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

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer survival analysis is crucial for patient prognosis.
• Integrating tumor microenvironment (TME) heterogeneity across diverse cancer types remains a challenge.
• Existing models often lack the granularity to capture complex TME interactions.

Purpose of the Study:

• To develop TMESurv, a biologically informed neural network for predicting cancer survival outcomes.
• To integrate hierarchical TME insights, including genes, cells, and their roles, into a predictive framework.
• To improve the accuracy and interpretability of cancer survival predictions by leveraging TME complexity.

Main Methods:

• Developed TMESurv, a neural network integrating TME data (genes, cellular components, cells, cell roles) with clinical outcomes.
• Utilized LLM-derived marker genes and biological knowledge to establish model connections.
• Assessed predictive performance using concordance index (C-index), AUC, and hazard ratio (HR).
• Benchmarked TMESurv against existing models on TCGA cohorts and validated on an independent SKCM immunotherapy cohort.

Main Results:

• TMESurv demonstrated superior survival prediction performance compared to five other models across eight TCGA cohorts.
• Achieved a C-index of 0.68 for KIRC and an AUC of 0.62 for SKCM, outperforming competitors.
• Effectively stratified patients into high- and low-risk groups (HR > 1).
• Integrating immunostimulatory and immunosuppressive cell roles improved SKCM prediction accuracy (C-index and AUC boosted by 6 points).
• Validation on the SKCM immunotherapy cohort confirmed the role of specific cell types in predicting treatment response.

Conclusions:

• TMESurv provides robust and interpretable cancer survival predictions by integrating TME heterogeneity.
• The study highlights the foundational role of the TME in tumor biology and patient outcomes.
• Integrating biological constraints and TME principles enhances the predictive accuracy and interpretability of survival models.