Interpretable Multiomics Machine Learning Identifies GSDMB-Associated Epigenetic Repression and Reduced Immune Activity in Metastatic Colorectal Cancer

Colorectal cancer (CRC) poses a significant threat to public health due to its high mortality rate, particularly in cases of distant metastasis. Researchers sought to improve the understanding of this disease by combining transcriptomic and epigenomic data using machine learning techniques. Analyzing 518 primary tumors from the TCGA-COAD/READ cohort, the study integrated RNA-seq and DNA methylation data from both non-metastatic (M0) and metastatic (M1) tumors.

Five machine learning classifiers were employed to distinguish between M0 and M1 tumors, with stratified nested cross-validation proving most effective. The strongest predictor of metastatic status was observed when combining these two data types, with a ROC-AUC score of 0.787 +/- 0.047 for the Support Vector Machine (SVM) classifier.

Six key features emerged as significant predictors across multiple models, including ARC, ASPDH, C13orf15, C4orf23, GPATCH3, and the DNA methylation site cg12040555. Notably, the methylation level of cg10057218 was inversely correlated with GSDMB expression, with a Spearman rho value of -0.589. Furthermore, statistical mediation analysis revealed a significant indirect association between M1 status and reduced GSDMB expression through cg10057218 methylation, accounting for 82.1% of the overall effect.

In addition to molecular differences, M1 tumors exhibited reduced immune-related transcriptional activity. A 20-feature molecular score based on these integrated features separated M1 from M0 tumors with an impressive ROC-AUC of 0.918. Importantly, higher scores were associated with shorter overall survival, with a log-rank p-value of 1.67 x 10^-7 in the Logistic Regression model.

These findings underscore the potential of a multiomics approach to identify epigenetic features of metastatic CRC, paving the way for novel diagnostic and therapeutic strategies.