AI-based TME characterization as a cost-efficient tool for precision oncology

Background: Scalable characterization of the tumor microenvironment (TME) remains a major barrier to clinical implementation due to reliance on costly molecular assays. We evaluated whether AI-driven analysis of H&E images can derive biologically meaningful prediction of TME signatures across tumor types without transcriptomic profiling.

Methods: We analyzed retrospective cohorts from SUNY Upstate Medical University, including non-small cell lung cancer (NSCLC; n=675) and prostate cancer (PCa; n=295). Two complementary AI approaches were applied: (1) HistoTME, a weakly supervised framework quantifying 29 functional TME signatures (academic collaborative project), and (2) human-interpretable histologic features (HIFs) capturing spatial cellular organization (industry partnership project). Unsupervised clustering (UMAP, k-means) and multivariable Cox regression assessed prognostic associations.

Results: In NSCLC, both HistoTME and HIF-based clustering identified two major phenotypes: an immune-enriched/TLS-like subtype and a stromal-dominant immune-excluded subtype. The immune-enriched phenotype was associated with significantly improved OS (HR≈0.67-0.70, p<=0.001), with ~60% higher 5-year survival. HistoTME demonstrated stronger independent prognostic value in joint models (HR=0.77, p=0.039), outperforming HIF clustering. Key protective and risk-associated signatures included Treg trafficking (HR=0.74) and proliferation (HR=1.31), respectively. In PCa, TME signatures predicted overall survival (C-index=0.706, p<0.001) and molecular phenotypes, including PTEN loss (AUC=0.76) and castration resistance (AUC=0.755), with distinct immune and angiogenic profiles.

Conclusions: AI-derived TME features from routine histology enable robust, interpretable prognostic stratification across tumor types without molecular testing. These findings support histology-based TME modeling as a scalable framework for biomarker development and clinical translation in precision oncology.

Learning Objective:

1. To assess the clinical utility of histology-derived TME signatures for predicting survival and key molecular phenotypes