VEP-eval: A Variant-Type-Stratified Benchmark for Pathogenicity Prediction

This dataset accompanies the paper "Genomic heterogeneity inflates the performance of variant pathogenicity predictions" (Lu, Liu, Lin & Brandes, bioRxiv, 2025). [Paper] [Code]

Existing evaluations of variant-effect-prediction (VEP) models often pool highly heterogeneous variant types (e.g., all noncoding variants) into a single group. Because pathogenicity rates differ drastically across these types (e.g., canonical splice sites are almost always pathogenic, while 5′ UTR variants rarely are), models can achieve deceptively high performance simply by learning coarse variant-type priors rather than genuinely separating pathogenic from benign variants within a type. This dataset provides a genome-wide, variant-type-stratified benchmark that makes this distinction explicit, together with pre-computed scores from nine DNA-sequence models and six protein-sequence models, so that per-variant-type performance can be recomputed and audited directly.

Dataset summary

Two single-nucleotide-variant (SNV) benchmark tables are provided:

File	Source	Variants	Columns	Description
clinvar_benchmark.csv	ClinVar	242,132 (69,515 pathogenic / 172,617 benign)	120	Primary benchmark: all ClinVar SNVs with a definitive (Pathogenic/Benign) germline classification, annotated by genomic context and scored by every evaluated model.
cosmic_benchmark.csv	COSMIC Cancer Mutation Census (CMC)	55,538 (2,113 Tier-1 pathogenic / 53,425 other)	60	Secondary, independent replication benchmark built from somatic cancer mutations, used to confirm that model rankings generalize beyond ClinVar.

Both files are variant-level tables (one row per SNV) with genomic coordinates, ClinVar/COSMIC labels, genomic-context annotations (which region/effect the variant falls in, per transcript), and one column per evaluated model's raw score for that variant. This is exactly the data used to produce the AUROC results and figures in the paper — it lets anyone reproduce, re-slice, or extend the variant-type-specific evaluation without re-running any model.

Dataset structure

clinvar_benchmark.csv

Identifiers & labels

• #CHROM, POS, ID, REF, ALT — genomic coordinates (GRCh38) and alleles. ID is the internal ClinVar variant ID.
• ClinVar_label — 1 = Pathogenic/Likely pathogenic, 0 = Benign/Likely benign (definitive labels only).
• ClinVar_gold_stars — ClinVar review-status star rating (1–4), used for the star-rating robustness analysis (≥2, ≥3 stars) in the paper.
• ClinVarName_* — amino-acid position/change and effect type parsed from the ClinVar variant name (e.g., ClinVarName_AAPOS, ClinVarName_AAREF, ClinVarName_AAALT, ClinVarName_missense, ClinVarName_synonymous, ClinVarName_stop_gain, ClinVarName_splice, ClinVarName_RNA_gene).

Genomic-context annotation (per MANE-Select transcript)

• MANE_transcript, MANE_mRNA*, MANE_coding_sequence, MANE_start_codon, MANE_stop_codon, MANE_five_prime_UTR, MANE_three_prime_UTR, MANE_mRNA_intron, MANE_mRNA_splice, MANE_snRNA*, MANE_snoRNA*, MANE_other — binary flags indicating which genomic feature the variant overlaps on the MANE-Select transcript.
• Strand, group: +, group: - — transcript strand.
• Union_stop_loss, Union_splice, Union_RNA_gene, MANE_intron — variant effect unioned across all overlapping transcripts (not just MANE-Select), used where effects can differ by transcript.
• group: * (36 columns, e.g. group: missense, group: noncoding, group: splice, group: 5'UTR + intron (non-splice), group: missense + 3'UTR) — the mutually-informative variant-type subgroup labels used for the stratified evaluation in Figure 1/3. Some variants belong to compound groups because they have different effects on different transcripts (e.g., missense in one transcript, 3′ UTR in another).

Model scores (one variant-level score per model; higher usually = more damaging, see paper Methods for score-direction conventions per model)

• DNA-sequence models: Evo2_7B, Evo2_40B, DNABERT2, PhyloP, GPN_MSA, PhyloGPN, gpnstar_v_llr (GPN-Star), AlphaGenome_quantile / AlphaGenome_raw, ntv3_pre_position_llr / ntv3_pre_seq_pllr / ntv3_post_position_llr / ntv3_post_seq_pllr / ntv3_post_log2fc_max (Nucleotide Transformer v3, pre- and post-trained variants).
• Protein-sequence models: ESM1b, ESM1v, ESM2, vesm_score (VESM++), AlphaMissense, PrimateAI_3D.
• Rule_based — the paper's rule-based baseline: each variant is scored by the empirical pathogenicity frequency of its variant-type group (used to show that variant-type heterogeneity alone yields strong apparent separation).
• variant_id — internal join key used when merging model outputs.

cosmic_benchmark.csv

Identifiers & labels

• #CHROM, POS, REF, ALT, gene (gene.1 duplicated for join convenience) — genomic coordinates and gene symbol.
• label — 1 = COSMIC Cancer Mutation Census (CMC) Tier-1 pathogenic mutation, 0 = other (background/non-Tier-1).
• label_source, cmc_tier — CMC tier annotation and stratified-sampling group the row was drawn from.
• subgroup — variant-effect subgroup used for evaluation (missense, stop_gain, synonymous, other).
• af_bin — gnomAD population allele-frequency bin (very_rare, rare, common, …); gnomad_af is the raw allele frequency.
• hgvsc, hgvsp, transcript, mut_type_raw — HGVS coding/protein notation, transcript ID, and raw Sequence Ontology mutation-type code.

Genomic-context annotation — the same style of binary region/feature flags as in the ClinVar table (mRNA, mRNA_promoter, mRNA_exon, coding_sequence, start_codon, stop_codon, five_prime_UTR, three_prime_UTR, mRNA_intron, mRNA_splice, and equivalent flags for lncRNA, snRNA, antisenseRNA, telomeraseRNA, RNaseMRPRNA, snoRNA, plus a catch-all other).

Model scores

• Evo2_40b_Var_minus_Ref, Evo2_7b_Var_minus_Ref, AlphaGenome_quantile_abs, AlphaGenome_raw_abs, PhyloGPN, NT, DB2 (DNABERT2), PhyloP, GPN_MSA_score, ESM1b_score, ESM1v-1_score, ESM2_score, am_score (AlphaMissense), pai_score (PrimateAI-3D), Rule_based.

Models evaluated

Model	Modality	Params	Input
Evo2 (7B / 40B)	DNA	7B / 40B	sequence
Nucleotide Transformer v3 (pre-/post-trained)	DNA	650M	sequence
AlphaGenome	DNA	—	sequence
GPN-MSA	DNA	86M	MSA
PhyloGPN	DNA	83M	sequence
GPN-Star	DNA	200M	MSA
DNABERT2	DNA	117M	sequence
PhyloP	DNA (baseline)	—	conservation score
ESM1b / ESM1v / ESM2	protein	650M	sequence
VESM++	protein	4.4B	sequence
AlphaMissense	protein	—	sequence, MSA
PrimateAI-3D	protein	—	3D structure, MSA

Full architecture/training details are in Table 1 of the paper.

Code

The full pipeline for extracting model predictions, computing AUROC per variant-type subgroup, and generating all figures is available at: 👉 https://github.com/Brandes-Lab/VEP-eval/

Citation

If you use this dataset, please cite:

@article{genomic2025biorxiv,
  author    = {Baiyu Lu and Xueshen Liu and Po-Yu Lin and Nadav Brandes},
  title     = {Genomic heterogeneity inflates the performance of variant pathogenicity predictions},
  journal   = {bioRxiv},
  year      = {2025},
  doi       = {10.1101/2025.09.05.674459},
  url       = {https://www.biorxiv.org/content/10.1101/2025.09.05.674459v2}
}

License

Released under the MIT License.