pabloarozarenad/vGUE-benchmark

vGUE: viral Genome Understanding Evaluation Benchmark

The viral Genome Understanding Evaluation (vGUE) benchmark is a dedicated evaluation suite probing viral genomic foundation models across five key biological dimensions:

1. Broad organism-level distinctions: Virus vs Non-Virus (5kb bacterial contigs & Illumina metagenomic short-reads).
2. Genome-wide signatures & evolutionary fingerprints: DNA vs RNA genome classification & 4-class broad host range prediction.
3. Intra-genus species distinction: HIV-1 vs HIV-2 species separation.
4. Variant/subtype typing: SARS-CoV-2 lineage subtyping across 7 clades.
5. Phenotypic context signal identification: HIV-1 tissue tropism (Brain vs Non-Brain).

Tasks & Subsets

Subset Config	Purpose	Data
virus_bacteria	Virus vs Non-Virus (5kb contigs)	Viral complete genomes & bacterial RefSeq contigs
metagenomic_reads	Short-read viral identification	150bp Illumina metagenomic reads
dna_rna	DNA vs RNA virus classification	5,000 DNA & 5,000 RNA complete genomes
hiv1_hiv2	Fine-grained intra-genus separation	7,248 HIV-1 & 94 HIV-2 complete genomes
sarscov2_subtyping	Fine-grained variant typing	1,000 genomes across 7 lineages
host_prediction	Broad host range classification	2,000 genomes (Vertebrates, Invertebrates, Plants, Bacteria)
hiv1_tropism	Phenotypic tissue adaptation	HIV-1 envelope CDS (Brain vs Plasma)

Schema

Every subset has two columns:

Column	Type	Description
sequence	string	Nucleotide sequence (genome, contig, or read)
label	string	Class label for the task (see subset description above)

Usage

from datasets import load_dataset

# Load a specific task configuration from vGUE
dna_rna_dataset = load_dataset("pabloarozarenad/vGUE-benchmark", "dna_rna", split="test")
print(dna_rna_dataset[0])