Super-enhancers no longer need epigenomic profiling: Borzoi-SE calls them from DNA sequence, and CRISPRi confirms HS2 as the β-globin hub
Synopsis
Super-enhancers could previously be defined only by stitching epigenomic peaks such as H3K27ac, so unprofiled cell states were out of reach and the underlying sequence grammar stayed unread. Borzoi-SE recasts super-enhancer discovery as cell-type-resolved genomic object detection over 524-kb sequences, returning super-enhancer probabilities and boundaries across 90 human cell and tissue lineages. On held-out chromosomes it reached a region-level PR-AUC of 0.643 and F1 of 0.498, against ceilings of about 0.34 and 0.15 for four fragment-based baselines; in silico ablation of the β-globin super-enhancer ranked HS2 as the dominant hub, and CRISPRi silencing of HS2 reduced chromatin accessibility across the locus by 35% to 85%.
Interpretation
Super-enhancer identification moves from experimental epigenomic peak stitching to direct prediction from primary DNA sequence, yielding super-enhancer probabilities and physical boundaries across 90 human cell and tissue types at once. ROSE-style methods require H3K27ac or MED1 ChIP-seq peaks, stitch them across empirical distance thresholds and keep domains above an empirical signal cutoff, so they cover only profiled cell states and cannot read out sequence grammar. Across five independent chromosome splits, end-to-end detection on continuous unfragmented held-out chromosomes gave Borzoi-SE a region-level PR-AUC of 0.643 and F1 of 0.498, versus baseline ceilings of about 0.34 and 0.15; at 80% recall, false-positive density in quiescent genomic deserts was 0.028±0.002 FP/Mb against 0.265–0.349 FP/Mb for the baselines.
A pretrained Borzoi backbone reads 524-kb input, downsamples it to 256-bp resolution and splits it into two branches, one preserving local sequence context and one driving a dilated Cluster head that models enhancer clustering over roughly 64.7 kb. The Cluster head's 90-channel cluster logits are concatenated with the context features and recalibrated by Squeeze-and-Excitation channel attention before a shallower SE head discriminates true super-enhancers. The output is a 90-cell-type by 1,600-bin prediction matrix over the central ~409.6 kb at 256-bp resolution; on K562, HepG2, GM12878 and WTC11, the matching channels reached PR-AUCs of 0.544–0.703, versus 0.131–0.369 for baselines retrained per lineage.
Ablating constituents of the β-globin locus control region in silico ranked HS2 as the strongest functional hub, with its removal dropping the predicted super-enhancer activity integral ratio for the whole domain to 0.833. Ablating individual promoters such as HBE1 (ratio 0.978) or secondary hypersensitive sites HS1, HS3, HS4 and HS5 (ratios 0.888–0.987) caused only focal perturbations, whereas ablating HS2 depleted probability across the entire domain and downstream promoters. In K562 cells, CRISPRi silencing of HS2 reduced relative chromatin accessibility at HBD, HBG1, HBG2, HBE1, HS1, HS3 and HS4 by 35% to 85%, while silencing HS1, HS3, HS4 or HS5 closed only its own site and left neighbors at 0.60–1.00 accessibility.
Perspective
The demonstrated scope is H3K27ac-derived ROSE annotations across 90 human cell and tissue types, detection and ranking on five held-out chromosome splits, and CRISPRi validation at the β-globin locus control region in K562 cells; the authors propose extending the model into a multi-task framework that jointly predicts one-dimensional functional readouts and 3D contacts such as Hi-C, Micro-C and HiChIP.
Global recall is about 0.40, and detection recall declines monotonically with experimental super-enhancer strength, from 0.61±0.20 in the strongest decile to 0.23±0.14 in the weakest, so false negatives concentrate among borderline elements near the empirical ROSE threshold; the authors also note that relying on one-dimensional sequence leaves open whether the model internalizes 3D genome architecture such as loop extrusion.