Decoding common and rare noncoding variant effects across cellular and developmental contexts.

Interpreting how noncoding variants act in specific cell types across human development is a major challenge. Here we generated 3 billion predictions from deep learning sequence models of chromatin accessibility across diverse fetal and adult cellular contexts. These prioritized functional variants and revealed a dichotomy: common variants are more cell-type-specific, whereas ultra-rare variants had larger and broader effects across cell types, with the strongest evidence of purifying selection in fetal neurons. Leveraging these insights, we developed FLARE (Functional Lasso Analysis of Regulatory Evolution), which integrates evolutionary constraint to prioritize noncoding variants with extreme regulatory effects. FLARE provided a general framework for studying regulatory variation, from de novo mutations in childhood disorders to rare variants underlying outlier adult brain expression and common variants enriched for schizophrenia heritability. Together, these results demonstrate how integrating single-cell chromatin accessibility, population genetics and deep learning can identify regulatory variants that influence human development and disease.