UKBBGym — Genebass setup

Pre-built tables for the genebass/ setup of UKBBGym, a population-scale benchmark for variant effect predictors. The genebass/ analysis notebooks fetch these files automatically on first run; see genebass/README.md for the method and how to rebuild them.

Task. Score a variant by the mean trait value across its heterozygous carriers (corrected for covariates and polygenic risk), then rank predictors by the per-gene–trait Spearman correlation between their scores and that quantity. Here, the carrier mean is approximated by the Genebass SAIGE single-variant BETA (mean_pheno_value), rank-equivalent to it for rare variants. Genebass single-variant summary statistics were downloaded from a publicly available Google Cloud Storage bucket gs://ukbb-exome-public/500k/results/variant_results.mt. Gene–trait pairs are selected by FDR-significant pLoF burden tests (one strongest trait per gene); all variant consequences in those genes are kept, filters applied downstream. No train/test split — evaluation is a correlation over held-out carrier phenotypes.

Please cite the preprint if you use this benchmark: Londhe et al., (2026). Phenotypes of ultra-rare variant carriers benchmark variant effect scores, bioRxiv

Files

File	Grain	Rows	Purpose
genebass_annotated.parquet	(id, region)	460,617	Master table. All variants in the 657 benchmark genes, joined to the BETA proxy, association stats, and ~150 annotations / predictor scores. Input to nearly every notebook.
genebass_annotated_all.parquet	(id, region, phenotype)	1,859,448	All-variants version, not collapsed to one trait per gene, wider set (658 genes, 99 traits). Used by expAssays_all_genes_correlations_pheno.ipynb.
genebass_unassociated.parquet	(id, region, phenotype)	1,029,471	Null-trait panel: benchmark-gene variants paired with traits the gene is not associated with. No annotation columns. Used by noise_ceiling/* for the noise floor.
other_benchmarks/clinvar_annotated.parquet	variant	1,859,698	ClinVar pathogenicity labels (17,683 genes), same annotation stack.
other_benchmarks/proteingym_snv_annotated.parquet	DMS measurement	122,127	Annotated ProteinGym substitution assays.
other_benchmarks/dms_mol_sys_bio_2023.parquet	DMS measurement	601,270	Extra deep-mutational-scanning assays (Mol Syst Biol 2023).
other_benchmarks/ldlr_science_2025.parquet	DMS measurement	41,772	LDLR saturation genome editing assay (Science 2025).

Key columns — genebass_annotated.parquet

Column	Description
id	chrom:pos:ref:alt on GRCh38. Join key.
chrom, pos, ref, alt	Parsed coordinates.
region, gene_name	Ensembl gene ID and symbol of the target gene.
phenotype, phenocode	Trait name (_int = inverse-normal transformed) and UKB phenocode.
mean_pheno_value	Target. Genebass SAIGE single-variant BETA (carrier-mean proxy).
mean_pheno_value_dircor	Direction-corrected target (× loftee_corr_dir).
SE, Pvalue, AF, n_cases	Genebass single-variant stats.
AC, AC_proxy	AF × 2 × 394,841; AF × 2 × n_cases (the minor-allele count the MAC ≤ 20 cut uses).
ac_ukb, mac_ukb, sc_ukb	UK Biobank allele / minor-allele / carrier counts.
pval_fdr	Gene–trait pLoF burden FDR (≤ 0.05 for inclusion).
loftee_corr, loftee_corr_dir	pLoF burden effect-direction correlation and its sign.
n_variants, appv_region	Variant count in the pair; gene the BETA was taken from.
is_cross_gene_beta	True where region != appv_region (5,516 rows).
consequence_*	One-hot VEP consequence flags.
clinical_significance	Raw ClinVar string.
am_pathogenicity, revel_score, esm1v, popeve, cadd_phred, delta_score, pangolin_score, plddt, priphylop, …	Predictor scores and annotations (~150 columns).
<name>_is_na	Missingness mask for the imputed column <name>.
uniprot_id, aa_ref, aa_alt, position	Protein-level join key.

Usage

import polars as pl

path = "hf://datasets/gagneurlab/ukbbgym/genebass_annotated.parquet"
df = pl.read_parquet(path)

# One gene–trait pair, ultra-rare missense; MAC<=20 is AF <= 20/(2*n_cases), i.e. AC_proxy<=20
sub = df.filter(
    (pl.col("AC_proxy") <= 20)
    & (pl.col("region") == "ENSG00000130164")
    & (pl.col("consequence_missense_variant") == 1)
)
rho = (
    sub.select("am_pathogenicity", "mean_pheno_value_dircor")
    .drop_nulls()
    .select(pl.corr("am_pathogenicity", "mean_pheno_value_dircor", method="spearman"))
    .item()
)

Or via the repo: git clone https://github.com/gagneurlab/ukbbgym && cd ukbbgym && uv sync && huggingface-cli login && genebass/run_all.sh (add --new-score my_scores.parquet to score your own predictor against every plot).