license: other
task_categories:
- other
pretty_name: BEND 2.0
license_name: bend2-source-specific
license_link: LICENSE.md
tags:
- biology
- genomics
- dna
- benchmark
- sequence-design
- bend2
configs:
- config_name: atlas_sequences
default: true
data_files:
- split: data
path: data/atlas-population.parquet
- config_name: atlas_source_activity
data_files:
- split: data
path: data/atlas-source-activities.parquet
- config_name: atlas_candidate_activity
data_files:
- split: data
path: data/atlas-candidate-activities.parquet
- config_name: sequence_prior
data_files:
- split: data
path: data/prior-panel.parquet
- config_name: siraj_cell_activity
data_files:
- split: data
path: data/siraj_selective.parquet
Paper · Code · Live leaderboard
We built BEND 2.0 to test whether genomic models can guide the design of DNA. Across eight tasks, we examine how they predict sequence function, account for cell type and genetic background, and choose sequences that produce a desired response.
This repository contains the data used by the benchmark: fixed sequence panels, prepared measurements and unchanged copies of the larger source files. Download only what you need, or keep the complete 1.39 GB release locally.
Browse files · File guide · Source terms · Citation
Start with the enhancer sequences
Atlas contains 7,657 sequences used by BEND 2.0. Models learn from 5,246 sequences, then choose among 2,411 candidates to produce a desired pattern of activity across twelve cell types.
import pyarrow.compute as pc
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download
path = hf_hub_download(
"yhsure/BEND2",
"data/atlas-population.parquet",
repo_type="dataset",
revision="v1.0",
)
sequences = pq.read_table(path)
candidates = sequences.filter(
pc.equal(sequences["split"], "candidate")
).select(["id", "sequence"])
The table keeps the training, validation and candidate assignments, plus 27 excluded sequences for traceability. The two activity tables contain the measured responses. The viewer above shows each table separately; its data label is not a training split.
Download the full release or use it with BEND 2.0
Install huggingface_hub to download files; the example above also uses pyarrow.
from huggingface_hub import snapshot_download
snapshot_download(
"yhsure/BEND2",
repo_type="dataset",
revision="v1.0",
local_dir="bend2-data",
allow_patterns=["data/*", "assets.json", "checksums.json", "LICENSE.md"],
)
If you already have the BEND 2.0 evaluation package, point it to this local copy:
export BEND2_ASSET_DIR="$PWD/bend2-data/data"
bend2 run my_model.py --tasks siraj --output results/my-model
The package checks the files against its existing manifest. Its small bundled datasets are unchanged. Model weights and embeddings are separate from this data release.
Eight tasks
| Task | What is evaluated | Data |
|---|---|---|
| Sequence prior | Recovering a hidden DNA base | 512 GRCh38 windows |
| Splicing effects | Predicting how mutations change splicing | 590,104 variants in 466 genes · OpenSplice |
| Cell differences | Predicting how a mutation's effect differs between two cells | 170,292 substitutions in 128 elements · Siraj |
| Tissue identification | Identifying tissues from predicted expression profiles | 5,782 held-out genes, 51 tissues · GTEx |
| Genetic interactions | Predicting how a mutation's effect depends on the surrounding alleles | 68,026 paired mutations in 114 elements · Siraj |
| Matching a response | Choosing an edit with a similar effect to a reference edit | 331 targets in 110 elements · Siraj |
| Cell-selective editing | Increasing expression in one cell while limiting changes in another | 167,958 substitutions in 127 loci · Siraj |
| Multi-cell design | Choosing enhancers for a desired pattern of activity across cells | 16 patterns, 2,411 candidates · Atlas |
BEND 2.0 keeps related genes, elements or sequence families together when splitting data. Preserve these assignments when comparing models. See the file guide for the stored tables, measurements and exclusions.
Data sources
The original studies provide the experiments; BEND 2.0 provides the benchmark selections and prepared files. Please credit both. Each source retains its own terms, including Siraj's noncommercial restriction.
| Source | Release | Terms |
|---|---|---|
| OpenSplice | Experimental master table, version 5 | CC BY 4.0 |
| Siraj et al. | Saturation mutagenesis | CC BY-NC 4.0 |
| Castillo-Hair et al. | Atlas of AI-designed enhancers | CC BY 4.0 |
| GTEx | v10 public expression summaries | GTEx data terms |
| Genome Reference Consortium / UCSC | GRCh38 / hg38 | Reference assembly terms |
Cell-selective measurements also use ENCODE data from Ryan Tewhey's laboratory; gene coordinates use GENCODE v39. Full acknowledgements and file-specific terms are in LICENSE.md.
The GTEx files are a frozen BEND 2.0 subset of GTEx v10 and do not necessarily reflect the most current or accurate data available from the GTEx Portal. They contain public aggregate expression, not individual donor records or genotypes.
The original asset manifest is preserved in assets.json. checksums.json records the size and SHA-256 of every data file in this release.
Citation
If you use BEND 2.0, cite the benchmark and the original datasets used in your analysis.
@misc{bjerregaard2026benchmarking,
title = {{BEND 2.0}: Benchmarking the Capabilities Needed to Design {DNA}},
author = {Bjerregaard, Andreas and DeVries, Rachael and
Brilli, Dionysia Danai and Antoniadis, Panagiotis and Li, Yan and
Krogh, Anders and Boomsma, Wouter and Winther, Ole and
Marin, Frederikke Isa},
year = {2026},
eprint = {2611.03806},
archivePrefix = {arXiv},
primaryClass = {q-bio.GN},
url = {https://arxiv.org/abs/2611.03806}
}