evals_eqtl

Variant-effect-prediction benchmark of GTEx v8 fine-mapped eQTLs (49 tissues, pooled) vs tested-but-low-PIP variants, gene-matched 1:1 within consequence categories, MAF bins, and distance bins.

Description

Positives	GTEx v8 SuSiE fine-mapped variants with max(PIP across tested tissues) > 0.9
Negatives	max(PIP) < 0.01 across all tested tissues; gene-matched 1:1 to positives
Source	eQTL Catalogue r7, study QTS000015 (GTEx v8), all 49 tissues × ge (gene-expression) quantification
Genome build	GRCh38 (Catalogue is hg38 native)
Variant type	SNVs only
Coordinates	1-based (pos is 1-based; ref / alt are single bases)

The negative pool is sourced from per-tissue full nominal sumstats (*.all.tsv.gz, ~3.5 GB per tissue), so it includes every variant tested in fine-mapping — including those that never reached a credible set. This is a structural improvement over earlier releases of this dataset (which sourced from a pre-filtered single-file release whose PIP floor at ~1e-4 capped the negative pool at ~1.3M variants); the current pool is ~10M tested variants.

Splits

Split	Variants (positives + matched negatives)	Chromosomes
train	4,612 (2,306 pos + 2,306 neg)	odd: 1, 3, …, X
test	3,854 (1,927 pos + 1,927 neg)	even: 2, 4, …, Y
total	8,466 (4,233 pairs)

Columns

Column	Type	Description
chrom, pos, ref, alt	str / int / str / str	Variant coordinates (1-based, GRCh38)
label	bool	True for high-PIP eQTL, False for tested-but-low-PIP matched negative
subset	str	Consequence-group label for stratified eval (distal, tss_proximal, non_coding_transcript_exon_variant, 3_prime_UTR_variant, 5_prime_UTR_variant, missense_variant, synonymous_variant, splicing)
match_group	int	Pos / neg pair ID (every group has 1 positive + 1 matched negative)
pip	float	Maximum PIP across the 49 tissues in which the variant was tested. For negatives this is < 0.01 (often 0 — see Source)
tissues	str	Comma-separated list of tissues with PIP > 0.9 (empty for negatives)
genes	str	Comma-separated list of eGene Ensembl IDs the variant regulates with PIP > 0.9 (empty for negatives)
biotype_classes	str	Comma-separated pc / nc classes of the eGenes (e.g. pc, nc, nc,pc). Empty for negatives. Genes missing from Ensembl's biotype table default to nc.
MAF	float	Cohort-matched GTEx donor MAF (mean across the 49 per-tissue donor cohorts; the cohorts overlap heavily so per-tissue MAFs are very close)
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 (0-based, half-open) to nearest protein-coding / non-protein-coding TSS, plus their min
tss_closest_pc_gene_id, tss_closest_nc_gene_id, tss_closest_gene_id	str	Ensembl gene IDs at those distances
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_tss_nc_bin, distance_exon_pc_bin, MAF_bin	str	Categorical bins used as exact-match keys during gene-matching

Per-subset retention

Subset	n_pos in dataset_all	matched (kept)	retention
distal	5,333	3,067	57.5%
tss_proximal	2,160	437	20.2%
non_coding_transcript_exon_variant	1,290	318	24.7%
5_prime_UTR_variant	544	86	15.8%
3_prime_UTR_variant	535	223	41.7%
splicing	333	9	2.7%
missense_variant	269	46	17.1%
synonymous_variant	230	47	20.4%
total	10,694	4,233	39.6%

(Two tiny subsets — mature_miRNA_variant n=1, stop_retained_variant n=1 — fail to match and are dropped.)

Splicing has the worst retention (~3%) — splicing variants live next to a specific small set of exons, so the gene-id-based categorical match starves neg supply. The phenomenon is structural to the matching framework, not specific to eqtl.

Matching design

Locked iter 33 from issue #156, with one round-2 tweak for eqtl (commit 1ad4fef7f7) that closes a leak that only became detectable after switching to the richer Catalogue negative pool.

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=1.

• Continuous features: distance_tss_pc, distance_tss_nc, distance_exon_pc, distance_exon_nc, MAF.
• Categorical features:
  • chrom, consequence_final
  • tss_closest_pc_gene_id, tss_closest_nc_gene_id, exon_closest_pc_gene_id, exon_closest_nc_gene_id
  • distance_tss_pc_bin (tss_proximal only; edges [0, 50, 100, 200, 500, 1000])
  • distance_tss_nc_bin — two-subset bin:
    • tss_proximal: edges [0, 50, 100, 200, 500, 1000] (iter 33)
    • non_coding_transcript_exon_variant: edges [0, 200, 1000, 5000] (round-2 widen-and-extend; ncRNA-exon variants span a much broader dist-to-nc-TSS range than tss_proximal, q=0.05 ≈ 9 bp, q=0.95 ≈ 8.7 kb)
  • distance_exon_pc_bin (splicing only; edges [0, 5, 20, 30])
  • MAF_bin — per-subset tiered scheme with local-quantile bins for distal (MAF_TIERED_LOG8_DISTAL_ONLY):
    • distal: local equal-width log10(MAF) bins computed per categorical match group, 8 buckets, joint pos+neg reference. Closes the asymptotic distal MAF residual leak that fixed global edges leave.
    • tss_proximal, non_coding_transcript_exon_variant: 20-bin (MAF_BIN_EDGES_20)
    • 3_prime_UTR_variant, 5_prime_UTR_variant, missense_variant: 10-bin (MAF_BIN_EDGES_10)
    • synonymous_variant, splicing, …: 5-bin (MAF_BIN_EDGES_5)

tissues / genes / biotype_classes are passthrough columns, not in the match key.

Matched-feature diagnostic (round 2-1): one Bonferroni-significant leak closed vs. the round-2 kickoff (ncRNA × distance_tss_nc, PA 0.401 → 0.483). Two small-effect residual leaks in distal (PA=0.470 on distance_tss_pc, PA=0.536 on MAF) left as below the "don't bother" effect-size threshold.

Full per-(subset, feature) PA / p-value table (round 2-1)

Format: PA / p. ** = Bonferroni-significant at α = 0.05 / 40 = 1.25e-3 (8 subsets × 5 matched features). * = p < 0.05. PA close to 0.5 = well-matched.

subset	n	distance_tss_pc	distance_tss_nc	distance_exon_pc	distance_exon_nc	MAF
3_prime_UTR_variant	223	0.502 / 1.0	0.457 / 0.23	0.487 / 0.24	0.439 / 0.03 *	0.547 / 0.18
5_prime_UTR_variant	86	0.360 / 0.013 *	0.395 / 0.066	0.494 / 1.0	0.442 / 0.25	0.535 / 0.59
distal	3,067	0.470 / 7.8e-4 **	0.494 / 0.49	0.476 / 0.009 *	0.494 / 0.49	0.536 / 8.2e-5 **
missense_variant	46	0.413 / 0.30	0.435 / 0.46	0.511 / 1.0	0.489 / 1.0	0.630 / 0.10
non_coding_transcript_exon_variant	318	0.544 / 0.13	0.483 / 0.57	0.487 / 0.69	0.502 / 1.0	0.553 / 0.064
splicing	9	0.556 / 1.0	0.444 / 1.0	0.667 / 0.51	0.611 / 0.73	0.333 / 0.51
synonymous_variant	47	0.404 / 0.24	0.340 / 0.040 *	n/a	0.489 / 1.0	0.574 / 0.38
tss_proximal	437	0.455 / 0.069	0.460 / 0.10	0.479 / 0.42	0.481 / 0.44	0.538 / 0.13

Source

Per-tissue files from the eQTL Catalogue r7 stable release FTP at ftp://ftp.ebi.ac.uk/pub/databases/spot/eQTL/, study QTS000015 (GTEx v8), all 49 GTEx tissue datasets × ge quantification. Two files per dataset:

• *.credible_sets.tsv.gz — SuSiE credible-set members + PIP.
• *.all.tsv.gz — full nominal sumstats (every tested variant × gene). Provides the "tested but never reached a credible set" negative pool.

Per-variant pip is max across the (variant, tissue, gene) rows that mention it. Variants present in all.tsv.gz but absent from every credible set are 0-filled (sentinel: "tested, no signal"). Per-variant MAF is the mean across the rows.

Provenance

Built by the bolinas-dna eval pipeline at commit 1ad4fef7f7. PR thread: #167.

• Curation pipeline: snakemake/evals/
• Catalogue parser: src/bolinas/evals/catalogue_parser.py
• Labeling cascade: src/bolinas/evals/labeling.py
• Matching algorithm: src/bolinas/evals/matching.py
• Design discussion: issue #156 (iter 33 + round 2 ncRNA bin); eqtl-specific motivation: issue #165