Taykhoom/regulatory-vep-dart-eval

DART-Eval Task 5: Regulatory Variant Effect Prediction

This repository contains the hg38 parquet release of Task 5 from DART-Eval. The benchmark measures whether a model can distinguish experimentally supported regulatory QTLs from controls and recover the direction and magnitude of their chromatin-accessibility effects.

The two source studies are exposed as separate Hugging Face configs:

• african-caqtl: chromatin accessibility QTLs measured across African population samples.
• yoruban-dsqtl: DNase sensitivity QTLs measured in Yoruba lymphoblastoid cell lines.

Both configs use hg38 coordinates and the same reference/alternate sequence convention.

Loading

from datasets import load_dataset

african = load_dataset(
    "Taykhoom/regulatory-vep-dart-eval",
    "african-caqtl",
    split="test",
)

yoruban = load_dataset(
    "Taykhoom/regulatory-vep-dart-eval",
    "yoruban-dsqtl",
    split="test",
)

Dataset size

Config	Rows	Positive labels
african-caqtl	208,872	11,068
yoruban-dsqtl	27,347	559

Common columns

Column	Description
split	Always test.
sequence	2,114 bp hg38 sequence carrying the reference allele.
label	Whether the source study labels the variant as a regulatory QTL.
pair_id	Stable identifier containing hg38 position and alleles.
alt_sequence	Matched 2,114 bp sequence carrying the alternate allele.
chrom	hg38 chromosome or contig name.
pos	One-based hg38 variant position.
ref, alt	Reference and alternate alleles in hg38 orientation.
effect_size	Experimental effect expressed as alternate minus reference.
rsid	dbSNP identifier when available.
source_index	Row index in the source benchmark table before filtering or liftover.

The variant is at zero-based sequence index 1,057 in both sequence and alt_sequence. Both columns use the ACGTN alphabet; ambiguous non-focal reference-context symbols are represented as N.

The African caQTL config also includes:

Column	Description
p_value	Experimental association p-value.
in_peaks	Whether the variant falls in the source ATAC-seq peak set.

Genome build and allele orientation

The African caQTL source already uses GRCh38. The Yoruba dsQTL source used hg19 and was lifted to hg38 before publication here. Reference and alternate sequences were rebuilt from hg38, and signed effects were flipped when the reference allele changed during liftover.

For African caQTLs, the literal current TSV beta and this Parquet effect_size are both alternate minus reference. The corrected intermediate HDF5 instead follows its stored sequence order and records allele 2 minus allele 1, negating 39,504 source rows where allele 2 is reference. This distinction and the Task 4 peak-set correction are documented in DART-Eval issue #14.

Yoruba source obs.estimate is allele 1 minus allele 2. Processing negates that source value before reference/alternate normalization, so every finite effect_size here is alternate minus reference. Published Yoruba benchmark metrics retain the exact hg19 row membership; this hg38 file is the separate model-input view.

Of 27,373 official-use Yoruba variants, 27,347 map uniquely to hg38. The 26 source intervals deleted in hg38 are not included in the yoruban-dsqtl config. One mapped variant lies on chr8_KI270821v1_alt.

The maintained processing code and task documentation are available at:

https://github.com/TaykhoomDalal/DART-Eval-Processing/tree/main/VEP

The release contract and artifact metadata are documented at:

https://github.com/TaykhoomDalal/DART-Eval-Processing/blob/main/manifest.json

The original benchmark code is available at:

https://github.com/kundajelab/DART-Eval

Sources and citation

African caQTLs:

DeGorter, M. K. et al. Transcriptomics and chromatin accessibility in multiple African population samples. bioRxiv (2023). https://doi.org/10.1101/2023.11.04.564839

Yoruba dsQTLs:

Degner, J. F. et al. DNase I sensitivity QTLs are a major determinant of human expression variation. Nature 482, 390-394 (2012). https://doi.org/10.1038/nature10808

Please also cite DART-Eval:

https://arxiv.org/abs/2412.05430

@inproceedings{patel2024darteval,
  title = {DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA},
  author = {Patel, Aman and Singhal, Arpita and Wang, Austin and Pampari, Anusri and Kasowski, Maya and Kundaje, Anshul},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {37},
  year = {2024},
  url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/71998bfc3217ffe1cca1ee084dfadadd-Abstract-Datasets_and_Benchmarks_Track.html}
}

License

This repository repackages data distributed with DART-Eval. It does not assign a new license to the underlying data. The African source dataset is described by DART-Eval as CC BY 4.0; the Yoruba data and all other materials retain their original terms. Users should review and follow the source-study and DART-Eval requirements.