Towards Scalable Molecular & Spatial Analysis of Tumor Microenvironment from Digital Pathology

Introduction:The tumor microenvironment (TME) is a key focus for biomarker discovery and therapeutic targeting in cancer. However, current TME profiling techniques such as spatial transcriptomics and proteomics are limited by high costs and technical complexity, restricting their scalability and wide access in routine clinical settings. We introduce HistoTMEv2-a robust and cost-effective deep learning tool to infer cellular, molecular and spatial characteristics of the TME directly from Hematoxylin & Eosin (H&E)-stained slides-markedly expanding on our previous work (Patkar et al, <em>NPJ Prec Onc</em>, 2024).Methods:HistoTMEv2 was trained in a weakly supervised fashion using attention-based multiple instance learning (ABMIL) to predict enrichment scores corresponding to 29 clinically relevant TME signatures (Bagaev et al, <em>Cancer Cell</em>, 2021). The training and internal validation cohort comprised 7,586 patients (24 cancer types) from The Cancer Genome Atlas (TCGA). External validation was performed on independent public and institutional datasets, (5,657 patients, 9 cancer types). HistoTMEv2's performance was compared against SEQUOIA, a recent state of the art H&E-based gene expression prediction tool (Pizurica et al <em>Nat. Commun</em>, 2024).Results:HistoTMEv2 outperforms SEQUOIA, achieving a median Pearson correlation of 0.61 across gene signatures in TCGA cross-validation and 0.51 in external validation, compared to SEQUOIA's 0.35. Predicted enrichment scores correlate strongly with single-cell-derived abundances of T cells (&rho;: 0.54-0.65), B cells (&rho;: 0.68-0.71), macrophages (&rho;: 0.38-0.42), endothelial cells (&rho;: 0.36-0.41), and fibroblasts (&rho;: 0.38-0.72)-across bulk, regional (TMA cores), and spot-level (Visium) resolutions. In high-risk prostate cancer (Gleason >4+4), HistoTMEv2-derived stromal signatures successfully predicted treatment failure after ADT + enzalutamide therapy at 6 months (AUC: 0.83, odds ratio: 7.5, <em>p</em> < 0.001), demonstrating its value for biomarker discovery.Conclusion/Discussion:HistoTMEv2 enables scalable, interpretable, and clinically actionable TME profiling from standard H&E slides, without requiring spatially registered multiplex imaging or manual cell-level annotations. Its strong performance and generalizability across diverse cancer types and datasets position it as a powerful tool for integration into digital pathology workflows and large-scale biomarker discovery initiatives. (Github: https://github.com/spatkar94/HistoTME/tree/main)

Learning Objectives: 

1. Understand the role of AI in characterizing the tumor microenvironment and driving biomarker discovery from routine pathology slides
2. Understand the crucial but often overlooked role of tumor adjacent stroma and matrix in driving tumor progression and its potential as a novel biomarker for castration resistance development in locally advanced hormone sensitive prostate cancer
3. Engage in discussions to further advance translational impact of AI-driven quantitative analysis tools harnessing digital pathology