DNA sequence-to-function model that predicts chromatin accessibility and expression for any cell state set by a single-cell atlas embedding.
No providers recorded yet. Browse all providers
A sequence-to-function model such as Borzoi reads hundreds of kilobases of DNA and predicts chromatin and expression tracks, but each output track is a fixed label chosen at training time: "L2/3 IT neuron" is one column, and a neuron from a different cortical region, a younger brain or a stimulated animal has no column at all. Single-cell foundation models have the opposite problem: they embed cells continuously but cannot read DNA. Bolero joins the two by making cell state an input the user sets rather than a category the model was trained on.
The mechanism is conditional weight generation. Bolero keeps the Flashzoi variant of Borzoi frozen and, for every adapted layer, uses a small network to turn a cell-state embedding into low-rank weight updates that are added to the backbone's weights. Each cell state therefore gets its own effective network, while what Borzoi learned about DNA stays shared. Because the cell state alters only the weights, predictions remain fully attributable to the DNA sequence, so the attribution and in silico mutagenesis used with DNA-only models still apply. Developmental age, tissue, species and a transcription-factor activity score can enter the same way as extra conditioning axes.
Bolero was developed by Hanqing Liu, Sheng Mao, Jingtian Zhou, Michael Greenberg and colleagues at the Arc Institute, Harvard Medical School, Harvard University, FutureHouse, the Broad Institute and Stanford University. It was posted to bioRxiv in October 2026 and has not yet been peer reviewed.
Bolero takes a 524,288 bp one-hot DNA sequence and predicts 16,384 bins at 32 bp resolution, following Borzoi's output geometry. The backbone (convolutional stem, residual tower with U-Net-style upsampling, eight transformer layers with 1,536 channels) is initialized from the Flashzoi fold-0 checkpoint and kept frozen; LoRA ranks range from 2 in the DNA stem to 60 in the final convolutions. Heads predict accessibility, paired differences between states, and gene expression. Training used Bolero-10M, an atlas of 10.7 million cells from 33 single-cell ATAC and multiome datasets across six mammals: human (5.4M cells), mouse (4.2M), macaque, marmoset, bonobo and opossum. Similar cells were merged into depth-matched pseudobulks. Accuracy rose with trainable parameters up to roughly one billion, then plateaued.
On ten benchmark datasets spanning four species, Bolero matched fine-tuned Borzoi on across-peak correlation and outperformed Borzoi and Scooby on across-sample correlation in every dataset; for expression it outperformed Decima on across-sample correlation. On cortical regional and developmental differentially accessible regions, median across-sample correlation reached 0.59–0.68 against 0.38–0.53 for Scooby. Predictions for held-out cell states beat the nearest training state in 17 of 18 cases. Bolero matched Scooby and exceeded ChromBPNet at ranking fine-mapped caQTLs, and matched Decima on OneK1K eQTLs. With all mouse stimulation data held out, Bolero-AP1-Score predictions tracked measured fear-engram accessibility changes with Pearson r = 0.78.
Bolero is aimed at researchers who need regulatory predictions for cell states that no experiment has profiled: assigning GTEx and GWAS variants to the cell types in which they act, scoring non-coding de novo variants in autism cohorts by developmental stage, nominating activity-regulated enhancers in human neurons where stimulation is impractical, and dissecting the motif identity and spacing rules that make an AP-1 response cell-type-specific. Across 84 Zoonomia mammal genomes, it predicted that AP-1-inducible responses at basal-ganglia regulatory elements are less conserved than their accessibility.
Bolero turns a reference cell atlas into a queryable model of how the genome is read in each cell state, and because its conditioning is independent of the backbone, the same construction could give other sequence-to-function models, such as AlphaGenome, cell-state resolution. Code, checkpoints (six-species, AP-1, and single-species human and mouse models) and the Bolero-10M corpus are released under MIT. Readouts are limited to accessibility and transcript abundance, not transcription-factor occupancy or 3D contacts; the atlas is weighted toward human, mouse and brain; accuracy falls for cell states far from training; and the AP-1 score proxies a single pathway.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.