Datasets:
license: apache-2.0
task_categories:
- text-classification
language:
- en
tags:
- dna
- genomics
- variant-effect-prediction
- clinvar
size_categories:
- 10K<n<100K
configs:
- config_name: coding
data_files:
- split: test
path: coding/*.parquet
- config_name: non_coding
data_files:
- split: test
path: non_coding/*.parquet
ClinVar VEP — coding + non-coding benchmark
A ClinVar variant-effect-prediction (VEP) benchmark with two complementary subsets — coding (39,473 variants) and non-coding (15,258 variants) — for evaluating DNA language models on clinical variant pathogenicity.
Each split is a balanced binary classification task: label = 1 for pathogenic / likely pathogenic and label = 0 for benign / likely benign.
Subsets
| Config | # variants | benign | pathogenic | Region scope |
|---|---|---|---|---|
coding |
39,473 | 17,231 | 22,242 | exonic protein-coding |
non_coding |
15,258 | 7,629 | 7,629 | intronic + 5′/3′ UTR |
from datasets import load_dataset
coding = load_dataset("hf-carbon/clinvar-vep-final", "coding", split="test")
non_coding = load_dataset("hf-carbon/clinvar-vep-final", "non_coding", split="test")
Coding subset — origin
For the coding ClinVar benchmark, we use the ClinVar VEP subset from GenerTeam/variant-effect-prediction, which was originally used in the GPN-MSA study (Benegas et al., 2025). After annotating this dataset with ClinVar molecular consequence / region information, we found that it is overwhelmingly coding: among matched variants, 39,473 are coding and only 1,503 are non-coding, with only 10 pathogenic non-coding variants in the entire non-coding split — far too few and far too imbalanced for meaningful AUROC. We therefore keep this dataset as our coding ClinVar benchmark.
The coding parquet preserves all upstream baseline scores (CADD, phyloP-100v, phyloP-241m, phastCons-100v, GPN-MSA, NT, NT-v2, HyenaDNA, Evo2-1B-Base, Evo2-7B-Base, Evo2-7B, GENERator-1B/3B, GENERator-v2-1B/3B) along with chrom, pos, ref, alt, label, type, refseq_id.
Non-coding subset — construction
To construct a separate non-coding ClinVar benchmark, we start from the original ClinVar VCF and restrict to single-nucleotide variants on chromosomes 1–22, X, and Y with binary clinical labels, mapping benign / likely benign to label = 0 and pathogenic / likely pathogenic to label = 1. We retain reviewed variants, annotate each variant using ClinVar consequence terms into broad region classes (coding, non_coding, splice, unknown) and finer subtypes (intronic, utr_5_prime, utr_3_prime, etc.), and then select the clean non-coding subset with coding_status == "non_coding".
This yields a balanced non-coding benchmark of 15,258 variants: 7,629 benign and 7,629 pathogenic, covering intronic (10,310) and UTR variants (4,948 total — 4,174 5′UTR + 774 3′UTR). Splice (250) and unknown-region variants (12,200) are kept separate from the main non-coding subset because they represent distinct or ambiguous categories, and live in the source repository hf-carbon/ClinVar-VEP.
Citation
If you use this benchmark, please cite GPN-MSA (Benegas et al., 2025) and the ClinVar database.