High-risk pancreatic cancer pathology reveals a B3GNT3-associated glycosylation program

Objective Pancreatic cancer has a poor prognosis and marked heterogeneity, whereas existing pathomics models generally lack biological interpretability. We aimed to construct the pathomics prognostic score GlyPath-Risk and explore the molecular states associated with the high-risk phenotype captured by the score. Methods We integrated H&E whole-slide images from TCGA-PAAD and Fujian Medical University Union Hospital (FMUUH), extracted handcrafted and deep-learning features using CellProfiler and an ImageNet-pretrained ResNet-50, and constructed GlyPath-Risk using a combined LASSO and random survival forest model. Prognostic performance was evaluated in training, internal validation, and FMUUH cohorts. Correlation analysis, weighted gene co-expression network analysis (WGCNA), gene set enrichment analysis (GSEA), and xCell immune-infiltration estimation were used to investigate associations among GlyPath-Risk, the glycosylation transcriptional program, and the tumor microenvironment. Single-cell transcriptomics, CellChat cell-cell communication analysis, Geneformer virtual perturbation, and B3GNT3 knockdown experiments were used to generate and evaluate candidate biological hypotheses. Results GlyPath-Risk comprised six pathological features and showed C-indices of 0.88, 0.68, and 0.83 in the training, internal validation, and FMUUH cohorts, respectively. Its association with overall survival remained after adjustment for age, sex, and clinical stage in the reported Cox model (multivariable HR = 1.230). The glycosylation gene set showed the strongest mean absolute correlation with pathological features among 20 post-translational modification classes. WGCNA identified six co-expression modules significantly associated with GlyPath-Risk (BH-FDR < 0.05), including a blue module containing B3GNT3. Immune-cell score analyses showed nominal associations with selected immune- and vascular-related scores, but none of 51 candidate scores reached BH-FDR < 0.05. Single-cell analysis suggested differences in predicted outgoing communication from B3GNT3-high and B3GNT3-low tumor cells, while virtual perturbation identified transcriptional states associated with perturbation sensitivity. B3GNT3 knockdown was associated with reduced proliferation and altered macrophage-associated markers in conditioned-medium experiments. Conclusion GlyPath-Risk showed prognostic associations across the evaluated cohorts. The score was associated with a B3GNT3-related glycosylation transcriptional program and may be related to immune- and vascular-related transcriptomic features. These findings provide a testable framework for investigating the biological correlates of pathomics scores and nominate B3GNT3 as a candidate molecular anchor requiring independent validation.