evals_mendelian_traits

Variant-effect-prediction benchmark of pathogenic Mendelian SNVs vs gnomAD common SNVs, 1:9 matched within consequence categories on (chrom, consequence_final) plus subset-targeted distance bins.

Description

Positives	OMIM ∪ Smedley et al. 2016 ∪ HGMD (latter via Sei, Chen et al. Nat Genet 2022), deduplicated, gnomAD AF<0.001
Negatives	gnomAD common: AN≥25 000 and AF>0.001, 1:9 matched per positive
Genome build	GRCh38
Variant type	SNVs only
Coordinates	1-based (pos is 1-based; ref/alt are single bases)
Matching ratio	1:9 (positive : matched negatives)

Splits

Split	Variants (pos + 9·neg)	Positives	Chromosomes
train	16,140	1,614	odd: 1, 3, …, X
test	9,490	949	even: 2, 4, …, Y
total	25,630	2,563

Columns

Column	Type	Description
chrom, pos, ref, alt	str / int / str / str	Variant coordinates (1-based, GRCh38)
label	bool	True for pathogenic positive, False for matched gnomAD-common negative
subset	str	Consequence-group label for stratified eval
match_group	int	Integer ID grouping each positive with its 9 matched negatives
source	str	Pathogenic source (omim, smedley_et_al, hgmd); positives only
clinvar_id, trait	int / str	ClinVar IDs (OMIM-derived positives) and/or phenotype text
AF, AC, AN	float / int / int	gnomAD allele frequency, count, total
consequence, consequence_cre, consequence_final, consequence_group	str	Ensembl VEP consequence + grouping used by the matcher
distance_tss_pc, distance_tss_nc, distance_tss	int	Distance to nearest protein-coding / non-protein-coding transcript TSS (and the minimum, used by the consequence_group recategorization)
tss_closest_pc_gene_id, tss_closest_nc_gene_id, tss_closest_gene_id	str	Ensembl gene IDs at those distances (passthrough metadata — gene-id was not used in matching)
distance_exon_pc, distance_exon_nc, distance_exon	int	Same shape, for nearest exon
exon_closest_pc_gene_id, exon_closest_nc_gene_id, exon_closest_gene_id	str	Same shape
distance_tss_pc_bin, distance_exon_pc_bin	str	Subset-prefixed bin labels used as exact-match keys during matching (e.g. "tss_prox:b0", "splicing:b2"); BIN_NA outside the binned subsets

Per-subset retention

Subset	n_pos in dataset_all	matched (kept)	retention
missense_variant	15,481	1,000	6.5%
splicing	668	550	82.3%
5_prime_UTR_variant	306	305	99.7%
tss_proximal	271	270	99.6%
non_coding_transcript_exon_variant	159	159	100.0%
3_prime_UTR_variant	110	110	100.0%
distal	99	99	100.0%
synonymous_variant	65	65	100.0%
mature_miRNA_variant	7	5	71.4%
coding_sequence_variant	1	0	0.0%
total	17,167	2,563	14.9%

The bulk loss is the deliberate missense cap (12.7k → 1k for VEP inference speed; cap is mendelian_traits.max_positives_per_subset.missense_variant: 1000 in the pipeline config). Beyond that, drops come from (chrom, consequence_final × bin) strata with fewer than 9 negatives, which forces _match_single_group to subsample positives.

Matching design

Matching is exact on every categorical key, then Euclidean-nearest on the (RobustScaler-scaled) continuous features as a within-group tie-breaker. Without replacement, k=9.

• Continuous features: distance_tss_pc, distance_tss_nc, distance_exon_pc, distance_exon_nc.
• Categorical features:
  • chrom, consequence_final
  • distance_tss_pc_bin — tss_proximal: edges [0, 100, 1000, ∞]; BIN_NA elsewhere
  • distance_exon_pc_bin —
    • tss_proximal: edges [0, 100, 1000, ∞]
    • splicing: edges [0, 5, 30, ∞]
    • distal: edges [0, 100, 1000, 5000, 10000, ∞]
    • BIN_NA elsewhere

Gene-ID columns are kept as passthrough metadata but not used as match keys — exact gene matching at 1:9 dropped too many positives. The bin columns and gene-IDs are still useful for downstream stratification.

Matched-feature AUPRC diagnostic

Each continuous matching feature f is scored as a single-feature predictor within each subset: {f}_auprc = max(AP(label, +f), AP(label, −f)). Baseline = 0.1 for 1:9 matching. Values near baseline mean the feature does not separate positives from negatives within the subset (matching worked); values well above baseline are residual leak.

Per-(subset, feature) AUPRC table

subset	n	distance_tss_pc	distance_tss_nc	distance_exon_pc	distance_exon_nc
missense_variant	1,000	0.102	0.103	0.101	0.101
splicing	550	0.119	0.108	0.100	0.114
5_prime_UTR_variant	305	0.106	0.106	0.101	0.106
tss_proximal	270	0.111	0.116	0.111	0.113
non_coding_transcript_exon_variant	159	0.111	0.135	0.108	0.100
3_prime_UTR_variant	110	0.111	0.131	0.105	0.102
distal	99	0.110	0.109	0.132	0.113
synonymous_variant	65	0.115	0.112	0.100	0.105
mature_miRNA_variant	5	0.370	0.137	0.177	0.100

Provenance

Built by the bolinas-dna eval pipeline at commit main.

• Curation pipeline: snakemake/evals
• Matching algorithm: src/bolinas/pipelines/evals/matching.py
• Diagnostic helper: src/bolinas/pipelines/evals/matching_qc.py

The curation is a from-scratch reimplementation of the TraitGym pipeline.

Companion datasets

• bolinas-dna/evals_mendelian_traits_harness_255 — same variants with 255 bp reference-genome windows materialized for direct use as eval-harness inputs.

Citation

If you use this benchmark, please cite the upstream sources:

• TraitGym — Benegas et al. 2025, bioRxiv 2025.02.11.637758
• gnomAD — Karczewski et al. (Nature 2020)
• OMIM — omim.org
• Smedley et al. — AJHG 99(3): 595–606 (2016)
• Sei (HGMD redistribution path) — Chen et al. Nat Genet 2022
• HGMD — Stenson et al. (Hum Genet 2017)