Improving Performance, Robustness and Interpretability of H&E-Based Models in Spatial Pathology
Spatial transcriptomics (ST) links tissue morphology with gene expression, enabling new opportunities for computational pathology and precision medicine. However, ST technologies remain costly and difficult to scale clinically. Predicting molecular and spatial phenotypes directly from routine H&E histology images offers a promising alternative for scalable molecular pathology, but current models remain limited in predictive performance, robustness, and biological interpretability.
Here, we present a unified framework for robust and interpretable prediction of spatial molecular features and clinical outcomes from H&E images by integrating pathology foundation models, uncertainty-aware learning, and biologically informed representation refinement. We develop methods that: (i) predict spatial gene expression and cell-type composition with quantified uncertainty, (ii) improve biological interpretability through pathway-guided prototype learning, and (iii) integrate complementary strengths of multiple pathology foundation models using adaptive gene-specific weighting.
Across multiple spatial transcriptomics platforms, cancer types, and independent datasets, our approaches consistently improve prediction accuracy for clinically relevant biomarkers, including PAM50 subtype markers and therapeutic targets. Importantly, predicted molecular features can stratify patient survival and infer drug response, while attribution analyses reveal histomorphological patterns associated with functional gene programs. Together, this work demonstrates how interpretable and robust foundation model integration can bridge histology and spatial genomics to advance scalable precision pathology.
Bio
Assoc. Prof. Quan Nguyen is Head of the Genomics and Machine Learning Laboratory, Scientific Director of the QIMRB National Centre for Spatial Tissue and AI Research (NCSTAR), and founder of an emerging QIMRB spin-out company (STARCO). He is a former Australian Research Council DECRA Fellow and currently holds an NHMRC Leadership Fellowship (EL2, L1). He serves as Co-Chair of the international SpatioTemporal Omics Consortium (STOC; members from 36 countries). His team develops and applies machine learning approaches to analyse single-cell and spatial multiomics data from patient samples and preclinical models. Their research aims to discover diagnostic biomarkers, identify drug targets involved in cell–cell interactions, predict disease progression, and stratify patient responses to therapies. His research focuses on developing scalable precision medicine approaches that can benefit broader patient populations.
