Inferring Clinically Relevant Molecular Subtypes of Pancreatic Cancer from Routine Histopathology Using Deep...

Area of Science:

• Computational pathology
• Digital pathology
• Precision oncology

Background:

• Molecular subtyping of pancreatic ductal adenocarcinoma (PDAC) into basal-like and classical subtypes has prognostic and predictive value.
• Current methods are limited by cost, time, and tissue requirements, hindering clinical application.
• PanSubNet offers a novel solution for accessible molecular subtyping.

Purpose of the Study:

• To introduce PanSubNet, an interpretable deep learning framework for predicting PDAC molecular subtypes from H&E-stained whole-slide images.
• To overcome the limitations of existing subtyping methods in clinical practice.

Main Methods:

• Developed PanSubNet using data from 1,055 patients across two cohorts (PANCAN, TCGA) with paired histology and RNA sequencing data.
• Utilized a dual-scale architecture with attention mechanisms for multi-scale representation learning and feature attribution.
• Ground-truth labels derived from the Moffitt 50-gene signature refined by GATA6 expression.

Main Results:

• Achieved an AUC of 88.5% on internal validation (PANCAN) and 84.0% on external validation (TCGA).
• Demonstrated robust generalizability and preserved/strengthened prognostic stratification compared to RNA-seq.
• Prediction uncertainty correlated with intermediate transcriptional states, not classification noise.

Conclusions:

• PanSubNet enables rapid, cost-effective molecular stratification of PDAC from routine H&E slides.
• It is a clinically deployable and interpretable tool for genetic subtyping.
• PanSubNet supports integration into digital pathology workflows, advancing precision oncology for PDAC.