Microbiome-transcriptome-histology triad enhances survival risk stratification in multiple cancers

Accurate prognostic stratification is essential for optimizing postoperative therapeutic strategies in oncology. While deep learning approaches have shown promise for survival prediction through unimodal analyses of histopathological images, transcriptomic profiles, and microbial signatures, their clinical utility remains limited due to fragmented biological insights. In this study, we introduce HMTsurv, a multimodal survival prediction framework that integrates digital histopathology, host transcriptomics, and tumor-associated microbiome features. Utilizing multi-omics datasets from four major malignancies-colorectal, gastric, hepatocellular, and breast cancers-our model exhibited superior prognostic accuracy (c-index: 0.68-0.72) when compared to single-modality benchmarks, as validated through rigorous cross-validation methods. Notably, our model achieved robust risk stratification (log-rank p < 0.001 across all cohorts) as demonstrated by Kaplan-Meier analysis, effectively distinguishing patients into distinct survival trajectories. Systematic examination of multimodal signatures identified 14 pan-cancer survival biomarkers, including MAGE family genes, which were consistently upregulated in high-risk subgroups. Additionally, we elucidated distinct histopathological patterns, dysregulated microbial communities, and altered gene-microbiota co-expression networks that were predictive of adverse outcomes. This study not only establishes a generalizable multimodal architecture for cancer prognosis but also elucidates the intricate interactions among histological, molecular, and ecological determinants of survival, providing a clinically actionable framework for precision oncology.