arcinstitute/bolero-models

Bolero models

Trained checkpoints for Bolero, a sequence-to-function model that predicts cell-state-specific chromatin accessibility and gene expression from DNA sequence. Each checkpoint holds LoRA weights trained on a Borzoi backbone.

Layout

One folder per model: <ModelGroup>/<model_key>/{checkpoint.pt, config.json}. -Gene models add a gene-expression head; +AP1 models also take an AP-1 motif score as input.

Model details

• Architecture: the Flashzoi variant of Borzoi (fold-0 checkpoint, frozen) with conditional low-rank adapters. An MLP generates each adapted layer's low-rank weights from a 30-dimensional cell-state embedding, plus optional conditions (tissue, AP-1 activity score).
• Input: 524,288 bp of one-hot DNA, a cell-state embedding and, optionally, a bulk reference accessibility track. Without the track the model runs in DNA-only mode.
• Output: chromatin accessibility (and, for -Gene models, transcript abundance) at 32 bp resolution over 16,384 bins.

Which checkpoint to use

Key	Model
Borzoi10M	Bolero, trained on all of Bolero-10M (6 species). The main model in the paper
Bolero10M+AP1+260104	Bolero-AP1-Score: Bolero also conditioned on the AP-1 activity score. The paper's AP-1 model
Bolero10M+AP1	An earlier AP-1 model, trained with an earlier version of the AP-1 score
WholeHuman, WholeMouse	Bolero trained on one species' datasets

The -Gene variant of a key adds the gene-expression head.

Intended use

Research use: predicting cell-state-specific accessibility and expression from sequence, scoring non-coding variant effects per cell state, sequence attribution and motif analysis, and cross-species prediction. The models are not validated for clinical use.

Training data

Bolero-10M: 33 single-cell ATAC and multiome datasets, 10.7 million cells, 6 species (human, mouse, macaque, marmoset, bonobo, opossum). The data are in arcinstitute/bolero-data. Genomic regions follow the Borzoi fold-0 split: training folds 2-7, validation fold 0, test fold 1.

Evaluation

The models were evaluated on held-out test regions for across-peak and across-sample accuracy against Scooby, fine-tuned Borzoi and Decima. Further evaluations covered held-out cell states, fine-mapped caQTLs and eQTLs, held-out stimulation paradigms and other species. See the paper for results.

Limitations

• The training atlas is weighted toward human and mouse, and toward brain.
• Predictions are accessibility and transcript abundance. They do not resolve transcription-factor binding or 3D genome contacts.
• Prediction for cell states far from the training states is less accurate.
• The AP-1 score is a proxy for one inducible transcription-factor pathway.

Usage

Load the models through bolerodata, which downloads the files on first use. Building a predictor also needs bolero and a GPU.

pip install bolerodata

from bolerodata import MODELS

model = MODELS["WholeHuman"]          # registry entry, nothing downloaded yet
model.dataset = "HumanBrainDev"       # dataset whose cell states to predict
predictor = model.create_predictor()  # downloads checkpoint, config and inputs

Inputs (pseudobulk records, peaks, reference bigwigs) come from arcinstitute/bolero-data.

Citation

Liu H, Mao S, Gallent T, Chen Y, Yan Z, Vogel I, Traunmuller L, Zhang R, Buenrostro JD, Zhou J, Greenberg ME. Predicting cell-state-specific gene regulation from DNA across species. bioRxiv (2026). doi:10.64898/2026.10.19.760026

@article{liu2026bolero,
  title   = {Predicting cell-state-specific gene regulation from DNA across species},
  author  = {Liu, H. and Mao, S. and Gallent, T. and Chen, Y. and Yan, Z. and
             Vogel, I. and Traunmuller, L. and Zhang, R. and Buenrostro, J. D. and
             Zhou, J. and Greenberg, M. E.},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.19.760026}
}