Skip to content

bolerodata

bolerodata is the dataset, model-zoo and metadata registry for Bolero — a model that predicts cell-state-specific gene regulation from DNA sequence. It is a thin, uniform Python API that maps short keys (dataset names, model keys, tissue / cell-type names) to the artifacts behind the paper — single-cell ATAC/RNA datasets, trained model checkpoints, QTL collections and differential-accessibility records — and returns them as pandas / AnnData objects and ready-to-use file paths.

What it provides

Object What it gives you
metadata Curated tables: dataset metadata, the model zoo, QTL collection, JASPAR motif clusters, DA records
DATASETS Per-dataset accessor (DATASETS.HumanBrainDev): cell/sample/cluster metadata, meta-cell AnnData, pseudobulks, reference bigwigs, motif scans
MODELS The model zoo (MODELS["<key>"]): resolves checkpoints + configs and builds a bolero predictor
GTEx, Zeng2024, GTEx_eqtl_catalog, Onek1k QTL evaluation collections (variant-effect eval sets)
diff_analysis.DA Differential-accessibility records + IGV browsing (optional diff extra) 
import bolerodata
from bolerodata import DATASETS, MODELS, metadata, GTEx

metadata.MODEL_ZOO.head()                 # browse the model zoo (no downloads)
ds = DATASETS["HumanBrainDev"]
ds.cell_metadata                          # auto-downloads the feather on first use
model = MODELS["<model_key>"]             # cheap; the checkpoint downloads lazily

How the data reaches you

bolerodata ships only small registry tables. The large artifacts (checkpoints, metadata feathers, QTL tables, …) are downloaded on demand from HuggingFace the first time you touch them and cached locally — see Data setup & caching. The curated subset that is published is decided by huggingface_digest.md; anything withheld raises a friendly error rather than a broken path.

Relationship to bolero

bolerodata is the data/registry half of the project; the modelling code lives in bolero. Browsing the registry and downloading data needs only bolerodata. Building and running a predictor (MODELS[...].create_predictor()), resolving a Genome, or using diff_analysis additionally needs bolero (and a GPU stack) installed — bolerodata imports it lazily. See Installation.

Next steps