Pancreatic ductal adenocarcinoma (PDAC) is not a single uniform disease at the molecular level. Gene expression profiling studies have identified distinct molecular subtypes with fundamentally different biological behaviors and responses to therapy.
The two primary tumor subtypes are Classical and Basal-like. Classical PDAC tends to have better prognosis and may respond differently to chemotherapy regimens compared to Basal-like PDAC, which is more aggressive and associated with shorter survival.
Current molecular subtyping requires bulk RNA sequencing of tumor tissue, a process that is expensive, time-consuming, and not routinely available in clinical pathology laboratories. A method to infer molecular subtypes directly from standard histology slides would make subtyping clinically accessible worldwide.
In addition to tumor cell subtypes, the surrounding stroma (connective tissue and non-cancerous cells around the tumor) also has distinct activation states. Stromal activity has independent prognostic value, and characterizing both tumor and stromal biology from histology could provide comprehensive molecular profiling from routine slides.
PACpAInt is a multi-step deep learning pipeline designed to predict PDAC molecular subtypes directly from digitized hematoxylin and eosin (H and E) stained histology slides. The pipeline processes whole-slide images by dividing them into small tiles and analyzing millions of tiles per slide.
The pipeline consists of four sequentially applied models. The Neo model identifies tiles containing neoplastic (cancer) cells versus stroma or normal tissue. The B/C model then classifies neoplastic tiles as Basal-like or Classical subtype. A separate Cell Type model identifies stromal cell populations.
Finally, the Comp (Composition) model integrates tile-level predictions across the entire slide to produce a patient-level molecular subtype call. Each step filters or refines the analysis from the previous step, progressively focusing on the most diagnostically informative tissue regions.
The model was trained on digitized whole-slide images from 202 patients with linked molecular subtyping data from bulk RNA sequencing. The molecular subtype labels used as training targets were derived from published transcriptomic analyses, providing a high-quality ground truth signal for supervised learning.
The model was validated on four independent cohorts totaling 598 patients from different institutions and countries. External validation on geographically and technically diverse datasets is essential for demonstrating that a model has learned clinically meaningful patterns rather than institution-specific artifacts.
Validation cohorts included patients from European and North American centers, ensuring that the model's performance was not limited to a specific geographic population or tissue processing protocol. The cohorts also varied in slide preparation and staining intensity, testing robustness to technical variation.
Concordance between PACpAInt predictions and ground truth RNA-based subtype assignments was used as the primary validation metric. High concordance across diverse cohorts would confirm that histological features visible to the model reliably reflect underlying molecular biology.
PACpAInt revealed that many PDAC tumors contain a mixture of Classical and Basal-like tiles rather than being uniformly one subtype. This intratumor heterogeneity, previously underappreciated because bulk RNA sequencing averages signal across the entire tumor, may have profound implications for treatment resistance and relapse.
The model identified a novel hybrid subtype category, characterized by tumors with substantial proportions of both Classical and Basal-like areas. Hybrid tumors showed intermediate survival outcomes between purely Classical and purely Basal-like tumors, suggesting that subtype admixture is clinically meaningful.
Spatial mapping of subtype predictions across tumor tiles revealed that Basal-like regions tend to cluster in specific areas of the tumor rather than being randomly distributed. This spatial organization suggests that Basal-like transformation may occur through focal clonal evolution rather than a uniform tumor-wide molecular shift.
Stromal subtype predictions showed that Active stroma (associated with inflammatory and fibroblast activation) correlated with worse prognosis independent of tumor subtype. Combining tumor and stromal subtype classifications provided stronger prognostic stratification than either alone.
PACpAInt's ability to perform molecular subtyping from standard H and E slides could enable routine subtype-informed treatment decisions without requiring additional molecular testing. If validated prospectively, this would make precision oncology for PDAC accessible in any pathology laboratory worldwide.
A key implication is that clinical trials testing subtype-specific therapies could use PACpAInt as an inexpensive and widely applicable companion diagnostic. Patient stratification based on predicted molecular subtype could improve trial design by enrolling only patients whose tumors are predicted to be subtype-responsive.
The discovery of intratumor heterogeneity also raises important questions about biopsy representativeness. Since EUS-FNA samples only a small portion of the tumor, a needle biopsy may sample a Classical region while the true aggressive Basal-like component remains elsewhere in the tumor. Whole-resection slide analysis by PACpAInt could provide more complete molecular characterization.