Fig. 1: Correlation between CCA molecular subtypes and clinical and anatomical features.

a Initial non-negative matrix factorization (NMF) clustering on transcriptomic data of all samples (N = 438) revealed four molecular classes. Molecular cluster & Anatomical site, p = 1.00E-06; Molecular cluster & Stage, p = 1.00E-07; Molecular cluster & Resection method, p = 0.0001; Molecular cluster & Hepatic contamination, p = 2.20E-16; Molecular cluster & Pancreatic contamination, p = 2.20E-16; Molecular cluster & Duodenal contamination, p = 4.20E-06; Molecular cluster & Lymphatic contamination, p = 0.0036; Molecular cluster & Neural contamination, p = 0.0017. b Sankey diagram representing the select rules of the “purified cohort” (N = 164) and “Verification cohort” (N = 274). First panel: all 438 CCA cases enrolled in this study; second panel: samples grouped according to whether the overall tissue contamination proportion >25%; third panel: samples grouped according to whether the NTP result < = 0.1; fourth column: samples with FDR < = 0.1 in NTP result and with an overall tissue contamination proportion >25% were selected into the “verification cohort”, while the other samples selected into the “purified cohort”. c Dendrogram generated based on expression matrix of the purified cohort (protein-coding genes only; N = 164). d Two molecular classes were determined through consensus clustering on the purified cohort (N = 164). Molecular cluster & Stage, p = 0.03; Molecular cluster & Liver percentage, p = 0.02; P values were calculated by two-sided Fisher’s exact test for categorical variables and Kruskal–Wallis rank sum test for categorical and continuous data. Red asterisks indicate variables significantly correlated with molecular classes (p < 0.05).