Clinical Variants · HuggingFaceBio
clinvar-vep
A ClinVar variant effect prediction benchmark for evaluating genomic language models on variant pathogenicity.
What clinvar-vep is
clinvar-vep is a benchmark dataset designed for evaluating DNA language models on clinical variant pathogenicity classification. Curated by HuggingFaceBio, it provides standardised binary classification tasks distinguishing pathogenic and likely pathogenic variants from benign and likely benign variants across human genomic sequences.
The dataset addresses evaluation imbalances commonly found in existing genomic benchmarks. While many clinical variant evaluations are predominantly composed of protein-coding alterations, this dataset structures evaluation into two distinct partitions: an exonic coding subset and an explicitly balanced non-coding subset covering intronic and untranslated regions.
This separation allows researchers in biology, genomics, and medicine to independently evaluate model performance on protein-coding mutations as well as non-coding regulatory sequences.
How it was made
The benchmark is split into two configurations: coding and non-coding. The coding subset is derived from the ClinVar VEP collection in GenerTeam/variant-effect-prediction (originally used in the GPN-MSA study), containing 39,473 exonic protein-coding variants.
The non-coding subset is constructed directly from ClinVar VCF files. It is restricted to single-nucleotide variants on chromosomes 1–22, X, and Y with verified binary clinical labels. Reviewed variants are annotated using ClinVar consequence terms to categorise them into broad regions (coding, non_coding, splice, unknown) and subtypes (such as intronic, utr_5_prime, and utr_3_prime). Only entries with coding_status set to non_coding are retained, yielding a strictly balanced set of 15,258 variants comprising 7,629 benign and 7,629 pathogenic examples across intronic and UTR regions.
Key facts
- Task
- Binary text and tabular classification
- Coding subset breakdown
- 17,231 benign, 22,242 pathogenic
- Non-coding subset breakdown
- 7,629 benign, 7,629 pathogenic
- Non-coding region types
- 10,310 intronic, 4,174 5′ UTR, 774 3′ UTR
- Genomic scope
- Single-nucleotide variants on chromosomes 1–22, X, and Y
- Data format
- Parquet
What you can use it for
- Benchmarking genomic language models
- Assessing the ability of DNA foundation models to classify human single-nucleotide variants as pathogenic or benign.
- Non-coding variant pathogenicity evaluation
- Testing predictive models specifically on intronic and UTR variants using a balanced set of pathogenic and benign cases.
- Coding variant effect prediction
- Measuring model accuracy on exonic protein-coding alterations derived from ClinVar.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from HuggingFaceBio/clinvar-vep at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Provides a fully balanced non-coding split with equal numbers of benign and pathogenic labels.
- Separates evaluation cleanly into exonic coding regions and non-coding intronic and UTR sequences.
- Restricts non-coding entries to reviewed single-nucleotide variants with binary clinical annotations across standard human chromosomes.
Watch out for
- The coding subset is slightly imbalanced, containing 17,231 benign and 22,242 pathogenic variants.
- The non-coding subset is restricted to single-nucleotide variants, excluding complex multi-nucleotide or structural variants.
- Excludes variants with ambiguous consequence annotations or those falling under splice and unknown classifications in the clean non-coding set.
Questions
Can I look inside clinvar-vep without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is clinvar-vep under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
How are clinical labels defined in the dataset?
Each split is a binary classification task where label 1 corresponds to pathogenic or likely pathogenic variants, and label 0 corresponds to benign or likely benign variants.
What genomic regions are represented in the non-coding subset?
The non-coding subset covers 10,310 intronic variants and 4,948 untranslated region variants (4,174 in the 5′ UTR and 774 in the 3′ UTR).
Why was a separate non-coding subset constructed?
Existing VEP subsets derived from prior studies were found to be predominantly coding (39,473 coding versus 1,503 non-coding) and suffered from extreme class imbalance, having only 10 pathogenic non-coding variants.
How to cite it
@article{allal2026carbon, title={Carbon: Decoding the Language of Life}, author={Allal, Loubna Ben and Li, Qiuyi and Fiusco, Maurizio and Tunstall, Lewis and Rasul, Kashif and Beeching, Ed and Aubakirova, Dana and Pati{\~n}o, Carlos and Frere, Thibaud and Lozhkov, Anton and others}, journal={bioRxiv}, pages={2026--05}, year={2026}, publisher={Cold Spring Harbor Laboratory} }
Written from clinvar-vep's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.