Signature based on molecular subtypes of deoxyribonucleic acid methylation predicts overall survival in gastric cancer

Abstract

BACKGROUND Gastric cancer (GC) ranks as the third leading cause of cancer-related death worldwide. Epigenetic alterations contribute to tumor heterogeneity in early stages. AIM To identify the specific deoxyribonucleic acid (DNA) methylation sites that influence the prognosis of GC patients and explore the prognostic value of a model based on subtypes of DNA methylation. METHODS Patients were randomly classified into training and test sets. Prognostic DNA methylation sites were identified by integrating DNA methylation profiles and clinical data from The Cancer Genome Atlas GC cohort. In the training set, unsupervised consensus clustering was performed to identify distinct subgroups based on methylation status. A risk score model was built based on Kaplan-Meier, least absolute shrinkage and selector operation, and multivariate Cox regression analyses. A test set was used to validate this model. RESULTS Three subgroups based on DNA methylation profiles in the training set were identified using 1061 methylation sites that were significantly associated with survival. These methylation subtypes reflected differences in T, N, and M category, age, stage, and prognosis. Forty-one methylation sites were screened as specific hyper- or hypomethylation sites for each specific subgroup. Enrichment analysis revealed that they were mainly involved in pathways related to carcinogenesis, tumor growth, and progression. Finally, two methylation sites were chosen to generate a prognostic model. The high-risk group showed a markedly poor prognosis compared to the low-risk group in both the training [hazard ratio (HR) = 2.24, 95% confidence interval (CI): 1.28-3.92, P < 0.001] and test (HR = 2.12, 95%CI: 1.19-3.78, P = 0.002) datasets. CONCLUSION DNA methylation-based classification reflects the epigenetic heterogeneity of GC and may contribute to predicting prognosis and offer novel insights for individualized treatment of patients with GC.

Machine learning in the identification of prognostic DNA methylation biomarkers among patients with cancer: a systematic review of epigenome-wide studies

“…Based on selected CpGs in the prior stage, eight studies performed additional feature selection, using methods including LASSO (38, 45, 55, 57, 69, 78, 80), stepwise selection (38, 39), and the Akaike information criterion (64, 77), to further refine the number of CpGs. In contrast, four studies (23, 33, 34)(47) directly trained classification models with clusters as labels using Bayesian networks or random forest, while 17 studies constructed prognostic models (37-39, 45, 46, 53-57, 64, 69, 70, 75, 77, 78, 80).…”

“…As a remedy semi-supervised approaches might be used, which could combine the best of both approaches. Alternatively, studies might use unsupervised and supervised approaches in conjunction at different stages to select features, as done in some of the studies (38, 45, 55, 57, 69, 78, 80).…”