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

• Authors: Chang Shu, Yale School of Medicine; Amy C Justice, Connecticut Veteran Healthcare System; Xinyu Zhang, Yale School of Medicine; Vincent Marconi, Emory University; Dana B Hancock, RTI International; Eric O Johnson, RTI International; Ke Xu, Yale School of Medicine
• Journal Title: EPIGENETICS
• Volume: Volume 16, Number 7
• Publisher: TAYLOR & FRANCIS INC | 2020-10-23, Pages 741-753
• Type of Work: Article | Final Publisher PDF
• Publisher DOI: http://doi.org/10.1080/15592294.2020.1824097
• Subjects: Health Sciences, Medicine and Surgery
• Keywords: Science & Technology; Life Sciences & Biomedicine; Biochemistry & Molecular Biology; Genetics & Heredity; DNA methylation; HIV; mortality risk; machine learning prediction; LIFE EXPECTANCY; ANTIRETROVIRAL THERAPY; AGING COHORT; INDEX; AGE; CLASSIFICATION; SMOKING; PACKAGE; IMPACT; CANCER

Abstract

Background: With the improved life expectancy of people living with HIV (PLWH), identifying vulnerable subpopulations at high mortality risk is important. Evidences showed that DNA methylation (DNAm) is associated with mortality in non-HIV populations. Here, we established a panel of DNAm biomarkers that can predict mortality risk among PLWH. Methods: 1,081 HIV-positive participants from the Veterans Ageing 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 and the Gene Ontology enrichment analysis of the predictive CpG sites were performed. Results: We selected a panel of 393 CpGs for the ensemble prediction model that 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 high mortality risk group was significantly associated with 10-year mortality (hazard ratio = 1.79, p = 4E-05) compared with low risk group. These 393 CpGs were located in 280 genes enriched in immune and inflammation response 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. Abbreviations: AUC: Area Under Curve; CI: Confidence interval; DMR: differentially methylated region; DNA: Deoxyribonucleic acid; DNAm: DNA methylation; DAVID: Database for Annotation, Visualization, and Integrated Discovery; EWA: epigenome-wide association; FDR: False discovery rate; FWER: Family-wise error rate; GLMNET: elastic-net-regularized generalized linear models; GO: Gene ontology; HIV: Human immunodeficiency virus; HM450K: Human Methylation 450 K BeadChip; k-NN: k-nearest neighbours; NK: Natural killer; PC: Principal component; PLWH: people living with HIV; QC: Quality control; SVM: Support Vector Machines; VACS: Veterans Ageing Cohort Study; XGBoost: Extreme Gradient Boosting Tree.