DNA Benchmarks is a collection of genomic datasets organized for benchmarking DNA foundation models, genomic representation learning methods, and multimodal genomic learning frameworks.

The collection covers a wide range of sequence scales, from short regulatory sequences of a few hundred base pairs to genome-scale tiling datasets derived from whole genomes.

This repository is designed as a general-purpose benchmark hub rather than a dataset tied to a single model. It contains both original downloaded files and standardized raw files that can be used by sequence-based models, visual genomic representation models, and multimodal genomic models.

OpticalDNA uses selected datasets from this repository, but this dataset collection itself is model-agnostic and can be used independently of OpticalDNA.

This repository contains both:

Official raw data downloaded from the original public sources without additional processing.
VisualDNA-compatible raw data converted into a unified format that can be directly used by VisualDNA-based pipelines and downstream genomic modeling frameworks.

Repository Structure

The repository contains two core directories:

dna_benchmarks/
├── official_data/
└── raw_data/

`official_data/`

The official_data directory stores files downloaded from official or original public sources without additional processing. These files are preserved to support data provenance and reproducibility.

Current structure:

official_data/
├── data_long_range_dna/
│   └── eQTL.zip
└── genomic_benchmarks/
    ├── README.md
    ├── demo_coding_vs_intergenomic_seqs.zip
    ├── demo_human_or_worm.zip
    ├── drosophila_enhancers_stark.zip
    ├── dummy_mouse_enhancers_ensembl.zip
    ├── human_enhancers_cohn.zip
    ├── human_enhancers_ensembl.zip
    ├── human_ensembl_regulatory.zip
    ├── human_nontata_promoters.zip
    └── human_ocr_ensembl.zip

`raw_data/`

The raw_data directory stores datasets in a format supported by VisualDNA-style data processing.

A dataset in this directory may be:

converted from the corresponding files in official_data, or
directly uploaded if it already follows a VisualDNA-supported raw format.

The standard layout is:

raw_data/
└── <dataset_name>/
    └── raw/
        ├── <dataset_name>.csv
        └── statistic.txt

or, for larger datasets:

raw_data/
└── <dataset_name>/
    └── raw/
        ├── <dataset_name>.parquet
        └── statistic.txt

Each statistic.txt file records basic dataset information, such as the number of samples, available columns, sequence statistics, label distribution, and split statistics when available.

Dataset Groups

1. Genome-scale pretraining and tiling datasets

These datasets are designed for large-scale genome representation learning and pretraining.

raw_data/
├── RiceSuperPIRdb-PRETRAIN_GENOME_TILING/
│   └── NIP-T2T_w2048_o1920_SeqCase.UPPER/
│       └── raw/
│           ├── NIP-T2T_w2048_o1920_SeqCase.UPPER.parquet
│           └── statistic.txt
└── hg38-2048/
    └── raw/
        ├── hg38-2048.parquet
        └── statistic.txt

These datasets can be used for:

genome-scale pretraining
masked sequence modeling
long-context genomic sequence modeling
cross-species genomic representation learning
visual or multimodal genomic modeling

2. Long-range DNA / eQTL datasets

The repository includes long-range DNA datasets for evaluating models on regulatory signals that may require extended genomic context.

raw_data/
└── data_long_range_dna/
    └── eqtl.zip

The eqtl.zip archive contains nine tissue-specific datasets:

eqtl/
├── Adipose_Subcutaneous/
│   └── raw/
│       ├── Adipose_Subcutaneous.csv
│       └── statistic.txt
├── Artery_Tibial/
│   └── raw/
│       ├── Artery_Tibial.csv
│       └── statistic.txt
├── Cells_Cultured_fibroblasts/
│   └── raw/
│       ├── Cells_Cultured_fibroblasts.csv
│       └── statistic.txt
├── Muscle_Skeletal/
│   └── raw/
│       ├── Muscle_Skeletal.csv
│       └── statistic.txt
├── Nerve_Tibial/
│   └── raw/
│       ├── Nerve_Tibial.csv
│       └── statistic.txt
├── Skin_Not_Sun_Exposed_Suprapubic/
│   └── raw/
│       ├── Skin_Not_Sun_Exposed_Suprapubic.csv
│       └── statistic.txt
├── Skin_Sun_Exposed_Lower_leg/
│   └── raw/
│       ├── Skin_Sun_Exposed_Lower_leg.csv
│       └── statistic.txt
├── Thyroid/
│   └── raw/
│       ├── Thyroid.csv
│       └── statistic.txt
└── Whole_Blood/
    └── raw/
        ├── Whole_Blood.csv
        └── statistic.txt

These datasets are suitable for:

eQTL prediction
long-range regulatory modeling
tissue-specific regulatory sequence prediction
long-context genomic benchmark evaluation

3. Genomic Benchmarks datasets

The genomic_benchmarks group contains standard DNA sequence classification benchmarks converted into the unified raw-data layout.

raw_data/
└── genomic_benchmarks/
    ├── demo_coding_vs_intergenomic_seqs/
    ├── demo_human_or_worm/
    ├── drosophila_enhancers_stark/
    ├── dummy_mouse_enhancers_ensembl/
    ├── human_enhancers_cohn/
    ├── human_enhancers_ensembl/
    ├── human_ensembl_regulatory/
    ├── human_nontata_promoters/
    └── human_ocr_ensembl/

Each dataset follows:

<dataset_name>/
└── raw/
    ├── <dataset_name>.csv
    └── statistic.txt

These datasets cover tasks such as:

coding versus intergenic sequence classification
species classification
enhancer prediction
promoter prediction
regulatory element prediction
open chromatin region prediction

4. Nucleotide Transformer downstream tasks

The nucleotide_transformer_downstream_tasks group contains downstream genomic prediction tasks commonly used for evaluating DNA language models and genomic foundation models.

raw_data/
└── nucleotide_transformer_downstream_tasks/
    ├── H3/
    ├── H3K14ac/
    ├── H3K36me3/
    ├── H3K4me1/
    ├── H3K4me2/
    ├── H3K4me3/
    ├── H3K79me3/
    ├── H3K9ac/
    ├── H4/
    ├── H4ac/
    ├── enhancers/
    ├── enhancers_types/
    ├── promoter_all/
    ├── promoter_no_tata/
    ├── promoter_tata/
    ├── splice_sites_acceptors/
    ├── splice_sites_all/
    └── splice_sites_donors/

Each task is stored as:

<task_name>/
└── raw/
    ├── <task_name>.csv
    └── statistic.txt

These datasets support tasks such as:

histone mark prediction
enhancer prediction
enhancer type classification
promoter prediction
TATA and non-TATA promoter classification
splice site prediction
chromatin-related sequence classification

File Format

CSV files

Most downstream benchmark datasets are provided as CSV files. A typical CSV file contains a DNA sequence column and one or more task-specific label or metadata columns.

Typical columns may include:

sequence
label
split

or equivalent dataset-specific names such as:

seq
fasta_seq
target
class

Please inspect the corresponding statistic.txt file for the exact schema of each dataset.

Parquet files

Large-scale genome tiling and pretraining datasets are provided in Parquet format for more efficient storage and loading.

Typical columns may include:

index
sequence / seq / fasta_seq
split
chromosome
start
end
species
additional metadata

The exact schema may vary by dataset.

`statistic.txt`

Each VisualDNA-compatible raw dataset includes a statistic.txt file. This file provides dataset-level information and should be checked before training or evaluation.

It may include:

number of samples
column names
sequence length statistics
label distribution
split distribution
metadata summary

Usage

Download the repository

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="hxxiang/dna_benchmarks",
    repo_type="dataset",
    local_dir="./dna_benchmarks",
)

Load a CSV dataset

import pandas as pd

path = "./dna_benchmarks/raw_data/genomic_benchmarks/human_nontata_promoters/raw/human_nontata_promoters.csv"
df = pd.read_csv(path)

print(df.head())
print(df.columns)

Load a Parquet dataset

import pandas as pd

path = "./dna_benchmarks/raw_data/hg38-2048/raw/hg38-2048.parquet"
df = pd.read_parquet(path)

print(df.head())
print(df.columns)

Inspect dataset statistics

with open(
    "./dna_benchmarks/raw_data/hg38-2048/raw/statistic.txt",
    "r",
    encoding="utf-8",
) as f:
    print(f.read())

Inspect eQTL datasets after extracting the archive

import zipfile
from pathlib import Path

zip_path = Path("./dna_benchmarks/raw_data/data_long_range_dna/eqtl.zip")
extract_dir = Path("./dna_benchmarks/raw_data/data_long_range_dna/eqtl")

with zipfile.ZipFile(zip_path, "r") as zf:
    zf.extractall(extract_dir)

for csv_path in sorted(extract_dir.glob("eqtl/*/raw/*.csv")):
    print(csv_path)

Usage with VisualDNA

The raw_data directory is organized to be compatible with VisualDNA-style data processing. Each dataset is expected to follow:

<dataset_name>/
└── raw/
    ├── <dataset_name>.csv or <dataset_name>.parquet
    └── statistic.txt

This format allows VisualDNA-based tools to locate and process genomic datasets in a consistent way.

VisualDNA can further convert these raw genomic sequences into model-specific representations, such as rendered visual genomic documents, pixel-based DNA images, or other supported formats.

Usage with OpticalDNA

OpticalDNA uses selected datasets from this repository for genomic visual representation learning and benchmark evaluation.

In this setting, datasets in raw_data can be used as input to the OpticalDNA / VisualDNA pipeline. The pipeline may convert DNA sequences into visual representations and then use them for pretraining, fine-tuning, or evaluation.

This repository is not limited to OpticalDNA. The same datasets can also be used by conventional DNA language models, CNN-based sequence models, Transformer-based genomic models, and other genomic representation learning methods.

Intended Use

This dataset collection is intended for research on:

DNA foundation models
genomic sequence modeling
genomic benchmark evaluation
regulatory genomics
promoter and enhancer prediction
chromatin profile prediction
splice site prediction
eQTL and long-range regulatory prediction
genome-scale pretraining
cross-species genomic representation learning
visual and multimodal genomic representation learning

Out-of-Scope Use

This dataset collection is not intended for:

clinical diagnosis
medical decision-making
identifying individuals from genomic data
inferring sensitive personal information
direct therapeutic recommendation
unvalidated biological or clinical interpretation

Any biological conclusions drawn from models trained on these datasets should be validated through appropriate downstream analyses and experimental evidence.

Data Sources and Provenance

This repository aggregates and reformats datasets from multiple public genomic benchmark sources, including:

Genomic Benchmarks datasets
Long-range DNA / eQTL-related datasets
Nucleotide Transformer downstream benchmark tasks
Genome-scale human and rice sequence resources

The official_data directory preserves selected original downloaded files when available. The raw_data directory provides converted or VisualDNA-compatible versions.

Users should cite the original dataset sources when using specific subsets.

License

This repository is released under the MIT License.

Please note that some files in official_data may originate from external public datasets. Users are responsible for following the licenses, terms of use, and citation requirements of the original data sources.

Dataset Maintenance

This dataset collection is expected to grow over time. Future releases may include additional genomic benchmarks, larger pretraining corpora, more species, and additional multimodal genomic annotations.

The current repository layout is designed to support future extensions:

official_data/
└── <source_dataset_collection>/

raw_data/
└── <benchmark_group>/
    └── <dataset_name>/
        └── raw/
            ├── <dataset_name>.csv or <dataset_name>.parquet
            └── statistic.txt

Limitations

The datasets in this repository are collected from multiple sources and may differ in sequence length, label definition, split strategy, preprocessing procedure, species, and task formulation.

The raw_data files provide a unified storage layout, but they do not imply that all datasets share the same biological task, label space, or evaluation protocol.

Users should inspect each dataset and its corresponding statistic.txt file before conducting experiments.

For fair benchmarking, users should avoid mixing train, validation, and test splits unless the specific task protocol explicitly allows it.

Citation

If you use this dataset collection, please cite this repository and the corresponding original data sources.

If you use this dataset with OpticalDNA, please also cite:

@inproceedings{xiang2026rethinking,
  title     = {Rethinking Genomic Modeling Through Optical Character Recognition},
  author    = {Xiang, Hongxin and Ma, Pengsen and Cao, Yunkang and Yu, Di and Chen, Haowen and Yang, Xinyu and Zeng, Xiangxiang},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=nggzekChuU}
}

For the arXiv version, please cite:

@article{xiang2026rethinking_arxiv,
  title   = {Rethinking Genomic Modeling Through Optical Character Recognition},
  author  = {Xiang, Hongxin and Ma, Pengsen and Cao, Yunkang and Yu, Di and Chen, Haowen and Yang, Xinyu and Zeng, Xiangxiang},
  journal = {arXiv preprint arXiv:2602.02014},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.02014}
}

Please also cite the original papers or repositories associated with the specific benchmark subsets you use.

Contact

For questions or issues, please contact:

xianghx@hnu.edu.cn

Downloads last month: 45

Total file size:

25.3 GB

Paper for hxxiang/dna_benchmarks

Rethinking Genomic Modeling Through Optical Character Recognition

Paper • 2602.02014 • Published Feb 2 • 1