DNA methylation biomarker selected by an ensemble machine learning approach predicts mortality risk in an HIV-positive veteran population

Abstract

Background: With the improved life expectancy of people living with HIV (PLWH), identifying vulnerable subpopulations at high risk of mortality is important for clinical care. Evidences showed that DNA methylation (DNAm) is associated with aging and mortality in non-HIV populations. Here, we aimed to establish a panel of DNAm biomarkers that can predict mortality risk among PLWH. Methods: 1,081 HIV-positive participants from the Veterans Aging Cohort Study (VACS) were divided into training (N=460), validation (N=114), and testing (N=507) sets. VACS index was used as a measure of mortality risk among PLWH. Model training and fine-tuning were conducted using the ensemble method in the training and validation sets and prediction performance was assessed in the testing set. The survival analysis comparing the predicted high and low mortality risk groups was conducted. The Gene Ontology enrichment analysis of the predictive CpG sites was performed. Results: We selected a panel of 393 CpGs for the ensemble prediction model. The prediction model showed excellent performance in predicting high mortality risk with an auROC of 0.809 (95%CI: 0.767-0.851) and a balanced accuracy of 0.653 (95%CI: 0.611, 0.693) in the testing set. The predicted high mortality risk group was significantly associated with 10-year mortality (hazard ratio=1.79, p=4E-05) compared with low mortality risk group. These 393 CpGs were located in 280 genes enriched in immune and inflammation responses pathways. Conclusions: We identified a panel of DNAm features associated with mortality risk in PLWH. These DNAm features may serve as predictive biomarkers for mortality risk among PLWH.