Datasets:

marin-dna
/

evals_dsqtl

chrom stringclasses 11 values	pos int64 58.4k 249M	ref stringclasses 4 values	alt stringclasses 4 values	label bool 2 classes	effect float64 -1.96 1.81 ⌀	chrombpnet_atac_ips float64 -0.48 0.38	chrombpnet_atac_logfc float64 -2.19 2.56	chrombpnet_dnase_ips float64 -0.21 0.13	chrombpnet_dnase_logfc float64 -1.55 1.51	enformer_dnase_local_logfc float64 -2.2 2.36
1	906,870	A	G	false	null	-0.007423	0.349371	-0.005049	0.249293	0.322042
1	910,558	G	A	false	null	0.000339	0.029383	0.000088	0.022997	0.00212
1	911,109	T	C	false	null	0.000018	-0.004395	0.000016	0.023751	0.008376
1	911,661	C	T	false	null	-0.000029	-0.01181	-0.000011	-0.002494	-0.001846
1	914,061	A	G	false	null	-0.000025	-0.027344	-0.000032	-0.02781	0.003367
1	914,991	G	T	false	null	-0.000015	-0.030552	-0.00002	-0.042853	-0.001422
1	917,584	T	G	false	null	0.000251	-0.096874	0.000158	-0.109916	0.00268
1	921,203	G	A	true	-0.9301	-0.039558	-0.728016	-0.001564	-0.20898	-0.207411
1	928,396	A	C	false	null	-0.000028	-0.013996	-0.000014	-0.01092	-0.003587
1	931,558	G	A	false	null	-0.000077	0.053353	-0.000161	0.050868	-0.002026
1	932,733	G	A	false	null	-0.000039	0.034793	-0.000006	0.009371	0.003991
1	933,653	C	T	false	null	-0.000052	0.01957	-0.000027	0.015163	-0.003571
1	943,937	C	T	false	null	-0.000259	-0.022812	-0.000046	0.001414	0.004688
1	956,632	C	T	false	null	-0.000119	0.03887	-0.00005	0.027026	0.001342
1	966,179	G	A	false	null	-0.004937	-0.227813	-0.00159	-0.107897	0.016725
1	967,369	C	G	false	null	-0.000079	-0.012255	-0.000496	-0.020827	0.035881
1	976,215	A	G	true	0.6285	-0.006116	0.07933	-0.001013	-0.082127	0.434146
1	986,007	G	A	false	null	-0.00002	0.032529	-0.000009	0.007114	-0.008184
1	993,218	C	G	false	null	-0	-0.006678	-0.000005	-0.025973	0.014872
1	995,187	A	G	false	null	0.000014	-0.022089	0.000026	-0.041987	-0.015263
1	995,786	A	G	false	null	0.000058	-0.02397	0.000134	-0.047281	0.032064
1	1,006,159	C	T	false	null	0.000064	0.006775	0.000071	0.05262	-0.004558
1	1,008,332	T	C	false	null	-0.000016	-0.133396	-0.00002	-0.047325	-0.003785
1	1,022,856	T	C	false	null	-0.000146	0.063706	-0.00009	0.042687	0.006437
1	1,045,707	A	G	false	null	0.000065	-0.050021	0.000043	-0.004303	0.00661
1	1,081,817	C	T	false	null	-0.000081	0.072705	-0.000008	0.032142	-0.001089
1	1,083,795	C	G	false	null	-0.000028	-0.021824	-0.000018	-0.020078	-0.000545
1	1,086,315	A	G	false	null	0.000234	-0.066311	0.000276	-0.076014	0.000259
1	1,125,786	T	C	false	null	-0.00001	-0.00751	-0.000001	0.009896	-0.000244
1	1,150,036	T	C	false	null	-0.00003	-0.003515	-0.000004	-0.005269	0.001857
1	1,153,154	T	C	false	null	-0	-0.033664	-0	-0.005294	-0.008539
1	1,156,085	G	A	false	null	-0.000047	-0.04261	-0.000034	-0.022304	-0.004164
1	1,157,209	G	A	false	null	-0.000497	-0.050296	-0.000143	-0.024374	-0.005923
1	1,159,750	T	C	false	null	-0.000072	-0.010502	-0.000008	0.003703	0.029254
1	1,206,007	C	T	false	null	-0.000156	0.010444	-0.000019	-0.008256	0.039049
1	1,210,386	C	G	false	null	-0.000031	0.01934	-0.000041	0.023859	-0.00635
1	1,222,167	T	C	false	null	-0.000482	-0.082479	-0.000268	-0.034939	0.009688
1	1,222,251	G	A	false	null	-0.000234	-0.006458	-0.000236	-0.021063	-0.003732
1	1,222,580	T	C	false	null	-0.000075	-0.01785	-0.000122	-0.011475	0.010603
1	1,222,640	A	C	false	null	-0.001104	-0.041352	-0.001681	-0.061332	-0.013569
1	1,226,586	C	A	false	null	-0.000054	0.01853	-0.000196	0.03749	-0.008301
1	1,226,734	G	A	false	null	-0.00016	0.066217	-0.000349	0.065177	-0.001485
1	1,228,393	C	T	false	null	-0.000327	-0.015758	-0.000093	-0.039286	-0.001537
1	1,228,424	C	T	false	null	-0.000191	0.098929	-0.000055	0.054111	-0.00793
1	1,235,869	T	C	false	null	-0.000006	0.000316	-0.000002	0.01331	-0.004658
1	1,238,902	T	C	false	null	0.000385	0.012744	0.000063	-0.008488	-0.011833
1	1,242,864	G	C	false	null	-0.000031	-0.012867	-0.000007	0.007389	0.003489
1	1,246,407	G	A	false	null	-0.00013	0.020684	-0.000102	0.021805	-0.003351
1	1,246,970	C	T	false	null	-0.000147	-0.004044	-0.000013	-0.003232	0.001545
1	1,251,122	A	T	true	0.7363	-0.035028	0.826351	-0.021172	0.61355	1.075684
1	1,260,213	G	A	false	null	-0.00009	-0.002409	-0.000249	0.023576	-0.004145
1	1,274,920	C	G	false	null	-0.000113	-0.069069	-0.000119	-0.097002	-0.023274
1	1,277,886	T	A	false	null	-0.000012	-0.045492	-0.000005	-0.026796	0.00134
1	1,280,469	A	T	false	null	-0.000003	-0.017447	-0.000007	0.01344	-0.001817
1	1,292,032	G	A	true	0.9236	-0.384122	1.661221	-0.088673	0.695564	1.162808
1	1,372,258	T	C	false	null	0.000858	-0.000999	0.00007	0.01197	0.038685
1	1,433,490	C	A	false	null	-0.000308	0.047083	-0.000077	0.009692	-0.007093
1	1,533,253	C	T	false	null	0.000011	-0.017932	0.000007	-0.020893	-0.004643
1	1,534,092	G	C	false	null	-0.000074	-0.009675	-0.000007	0.006005	-0.001084
1	1,594,618	C	G	false	null	-0.000015	-0.038087	-0.000001	-0.017602	-0.001011
1	1,600,591	C	T	false	null	-0.000094	0.020978	-0.000027	-0.00697	-0.003108
1	1,603,213	C	T	false	null	-0.000013	-0.002982	-0.000012	-0.002676	-0.007915
1	1,607,931	G	A	false	null	-0	0.000288	-0.000028	0.035213	-0.004621
1	1,620,609	G	A	false	null	-0.000007	0.00051	-0.000003	0.028768	-0.009104
1	1,631,911	C	T	false	null	-0.000296	0.01951	-0.000041	0.006369	0.015512
1	1,685,405	G	A	false	null	-0.000008	0.041544	-0.000009	0.033675	0.000807
1	1,724,382	T	G	false	null	-0.000579	-0.043844	-0.00081	-0.046905	0.000874
1	1,753,735	C	T	false	null	-0.000085	-0.003554	-0.000043	0.009463	-0.016177
1	1,753,774	C	G	false	null	-0.000081	0.008152	-0.00005	0.00464	-0.001219
1	1,805,214	T	G	false	null	-0.000042	-0.028886	-0.00007	-0.011144	-0.023266
1	1,864,470	T	C	false	null	-0.00008	0.034193	-0.000945	0.101014	0.062708
1	1,891,375	G	A	false	null	-0.000259	0.006728	-0.000124	0.005806	-0.013286
1	1,924,659	C	T	false	null	-0.000097	0.051931	-0.000132	0.086053	0.000472
1	1,929,687	C	G	false	null	-0.000098	0.017849	-0.00009	0.008015	0.007863
1	1,942,075	A	G	false	null	0.000037	0.010659	0.000071	-0.014123	-0.002677
1	1,942,260	G	C	false	null	0.000043	0.017735	0.000058	0.01111	0.000206
1	1,947,487	T	A	false	null	-0.00409	0.309012	-0.004284	0.285512	0.005278
1	1,958,353	A	T	false	null	0.000244	0.032974	0.000189	0.02653	-0.031017
1	1,959,717	C	T	false	null	-0.000126	0.044419	-0.000063	0.019492	-0.007324
1	1,976,473	C	T	false	null	-0.000009	0.003861	-0.000029	0.004387	-0.006139
1	1,978,114	T	A	false	null	-0.000006	0.060238	-0.000004	-0.000038	-0.000971
1	1,987,096	A	T	false	null	-0.000035	-0.022982	-0.000044	0.000905	0.001716
1	2,005,104	A	G	false	null	-0.000128	-0.04968	-0.000059	0.026098	0.007151
1	2,019,182	G	A	false	null	0.000418	-0.05341	0.00011	-0.026145	-0.006622
1	2,021,343	C	A	false	null	0.000005	-0.010013	0.000006	-0.024986	-0.001347
1	2,031,736	A	G	false	null	0.000018	-0.050181	0.000001	-0.023945	0.001557
1	2,033,837	C	T	false	null	-0.000043	0.03626	-0.000053	0.034799	-0.004578
1	2,042,324	C	A	false	null	-0	-0.006612	-0.000005	0.015357	-0.002207
1	2,044,405	C	T	false	null	-0.000104	-0.000014	-0.000267	-0.051896	-0.002004
1	2,067,567	G	A	false	null	-0.000595	-0.110189	-0.000644	-0.065647	-0.00118
1	2,068,401	C	T	false	null	-0.000025	0.007602	-0.000005	-0.005124	-0.000148
1	2,075,464	G	A	false	null	-0.000332	0.115023	-0.000113	0.034144	-0.005082
1	2,075,725	A	C	false	null	-0.000143	-0.089451	-0.000107	-0.056293	0.001778
1	2,076,945	T	C	false	null	0.000076	-0.082065	0.000032	-0.026904	0.002886
1	2,079,172	G	A	false	null	-0.000089	-0.03819	-0.000034	0.001027	-0.001271
1	2,079,230	A	G	false	null	0.000127	0.045023	0.000284	0.03416	0.002241
1	2,094,081	G	A	false	null	-0.000055	0.006946	-0.000045	0.023981	-0.001834
1	2,096,534	C	T	false	null	-0.000553	0.136196	-0.000027	0.002688	-0.003655
1	2,112,343	C	T	false	null	-0.000074	0.00562	-0.00014	-0.001145	-0.145587
1	2,121,579	T	C	false	null	0.000146	-0.056528	0.00049	-0.047761	0.002337

End of preview. Expand in Data Studio

evals_dsqtl

Supervised variant-effect-prediction benchmark of dsQTLs (DNase-seq chromatin accessibility): statistically significant chromatin-accessibility QTLs vs control variants. These are the standardized ChromBPNet benchmark variant sets (Synapse syn64126763) used by both the ChromBPNet and AlphaGenome papers, brought into the marin-dna evals pipeline with chromosome-parity train/test splits. Each row carries precomputed ChromBPNet/Enformer baseline scores, so those baselines reproduce without running any model.

No matching and no subsampling — every positive and negative is kept at its natural ratio (≈1:48 positive:negative).

Description


Positives	Significant dsQTLs (`label = True`; −log10p > 5)
Negatives	Control variants (`label = False`; −log10p < 3)
Assay	DNase-seq chromatin accessibility (GM12878 / LCL)
Source study	Yoruba dsQTLs (Degner et al. Nature 2012)
Source data	Standardized ChromBPNet benchmark, Synapse `syn64126779`
Genome build	GRCh38 (lifted from hg19)
Variant type	SNVs only
Coordinates	1-based (`pos` is 1-based; `ref`/`alt` are single bases)
Matching	none (natural class ratio)

Splits

Chromosome-parity split (same convention as the other evals_* datasets). Chromosome subsets used by the official benchmark — all chroms, the AlphaGenome test chroms {3, 6, 9, 12, 16, 18, 19, 21}, or this train/test parity — are a scoring-time slice of the full variant set, not a re-score.

Split	Variants	Positives	Negatives	Chromosomes
`train`	15,018	309	14,709	odd: 1, 3, …, X
`test`	12,328	250	12,078	even: 2, 4, …, Y
total	27,346	559	26,787

Columns

Column	Type	Description
`chrom`, `pos`, `ref`, `alt`	str / int / str / str	Variant coordinates (1-based, GRCh38). `ref`/`alt` are genome-oriented (`alt` ≠ reference).
`label`	bool	`True` for a significant dsQTL, `False` for a control variant
`effect`	float	Signed study effect (`obs.estimate`), oriented to the `alt` allele — positive ⇒ `alt` increases accessibility; sign-flipped for variants whose ref/alt were swapped to match the reference. Present for every positive; controls may be null.
`chrombpnet_atac_ips`	float	ChromBPNet GM12878 ATAC IPS (= logFC × JSD × AAQ) — the recommended causality score
`chrombpnet_atac_logfc`	float	ChromBPNet GM12878 ATAC log2 fold-change — the recommended direction score
`chrombpnet_dnase_ips`	float	ChromBPNet GM12878 DNase IPS
`chrombpnet_dnase_logfc`	float	ChromBPNet GM12878 DNase log2 fold-change
`enformer_dnase_local_logfc`	float	Enformer GM12878 DNase recomputed local-2 kb log2 fold-change

The carried chrombpnet_* / enformer_* columns are the papers' precomputed per-variant baseline scores (signed, oriented to alt) — supplied so the ChromBPNet and Enformer baselines need no model run.

Evaluation

Two official metrics (the ChromBPNet/AlphaGenome accessibility-QTL protocol; see issue #262), each over a different variant set:

Causality — AUPRC over all variants on |score|: significant dsQTLs (label = True) vs controls (label = False).
Direction — Pearson over the positive variants only (label = True): between effect (signed, oriented to alt) and the model's signed alt-vs-ref score. Controls are not used in the correlation.

ChromBPNet uses IPS for causality and logFC for direction; AlphaGenome / Enformer use signed accessibility log-fold-change. Metrics are computed by marin_dna.pipelines.chrombpnet_eval.metrics.compute_supervised_qtl_metrics.

Provenance

Built by the marin-dna eval pipeline at commit 17c0377.

Curation pipeline: snakemake/evals
Rules: snakemake/evals/workflow/rules/chrombpnet_qtl.smk
Build + parsing: src/marin_dna/pipelines/evals/chrombpnet_qtl.py

License

Released under the terms of its upstream sources. The variant set is redistributed from the standardized ChromBPNet benchmark (Synapse syn64126763) and derives from the original QTL study (Yoruba dsQTLs (Degner et al. Nature 2012)); consult those sources for redistribution and commercial-use terms.

Citation

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

ChromBPNet — Pampari et al. 2024, bioRxiv 2024.12.25.630221
AlphaGenome — Avsec et al. 2025
dsQTL study — Yoruba dsQTLs (Degner et al. Nature 2012)

Downloads last month: 23