biodatageeks/vepyr_116_GRCh38_plugin_spliceai

vepyr plugin cache — SpliceAI (GRCh38, VEP 116)

A prebuilt, frequency-tiered Parquet cache of SpliceAI splice-altering predictions for use with vepyr, the Rust/DataFusion VEP-compatible variant annotation engine. It reproduces the CSQ output of Ensembl VEP 116's --plugin SpliceAI without requiring the upstream 400 GB-class VCF or the Perl plugin at annotation time.

Source version

This is the fact you most likely came here for.

Source file	spliceai_scores.masked.snv.ensembl_mane.grch38.110.vcf.gz
Source URL	https://ftp.ensembl.org/pub/data_files/homo_sapiens/GRCh38/variation_plugins/
Upstream release	Ensembl-distributed SpliceAI scores, Ensembl release 110 — masked, SNV-only, MANE transcript set
Model version	SpliceAI v1.3.1
Source MD5	66d9183dcaadd7a35c186dee61983773 (computed on the downloaded copy, 28,512,375,517 bytes; Ensembl publishes no checksum for this directory)
Source retrieved	2026-07-06
Genome build	GRCh38, 1-based, bare (non-chr-prefixed) contigs in source
Cache built	2026-09-05
Target VEP version	Ensembl VEP 116 (VEP_plugins release/116 SpliceAI.pm)
Build manifest	plugins/spliceai/spliceai.source.toml
vepyr-plugins tag	v0.1.1 — recorded in manifest.json as cache_source_version: v0.1.1@3e1c039
Source verification	the source file and its .tbi were MD5-verified against the manifest before the build (verified_md5 in manifest.json)

Three limits are inherited from this particular upstream distribution and matter for interpretation:

• SNV-only. Indels are not scored. The Ensembl masked.snv release covers single nucleotide variants exclusively; an indel query returns no annotation, which is absence of data, not a negative prediction.
• Masked. In the masked release, scores for splice-site losses at positions that are already annotated splice sites, and gains elsewhere, are the retained signal — masked scores are set to 0 where the prediction reinforces the existing annotation. Use the raw release instead if you need unmasked deltas.
• MANE. Predictions are made against the MANE transcript set, so genes without a MANE transcript are absent.

Provenance

Rebuilt on 2026-09-05 from the complete 28.5 GB source file (not per-contig slices), verified against the v0.1.1 manifest, with chrX and chrY added. The sources block in manifest.json records url, declared and verified MD5, size and index digest for the input. Shard bytes are reproducible: the tier stage orders rows totally by the probe key (tier, start, allele_string, symbol), so a second build from the same verified source yields identical MD5s.

Contents

chr1.parquet … chr22.parquet, chrX.parquet, chrY.parquet    24 per-contig shards
manifest.json                   schema, CSQ field mapping, per-shard row/tier counts, source provenance

Covers chr1–chr22, chrX and chrY. The source has no mitochondrial records, so chrMT is listed in manifest.json with rows: 0 and has no shard file; an MT variant gets empty SpliceAI fields. Total ≈ 25 GB, 3,393,685,728 rows (8,048,804 warm / 3,385,636,924 cold). Largest shard chr1.parquet ≈ 2.2 GB.

Schema

column	type	CSQ field
chrom	string	— (contig)
start	uint32	— (1-based position)
end	uint32	—
allele_string	string	— (REF/ALT, exact)
symbol	string	— (per-transcript match discriminator: gene symbol)
symbol_out	string	SpliceAI_pred_SYMBOL
ds_ag	string	SpliceAI_pred_DS_AG — delta score, acceptor gain
ds_al	string	SpliceAI_pred_DS_AL — delta score, acceptor loss
ds_dg	string	SpliceAI_pred_DS_DG — delta score, donor gain
ds_dl	string	SpliceAI_pred_DS_DL — delta score, donor loss
dp_ag	int32	SpliceAI_pred_DP_AG — delta position (bp), acceptor gain
dp_al	int32	SpliceAI_pred_DP_AL — delta position (bp), acceptor loss
dp_dg	int32	SpliceAI_pred_DP_DG — delta position (bp), donor gain
dp_dl	int32	SpliceAI_pred_DP_DL — delta position (bp), donor loss
tier	int8	— (frequency tier: 0 = warm, 1 = cold)

symbol and symbol_out hold the same value under two names: one is the join key, one is an emitted field, and the builder's projection rejects two output columns sharing a source name.

Delta scores (DS_*) range 0–1 and give the maximum change in predicted splice-site probability within ±50 bp. Illumina's suggested cutoffs: <0.01 none, ≥0.2 high-recall, ≥0.5 recommended, ≥0.8 high-precision. Delta positions (DP_*) give the position of that maximum change relative to the variant (positive = downstream, negative = upstream).

Matching semantics

SpliceAI is a per-transcript annotation: VEP gates it by matching the current transcript consequence's gene SYMBOL against the row's own symbol (_join_transcript in SpliceAI.pm). The lookup key is therefore (chrom, start, end, allele_string) plus the symbol discriminator.

Alleles are stored exact (allele_match = "exact") — no minimisation, matching the plugin, and moot in practice for an SNV-only source.

The masked release genuinely repeats a bare (chrom, pos, ref, alt) key at overlapping-gene loci — one row per gene, same variant. Uniqueness holds only once symbol joins the key, which is why the full key is (start, allele_string, symbol).

Frequency tiering

Each shard is sorted by (tier, start, allele_string, symbol). A row's tier is inherited row-for-row from the release-116 GRCh38 variation cache this plugin cache was built against: a plugin row takes the tier of its matching variation row, and a plugin row with no match there is cold. tier = 0 is warm — 8,048,804 rows, 0.24% of the cache; tier = 1 is cold. Because warm rows are physically contiguous at the front of the file, a warm-only probe touches a handful of row groups instead of scanning a multi-hundred-million-row shard. Per-shard warm/cold counts are in manifest.json.

Quality profile

Generated 2026-09-05 by profile_plugin_cache.py (vepyr 0.4.0, Polars 1.39.3) from the shards in this commit; machine-readable copy in qa_profile.json.

Invariants

check	status	detail
schema	✅ pass	24 shards match the manifest
contig	✅ pass	0 foreign-contig rows in 24 shards
order	✅ pass	0 descending steps in 24 shards
tier_domain	✅ pass	0 rows with tier outside {0,1} in 24 shards
manifest_counts	✅ pass	rows/warm/cold match in 24 shards
manifest_files	✅ pass	25 manifest contigs, no stray shards
positions	✅ pass	0 rows with start < 1 or end < start - 1 in 24 shards
allele_form	✅ pass	0 malformed allele strings in 24 shards
duplicates	✅ pass	0 duplicate probe keys in 24 shards (manifest assume_unique=true)

Contigs

contig	rows	warm	cold	warm %	size
chr1	300,634,726	689,526	299,945,200	0.2%	2.2 GB
chr10	170,831,107	428,091	170,403,016	0.3%	1.3 GB
chr11	170,342,881	397,892	169,944,989	0.2%	1.3 GB
chr12	168,273,116	398,032	167,875,084	0.2%	1.3 GB
chr13	87,549,059	208,829	87,340,230	0.2%	649 MB
chr14	103,964,964	245,674	103,719,290	0.2%	774 MB
chr15	113,571,747	273,533	113,298,214	0.2%	846 MB
chr16	97,811,562	266,254	97,545,308	0.3%	732 MB
chr17	121,393,500	295,481	121,098,019	0.2%	909 MB
chr18	80,740,085	195,864	80,544,221	0.2%	604 MB
chr19	83,348,379	248,581	83,099,798	0.3%	639 MB
chr2	282,792,329	642,961	282,149,368	0.2%	2.1 GB
chr20	78,568,878	190,655	78,378,223	0.2%	593 MB
chr21	31,683,675	88,642	31,595,033	0.3%	239 MB
chr22	52,715,028	140,590	52,574,438	0.3%	400 MB
chr3	249,753,459	574,485	249,178,974	0.2%	1.9 GB
chr4	187,041,586	441,600	186,599,986	0.2%	1.4 GB
chr5	193,324,333	429,720	192,894,613	0.2%	1.4 GB
chr6	189,736,389	456,749	189,279,640	0.2%	1.4 GB
chr7	201,474,868	502,005	200,972,863	0.2%	1.5 GB
chr8	153,725,275	396,184	153,329,091	0.3%	1.1 GB
chr9	134,811,523	337,501	134,474,022	0.3%	1.0 GB
chrMT	0	0	0	0.0%	0 B
chrX	132,535,241	198,772	132,336,469	0.1%	990 MB
chrY	7,062,018	1,183	7,060,835	0.0%	53 MB
total	3,393,685,728	8,048,804	3,385,636,924	0.2%	25 GB

Columns

column	role	type	null %	empty %	distinct	numeric (min / p50 / p95 / max)	top values
symbol	match	String	0.00	0.00	~19K	—	—
symbol_out	value	String	0.00	0.00	~19K	—	—
ds_ag	value	String	0.00	0.00	102	-0.000 / 0.000 / 0.000 / 1.000	—
ds_al	value	String	0.00	0.00	102	-0.000 / 0.000 / 0.000 / 1.000	—
ds_dg	value	String	0.00	0.00	102	-0.000 / 0.000 / 0.000 / 1.000	—
ds_dl	value	String	0.00	0.00	102	-0.000 / 0.000 / 0.000 / 1.000	—
dp_ag	value	Int32	0.00	—	101	-50.000 / 0.000 / 44.000 / 50.000	—
dp_al	value	Int32	0.00	—	101	-50.000 / 0.000 / 44.000 / 50.000	—
dp_dg	value	Int32	0.00	—	101	-50.000 / 0.000 / 44.000 / 50.000	—
dp_dl	value	Int32	0.00	—	101	-50.000 / 0.000 / 44.000 / 50.000	—

Usage

# whole cache (~25 GB)
hf download biodatageeks/vepyr_116_GRCh38_plugin_spliceai \
  --repo-type dataset --local-dir ~/vepyr_plugin_cache/plugin/spliceai

# or a single chromosome
hf download biodatageeks/vepyr_116_GRCh38_plugin_spliceai \
  chr21.parquet manifest.json --repo-type dataset --local-dir ~/vepyr_plugin_cache/plugin/spliceai

The files are plain Parquet — usable directly from DuckDB, Polars or DataFusion independently of vepyr:

SELECT start, allele_string, symbol_out, ds_ag, ds_al, ds_dg, ds_dl
FROM 'chr21.parquet'
WHERE greatest(
        CAST(ds_ag AS DOUBLE), CAST(ds_al AS DOUBLE),
        CAST(ds_dg AS DOUBLE), CAST(ds_dl AS DOUBLE)) >= 0.5;

Licence and intended use

SpliceAI is developed by Illumina and released under the Illumina SpliceAI licence — free for academic and non-commercial research use; commercial use requires a separate licence from Illumina. The precomputed scores redistributed here were obtained from Ensembl's variation_plugins distribution and inherit those terms. This cache is a format conversion; the predictions are unmodified.

SpliceAI is a predictive model for research use, not a validated clinical test.

Citation

Jaganathan, K., Kyriazopoulou Panagiotopoulou, S., McRae, J. F., et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535–548.e24 (2019). doi:10.1016/j.cell.2018.12.015