Single-cell and spatial multi-omics combined with machine learning to unravel the heterogeneity of cholangiocarcinoma and the WFDC3-mediated prognostic regulatory mechanisms

Background Intrahepatic cholangiocarcinoma (ICC) exhibits substantial cellular heterogeneity and a complex tumor microenvironment (TME), contributing to poor clinical outcomes. Effective prognostic biomarkers and a deeper understanding of the cellular ecosystems associated with high-risk disease remain urgently needed. Methods We integrated five scRNA-seq datasets from GEO with four publicly available spatial transcriptomic sections and bulk transcriptomic data from the TCGA, fu-ICCA, and GEO cohorts. A multidimensional framework combining single-cell analysis, spatial transcriptomics, and machine learning was used to characterize cellular heterogeneity and identify prognostically relevant features. Hotspot was used to characterize T- and B-cell transcriptional modules, Scissor was applied to identify prognosis-associated malignant epithelial cells, and 101 machine-learning pipelines were evaluated to construct the ICCS prognostic signature. Cell–cell communication analysis and exploratory spatial co-density analysis were further performed to characterize ICCS-associated stromal–epithelial interactions, while in silico perturbation and functional assays were used to investigate WFDC3. Results The single-cell atlas identified 11 major cell populations and prognosis-associated T- and B-cell transcriptional programs. A six-gene ICCS comprising CEP55, EGLN3, GPRC5A, PRR15, SFN, and WFDC3 showed reproducible prognostic discrimination across the analyzed cohorts. ICCS-High displayed a markedly epithelial-dominant cellular composition, with epithelial cells increasing from 15.47% in ICCS-Low to 78.44% in ICCS-High, whereas T/NK cells decreased from 44.28% to 3.95%. Comparative cell–cell communication analysis revealed prominent mesenchymal-to-epithelial signaling enriched for collagen-related ligand–receptor interactions. COL4A1–CD44 showed the highest relative contribution and a more prominent mesenchymal–epithelial communication pattern in ICCS-High. Exploratory spatial density maps further identified local co-occurrence of mesenchymal and epithelial signals, supporting an ECM-associated microenvironmental context in the high-risk state. Separately, in silico WFDC3 perturbation implicated tumor-related pathways, and WFDC3 knockdown reduced colony formation, migration, and invasion of ICC cells. Conclusion ICCS stratification captures distinct cellular and molecular states of the ICC microenvironment. The ICCS-High state is characterized by epithelial predominance, reduced T/NK-cell representation, and stronger inferred mesenchymal-to-epithelial collagen-related communication. COL4A1-CD44 and WFDC3 should be regarded as hypothesis-generating biological features requiring further mechanistic validation. ICCS remains a research-stage prognostic signature that requires standardized assays and independent prospective validation before clinical application.