Application of an interpretable graph neural network to predict gene expression signatures associated with tertiary lymphoid structures in histopathological images

Tertiary lymphoid structures (TLS) are vascularized lymphocyte aggregates in the tumor microenvironment (TME) that correlate with better patient outcomes. Previous studies identified a 12-chemokine gene expression signature associated with disease progression and the type and degree of TLS1 . These signatures could provide insight important for clinical decision making during pathologic evaluation.

Recently, deep learning (DL) approaches in digital pathology have been successfully piloted for cancer diagnosis; however, DL-based imputation of molecular phenotypes from pathology images has had mixed success. Predicting gene expression from whole slide images (WSI) may be impeded by low prediction accuracy and lack of interpretability.

To address these limitations, we developed an artificial intelligence (AI)-based, state-of-the-art workflow to predict the 12-chemokine TLS gene signature as well as the patient outcomes from lung and breast cancer WSIs, and to identify histological features relevant to model predictions. We also show that stratification based on GNN-inferred TLS gene expression has significant prognostic value for survival outcome.