Abstract
Noncoding genetic variation contributes to brain disorder risk, but the mechanisms through which it acts in specific brain cell types remain unclear. DNA methylation (DNAm), a highly cell type-specific regulatory layer in the brain, may mediate noncoding genetic risk, yet whether methylation at CG (mCG) and neuron-enriched non-CG (mCH) dinucleotides contribute differently to that risk remains unknown. Here we develop a deep learning framework that predicts DNAm from DNA sequence and estimates variant effects across 186 brain cell subtypes in both mCG and mCH, leveraging single-nucleus DNAm profiles from 46 brain regions. The models reveal distinct transcription factor (TF) programs underlying the two methylation contexts, with mCH-associated TFs showing stronger evolutionary constraint. Predicted variant effects agree closely with cell type-matched mQTLs in both direction and magnitude. Common variants predicted to affect mCG, particularly in excitatory neurons, show substantially greater heritability enrichment for brain-related traits than variants affecting mCH. By contrast, noncoding de novo mutations in autism preferentially perturb mCH, but not mCG, at conserved neuronal regulatory regions. This pattern is replicated across two independent cohorts totaling 5,782 probands and 4,053 unaffected siblings. Together, these findings indicate that common and rare noncoding variants contribute to brain disorders through distinct DNA methylation mechanisms.
Full Text
The Full Text of this preprint is available as a PDF (8.8 MB). The Web version will be available soon.