Published April 23, 2026 | Version v1
Dataset Open

EVEE: Interpretable variant effect prediction from genomic foundation model embeddings

Authors/Creators

  • 1. Goodfire
  • 2. Mayo Clinic

Description

This dataset contains the precomputed variant effect predictions and interpretability features that power the Evo Variant Effect Explorer (EVEE) web application, accompanying the preprint "EVEE: Interpretable variant effect prediction from genomic foundation model embeddings" (Pearce et al., 2026, doi:10.64898/2026.04.10.717844). Each row is one ClinVar variant (4,252,870 total) and carries its genomic coordinates, gene and consequence annotations, ClinVar clinical significance, an Evo 2 embedding-based pathogenicity score, and roughly 4,900 additional probe outputs covering protein-level disruption features (InterPro domains, post-translational modifications, secondary structure, active/binding sites, disorder, etc.), regulatory-track predictions (ChromHMM states, ATAC-seq and ChIP-seq peaks across multiple cell types, CCRE annotations), amino-acid and consequence classifiers, and per-variant reference-predictor scores (AlphaMissense, REVEL, CADD, PrimateAI, SpliceAI, and others). The table is released as five chromosome-balanced Parquet shards (clean_shard_0.parquet through clean_shard_4.parquet, each 6.8–7.3 GB) plus a manifest.json describing which chromosomes live in each shard. Consumers can read all shards as a single logical table with polars.scan_parquet("clean_shard_*.parquet") or duckdb.read_parquet. This is the exact artifact used to build the EVEE variants.duckdb served at https://evee.goodfire.ai.

Notes

Per-variant flat table (builds/v5/clean.parquet) used to populate the EVEE web app (https://evee.goodfire.ai). One row per ClinVar variant; columns include variant_id, gene_name, consequence, ClinVar significance and label, an Evo 2 probe pathogenicity score, and ~4,900 additional probe heads (disruption, effect, and annotation categories). The variants.duckdb artifact served by the website is derived from this parquet via scripts in https://github.com/goodfire-ai/variant-viewer.

Files

manifest.json

Files (35.0 GB)

Name Size Download all
md5:6b25a426954b7590c04e93b42bd0624f
6.8 GB Download
md5:5f9fe5fb06689b0b7b6f412d86975e05
6.8 GB Download
md5:ffbf39a9d047897932bbd4c7d43c05a2
7.2 GB Download
md5:4714443ef1ac68edf7cc92fb49587df7
7.0 GB Download
md5:fa19cce9b5d64235eedce017efb62074
7.3 GB Download
md5:9eae4cd35bb87ce89b7ebdef8735ebe2
1.4 kB Preview Download

Additional details

Related works

Is supplement to
Preprint: 10.64898/2026.04.10.717844 (DOI)

Dates

Available
2026-04-22