Evaluating ancestry adjustment in multi-ancestry epigenome-wide analysis

Proper adjustment for population substructure is essential in epigenome-wide association studies (EWAS), particularly in cohorts with diverse ancestries. EPISTRUCTURE offers a genotype-free approach to ancestry inference, originally developed using a European reference population from the Cooperative Health Research in the Region of Augsburg (KORA) study. However, its effectiveness in genetically diverse, multi-ancestry cohorts remains insufficiently evaluated. For EWAS using cord-blood samples from the Hyperglycemia and Adverse Pregnancy Outcome (HAPO) study, we systematically assessed the ancestry adjustment performance of EPISTRUCTURE principal components (PCs) derived from the widely used KORA-based reference set versus new reference sets generated from genotyping data of the multi-ancestry HAPO cohort. HAPO-based reference sets were defined by varying SNP - CpG $R^2$ thresholds (e.g. RS30: $R^2>0.3$) to identify ancestry-informative CpGs. We applied these reference sets for population substructure adjustment in EWAS of three newborn adiposity traits, birthweight, cord C-peptide, and sum of skinfolds, to evaluate their impact on association detection and biological interpretation. Compared to the KORA reference, the HAPO RS30 reference consistently produced lower genomic inflation and identified more biologically relevant associations for birthweight and cord blood C-peptide in EWAS of HAPO cord blood samples (n = 3,116). Pathway enrichment analyses revealed strong immune and metabolic signals, including pathways uniquely captured by EWAS when using the HAPO-derived reference for ancestry adjustment. Trait enrichment using the EWAS Catalog further confirmed associations with fetal growth, maternal metabolic traits, and glucose regulation. Our findings demonstrate that reference sets derived from multi-ancestry cohorts like HAPO better capture underlying population substructure and improve ancestry adjustment in diverse EWAS settings.