Datasetpublic

biodatageeks/vepyr_116_GRCh38_plugin_alphamissense

vepyr plugin cache — AlphaMissense (GRCh38, VEP 116) A prebuilt, frequency-tiered Parquet cache of AlphaMissense pathogenicity predictions for use with vepyr, the Rust/DataFusion VEP-compatible variant annotation engine. It reproduces the CSQ output of Ensembl VEP 116's --plugin AlphaMissense without requiring the upstream TSV or the Perl plugin at annotation time. Source version This is the fact you most likely came here for. Source file… See the full description on the dataset page: https://huggingface.co/datasets/biodatageeks/vepyr_116_GRCh38_plugin_alphamissense.

sourceHugging Facecc-by-nc-sa-4.0updated 19d agoView on Hugging Face

0likes300downloads

Dataset Card

vepyr plugin cache — AlphaMissense (GRCh38, VEP 116)

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

Source version

This is the fact you most likely came here for.


Source file	`AlphaMissense_hg38.tsv.gz` (canonical transcripts)
Source URL	`https://storage.googleapis.com/dm_alphamissense/AlphaMissense_hg38.tsv.gz`
Upstream release	DeepMind AlphaMissense, 2023 release (Cheng et al., Science 2023)
Source MD5	`9fd167735f16a1b87da6eb3e4c25fcb5` (upstream gzip, from GCS object metadata)
Build input MD5	`46d0028375cf95088bd014ff6855cffd` (`AlphaMissense_hg38.bgz.tsv.gz`, the BGZF+tabix re-compression of the upstream file that was actually built from — declared as `path_md5` in the manifest)
Source retrieved	2026-07-06
Genome build	GRCh38 / hg38, 1-based, `chr`-prefixed contigs in source
Cache built	2026-09-05
Target VEP version	Ensembl VEP 116 (`VEP_plugins` release/116 `AlphaMissense.pm`)
Build manifest	`plugins/alphamissense/alphamissense.source.toml`
vepyr-plugins tag	`v0.1.1` — recorded in `manifest.json` as `cache_source_version: v0.1.1@3e1c039`
Source verification	the BGZF re-compression used as build input was hashed (`verified_md5` in `manifest.json`, differs from the upstream `md5` by design, see the plugin README in vepyr-plugins)

AlphaMissense's distribution carries no internal version string — the file header is only the DeepMind copyright/licence banner — so the retrieval date above is the precise provenance marker for this build.

Provenance

Rebuilt on 2026-09-05 from sources verified against the v0.1.1 manifest; shard bytes are reproducible (a second build yields identical MD5s) since the tier stage orders rows totally. The sources block in manifest.json records url, declared and verified MD5, size and index digest for each input.

chr1.parquet … chr22.parquet, chrX/chrY/chrMT.parquet  25 per-contig shards
manifest.json                   schema, CSQ field mapping, per-shard row/tier counts, source provenance

Covers chr1–chr22, chrX, chrY and chrMT. Total ≈ 567 MB, 71,111,240 rows (84,844 warm / 71,026,396 cold).

Schema

column	type	role
`chrom`	`string`	contig
`start`	`uint32`	1-based position
`end`	`uint32`	1-based position
`allele_string`	`string`	`REF/ALT`, VEP-minimised
`protein_variant`	`string`	per-transcript match discriminator, `{ref_aa}{Protein_position}{alt_aa}`
`am_class`	`string`	→ CSQ field `am_class`
`am_pathogenicity`	`float`	→ CSQ field `am_pathogenicity`
`tier`	`int8`	frequency tier — `0` = warm, `1` = cold

Emitted CSQ fields, in VEP's own order: am_class, am_pathogenicity.

am_class thresholds are AlphaMissense's own: likely benign if score < 0.34, likely pathogenic if score > 0.564, ambiguous otherwise.

Matching semantics

AlphaMissense is a per-transcript annotation: VEP matches each transcript consequence's amino-acid change against the row's protein_variant. The lookup key is therefore (chrom, start, end, allele_string) plus the protein_variant discriminator, which vepyr builds at runtime from the engine attributes ref_aa, Protein_position and alt_aa.

Alleles are stored minimised (allele_match = "minimised"), matching AlphaMissense.pm, which calls get_matched_variant_alleles() before comparing rows.

Frequency tiering

Each shard is sorted by (tier, start). 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 — 83,141 rows, 0.12% 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 the 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	25 shards match the manifest
contig	✅ pass	0 foreign-contig rows in 25 shards
order	✅ pass	0 descending steps in 25 shards
tier_domain	✅ pass	0 rows with tier outside {0,1} in 25 shards
manifest_counts	✅ pass	rows/warm/cold match in 25 shards
manifest_files	✅ pass	25 manifest contigs, no stray shards
positions	✅ pass	0 rows with start < 1 or end < start - 1 in 25 shards
allele_form	✅ pass	0 malformed allele strings in 25 shards
duplicates	✅ pass	0 duplicate probe keys in 25 shards (manifest assume_unique=false)

Contigs

contig	rows	warm	cold	warm %	size
chr1	7,151,767	9,026	7,142,741	0.1%	57 MB
chr10	2,800,650	3,255	2,797,395	0.1%	22 MB
chr11	4,245,583	5,796	4,239,787	0.1%	34 MB
chr12	3,605,531	3,926	3,601,605	0.1%	29 MB
chr13	1,261,913	1,123	1,260,790	0.1%	10 MB
chr14	2,280,519	2,913	2,277,606	0.1%	18 MB
chr15	2,511,746	2,950	2,508,796	0.1%	20 MB
chr16	2,964,035	3,633	2,960,402	0.1%	24 MB
chr17	4,101,178	4,688	4,096,490	0.1%	33 MB
chr18	1,130,325	1,348	1,128,977	0.1%	9.0 MB
chr19	4,626,365	7,049	4,619,316	0.2%	37 MB
chr2	5,274,879	5,424	5,269,455	0.1%	42 MB
chr20	1,694,533	2,099	1,692,434	0.1%	13 MB
chr21	698,535	1,059	697,476	0.2%	5.6 MB
chr22	1,466,988	2,086	1,464,902	0.1%	12 MB
chr3	4,058,299	3,905	4,054,394	0.1%	32 MB
chr4	2,834,254	3,118	2,831,136	0.1%	23 MB
chr5	3,337,872	3,435	3,334,437	0.1%	27 MB
chr6	3,532,055	4,492	3,527,563	0.1%	28 MB
chr7	3,380,254	5,359	3,374,895	0.2%	27 MB
chr8	2,399,876	2,939	2,396,937	0.1%	19 MB
chr9	2,884,038	3,518	2,880,520	0.1%	23 MB
chrMT	24,074	0	24,074	0.0%	200 KB
chrX	2,670,963	1,694	2,669,269	0.1%	21 MB
chrY	175,008	9	174,999	0.0%	1.4 MB
total	71,111,240	84,844	71,026,396	0.1%	567 MB

Columns

column	role	type	empty %	distinct	numeric (min / p50 / p95 / max)	top values
protein_variant	match	String	0.00	~902K	—	—
am_class	value	String	0.00	3	—	likelybenign (41M), likelypathogenic (23M), ambiguous (7.9M)
am_pathogenicity	value	Float32	—	9,892	0.000 / 0.250 / 0.993 / 1.000	—

Usage

bash

hf download biodatageeks/vepyr_116_GRCh38_plugin_alphamissense \
  --repo-type dataset --local-dir ~/vepyr_plugin_cache/plugin/alphamissense

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

sql

SELECT start, allele_string, protein_variant, am_class, am_pathogenicity
FROM 'chr21.parquet'
WHERE start BETWEEN 33000000 AND 33100000;

Licence

AlphaMissense data is © 2023 DeepMind Technologies Limited and licensed [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) — non-commercial use only, share-alike. This cache is a format conversion of that data and inherits those terms. The predictions themselves are unmodified.

AlphaMissense is intended for research use; it is not validated for direct clinical application.

Citation

Cheng, J., Novati, G., Pan, J., et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381, eadg7492 (2023). doi:10.1126/science.adg7492