@genebe/popeve

popEVE proteome-wide missense pathogenicity score

popEVE ranks missense variants across the whole human proteome on a single, calibrated pathogenicity scale. It combines two unsupervised evolutionary models — EVE (a variational autoencoder trained on multi-sequence alignments) and ESM-1v (a protein language model) — through a population-constraint layer fit on UK Biobank and gnomAD allele frequencies, so scores are comparable between genes rather than only within one gene.

Only the popEVE score is retained from the source file. Lower values indicate a more severe/deleterious prediction. The source file lists one row per overlapping transcript/protein for a given genomic position, so the same (chromosome, position, ref, alt) can appear with several different popEVE values; this database keeps the minimum (most severe) value for each genomic SNV.

Source

grch38_popEVE_ukbb_20250715.vcf.gz, GRCh38, UK Biobank-calibrated release dated 2025-07-15.

Citation

Orenbuch R, Shearer CA, Kollasch AW, et al. Proteome-wide model for human disease genetics. Nature Genetics. 2025;57(12):3165-3174. doi:10.1038/s41588-025-02400-1.

Website: https://pop.evemodel.org — Code: https://github.com/debbiemarkslab/popEVE