Document resource
Background With a five-year survival rate under 12%, PDAC remains one of the most aggressive and fatal cancers. The presenting symptoms of PDAC, such as abdominal pain or jaundice, overlap significantly with those of benign hepatopancreatobiliary diseases, making accurate differentiation challenging. Furthermore, the biomarker CA19-9 shows limited sensitivity and specificity in distinguishing PDAC from benign diseases. As bile directly reflects the pancreatobiliary microenvironment, its microbiome may offer novel diagnostic signals. On this basis, this study aimed to develop and validate the BileBiomeScreen model, a machine learning–based classifier based on bile microbial features for high-confidence PDAC detection.Methods A total of 295 bile samples were collected between June 2022 and October 2024, comprising 65 PDAC and 54 benign controls in the discovery cohort, 43 PDAC and 15 benign controls in the validation cohort, and 118 other hepatobiliary malignancies ( IDDF2026-ABS-0152 Figure 1). Microbial profiles were analyzed by 16S rRNA sequencing. Model development involved evaluating eight machine-learning algorithms for feature selection and classifier construction, after which a random forest–based framework was selected as the best-performing approach. Diagnostic accuracy was evaluated by ROC analysis against CA19-9 and across these malignancies, and prognostic value was assessed through multivariate Cox regression to derive a bile microbiome–based risk score.Results Compared with benign controls, PDAC patients exhibited significantly reduced α- and β-diversity, with enrichment of pro-inflammatory taxa and depletion of probiotics ( IDDF2026-ABS-0152 Figure 2, IDDF2026-ABS-0152 Figure 3). The BileBiomeScreen model demonstrated high diagnostic accuracy (AUC 0.959 in discovery, 0.947 in validation), outperforming CA19-9 and showing the highest specificity for PDAC compared with other hepatobiliary malignancies (IDDF2026-ABS-0152 Figure 4). A bile microbiome–based Cox model further stratified patients by risk and independently predicted overall survival in both the discovery (HR 5.3, 95% CI 1.51–18.64, p=0.0093) and validation cohorts (HR 3.24, 95% CI 1.24–8.45, p=0.0162), beyond CA19-9 and other variables (IDDF2026-ABS-0152 Figure 5).Conclusions The BileBiomeScreen model demonstrates superior diagnostic performance compared with CA19-9, positioning bile microbiome profiling as a promising adjunctive tool for PDAC diagnosis. Additionally, alterations in the bile microbiome were found to serve as independent prognostic factors, suggesting broader clinical utility.Abstract IDDF2026-ABS-0152 Figure 1Abstract IDDF2026-ABS-0152 Figure 2Abstract IDDF2026-ABS-0152 Figure 3Abstract IDDF2026-ABS-0152 Figure 4Abstract IDDF2026-ABS-0152 Figure 5