evals_dart_task3

Cell-type-specific chromatin-accessibility peak dataset from DART-Eval Task 3 ("Discriminating Cell-Type-Specific Elements"). Each row is a 500 bp ATAC-seq consensus-peak window (±250 bp around the summit, GRCh38), labeled by the cell type it is differentially accessible in. The benchmark question: can a model embedding distinguish cell types from the sequence alone?

Interval dataset, not variants — no ref/alt, no consequence annotation, no matching and no subsampling. The window set is DART-Eval's input_data/top_5000_deseq_peaks.tsv — the top 5,000 DESeq2 differentially-accessible peaks per cell type (the subset they feed their zero-shot clustering / UMAP), 25,000 windows, 5,000 balanced per cell type, with unique peak coordinates.

Description

Element	500 bp ATAC-seq consensus peak (±250 bp around the summit; window midpoint = summit)
Label	One of 5 cell lines: GM12878, H1ESC, HEPG2, IMR90, K562
Selection	Top 5,000 DESeq2 differentially-accessible peaks per cell type (25,000 total, balanced)
Assay	ATAC-seq chromatin accessibility (ENCODE)
Source data	DART-Eval, Synapse syn62161401 (top_5000_deseq_peaks.tsv), project syn60581042
Genome build	GRCh38
Coordinates	0-based, half-open (end - start == 500)
Matching	none (no subsampling)

The full 500 bp peak is stored (never pre-cropped to a model's context window), so the embedding context / pooling choice stays an open downstream decision.

Splits

DART-Eval's canonical 3-way chromosome holdout (verbatim from their Task-3 training scripts), shipped one file per split. Per-split counts follow the peaks' genomic distribution.

File	Windows	Chromosomes
train.parquet	17,965	1, 2, 3, 4, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, X, Y
validation.parquet	1,958	6, 21
test.parquet	5,077	5, 10, 14, 18, 20, 22
total	25,000

Windows per cell type

Cell type	train	validation	test	total
GM12878	3,569	382	1,049	5,000
H1ESC	3,699	318	983	5,000
HEPG2	3,484	376	1,140	5,000
IMR90	3,572	325	1,103	5,000
K562	3,641	557	802	5,000

Columns

Column	Type	Description
chrom	str	Chromosome (no chr prefix), GRCh38
start	int	Window start — 0-based, inclusive
end	int	Window end — 0-based, exclusive (end - start == 500)
label	str	Cell type the peak is differentially accessible in

Provenance

Built by the marin-dna eval pipeline at commit e3bccd1.

• Curation pipeline: snakemake/evals
• Rules: snakemake/evals/workflow/rules/dart_task3.smk
• Parsing + splitting: src/marin_dna/pipelines/evals/dart_task3.py

License

Released under the terms of its upstream sources. The peak set is redistributed from DART-Eval (Task 3) and derives from ENCODE ATAC-seq in the 5 cell lines (freely redistributable under the ENCODE data-use policy); DART-Eval ships no explicit license, so consult it and ENCODE for redistribution and commercial-use terms.

Citation

If you use this benchmark, please cite the upstream sources:

• DART-Eval — Patel et al. 2024, arXiv 2412.05430 (NeurIPS D&B 2024)
• ENCODE Project Consortium — the ATAC-seq data for GM12878, H1ESC, HEPG2, IMR90, K562