TMESurv: a Tumor Microenvironment-informed Interpretable Neural Network for Survival Analysis across Multiple Cancer...

Cancer survival analysis is critical for prognosis; however, integrative analyses across multiple cancer types remain limited, particularly in capturing tumor microenvironment (TME) heterogeneity. To address this, we developed TMESurv, a biologically informed neural network that integrates hierarchical insights from the TME to predict survival outcomes across cancers. TMESurv employs a structured framework linking genes, cellular components, cells, cell roles, and clinical outcomes, with key connections established based on LLM-derived marker genes and established biological knowledge. Predictive performance was assessed using the concordance index (C-index), the area under the time-dependent ROC curve (AUC), and the hazard ratio (HR). When benchmarked against five survival models across eight TCGA cohorts and validated on an independent SKCM immunotherapy cohort, TMESurv demonstrated superior performance, achieving a C-index of 0.68 for KIRC and an AUC of 0.62 for SKCM, outperforming its closest competitors by 2-3 points. It effectively stratified patients into high- and low-risk groups (HR > 1) and improved predictive accuracy by integrating immunostimulatory and immunosuppressive cell roles, boosting the C-index and AUC for SKCM by 6 points when incorporated into the loss function. Validation on the SKCM immunotherapy cohort further highlighted the distinct roles of these cells in differentiating responders from non-responders. Guided by TME principles, TMESurv offers robust survival predictions and interpretable insights, emphasizing the foundational role of the TME in tumor biology and showcasing the benefits of integrating biological constraints to enhance predictive accuracy and interpretability.