Abstract
The process of diagnosing hazardous alcohol drinking (HAD) is based on self-reported data and is thereby vulnerable to bias. There has been an interest in developing epigenetic biomarkers for HAD that might complement clinical assessment. Because alcohol consumption has been previously linked to DNA methylation (DNAm), we aimed to select DNAm signatures in blood to predict HAD from two demographically and clinically distinct populations (N(total) = 1,549). We first separately conducted an epigenome-wide association study (EWAS) for phosphatidylethanol (PEth), an objective measure of alcohol consumption, and for self-reported alcohol consumption in Cohort 1. We identified 83 PEth-associated CpGs, including 23 CpGs previously associated with alcohol consumption or alcohol use disorder. In contrast, no CpG reached epigenome-wide significance on self-reported alcohol consumption. Using a machine learning approach, two CpG subsets from EWAS on PEth and on self-reported alcohol consumption from Cohort 1 were separately tested for the prediction of HAD in Cohort 2. We found that a subset of 143 CpGs selected from the EWAS on PEth showed an excellent prediction of HAD with the area under the receiver operating characteristic curve (AUC) of 89.4% in training set and 73.9% in validation set of Cohort 2. However, CpGs preselected from the EWAS on self-reported alcohol consumption showed a poor prediction of HAD with AUC 75.2% in training set and 57.6% in validation set. Our results demonstrate that an objective measure for alcohol consumption is a more informative phenotype than self-reported data for revealing epigenetic mechanisms. The PEth-associated DNAm signature in blood could serve as a robust biomarker for alcohol consumption.