ClinVar VEP: A Benchmark for Clinical Variant Pathogenicity Prediction

A ClinVar variant-effect-prediction benchmark for evaluating DNA language models, containing two balanced binary classification subsets. The dataset includes 39,473 coding variants and 15,258 non-coding variants, labeled as pathogenic/likely pathogenic or benign/likely benign. It was created by HuggingFaceBio and was last updated on 2026-05-06.

Use Cases

• Benchmarking DNA language model performance on clinical variant pathogenicity prediction based on the described binary classification task.
• Training classifiers to distinguish pathogenic from benign genetic variants based on the labeled coding and non-coding subsets.
• Evaluating model generalizability across coding and non-coding genomic regions based on the two defined subsets.

Strengths

• Contains 54,731 total variants across two distinct subsets.
• Provides a balanced binary classification task with clear pathogenic and benign labels.
• Offers separate subsets for coding (39,473 variants) and non-coding (15,258 variants) regions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

HuggingFaceBio

Collection Method

Likely derived from the ClinVar database for variant classification.

Freshness

Last updated 2026-05-06 10:41:46; freshness should be verified.