Hematoxylin and Eosin Architecture Uncovers Clinically Divergent Niches in Pancreatic Cancer

Abstract

Pancreatic ductal adenocarcinoma (PDAC) represents one of the only cancers with an increasing incidence rate and is often associated with intra- and peri-tumoral scarring, referred to as desmoplasia. This scarring is highly heterogeneous in extracellular matrix (ECM) architecture and plays complex roles in both tumor biology and clinical outcomes that are not yet fully understood. Using hematoxylin and eosin (H&E), a routine histological stain utilized in existing clinical workflows, we quantified ECM architecture in 85 patient samples to assess relationships between desmoplastic architecture and clinical outcomes such as survival time and disease recurrence. By utilizing unsupervised machine learning to summarize a latent space across 147 local (e.g., fiber length, solidity) and global (e.g., fiber branching, porosity) H&E-based features, we identified a continuum of histological architectures that were associated with differences in both survival and recurrence. Furthermore, we mapped H&E architectures to a CO-Detection by indEXing (CODEX) reference atlas, revealing localized cell- and protein-based niches associated with outcome-positive versus outcome-negative scarring in the tumor microenvironment. Overall, our study utilizes standard H&E staining to uncover clinically relevant associations between desmoplastic organization and PDAC outcomes, offering a translatable pipeline to support prognostic decision-making and a blueprint of spatial-biological factors for modeling by tissue engineering methods.

Impact Statement

Pancreatic ductal adenocarcinoma (PDAC) may become the second-leading cause of cancer mortality in the next decade and exhibits scarring with complex, yet unclear, roles in differentiating patient outcomes. We utilized hematoxylin and eosin, a standard histological stain used in clinical workflows, to quantify the extracellular matrix (ECM) architecture of PDAC-associated scarring. Unsupervised machine learning identified a continuum of histological architectures based on divergence in 147 ECM fiber features, which correlated with survival time, disease recurrence, and divergent cellular/protein niches. These findings offer a clinically accessible prognostic signature based on patient-specific histology and suggest spatial-biological targets for tumor modeling.