GenoAdv

Adversarial + clean genomic sequence classification benchmark, built on top of the GUE benchmark. Each subset merges the original clean split with adversarial examples produced by attackers targeting DNA foundation models (DNABERT-2, DNABERT-1, NT-1/2, HyenaDNA).

Companion to the GenoArmory paper and the GenoArmory codebase.

Usage

from datasets import load_dataset

# Pick any task below as the config name
ds = load_dataset("robinzixuan/GenoAdv", "prom_300_all")
print(ds)
# DatasetDict({train, validation, test})

Tasks

Epigenetic Marks Prediction (Histone)

Config	Train	Validation	Test	Labels	Seq length
H3	19,232	2,405	1,497	0/1	12–1060
H3K14ac	36,727	4,591	3,305	0/1	18–1129
H3K36me3	41,404	5,175	3,488	0/1	54–1127
H3K4me1	36,789	4,599	3,168	0/1	36–1569
H3K4me2	34,906	4,364	3,069	0/1	12–1444
H3K4me3	39,881	4,985	3,680	0/1	18–1445
H3K79me3	32,660	4,082	2,884	0/1	6–1025
H3K9ac	31,347	3,919	2,779	0/1	290–1080
H4	18,069	2,260	1,461	0/1	6–1231
H4ac	37,715	4,715	3,410	0/1	36–1110

Promoter Detection

Config	Train	Validation	Test	Labels	Seq length
prom_300_all	67,336	8,417	5,920	0/1	6–1068
prom_300_notata	56,770	7,096	5,307	0/1	42–1448
prom_300_tata	9,242	1,155	613	0/1	6–943
prom_core_all	69,226	8,653	5,920	0/1	24–211
prom_core_notata	61,137	7,642	5,307	0/1	42–224
prom_core_tata	8,808	1,100	613	0/1	31–251

Transcription Factor Prediction (Human)

Config	Train	Validation	Test	Labels	Seq length
tf0	44,095	1,361	1,000	0/1	13–185
tf1	41,335	1,347	1,000	0/1	11–465
tf2	26,421	1,390	1,000	0/1	24–180
tf3	35,886	1,314	1,000	0/1	21–233
tf4	28,104	1,479	1,000	0/1	12–542

Splice Site Prediction / Mouse TF

Config	Train	Validation	Test	Labels	Seq length
0	15,701	1,963	810	0/1	18–165
1	75,471	9,435	6,745	0/1	46–412
2	6,277	785	328	0/1	32–168
3	5,308	666	239	0/1	86–195
4	22,940	2,867	1,883	0/1	84–536

COVID Variant Classification

Config	Train	Validation	Test	Labels	Seq length
covid	83,609	10,450	9,168	0/1/2/3/4/5/6/7/8	946–1039

Reconstructed

Config	Train	Validation	Test	Labels	Seq length
reconstructed	39,714	4,964	4,562	0/1/2	354–435

Data format

Each CSV has two columns:

• sequence: nucleotide string (A/C/G/T, sometimes N)
• label: integer class id

Citation

@article{luo2025genoarmory,
  title   = {GenoArmory: A Unified Adversarial Framework for Genomic Foundation Models},
  author  = {Luo, Haozheng and others},
  journal = {arXiv preprint arXiv:2505.10983},
  year    = {2025}
}

License

MIT