Machine learning–based identification of a metabolic gene prognostic signature (MALMPS) model to investigate the clinical and molecular heterogeneity of stage II/III colorectal cancer.

Abstract

e15548 Background: Colorectal cancer (CRC) is a frequently lethal disease with heterogeneous outcomes, and stage II/III CRC accounts for nearly 70%. Metabolic reprogramming was associated with cancer progression. Methods: We utilized nine datasets and a machine learning framework with 83 algorithm combinations to develop a robust prognostic signature. The in-house SDPH dataset were used for transcriptomic, metabolomics and spatial metabolomics analyses. Results: The combination of the random survival forest and generalized boosted regression modeling machine learning algorithms yielded the MAchine Learning-based Metabolic gene Prognostic Signature (MALMPS) model, comprising 21 metabolic gene as independent variables. The MALMPS model outperformed traditional clinical traits and molecular features in prognostic prediction, achieving an average C-index of 0.862 ( 95%CI : 0.841-0.882) in the training dataset and 0.687 ( 95%CI : 0.677-0.707) in the validation datasets. We stratified the CRC patients into the high- and low-risk groups based on the median cutoff of the MALMPS. Notably, the high-risk subgroup exhibited poor prognosis, inflammation activation, and enriched carbohydrate, glycosaminoglycan, and lipid metabolism. In contrast, the low-risk group displayed a higher mutation frequency in TGF-β pathway, and enrichment in nucleotides, cofactors, and amino acids metabolism. Metabolites profiling in the SDPH dataset further validated the distinct metabolic alterations between the high- and low-risk groups. COX7B, identified as playing a vital role in the MALMPS, was demonstrated to promote the malignant behavior of CRC by multi-omics analysis and in vitro assays. Drug sensitivity analysis revealed the potential of targeting the IGF-1R and Wnt/β-catenin for treating high-risk samples. Finally, the nomogram constructed incorporating MALMPS and molecular features shows strong performance in the validation datasets. Conclusions: The findings indicate that MALMPS might be a valuable instrument for predicting the recurrence risk of stage II/III colorectal cancer, particularly for identifying individuals at high risk.