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

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Molecular diagnostics

Background:

• Molecular subtyping of pancreatic ductal adenocarcinoma (PDAC) into basal-like and classical subtypes has prognostic and predictive value.
• Clinical application of molecular subtyping is limited by cost, time, and tissue requirements.
• There is a need for accessible methods to stratify PDAC patients.

Purpose of the Study:

• Introduce PanSubNet, an interpretable deep learning framework to predict therapy-relevant molecular subtypes of PDAC.
• Enable molecular subtyping directly from standard hematoxylin and eosin (H&E)-stained whole-slide images.
• Facilitate clinical decision-making in PDAC management.

Main Methods:

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

Main Results:

• Internal validation (PANCAN) achieved an AUC of 88.5% with balanced sensitivity and specificity.
• External validation (TCGA) demonstrated robust generalizability with an AUC of 84.0%.
• PanSubNet preserved and strengthened prognostic stratification compared to RNA-seq-based labels, with predictions aligned to transcriptomic programs and DNA damage repair signatures.

Conclusions:

• PanSubNet provides a rapid, cost-effective, and clinically deployable tool for molecular stratification of PDAC from routine H&E slides.
• The interpretable framework supports integration into digital pathology workflows for precision oncology.
• Ongoing validation at two institutions aims to assess real-world performance and clinical utility.