Abstract
Abstract
Super-enhancers (SEs) drive cell-identity gene expression, yet their identification relies on epigenomic profiling (e.g., H3K27ac ChIP-seq), precluding discovery in unprofiled cellular states and leaving their underlying sequence grammar unknown. Here, we present Borzoi-SE, a deep learning framework built upon the genomic transformer Borzoi that formulates SE discovery as hierarchical, cell-type-resolved genomic object detection directly from long-range sequence across 90 human lineages. Base-resolution attribution recovers lineage-defining master transcription factors across cell types. Within SEs shared across lineages, it reveals lineage-dependent motif usage, suggesting that the same regulatory regions can support distinct cell-type-specific programs. Using
in silico
ablation, Borzoi-SE decoded the functional hierarchy of the β-globin super-enhancer, prioritizing hypersensitive site 2 (HS2) as the dominant constituent, which we experimentally verified via CRISPR interference. Trained without variant-effect labels, Borzoi-SE identified gain-of-super-enhancer variants in different tissues. At the
CHRM2
locus, a single-nucleotide variant created a
de novo
glucocorticoid receptor (GR) motif that recruited activated GR and potentiated adjacent MEF2 regulatory syntax, nucleating a cardiac super-enhancer. Orthogonal multi-omics and population-genetic evidence linked this variant to increased
CHRM2
expression and variation in resting heart rate and PR-interval duration. Borzoi-SE establishes a general framework for decoding SE regulatory grammar and prioritizing causal non-coding variants directly from sequence.