Pancreatic ductal adenocarcinoma (PDAC) exhibits five-year survival rates below 10% with profound therapeutic resistance mediated by stromal barriers and immunosuppression (Rahib et al. 2014; Siegel et al. 2023). Recent evidence suggests intratumoral mycobiome alterations may influence disease progression (Aykut et al. 2019; Riquelme et al. 2019). We performed comprehensive network pharmacology analysis to elucidate potential therapeutic mechanisms of Saccharomyces boulardii (SB) and Clostridium histolyticum collagenase (CHC) in PDAC. Gene sets representing immune activation, metabolic competition, and stromal remodeling were analyzed using protein–protein interaction networks (STRING v11.5), functional enrichment (DAVID v2021), and pathway databases (KEGG, Reactome, WikiPathways). Structural assessment employed blind molecular docking (CB-Dock2) and coarse-grained molecular dynamics (CABS-flex). Computational analysis was performed on TCGA-PAAD data (n = 177 patients, 93 events) using Cox regression and five machine learning classifiers (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, LightGBM) with Q1/Q3 survival stratification and strict fivefold nested cross-validation with within-fold feature selection (SelectKBest, k = 5). Random survival forest (RSF) incorporating 67 high-quality genes from four mechanistically curated modules was developed for time-to-event prediction, evaluated by AUC-ROC and Harrell’s C-index. Permutation-based feature selection identified prognostic biomarkers. Network analysis identified TNF, IL6, IFNG, and TLR4 as central hubs (degree > 15) with dense interconnectivity (clustering coefficient = 0.78). Pathway enrichment revealed over-representation in IL-17, TLR, and IL-10 signaling (BH-FDR < 0.05). Molecular docking revealed high-affinity interactions with AKT1 (ΔG = − 7.2 kcal/mol), TLR2 (− 7.6 kcal/mol), TLR4 (− 6.9 kcal/mol), and KRAS-SOS1 interface (− 6.9 kcal/mol). Kaplan–Meier analysis demonstrated significant survival differences by composite immune-metabolic signature (log-rank p = 0.0001) and immune sub-signature (CXCL9, CXCL10, CXCL11; p = 0.0028). The metabolic sub-signature (LDHA, ALDH9A1) demonstrated a significant continuous prognostic effect by Cox regression (HR = 1.431, p = 0.0292), with median dichotomisation not reaching significance (log-rank p = 0.2164), indicating a gradient rather than threshold effect. Cox modelling of the composite five-gene score identified it as a strong negative prognostic factor (HR = 1.897, p < 0.0001). Machine learning classifiers achieved good-to-strong discrimination of survival extremes under strict nested cross-validation (AUC range: 0.631–0.745; Logistic Regression: AUC = 0.745 ± 0.036), confirmed by permutation testing (n = 1000 permutations, empirical p < 0.01). RSF modelling on the full cohort achieved C-index = 0.649 ± 0.040 using the top-5 gene signature, representing acceptable discrimination for PDAC. Permutation-based feature selection identified CXCL11, CXCL10, CXCL9, LDHA, and ALDH9A1 as top prognostic biomarkers. Tertile-based risk stratification demonstrated significant survival separation between Low and High Risk groups (log-rank p = 0.0005). Kinetic modeling showed 4.2-fold SB competitive advantage under tumor hypoglycemia (< 1 mM glucose). Network analysis, structural modeling, and machine learning support biological plausibility of SB anti-tumor activity through immune activation and stromal remodeling. RSF modelling (C-index = 0.649) and Cox regression (HR = 1.897, p < 0.0001) provide a computational framework for patient stratification. Identification of CXCL11, CXCL10, CXCL9, LDHA, and ALDH9A1 as prognostic biomarkers spanning immune and metabolic modules provides targets for clinical monitoring. This systems-level framework provides mechanistic rationale for clinical investigation of intratumoral S. boulardii as adjuvant PDAC therapy.
