCK-8 reagent to each well, followed by incubation at 37 °C for 2 h. Absorbance was measured at 450 nm using a microplate reader. Blank-corrected OD values were normalized to vehicle-treated controls to calculate relative cell viability. Drug synergy was analyzed using SynergyFinder (https://synergyfinder.org/). Normalized viability values were uploaded to SynergyFinder, and synergy scores were calculated using four complementary reference models: Bliss independence, highest single agent (HSA), zero interaction potency (ZIP), and Loewe additivity. Synergy scores exceeding 10 were considered indicative of strong synergism, and synergy heatmaps were generated using SynergyFinder. The statistical significance (p-values) of the synergy scores for all four reference models (Bliss, HSA, ZIP, and Loewe) was determined using a standard Student’s t-test within the SynergyFinder platform. This test evaluates whether the mean synergy scores calculated across three independent biological replicates significantly deviate from the null hypothesis of zero interaction (Synergy Score = 0). Western blot analysis RT112 and CCLP1 cells were seeded in 6-well plates and allowed to attach overnight. Cells were then treated with vehicle, AZD4547, osimertinib, or their combination for the indicated time points. For signaling pathway analysis, cells were harvested after 6 h of treatment, whereas apoptosis-related proteins were assessed after 24 h of treatment. After treatment, cells were washed twice with ice-cold PBS and lysed on ice using RIPA lysis buffer supplemented with protease and phosphatase inhibitor cocktails. Whole-cell lysates were collected, centrifuged at 12,000 × g for 15 min at 4 °C, and protein concentrations were determined using a BCA protein assay. Equal amounts of protein were mixed with loading buffer, denatured at 95 °C for 10 min, separated by SDS-PAGE, and transferred onto PVDF membranes. The membranes were blocked with 5% non-fat milk in TBST and incubated overnight at 4 °C with primary antibodies against EGFR, phospho-EGFR, ERK1/2, phospho-ERK1/2, AKT, phospho-AKT, PARP, cleaved PARP, and GAPDH. After washing with TBST, membranes were incubated with HRP-conjugated secondary antibodies at room temperature. Protein bands were visualized using an enhanced chemiluminescence detection system and quantified using ImageJ software. Phosphorylated protein levels were normalized to their corresponding total protein levels, whereas total protein and apoptosis-related proteins were normalized to GAPDH. Statistical analysis Statistical analysis was performed with R software (version 4.3.2). Spearman’s correlation coefficients were utilized to assess the associations between two continuous variables. For categorical data, Fisher’s exact test was applied, and for continuous variables, the Wilcoxon rank-sum test or the t-test was used for comparison. The p-value correction was performed using the Benjamini-Hochberg method, with statistical significance defined as an adjusted p < 0.05. Ethical consideration This study utilized publicly available, de-identified datasets from the GEO (GSE157220), DepMap, GDSC2, PRISM, TCGA, and CPTAC repositories. All original data collection and sharing protocols were conducted in accordance with the ethical standards and informed consent requirements of the respective source institutions. Since this research involved only the secondary analysis of anonymized public data, further institutional review board (IRB) approval was not required. Results Study cohort and multi-omics profiling To systematically dissect the molecular mechanisms underlying FGFRi resistance, we constructed a comprehensive multi-omics framework integrating pharmacogenomic data from the GDSC2 and PRISM databases (Fig. 1A). Our analysis encompassed somatic mutations, DNA methylation, CNVs, transcriptomics, and proteomics profiles across 312 cell lines spanning the eight cancer lineages (Fig. 1B, C). To provide clinical and prognostic context, we leveraged patient cohorts from TCGA and CPTAC to validate the prognostic significance of candidate biomarkers identified from these bulk profiles. Beyond bulk analysis, we incorporated scRNA-seq data to capture resistance heterogeneity at cellular resolution. Furthermore, to translate these multi-omics insights into therapeutic strategies, we developed a machine learning-based predictive model and performed systematic synthetic lethality screening. The overall study design and the multi-dimensional data availability are illustrated in Fig. 1A–C.Fig. 1Study design and multi-omics landscape of FGFRi resistance. (A) a systematic framework integrating pharmacogenomic profiling (GDSC2 and PRISM) and scRNA-seq was established to dissect FGFRi resistance. The study follows three parallel streams: (left) development of a machine learning-based predictive model for drug sensitivity; (middle) construction of a multi-dimensional resistance atlas, highlighting EGFR pathway activation and metabolic reprogramming (UCP2) as key drivers, validated in TCGA and CPTAC clinical cohorts; (right) Identification of synthetic lethal vulnerabilities to propose actionable combination therapies. The dark blue box at the bottom illustrates the proposed precision medicine framework: 1. Stratification via the LASSO-derived transcriptomic signature model; 2. Molecular profiling to identify resistance mechanisms (Rtk/metabolic); and 3. Selection of rational combination therapies. (B, C) Availability of multi-omics profiles (mutations, CNVs, methylation, mRNA transcripts, and proteomic data) across cancer cohorts in the GDSC2 (B) and PRISM (C) databases Study design and multi-omics landscape of FGFRi resistance. (A) a systematic framework integrating pharmacogenomic profiling (GDSC2 and PRISM) and scRNA-seq was established to dissect FGFRi resistance. The study follows three parallel streams: (left) development of a machine learning-based predictive model for drug sensitivity; (middle) construction of a multi-dimensional resistance atlas, highlighting EGFR pathway activation and metabolic reprogramming (UCP2) as key drivers, validated in TCGA and CPTAC clinical cohorts; (right) Identification of synthetic lethal vulnerabilities to propose actionable combination therapies. The dark blue box at the bottom illustrates the proposed precision medicine framework: 1. Stratification via the LASSO-derived transcriptomic signature model; 2. Molecular profiling to identify resistance mechanisms (Rtk/metabolic); and 3. Selection of rational combination therapies. (B, C) Availability of multi-omics profiles (mutations, CNVs, methylation, mRNA transcripts, and proteomic data) across cancer cohorts in the GDSC2 (B) and PRISM (C) databases The genomic landscape of FGFRi resistance reveals context-specific alterations To characterize the genomic features associated with FGFRi sensitivity, we systematically profiled somatic mutations, DNA methylation patterns, and CNVs across diverse cancer lineages. Profiling of the GDSC2 MEL cohort revealed that cell lines harboring mutations in ARHGAP33, PDZD2, and UGT3A2 exhibited significantly reduced sensitivity (higher AUC values) to AZD4547 (Fig. 2A). Subsequent validation in the PRISM dataset corroborated this association, demonstrating a significant enrichment of these mutant genotypes specifically within the resistant subgroup (Fig. 2B). Given the known roles of ARHGAP33 in Rho signaling and UGT3A2 in drug metabolism, these alterations suggest potential mechanisms involving cytoskeletal dynamics and metabolic clearance, warranting further validation.Fig. 2The genomic landscape of FGFRi resistance reveals context-specific alterations. (A) Box plots illustrating the differences in AZD4547 AUC values between wild-type and mutant cell lines of target genes in specific cancer types from the GDSC2 database (MEL-ARHGAP33: n (WT) = 19, n (mut) = 3; MEL-PDZD2: n (WT) = 19, n (mut) = 3; MEL-UGT3A2: n (WT) = 18, n (mut) = 4; LUAD-TENM4: n (WT) = 28, n (mut) = 3; ESCC-APC2: n (WT) = 9, n (mut) = 3). (B) Stacked bar charts showing the distribution frequency discrepancies of wild-type and mutant genotypes of specific genes in FGFRi-sensitive (S) versus resistant (R) cell line groups across cancer types in the PRISM database. (C) Correlation scatter plots demonstrating the association between gene methylation levels and AZD4547 AUC values in designated cancer types from the GDSC2 database. (D) Box plots comparing the methylation levels of target genes between FGFRi-sensitive and resistant cell line cohorts of specific cancer types in the PRISM database (BRCA-EN1, HAAO, HOXC6, UCHL1: n(S) = 4, n(R) = 8; PAAD-KRTCAP3: n(S) = 6, n(R) = 6). (E) Analyses of CNV levels: Student’s t-test was applied to quantify CNV differences between sensitive and resistant groups for each cancer type in the PRISM database; Spearman’s rank correlation analysis was conducted to evaluate the association between CNV levels and AZD4547 AUC values in the GDSC2 database The genomic landscape of FGFRi resistance reveals context-specific alterations. (A) Box plots illustrating the differences in AZD4547 AUC values between wild-type and mutant cell lines of target genes in specific cancer types from the GDSC2 database (MEL-ARHGAP33: n (WT) = 19, n (mut) = 3; MEL-PDZD2: n (WT) = 19, n (mut) = 3; MEL-UGT3A2: n (WT) = 18, n (mut) = 4; LUAD-TENM4: n (WT) = 28, n (mut) = 3; ESCC-APC2: n (WT) = 9, n (mut) = 3). (B) Stacked bar charts showing the distribution frequency discrepancies of wild-type and mutant genotypes of specific genes in FGFRi-sensitive (S) versus resistant (R) cell line groups across cancer types in the PRISM database. (C) Correlation scatter plots demonstrating the association between gene methylation levels and AZD4547 AUC values in designated cancer types from the GDSC2 database. (D) Box plots comparing the methylation levels of target genes between FGFRi-sensitive and resistant cell line cohorts of specific cancer types in the PRISM database (BRCA-EN1, HAAO, HOXC6, UCHL1: n(S) = 4, n(R) = 8; PAAD-KRTCAP3: n(S) = 6, n(R) = 6). (E) Analyses of CNV levels: Student’s t-test was applied to quantify CNV differences between sensitive and resistant groups for each cancer type in the PRISM database; Spearman’s rank correlation analysis was conducted to evaluate the association between CNV levels and AZD4547 AUC values in the GDSC2 database Across other cancer lineages, resistant phenotypes were characterized by distinct genomic signatures, such as TENM4 mutations in LUAD and APC2 mutations in ESCC, which were similarly associated with diminished drug response (Fig. 2A, B). Notably, as APC2 is a component of the Wnt destruction complex, its mutational enrichment implicates aberrant Wnt signaling as a potential feature of the resistant phenotype. Beyond point mutations, epigenetic and structural analyses revealed multi-layered heterogeneity. In BRCA, resistance was associated with higher methylation levels of homeobox genes like EN1 and HOXC6 (Fig. 2C, D), suggesting the epigenetic silencing of these developmental regulators. Conversely, the resistant landscape of GBM was marked by recurrent copy number amplifications. We identified 403 resistance-correlated CNVs in GBM, including the amplification of ELF4 and ARHGEF6 (Fig. 2E). Previous studies have linked ELF4 to stemness maintenance in GBM, and ARHGEF6 to tumor progression in various cancers [27–30], suggesting that the amplification-driven upregulation of these oncogenic drivers may support the aggressive resistant state. Transcriptomic and proteomic signatures of FGFRi response To identify high-confidence molecular signatures associated with FGFRi response, we performed an integrated analysis of pharmacogenomic data from the GDSC2 and PRISM cohorts. We specifically screened for mRNA signatures that demonstrated consistent drug response associations between the two datasets within each specific cancer lineage, thereby yielding a high-confidence panel of biomarkers (Fig. 3A, Figure S2A, B). To ensure our results were independent of the specific grouping thresholds in PRISM, we applied three alternative classification strategies: cancer-type-specific median thresholds, top/bottom 25% quantile thresholds, and continuous LFC values (unbinned). The results showed high overlap across these strategies (Figure S3, Table S4), supporting the reproducibility of our biomarker screening.Fig. 3Identification of FGFRi response-related mRNAs and proteins. (A) Scatter plot displaying representative FGFRi response-related mRNAs consistently identified from the GDSC2 and PRISM databases. (B) Heatmap illustrating the correlation between specific mRNA expression levels and AZD4547 AUC values (GDSC2 database), as well as mRNA expression Fold changes between FGFRi-resistant and sensitive groups across cancer types (PRISM database). Statistical significance is indicated: *p < 0.05; **p < 0.01; ***p < 0.001. The color scale represents mRNA expression Fold change (resistant vs. sensitive) and the correlation coefficient between mRNA expression and AZD4547 AUC values. (C, D) Volcano plots depicting FGFRi response-related proteins identified from the GDSC2 (C) and PRISM (D) databases, respectively. (E) Heatmap showcasing the correlation between specific protein abundances and AZD4547 AUC values (GDSC2 database), along with protein abundance Fold changes between FGFRi-resistant and sensitive groups across cancer types (PRISM database). Statistical significance is denoted: *p < 0.05; **p < 0.01; ***p < 0.001. The color scale indicates protein abundance Fold change (resistant vs. sensitive) and the correlation coefficient between protein abundance and AZD4547 AUC values Identification of FGFRi response-related mRNAs and proteins. (A) Scatter plot displaying representative FGFRi response-related mRNAs consistently identified from the GDSC2 and PRISM databases. (B) Heatmap illustrating the correlation between specific mRNA expression levels and AZD4547 AUC values (GDSC2 database), as well as mRNA expression Fold changes between FGFRi-resistant and sensitive groups across cancer types (PRISM database). Statistical significance is indicated: *p < 0.05; **p < 0.01; ***p < 0.001. The color scale represents mRNA expression Fold change (resistant vs. sensitive) and the correlation coefficient between mRNA expression and AZD4547 AUC values. (C, D) Volcano plots depicting FGFRi response-related proteins identified from the GDSC2 (C) and PRISM (D) databases, respectively. (E) Heatmap showcasing the correlation between specific protein abundances and AZD4547 AUC values (GDSC2 database), along with protein abundance Fold changes between FGFRi-resistant and sensitive groups across cancer types (PRISM database). Statistical significance is denoted: *p < 0.05; **p < 0.01; ***p < 0.001. The color scale indicates protein abundance Fold change (resistant vs. sensitive) and the correlation coefficient between protein abundance and AZD4547 AUC values To provide clinical and prognostic context for the identified candidate markers, we mapped these candidate genes to patient data from the TCGA and CPTAC cohorts (Fig. 3B). This comprehensive landscape revealed that many identified markers exhibited tumor-specific differential expression and significant prognostic value. Specifically, a subset of resistance-associated transcripts—highlighted by UCP2, COL11A1, PRAME, and UNC5A—demonstrated a striking multi-dimensional consistency. These genes were not only correlated with AZD4547 resistance but were also identified as prognostic risk factors in patient cohorts. Notably, markers such as COL11A1 and PRAME were also significantly upregulated in tumor tissues compared to normal controls. Conversely, we identified a panel of sensitivity-associated markers, such as ATOH8, ATP8B1, LIMCH1, and NTN4. These genes formed a distinct cluster characterized by negative drug response correlations and favorable patient prognosis. We further performed survival analyses in patient subsets with FGFR amplification. Consistently, higher expression of COL9A3 (TCGA-BLCA), UNC5A (TCGA-BRCA), and DKK1 (TCGA-LUAD and CPTAC-LUAD) was associated with poorer prognosis in FGFR-amplified patients (Figure S2C). Complementing the transcriptomic landscape, we further leveraged proteomic profiling to capture functional alterations. Notably, the identified proteomic signatures showed minimal overlap with the transcriptomic markers (Fig. 3C, D, Figure S2D), indicating the prevalence of post-transcriptional regulation in driving resistance. This analysis highlighted distinct protein-level markers, such as TM9SF4 and TOM1L1, which were consistently upregulated in resistant phenotypes across different databases and thresholds (Fig. 3C, D; Figure S4; Table S5). Clinical validation in the CPTAC dataset confirmed these candidates were significantly overexpressed in tumor tissues (Fig. 3E). Crucially, our proteomic screen also highlighted ADAM17 as a resistance marker. Given its established role in shedding EGFR ligands, the upregulation of ADAM17 provides proteomic evidence corroborating the putative compensatory EGFR activation mechanism toward EGFR signaling. Furthermore, phosphoproteomic analysis showed significantly increased phosphorylation of EGFR (Tyr1068) and AKT (Ser473) in AZD4547-resistant BRCA and COAD cell lines compared with sensitive cells, supporting functional activation of EGFR bypass signaling in drug resistance (Figure S2E, F). Lineage-specific and convergent functional landscapes of FGFRi resistance To transition from discrete molecular markers to systems-level resistance mechanisms, we performed GO and Reactome functional enrichment analyses on the mRNA and protein signatures identified from our integrative screen. This initial profiling revealed that the molecular landscape of resistance is characterized by lineage-dependent adaptive programs rather than a uniform pan-cancer mechanism (Figure S5A-D). For instance, transcriptomic signatures in COAD were predominantly associated with metabolic pathways, such as ‘Respiratory electron transport’ (Figure S5B), whereas BLCA exhibited proteomic enrichment in structural modules like ‘Assembly of collagen fibrils’ (Figure S5D). To ensure the reliability of these lineage-specific findings and mitigate potential platform- or omics-level discordance, we performed GSEA across four independent discovery layers: GDSC2-mRNA/protein and PRISM-mRNA/protein. To prioritize high-confidence biological modules, we implemented a stringent consensus criterion, requiring significant enrichment in at least two data layers with consistent functional polarity. This integrative approach identified six convergent functional modules associated with FGFRi resistance: metabolic reprogramming, cell-cycle dysregulation, estrogen response, immunity & inflammation, signaling transduction, and invasion & migration (Fig. 4A; Table S6). Using three alternative cutoff strategies (top/bottom 25% quantile thresholds, cancer-type-specific median LFC thresholds, and continuous LFC values) for PRISM data, the GSEA consistently converged on the same six functional modules across all classification approaches (Figure S5E-F), confirming that the identified resistance programs are robust and independent of the initial binary threshold.Fig. 4Identification of FGFRi response-related pathways. (A) Heatmap of GSEA results from four datasets (GDSC2-mRNA, GDSC2-protein, PRISM-mRNA, PRISM-protein). *FDR < 0.05; **FDR < 0.01; ***FDR < 0.001. Color gradient indicates Normalized Enrichment Score (NES). (B) Left: correlation scatter plots of ATP1B3 and UCP2 mRNA levels vs. AZD4547 AUC in the COAD cohort (GDSC2). Right: box plots of ATP1B3 and UCP2 mRNA levels in FGFRi-sensitive (n = 10) vs. resistant (n = 4) groups (COAD, PRISM). (C) Survival analyses of UCP2 high- vs. low-expression subgroups in the CPTAC-COAD (left) and TCGA-COAD (right) cohorts. (D) Left: correlation scatter plot of FSCN1 mRNA vs. AZD4547 AUC in the BLCA cohort (GDSC2). Right: box plot of FSCN1 mRNA in FGFRi-sensitive (n = 9) vs. resistant (n = 6) groups (BLCA, PRISM). (E) Survival analysis of FSCN1 high- vs. low-expression subgroups in the TCGA-BLCA cohort. (F) Left: correlation scatter plot of MCM7 mRNA vs. AZD4547 AUC in the PAAD cohort (GDSC2). Right: box plot of MCM7 mRNA in FGFRi-sensitive (n = 9) vs. resistant (n = 9) groups (PAAD, PRISM). (G) Survival analyses of MCM7 high- vs. low-expression subgroups in the CPTAC-PAAD (left) and TCGA-PAAD (right) cohorts. (H) Left: correlation scatter plots of HPRT1 and MCM4 mRNA levels vs. AZD4547 AUC in the GBM cohort (GDSC2). Right: box plots of HPRT1 and MCM4 mRNA in FGFRi-sensitive (n = 10) vs. resistant (n = 7) groups (GBM, PRISM). (I) Survival analysis of MCM4 high- vs. low-expression subgroups in the TCGA-GBM cohort. (J) Survival analyses of HPRT1 high- vs. low-expression subgroups in the CPTAC-GBM (left) and TCGA-GBM (right) cohorts Identification of FGFRi response-related pathways. (A) Heatmap of GSEA results from four datasets (GDSC2-mRNA, GDSC2-protein, PRISM-mRNA, PRISM-protein). *FDR < 0.05; **FDR < 0.01; ***FDR < 0.001. Color gradient indicates Normalized Enrichment Score (NES). (B) Left: correlation scatter plots of ATP1B3 and UCP2 mRNA levels vs. AZD4547 AUC in the COAD cohort (GDSC2). Right: box plots of ATP1B3 and UCP2 mRNA levels in FGFRi-sensitive (n = 10) vs. resistant (n = 4) groups (COAD, PRISM). (C) Survival analyses of UCP2 high- vs. low-expression subgroups in the CPTAC-COAD (left) and TCGA-COAD (right) cohorts. (D) Left: correlation scatter plot of FSCN1 mRNA vs. AZD4547 AUC in the BLCA cohort (GDSC2). Right: box plot of FSCN1 mRNA in FGFRi-sensitive (n = 9) vs. resistant (n = 6) groups (BLCA, PRISM). (E) Survival analysis of FSCN1 high- vs. low-expression subgroups in the TCGA-BLCA cohort. (F) Left: correlation scatter plot of MCM7 mRNA vs. AZD4547 AUC in the PAAD cohort (GDSC2). Right: box plot of MCM7 mRNA in FGFRi-sensitive (n = 9) vs. resistant (n = 9) groups (PAAD, PRISM). (G) Survival analyses of MCM7 high- vs. low-expression subgroups in the CPTAC-PAAD (left) and TCGA-PAAD (right) cohorts. (H) Left: correlation scatter plots of HPRT1 and MCM4 mRNA levels vs. AZD4547 AUC in the GBM cohort (GDSC2). Right: box plots of HPRT1 and MCM4 mRNA in FGFRi-sensitive (n = 10) vs. resistant (n = 7) groups (GBM, PRISM). (I) Survival analysis of MCM4 high- vs. low-expression subgroups in the TCGA-GBM cohort. (J) Survival analyses of HPRT1 high- vs. low-expression subgroups in the CPTAC-GBM (left) and TCGA-GBM (right) cohorts Further analysis revealed that these resistance programs are highly context-dependent, manifesting as distinct molecular signatures across specific cancer lineages. In COAD, the resistant phenotype displayed a multidimensional functional profile encompassing metabolic reprogramming (e.g., adipogenesis), signaling transduction (e.g., mTORC1 signaling), and immunity & inflammation (e.g., interferon-gamma response) (Fig. 4A, Figure S6A). Within these pathways, ATP1B3 and UCP2 were prioritized as high-confidence regulatory components representing the mTORC1 signaling and adipogenesis programs, respectively. These markers exhibited concordant associations with resistance, characterized by positive drug-response correlations in GDSC2 and significantly elevated expression in the PRISM resistant cohort (Fig. 4B). Furthermore, UCP2 was validated as a significant prognostic risk factor in both CPTAC and TCGA datasets (Fig. 4C), aligning with its documented role in modulating mitochondrial bioenergetics and reinforcing the clinical relevance of mitochondrial metabolic adaptation [31, 32]. Collectively, these findings suggest that resistant cells in COAD may exploit metabolic plasticity and signaling rewiring to elevate the threshold for drug-induced cell death. Conversely, the resistance landscape in BLCA was marked by an invasion & migration signature, primarily involving Epithelial-Mesenchymal Transition (EMT) and apical junctions (Fig. 4A, Figure S6B). The actin-bundling protein FSCN1 was identified as a core component of this phenotype; its expression was positively correlated with FGFRi resistance and was identified as a significant prognostic risk factor in bladder cancer patients (Fig. 4D, E). The prominence of FSCN1 suggests that resistant cells may leverage cytoskeletal remodeling to enhance cellular plasticity and invasive potential, thereby facilitating therapeutic evasion. Additionally, the resistant phenotype in BRCA was significantly linked to estrogen response pathways and elevated MLPH expression (Fig. 4A, Figure S6C, D). Despite these lineage-specific variations, a convergent reliance on cell-cycle progression was identified as a shared hallmark of resistance in PAAD and GBM. These resistant phenotypes were significantly enriched for MYC targets, E2F targets, and G2M checkpoint programs (Fig. 4A, Figure S6E, F). In PAAD, MCM7 expression was positively correlated with FGFRi resistance and identified as a significant prognostic risk factor (Fig. 4F, G). Similarly, the resistant state in GBM was characterized by the upregulation of genes involved in cell-cycle progression and nucleotide metabolism, most notably MCM4 and HPRT1 (Fig. 4H, Figure S6G, H). Both markers exhibited striking associations with unfavorable outcomes in CPTAC and TCGA-GBM cohorts (Fig. 4I, J). Given the fundamental role of the Minichromosome Maintenance (MCM) family in DNA replication initiation [33], these findings point toward a link between FGFRi resistance and a hyper-proliferative state in PAAD and GBM, likely supported by the coordinated activation of cell-cycle and metabolic programs. Distinct multi-omics regulatory landscapes of FGFRi resistance To elucidate how genomic aberrations propagate across molecular hierarchies to shape the FGFRi resistance phenotype, we systematically analyzed the correlations between high-confidence somatic alterations (mutations, CNVs) and epigenetic features (DNA methylation) with downstream transcriptome and proteome profiles. This analysis revealed highly heterogeneous regulatory patterns, where distinct modes of signaling rewiring explain the lineage-specific phenotypes observed earlier. In GBM, we uncovered an extensive regulatory network integrating both CNV effects and TF associations. The analysis detected 410 CNV-mRNA and 1,677 CNV-protein regulatory pairs (Table S7), underscoring the pervasiveness of CNV impact in GBM. Beyond this genomic baseline, the network was further orchestrated by specialized transcriptional programs. Notably, the TFs CTCF, HDAC1, and RBMX emerged as central regulatory hubs. Among them, RBMX exhibited strong positive correlations with both prognostic risk factors MCM4 and HPRT1, whereas CTCF and HDAC1 were primarily linked to the regulation of MCM4 (Fig. 5A). This hierarchical flow suggests that CNV-TF coordination may sustain the hyper-proliferative state and nucleotide metabolic reprogramming associated with the FGFRi-resistant phenotype in GBM.Fig. 5Distinct multi-omics regulatory landscapes governing FGFRi resistance. (A) Heatmap exhibiting the transcriptional regulatory network in GBM (resistant phenotype), showing Spearman correlation coefficients between TF (columns) and target mRNA expression (rows). (B) Heatmap exhibiting the multi-omics associations in BLCA, including the impact of CNVs on protein and mRNA levels, as well as TF-mRNA regulatory pairs. (C) Epigenetic regulation in BRCA, highlighting the correlation between DNA methylation and gene expression. (D) Transcriptional regulatory patterns in PAAD, showing associations between TFs and downstream mRNA targets. (E) Integrated multi-omics interaction network visualizing core regulatory pathways. For all heatmaps, the color gradient reflects the correlation coefficient (Pink: positive correlation; Green: negative correlation). Statistical significance is denoted by asterisks: *p < 0.05; **p < 0.01; ***p < 0.001 Distinct multi-omics regulatory landscapes governing FGFRi resistance. (A) Heatmap exhibiting the transcriptional regulatory network in GBM (resistant phenotype), showing Spearman correlation coefficients between TF (columns) and target mRNA expression (rows). (B) Heatmap exhibiting the multi-omics associations in BLCA, including the impact of CNVs on protein and mRNA levels, as well as TF-mRNA regulatory pairs. (C) Epigenetic regulation in BRCA, highlighting the correlation between DNA methylation and gene expression. (D) Transcriptional regulatory patterns in PAAD, showing associations between TFs and downstream mRNA targets. (E) Integrated multi-omics interaction network visualizing core regulatory pathways. For all heatmaps, the color gradient reflects the correlation coefficient (Pink: positive correlation; Green: negative correlation). Statistical significance is denoted by asterisks: *p < 0.05; **p < 0.01; ***p < 0.001 In contrast, other cancer cohorts exhibited more focused regulatory profiles. In BLCA, we observed a discrete multi-dimensional landscape involving 14 CNV-mRNA, 9 CNV-protein, and 8 TF-mRNA associations. A notable genomic-transcriptomic link was the significant negative correlation between CHEK2P2 copy number and the expression of the resistance marker NT5E (Fig. 5B). Furthermore, we identified the transcription factor EHMT2 as a potential regulator of the actin-bundling protein FSCN1 (Fig. 5B). This EHMT2–FSCN1 axis may provide a mechanistic basis for the EMT and pro-invasive phenotype previously linked to FGFRi resistance in bladder cancer patients. Meanwhile, BRCA resistance was uniquely characterized by epigenetic modulation. Specifically, the expression of the resistance-associated marker MLPH exhibited significant positive correlations with the methylation levels of HAAO and EN1 (Fig. 5C). In PAAD, a convergent regulatory mode was identified, featuring 10 TF-mRNA associations. Notably, the transcriptional co-activator TRRAP emerged as a pivotal regulator, significantly orchestrating the mRNA expression of ABL1 and MCM7 (Fig. 5D). Finally, these core regulatory molecules were synthesized into a unified multi-omics interaction network (Fig. 5E), providing a holistic view of the coordinated rewiring of genomic, epigenetic, and transcriptional landscapes in FGFRi resistance. Single-cell profiling reveals heterogeneous transcriptomic metaprograms underlying lineage-specific FGFRi resistance To characterize the cellular heterogeneity of AZD4547 resistance, we transitioned from bulk-level profiling to scRNA-seq analysis across a pan-cancer atlas. Following a stringent quality control pipeline, a single-cell landscape of 9,997 individual cells was established, encompassing diverse lineages and distinct drug-response phenotypes (Fig. 6A–C). To identify conserved biological states across these heterogeneous cell lines, we utilized NMF to deconvolve the transcriptomic data into consensus gene metaprograms (MPs). This analysis derived four recurrent MPs representing distinct oncogenic processes: inflammatory immune response (MP1), cell cycle (MP2), protein modification homeostasis (MP3), and epithelial mesenchymal transition (MP4) (Fig. 6D, E).Notably, the activity landscapes of these MPs exhibited lineage-dependent characteristics, with MP2 preferentially activated in MEL and GBM cells, and MP3 showing prominent activation in COAD and BRCA lineages at the baseline state (Fig. 6F). Importantly, the enrichment of these MPs within resistant phenotypes was highly context-dependent as assessed by two independent scoring algorithms (AUCell and AddModuleScore). Specifically, MP1 activity was overrepresented in resistant COAD populations, MP2 was enriched in resistant subcohorts of PAAD and LUAD, MP3 activity was overrepresented in resistant MEL populations, and MP4 activity was overrepresented in resistant MEL and GBM populations (Fig. 6G, H). Given the pivotal role of cell-cycle deregulation in cancer drug resistance, we next focused specifically on the intra-lineage heterogeneity of MP2 to dissect its subpopulation dynamics. MP2 scores varied significantly across cell lines (Figure S7A, B), with MEL cell lines showing globally elevated MP2 activity. The coexistence of MP2-high and MP2-low subpopulations within the same cell line was supported by the fraction of MP2-high cells observed within individual cell lines (Figure S7C, D) and single-cell level visualization of MP2 score distributions (Figure S7E). The enrichment of MP2 within resistant PAAD and LUAD subcohorts, together with the coexistence of MP2-high and MP2-low subpopulations within individual cell lines, points to a clear baseline heterogeneity in cell-cycle states. Although these cross-sectional data cannot definitively differentiate whether this heterogeneity reflects the clonal selection of pre-existing subpopulations, dynamic transcriptional plasticity, or asynchronous cell-cycle variation in vitro, they underscore intrinsic cell-cycle stratification as a notable feature linked to FGFRi resistance in specific contexts. Furthermore, the upregulation of MP3 in resistant MEL suggests that protein-modification homeostasis may constitute a key survival adaptation under therapeutic stress.Fig. 6Identification of gene MPs linked to FGFRi response. (A) t-SNE visualization of cell lines from distinct cancer types (colored by cancer type). (B) Pie chart depicting the proportion of cell counts across various cancer types. (C) Bar chart showing the distribution of cell counts in FGFRi-resistant and sensitive groups across cancer types. (D) NMF results of pan-cancer cell line genes and heatmap analysis of four gene MPs. The number of core genes for each metaprogram: MP1: n = 3; MP2: n = 32; MP3: n = 68; MP4: n = 52. (E) GO enrichment analysis of genes in each metaprogram. (F) Bubble chart of metaprogram scores computed via the AUCell algorithm; (G, H) Heatmaps of average metaprogram scores in FGFRi-sensitive and resistant groups across cancer types, calculated using the AUCell algorithm (G) and AddModuleScore algorithm (H), respectively. BLCA: n(S) = 284, n(R) = 641; BRCA: n(S) = 245, n(R) = 1420); COAD: n(S) = 529, n(R) = 166; ESCC: n(S) = 886, n(R) = 179; GBM: n(S) = 789, n(R) = 226; LUAD: n(S) = 985, n(R) = 376; MEL: n(S) = 2547, n(R) = 162; PAAD: n(S) = 222, n(R) = 340 Identification of gene MPs linked to FGFRi response. (A) t-SNE visualization of cell lines from distinct cancer types (colored by cancer type). (B) Pie chart depicting the proportion of cell counts across various cancer types. (C) Bar chart showing the distribution of cell counts in FGFRi-resistant and sensitive groups across cancer types. (D) NMF results of pan-cancer cell line genes and heatmap analysis of four gene MPs. The number of core genes for each metaprogram: MP1: n = 3; MP2: n = 32; MP3: n = 68; MP4: n = 52. (E) GO enrichment analysis of genes in each metaprogram. (F) Bubble chart of metaprogram scores computed via the AUCell algorithm; (G, H) Heatmaps of average metaprogram scores in FGFRi-sensitive and resistant groups across cancer types, calculated using the AUCell algorithm (G) and AddModuleScore algorithm (H), respectively. BLCA: n(S) = 284, n(R) = 641; BRCA: n(S) = 245, n(R) = 1420); COAD: n(S) = 529, n(R) = 166; ESCC: n(S) = 886, n(R) = 179; GBM: n(S) = 789, n(R) = 226; LUAD: n(S) = 985, n(R) = 376; MEL: n(S) = 2547, n(R) = 162; PAAD: n(S) = 222, n(R) = 340 Together, these findings highlight a complex, lineage-specific transcriptional landscape associated with FGFRi resistance. Distinct cancer lineages employ different survival strategies—hyper-proliferation, inflammatory/immune responses, protein-modification homeostasis, and EMT—to collectively elevate the cellular survival threshold under FGFR inhibition. Development of a biologically grounded machine learning model for predicting FGFRi response To translate our discovery into clinical utility, we developed a transcriptome-based predictive model to stratify FGFRi sensitivity. The 10-times repeated nested holdout validation framework was used to benchmark model efficacy under a limited cohort size. In each of the ten independent runs, the PRISM discovery cohort was partitioned into a training set (70%), a validation set (20%), and a holdout test set (10%) via stratified random sampling (Fig. 7A). Within each training partition, high-dimensional feature reduction was implemented utilizing LASSO regression to determine an optimized candidate feature set. Subsequently, hyperparameter tuning for seven machine learning algorithms was executed via an inner 3-fold cross-validation loop. Benchmarking of seven algorithms identified random forest as the top-performing classifier, achieving a mean AUC of 0.69 (95% CI [0.57, 0.81]) in the internal validation set and 0.73 (95% CI [0.63, 0.80]) in the independent test set (Fig. 7B, C). In addition to the AUC, the model demonstrated acceptable performance across other evaluation metrics, including precision, recall, accuracy, and NPV, within both cohorts (Fig. 7D, E). The generalizability of this framework was further corroborated via external validation in the GDSC2 database, where cell lines predicted as resistant by the model exhibited significantly higher drug response values than those predicted as sensitive (Fig. 7F). Calibration curves aggregated across runs showed modest deviation from the ideal diagonal (Figure S8A), and the model achieved a mean test-set area under the precision-recall curve (AUPR) of 0.67 (Figure S8B), indicating acceptable but not excellent precision-recall performance.Fig. 7Development of a machine learning model for predicting FGFRi response. (A) Workflow of the prediction model construction. (B, C) Mean ROC curves demonstrating the predictive performance of each model in the validation set (B) and test set (C). (D, E) Heatmaps showing averaged performance metrics from 10 repeated modeling runs of seven models on the validation dataset. (D) and test set (E). (F) Box plot comparing the AZD4547 AUC values of cell lines predicted as FGFRi-sensitive or resistant by the RF model in the GDSC2 database across 10 repeated modeling runs. (G) Top 10 genes with the highest LASSO selection frequency (left) and top 10 genes sorted by averaged random forest feature importance (right) from ten repeated modeling runs Development of a machine learning model for predicting FGFRi response. (A) Workflow of the prediction model construction. (B, C) Mean ROC curves demonstrating the predictive performance of each model in the validation set (B) and test set (C). (D, E) Heatmaps showing averaged performance metrics from 10 repeated modeling runs of seven models on the validation dataset. (D) and test set (E). (F) Box plot comparing the AZD4547 AUC values of cell lines predicted as FGFRi-sensitive or resistant by the RF model in the GDSC2 database across 10 repeated modeling runs. (G) Top 10 genes with the highest LASSO selection frequency (left) and top 10 genes sorted by averaged random forest feature importance (right) from ten repeated modeling runs Next, we validated the RF models against three other FGFR inhibitors: erdafitinib, pemigatinib, and BGJ398 (Figure S8C). The models reliably predicted erdafitinib sensitivity in BLCA, BRCA and ESCC, (AUC > 0.7). In contrast, limited predictive performance was observed for pemigatinib and BGJ398, likely attributable to their distinct pharmacological profiles relative to AZD4547: pemigatinib exhibits preferential selectivity for FGFR2 over FGFR1/3/4, potentially engaging a distinct adaptive resistance program compared to the pan-FGFR inhibition context in which the signature was developed, while BGJ398 differs from AZD4547 in FGFR4 inhibitory activity and off-target kinase inhibition profiles, which may contribute to a divergent resistance transcriptomic landscape. Multi-omics analyses for erdafitinib further revealed both conserved and inhibitor-specific resistance programs across cancer types (Figure S8D-G). The EMT/invasion signature in BLCA and the cell-cycle dysregulation programs (e.g., MYC targets, E2F targets, and G2M checkpoint) in PAAD were consistently enriched in erdafitinib-resistant phenotypes, consistent with the patterns identified for AZD4547. In COAD, interferon-alpha/gamma response pathways were similarly enriched; however, neither the metabolic reprogramming (e.g., adipogenesis) nor the signaling transduction programs (e.g., mTORC1 signaling) prominent under AZD4547 were recapitulated. These results indicate that EMT-driven plasticity and cell-cycle dysregulation constitute convergent resistance hallmarks shared across mechanistically related FGFR inhibitors, while certain lineage-specific metabolic adaptations may be more selectively associated with AZD4547. We calculated the per-gene LASSO selection frequency across the ten modeling runs to evaluate feature-selection reproducibility, and averaged the random forest feature-importance values across runs to rank candidate genes (Fig. 7G). Combined analyses of LASSO selection frequency and mean random forest importance pinpointed DYRK2, MUCL1, CD34, TEX29, and SMAD9 as consensus features, which demonstrated stable screening of core predictors across resampling. Several of these consensus genes are consistent with the resistance programs highlighted in this study. DYRK2 is a kinase that regulates cell-cycle progression and has been reported to restrain EMT through SNAI1/Snail turnover. SMAD9 is an R-SMAD effector of BMP/TGF-β-superfamily signaling, align with the cell-cycle dysregulation and EMT axes we identified as convergent resistance hallmarks. Whereas CD34 and the mucin-like gene MUCL1 may reflect stem-like or secretory phenotypic states linked to drug tolerance. These features should therefore be regarded as reproducible, hypothesis-generating candidate markers rather than validated drivers, and warrant dedicated functional studies. Identification of synthetic lethal vulnerabilities linked to the FGF/FGFR axis To uncover potential therapeutic targets that synergize with FGFR inhibition, we implemented a systematic computational framework to screen for synthetic lethal (SL) partners of the FGF/FGFR axis (comprising 4 FGFRs and 22 FGF ligands) across eight cancer types. By aggregating the significant pairs identified independently across the cancer types, we constructed a comprehensive pan-cancer SL landscape for the FGF/FGFR family (Table S8). For instance, in ESCC, we identified 25 SL pairs, while 39 pairs were identified in PAAD. To elucidate the biological implications of these SL interactions, we performed an integrated pathway enrichment analysis combining both GO and KEGG databases. This multi-dimensional annotation revealed a convergent functional landscape focused on five core biological modules: signaling transduction, metabolic reprogramming, cell cycle and proliferation, invasion and migration, and immunity (Fig. 8A, B).Fig. 8Functional characterization of synthetic lethal vulnerabilities associated with the FGF/FGFR axis. (A) Gene Ontology (GO) enrichment analysis of identified synthetic lethal candidates. The bar plot illustrates significantly enriched biological processes, systematically categorized into five core functional modules: signaling transduction, metabolic reprogramming, cell cycle & proliferation, invasion, and immunity. The x-axis represents the count of genes involved in each term, and the color gradient indicates statistical significance (–log10 adjusted p-value). (B) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis. The analysis highlights the reactivation of canonical survival cascades (PI3K-Akt, MAPK) and specifically identifies the “EGFR tyrosine kinase inhibitor resistance” pathway, supporting the compensatory RTK pathway activation mechanism Functional characterization of synthetic lethal vulnerabilities associated with the FGF/FGFR axis. (A) Gene Ontology (GO) enrichment analysis of identified synthetic lethal candidates. The bar plot illustrates significantly enriched biological processes, systematically categorized into five core functional modules: signaling transduction, metabolic reprogramming, cell cycle & proliferation, invasion, and immunity. The x-axis represents the count of genes involved in each term, and the color gradient indicates statistical significance (–log10 adjusted p-value). (B) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis. The analysis highlights the reactivation of canonical survival cascades (PI3K-Akt, MAPK) and specifically identifies the “EGFR tyrosine kinase inhibitor resistance” pathway, supporting the compensatory RTK pathway activation mechanism First, Signal Transduction Reactivation. Both databases consistently highlighted a heavy reliance on canonical survival cascades, specifically “Positive regulation of MAPK/ERK cascade” (GO) and “PI3K-Akt signaling pathway” (KEGG). These findings resonate with our earlier multi-omics GSEA results (Fig. 4), where the KRAS and mTORC1 signaling pathways were significantly enriched in resistant phenotypes. This convergence suggests that when the FGFR axis is compromised, tumor cells strategically rewire their signaling networks to exploit these parallel hubs. Second, Cell Cycle Dysregulation. Terms related to “Positive regulation of cell cycle” and “G1/S transition” were prominent, which aligns seamlessly with our single-cell analysis that pinpointed the aberrant activation of the cell-cycle metaprogram (MP2) in resistant subpopulations (Fig. 6). This indicates that targeting SL partners involved in cell cycle checkpoints (e.g., CDKs) could specifically eliminate the hyper-proliferative resistant clones. Third, Compensatory RTK Pathway Activation. Most notably, we observed a significant enrichment of the ‘EGFR tyrosine kinase inhibitor resistance’ pathway (KEGG). This finding provides computational support for the RTK pathway remodeling hypothesis introduced earlier, confirming that FGFR inhibition triggers a compensatory upregulation of alternative RTKs, such as EGFR, to sustain downstream signaling. At the cell-line level, FGFR-inactive cells in GBM, MEL and BLCA consistently exhibited lower EGFR CRISPR dependency scores compared to their FGFR-active counterparts (Figure S9A-C), directly reflecting this compensatory reliance on EGFR when FGFR activity is lost. Consequently, dual blockade of FGFR and these identified SL targets (e.g., EGFR or MEK/ERK pathway components) could represent a promising synthetic lethal strategy to overcome resistance. To further evaluate whether the computationally identified SL partners have prior experimental support, we cross-validated our representative candidates against SynLethDB 2.0, a curated database of experimentally reported SL interactions (Table S9). This cross-validation identified eight shared SL gene pairs across six cancer types. Among these overlaps, four recurrent partner genes were particularly notable. EGFR emerged as the most recurrent SL partner in our overall landscape, appearing in four cancer types (MEL, BLCA, GBM, and LUAD), with three of these (MEL, BLCA, and GBM) also independently recorded in SynLethDB 2.0, highlighting a repeated dependency on remodeling RTK signaling converging on the MAPK/ERK and PI3K/AKT cascades [15, 16]. In parallel, ALK recurred in both COAD and PAAD, nominating an additional RTK-centered bypass route that likewise feeds into PI3K/AKT and MAPK signaling [34]. HCK was repeatedly identified in MEL and LUAD, suggesting a non-canonical Src-family kinase dependency linked to JAK/STAT, MEK/ERK, and PI3K/AKT signaling [35]. MAP2K1 was additionally identified across MEL and BLCA, consistent with prior evidence that MAPK pathway reactivation is a major route of escape from FGFR inhibition and providing a computational rationale for combined FGFR/MEK blockade [17, 36]. Beyond these multi-cancer recurrent partners, several cancer type-specific SL pairs also emerged. In COAD, the FGFR3–DDR2 pair pointed to a collagen-sensing receptor kinase axis associated with extracellular matrix remodeling and invasion [37]. In PAAD, FGFR4-SMO implicated Hedgehog signaling, which is notable because FGFR-SMO crosstalk has been reported in medulloblastoma [38]. In GBM, FGFR4–MAPK12 pointed to a stress-MAPK branch involving p38γ/MAPK12, a kinase increasingly linked to tumor aggressiveness [39]. Notably, the FGFR-EGFR axis is well supported by prior literature and fully consistent with the RTK remodeling framework emphasized in our study. In contrast, FGFR3–DDR2, FGFR4-SMO, and FGFR4–MAPK12, although recorded in SynLethDB 2.0, lack dedicated experimental investigation specifically in the context of FGFR inhibitor resistance, and may therefore represent novel SL vulnerabilities warranting future experimental validation. Taken together, the identified resistance mechanisms are summarised in Table S10, stratified by their recurrence across cancer types into pan-cancer convergent, recurrent multi-cancer, and cancer type-specific categories. Experimental validation of FGFR-EGFR synthetic lethality in RT112 and CCLP1 cells To experimentally test the computationally predicted FGFR-EGFR synthetic lethality, we performed in vitro drug combination experiments in two FGFR-dependent cell line models: RT112, a bladder cancer line harboring the FGFR3–TACC3 fusion with established susceptibility to bypass-mediated FGFRi resistance [26], and CCLP1, a cholangiocarcinoma line with documented feedback EGFR activation upon FGFR inhibition [15]. Cells were treated with AZD4547 (0, 0.25, 0.5, 1, and 2 µM) and osimertinib (0, 0.5, 1, 2, and 4 µM), either alone or in a 5 × 5 dose-response combination matrix for 48 h, and cell viability was assessed by CCK-8 assay. The raw 5 × 5 cell viability matrices for both cell lines, together with the corresponding Bliss and HSA synergy-score matrices, are provided in Table S11. Drug synergy was quantified using four complementary models (HSA, Bliss, ZIP, and Loewe) implemented in SynergyFinder, with scores exceeding 10 considered indicative of strong synergism. Prior to evaluating combination efficacy, the single-agent activity of each drug was characterized. Over the tested concentration ranges, both agents produced only limited growth inhibition as monotherapy, and IC50 values were not reached (IC50 > 2 μM for AZD4547 and >4 μM for osimertinib in both lines). Strikingly, despite this limited single-agent activity, the combination of AZD4547 and osimertinib demonstrated potent synergistic cytotoxicity in both cell lines. Specifically, in RT112 cells, all four synergy models returned scores substantially above the threshold: HSA = 34.43 (p = 5.63 × 10−6), Bliss = 32.94 (p = 7.54 × 10−6), ZIP = 32.02 (p = 1.66 × 10−5), and Loewe = 31.83 (p = 2.35 × 10−5), confirming strong synergistic interaction between AZD4547 and osimertinib (Fig. 9A). Consistent results were observed in CCLP1 cells, where HSA = 35.63 (p = 2.34 × 10−3), Bliss = 86.92 (p = 6.58 × 10−7), ZIP = 65.84 (p = 4.85 × 10−7), and Loewe = 21.74 (p = 2.66 × 10−2) (Fig. 9B). Notably, the two models displayed distinct dose-dependent synergy profiles. In RT112 cells, pronounced synergy was already evident at 0.5 μM osimertinib, a patient-achievable concentration, whereas the strong synergistic interactions for CCLP1 cells emerged only at higher concentrations (up to 2 μM). Thus, although both models were highly responsive to the combination, they exhibited distinct threshold sensitivities to EGFR feedback inhibition.Fig. 9Synergistic lethality induced by combined AZD4547 and osimertinib treatment in RT112 and CCLP1 cells. (A, B) Synergistic effects of AZD4547 and osimertinib in RT112 and CCLP1 cells were evaluated using the HSA, Bliss, ZIP, and Loewe models. Data are derived from three independent biological replicates. (C, D) Western blot analysis of p-EGFR, p-AKT, p-ERK, and cleaved PARP expression in RT112 and CCLP1 cells following treatment with AZD4547, osimertinib, or their combination Synergistic lethality induced by combined AZD4547 and osimertinib treatment in RT112 and CCLP1 cells. (A, B) Synergistic effects of AZD4547 and osimertinib in RT112 and CCLP1 cells were evaluated using the HSA, Bliss, ZIP, and Loewe models. Data are derived from three independent biological replicates. (C, D) Western blot analysis of p-EGFR, p-AKT, p-ERK, and cleaved PARP expression in RT112 and CCLP1 cells following treatment with AZD4547, osimertinib, or their combination To delineate the mechanistic basis underlying this synergy, western blot analysis was performed following treatment with AZD4547, osimertinib, or their combination. The results revealed distinct signaling responses in the two cell lines (Fig. 9C, D). In RT112 cells, AZD4547 alone triggered a compensatory increase in p-EGFR and p-ERK levels, suggesting rapid bypass activation of the EGFR-MAPK axis upon FGFR inhibition. This compensatory upregulation was effectively abrogated when combined with osimertinib. In CCLP1 cells, AZD4547 monotherapy led to a modest reduction in p-EGFR, p-ERK, and p-AKT levels. The dual inhibition of FGFR and EGFR was required to achieve maximal suppression of the MAPK (p-ERK) and PI3K/AKT (p-AKT) signaling axes. Moreover, the combination produced substantially greater cleaved PARP expression, indicating enhanced activation of apoptosis. Taken together, these results provide in vitro evidence that co-inhibition of FGFR and EGFR produces strong synergistic lethality through coordinated suppression of shared downstream survival signaling, and support the translational potential of the FGFR-EGFR co-dependency identified by our computational screening. Discussion Despite the initial clinical efficacy of FGFRi such as AZD4547, the inevitable emergence of acquired resistance remains a major bottleneck in precision oncology [2]. Current research has largely been confined to isolated resistance mechanisms within specific cancer types, failing to capture the systemic network rewiring that drives therapeutic failure. In this study, we bridged this gap by constructing a comprehensive pan-cancer multi-omics atlas of FGFRi resistance. By integrating data from 312 cell lines across the GDSC2 and PRISM databases, we deciphered the multi-layered heterogeneity of resistance and identified actionable synthetic lethal vulnerabilities that pave the way for next-generation combination therapies. A key finding of our study is the context-dependent reliance of resistant cells on metabolic reprogramming. In COAD, we observed a striking activation of the adipogenesis pathway and upregulation of UCP2. UCP2 is known to reduce mitochondrial oxidative stress and inhibit apoptosis, thereby raising the threshold for drug-induced cell death [31, 40]. Consistent with this, our synthetic lethality screening further highlighted “Choline metabolism” and “Lipid and atherosclerosis” pathways as candidate metabolic vulnerabilities. Corroborating this metabolic dependency at the functional genomics level, FGFR4-inactive COAD cells exhibited significantly greater CRISPR-based dependence on UCP2 compared to FGFR4-active cells (Figure S9D), further supporting UCP2-mediated metabolic adaptation as a functional vulnerability upon FGFR pathway loss. Together, these findings suggest that under FGFRi blockade, tumor cells may undergo a metabolic switch—shifting from glycolysis to fatty acid oxidation or altering membrane lipid composition—to sustain survival signals. This aligns with the “Warburg effect” plasticity and raises the hypothesis that combining FGFRi with metabolic inhibitors (e.g., inhibitors of fatty acid synthesis) might be beneficial in specific tumor subtypes [41, 42]. We emphasize, however, that while the FGFR-EGFR axis was experimentally tested in vitro, this metabolic co-targeting strategy remains a computationally and functionally derived hypothesis. Dedicated experiments (e.g., FGFRi combined with UCP2 or fatty-acid-synthesis inhibitors) will be required to establish its therapeutic relevance. Since UCP2-mediated metabolic adaptation is often tumor-specific and less prominent in normal tissues [43, 44], targeting this vulnerability could, if experimentally validated, offer a therapeutic window with potentially reduced toxicity. Beyond bulk omics, our single-cell analysis provided finer-grained insights into the baseline heterogeneity of resistant phenotypes. We identified a specific cell-cycle metaprogram (MP2) enriched in resistant subpopulations of PAAD and LUAD, characterized by core replication drivers such as CCNA2. This enrichment could stem from three plausible scenarios: the clonal selection of pre-existing cell-cycle-dysregulated subpopulations, dynamic transcriptional plasticity, or culture-associated cell-cycle variation. Disentangling these overlapping possibilities will require matched pre- and post-treatment single-cell profiling or functional lineage-tracing experiments. Our single-cell observations establish an instructive, hypothesis-generating framework to guide these future investigations. Perhaps the most translationally significant finding of our study is the convergent computational and experimental evidence implicating compensatory EGFR pathway activation as an adaptive resistance mechanism. Our GSEA analysis initially revealed that resistant cells maintain high activity of downstream KRAS and mTORC1 signaling despite FGFR blockade, hinting at the activation of alternative upstream drivers. This hypothesis was supported by our synthetic lethality screening, which identified the “EGFR tyrosine kinase inhibitor resistance” pathway as a top-ranked vulnerability. This indicates that upon FGFR inhibition, tumor cells dynamically rewire their signaling networks to exploit EGFR as a compensatory bypass route to reactivate the MAPK/PI3K cascade. To experimentally test this computational prediction, we assessed the combined inhibition of FGFR and EGFR in the FGFR-dependent RT112 and CCLP1 cell lines with AZD4547 and osimertinib. Synergy analysis across four complementary reference models (HSA, Bliss, ZIP, and Loewe) consistently yielded scores exceeding 10, indicating strong synergism. Western blot analysis corroborated the suppression of downstream signaling upon co-inhibition. Beyond EGFR-mediated bypass, our functional enrichment further revealed potential engagement of autophagy and PD-L1 checkpoint pathways, suggesting that this network rewiring extends beyond kinases to encompass survival-mediated autophagy and immune evasion. These results provide a compelling rationale for dual-blockade strategies (e.g., FGFRi + EGFRi or FGFRi + anti-PD-1) to abrogate the adaptive resistance mechanisms. Beyond establishing a mechanistic rationale, a critical consideration for clinical translation is whether the concentrations driving this synergy are pharmacologically achievable in patients. In RT112 cells, robust synergistic cytotoxicity was already evident at 0.5 μM osimertinib, a concentration comparable to the steady-state plasma peak (Cmax) reported in patients receiving the approved 80 mg once-daily regimen [45]. In CCLP1 cells, by contrast, the strong synergistic interactions emerged only at higher concentrations (up to 2 μM). These higher doses were included primarily to delineate the full pharmacological synergy landscape rather than to define a clinically required exposure. Whether these specific thresholds are fully attained at the tumor site remains an open question. Although osimertinib is characterized by extensive tissue distribution, direct intratumoral measurements are currently lacking, and the plasma-to-tissue relationship therefore cannot be assumed. Collectively, the clinical-dose synergy in RT112 supports the translational potential of this dual-blockade strategy. The distinct threshold sensitivities of cell lines underscore a model-specific exposure-response relationship that warrants future tumor-type-resolved pharmacokinetic validation. The biological plausibility of FGFR-EGFR co-dependency identified in our study is also well-supported by previous experimental and translational evidence from prior literature (Table S12). Quintanal-Villalonga et al. demonstrated that FGFR1 physically cooperates with EGFR to drive oncogenesis in non-small cell lung cancer, and that their combined inhibition produced markedly superior anti-tumor efficacy compared with either agent alone, providing direct experimental proof that these two RTKs function as cooperative oncogenic partners rather than independent parallel drivers [16]. Complementary translational evidence was subsequently provided by Wu et al., who showed that EGFR inhibition not only potentiates the efficacy of FGFR inhibitor therapy but also overcomes acquired resistance in patient-derived models of FGFR2 fusion-positive cholangiocarcinoma, establishing a mechanistic and therapeutic precedent for this adaptive bypass route in a clinically relevant context [15]. Concordantly, a systematic survey of publicly available drug combination data in the DrugComb database revealed consistent synergistic interactions between FGFR and EGFR inhibitors across multiple independent experiments, further corroborating the cooperative vulnerability of this RTK co-dependency at a pharmacological level (Table S13). Of particular relevance to our own findings, the upregulation of ADAM17 identified in our proteomic screen provides a molecular bridge linking these prior observations to our data: as the primary sheddase responsible for releasing membrane-bound EGFR ligands such as amphiregulin and epiregulin [46], elevated ADAM17 activity in resistant cells offers a plausible mechanism through which FGFR blockade triggers autocrine EGFR reactivation. Taken together, the convergence of our pan-cancer computational predictions, in vitro experimental support in two distinct cell line models, and published translational studies across diverse tumor lineages substantially enhances the reliability of our findings, reinforcing the case for FGFR-EGFR co-dependency as a candidate clinically actionable vulnerability. To translate our molecular mechanistic findings into clinically actionable diagnostic and therapeutic strategies, we conducted ten repeated modeling iterations, ultimately establishing the random forest model as the optimal classifier among the candidates. The model achieved a mean validation-set AUC of 0.69, a mean test-set AUC of 0.73, and a mean test-set AUPR of 0.67, demonstrating an acceptable predictive capacity. Nevertheless, we acknowledge that the current classifier presents several avenues for further optimization. First, the model training and validation were conducted using a limited number of samples retrieved exclusively from public cell line databases. Without external validation in independent clinical cohorts, the clinical applicability and generalizability of this model remain to be further verified. Second, the model was constructed solely based on transcriptomic features, without integrating multi-omic data such as genomic and proteomic profiles, which partially restricts the improvement of its predictive performance. In future studies, we intend to recruit an independent clinical tumor cohort, including paired samples collected before and after FGFR inhibitor treatment, to detect the expression levels of signature genes and validate the model’s ability to predict response to AZD4547 in real clinical cohorts. Furthermore, multi-dimensional features, including genomic status (e.g., FGFR mutation), core proteomic expression levels, and clinical indicators (e.g., tumor stage and prior treatment history), will be integrated to refine the model structure and further improve its predictive accuracy and clinical generalization ability. Although our analyses were primarily conducted in cell-line systems, the resistance programs identified here may have important implications for remodeling the tumor microenvironment (TME). First, the EGFR-centered bypass program highlighted by ADAM17 upregulation and enrichment of the “EGFR tyrosine kinase inhibitor resistance” pathway may extend beyond mere restoration of MAPK/PI3K survival signaling. Prior studies have shown that aberrant EGFR activation can promote PD-L1 expression through JAK/STAT3- and NF-κB-related signaling, thereby facilitating immune evasion [47, 48]. Second, the UCP2-high metabolic state observed in resistant cells may also reshape the metabolic milieu of the TME. UCP2 has been implicated in limiting mitochondrial ROS production and lipid/FAO-biased metabolic rewiring, which is increasingly linked to immunosuppressive myeloid states and impaired antitumor immunity [49, 50]. We therefore speculate that FGFRi resistance, RTK pathway remodeling, and immune-metabolic reprogramming may form a positive feedback loop: FGFR blockade selects for EGFR/UCP2-adapted clones, these clones reinforce an immune-evasive TME, and the remodeled microenvironment in turn stabilizes resistant states. This model provides a rationale for exploring FGFRi plus immune checkpoint blockade combinations, as reflected by ongoing or recent clinical investigations such as pemigatinib plus pembrolizumab (NCT04003610) [51], erdafitinib plus cetrelimab (NCT03473743) [52], and other FGFR/immunotherapy combination studies. Our study has several limitations. First, in vitro models lack the complex tumor microenvironment (TME), which may influence the translatability of immune-related findings, such as PD-L1 associations. Second, the lack of FGFRi-treated clinical cohorts limits direct validation of our resistance signatures. The TCGA and CPTAC datasets represent treatment-naïve populations, and therefore these analyses provide clinical and prognostic context rather than direct evidence of FGFRi resistance relevance. Prospective validation of the identified resistance markers in FGFRi-treated patient cohorts remains an important direction for future work. Third, our single-cell analyses are cross-sectional and were performed on baseline, untreated cancer cell lines. Although our lineage-corrected NMF analysis suggests that MP2 represents an intrinsic baseline feature, these data cannot definitively distinguish among clonal selection, dynamic transcriptional plasticity, or culture-associated cell-cycle variation. Resolving these distinct possibilities will require paired pre- and post-treatment single-cell datasets from future longitudinal studies. Fourth, the in vitro validation of the predicted FGFR-EGFR synthetic lethality was conducted in two FGFR-dependent cell line models (RT112 and CCLP1) that recapitulate the early adaptive or feedback EGFR bypass upon FGFR inhibition. Notably, both lines are not chronically drug-selected, acquired FGFRi-resistant derivatives. Our experiments therefore support the predicted co-dependency and feedback mechanisms but do not establish reversal of acquired resistance. Confirmation in chronically induced resistant derivatives and in more physiologically relevant systems, including patient-derived organoids (PDOs) and in vivo patient-derived xenograft (PDX) models, remains essential to substantiate preclinical efficacy and support clinical translation. Fifth, the pharmacogenomic analyses and machine learning model were developed primarily using AZD4547 drug response data. Cross-drug validation in erdafitinib, pemigatinib, and BGJ398 (Figure S8) revealed limited transferability of the LASSO-derived transcriptomic signature to pemigatinib and BGJ398, indicating our findings may not fully generalize to other FGFRi with distinct selectivity profiles. Future studies incorporating drug response data from clinically approved FGFRi are warranted to broaden the translational scope. Sixth, the limited sample size of our predictive modeling cohort (136 cell lines across eight lineages) precludes reliable performance stratification by individual cancer types. Due to sample sparsity and class imbalance within individual lineages, our framework was optimized and evaluated primarily as a pooled pan-cancer predictor. Future validation in larger, lineage-specific consortia is necessary to systematically define tissue-specific predictive boundaries. Conclusions In summary, this study establishes a multi-omics pan-cancer atlas of resistance to the FGFR inhibitor AZD4547 by integrating large-scale pharmacogenomics with single-cell transcriptomics and systematic synthetic lethality screening. We delineate a convergent resistance landscape involving compensatory EGFR pathway activation, UCP2-mediated metabolic rewiring, and enrichment of cell-cycle-dysregulated subpopulations. Although our mechanistic insights and machine-learning framework are centered on AZD4547, they clarify the heterogeneous molecular architecture of FGFRi resistance and offer a strong rationale for targeted combination strategies—experimentally supported for FGFR-EGFR co-inhibition and hypothesis-generating for metabolic targeting. Because distinct selectivity profiles may engage partly divergent resistance programs, extending this framework to additional clinically approved FGFR inhibitors remains an essential next step to distinguish AZD4547-specific traits from shared class vulnerabilities, ultimately guiding personalized interventions for patients with FGFRi-refractory malignancies. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2