Integration of radiomics, deep learning, transcriptomics, and metabolomics reveals prognostic risk stratification and underlying biological mechanisms in colorectal cancer.

Colorectal cancer (CRC) is the third most common malignancy and the second leading cause of cancer-related death worldwide, yet current prognostic stratification is hindered by tumor heterogeneity. Here, we developed a deep learning radiomics model (DLRM), optimized through systematic evaluation of ten machine learning algorithms across 117 combinations, using venous-phase computed tomography (CT) images of 1183 patients from four centers. The resulting risk stratification stratified patients into high- and low-risk groups with distinct survival outcomes, and integration with clinical factors further improved prediction. Integrative transcriptomic and metabolomic analyses revealed that high-risk tumors were enriched for extracellular matrix (ECM)-related pathways associated with tumor progression, whereas low-risk tumors exhibited immune-related signatures, including higher CD8⁺ T-cell infiltration. Both omics consistently identified butanoate metabolism and nitrogen metabolism as protective pathways, validated in an independent public cohort (n = 417). This integrative analytic framework provides robust risk stratification and uncovers biological processes with potential therapeutic relevance.