Epigenome-wide association analysis

Epigenome-wide association studies (EWAS) were carried out to agnostically discover differentially methylated CpG sites between case and control newborns. Linear regression models (by the function ‘lmfit’ in the ‘limma’ R package) were carried out, with DNA methylation beta-values (normalized into M-values) as the outcome, and case/control status as the variable of interest. Adjustments were made to account for the following potential confounding factors: 1) gestational age, 2) gender, 3) estimated cell-mixture (by the Refactor method which uses an unsupervised feature selection step followed by a sparse principal component analysis [⁴³]), 4) an estimation of ancestry (by the Epistructure method which uses the principal components of CpG sites known to be influenced by genetic variation [⁴⁴]), 5) batch effects, and 6) ethnicity as reported on the birth certificates. The lambda genomic inflation factor was calculated on p values using the ‘estlambda’ function of the ‘GenABEL’ R package. T-statistics resulting from the linear regressions were divided by the lambda genomic inflation factor to obtain p values corrected for multiple testing. The Bonferroni-corrected p = 1.57 × 10⁻⁷ was used as the genome-wide level of statistical significance. Manhattan plots and quantile-quantile (q-q) plots were drawn according to the ‘qqman’ R package. Additional analyses stratified by ethnicity were carried out. Additionally 1,000 bootstrap EWAS analyses were performed in both ethnic groups separately to decrease potential influences of outliers. Bootstrap-resulting p values were calculated by taking the geometric mean of the distribution of bootstrap p values for each association. A bootstrap p value below 0.05 and a concordance in the direction of the association were the criteria that were applied to cross-validate CL/P-associated sites.