biodatageeks/vepyr_116_GRCh38_plugin_cadd

vepyr plugin cache — CADD v1.7 (GRCh38, VEP 116)

A prebuilt, frequency-tiered Parquet cache of CADD deleteriousness scores for use with vepyr, the Rust/DataFusion VEP-compatible variant annotation engine. It reproduces the CSQ output of Ensembl VEP 116's --plugin CADD without requiring the upstream whole-genome TSVs or the Perl plugin at annotation time.

⚠️ Non-commercial use only. See Licence below.

Source version

This is the fact you most likely came here for.

Upstream release	CADD v1.7, GRCh38
Source URL	https://krishna.gs.washington.edu/download/CADD/v1.7/GRCh38/
SNV source file	whole_genome_SNVs.tsv.gz (all possible SNVs)
Indel source file	gnomad.genomes.r4.0.indel.tsv.gz (gnomAD genomes r4.0 indels)
SNV source MD5	88577a55f1cd519d44e0f415ba248eb9 (upstream MD5SUMs)
Indel source MD5	4b9c685c96d396af4d001c2f7dd9d8f9 (upstream MD5SUMs)
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 CADD.pm)
Build manifest	plugins/cadd/cadd.source.toml
vepyr-plugins tag	v0.1.1 — recorded in manifest.json as cache_source_version: v0.1.1@3e1c039
Source verification	both source files (SNV and indel parts) and their .tbi indexes were MD5-verified against the manifest before the build (verified_md5 in manifest.json)

Both source files are combined into one cache, exactly as CADD.pm takes snv=…,indels=… and queries the two together.

Indel coverage is limited to gnomAD genomes r4.0 calls — unlike SNVs, which are precomputed exhaustively. A novel indel absent from gnomAD r4.0 has no precomputed score here; scoring it requires the CADD web service or a local CADD install. That is absence of data, not a benign call.

Provenance

Rebuilt on 2026-09-05 from the complete source files (the 87 GB whole-genome SNV table and the gnomAD indel table, 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 each input. Shard bytes are reproducible: the tier stage orders rows totally by the probe key (tier, start, allele_string), so a second build from the same verified sources yields identical MD5s. The manifest records assume_unique: true: the source claims one row per allele, so the build skipped the exhaustive dedup and sampled that claim instead.

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 sources have no mitochondrial records, so chrMT is listed in manifest.json with rows: 0 and has no shard file; an MT variant gets empty CADD fields. Total ≈ 76 GB, 8,918,413,601 rows (25,038,237 warm / 8,893,375,364 cold). Largest shard chr2.parquet ≈ 6.2 GB.

Schema

column	type	CSQ field
chrom	string	— (contig)
start	uint32	— (1-based position)
end	uint32	—
allele_string	string	— (REF/ALT, VEP-minimised)
cadd_raw	string	CADD_RAW — raw CADD score
cadd_phred	string	CADD_PHRED — PHRED-like scaled score
tier	int8	— (frequency tier: 0 = warm, 1 = cold)

Scores are kept as strings to preserve upstream decimal formatting byte-for-byte, which is what VEP 116 CSQ parity requires; cast on read if you want numerics.

CADD_PHRED is the scaled rank: 10 = top 10% most deleterious of all possible SNVs, 20 = top 1%, 30 = top 0.1%. CADD_RAW is the unscaled model output and is only comparable within a single CADD release — do not compare raw scores across CADD versions.

Matching semantics

CADD is a per-variant annotation: CADD.pm matches on position and allele only, with no transcript or amino-acid discriminator, so the same score applies to every transcript line of a variant. The lookup key is (chrom, start, end, allele_string).

Alleles are stored minimised (allele_match = "minimised"), matching CADD.pm, which calls get_matched_variant_alleles() before comparing rows. Indel rows in the source carry a shared leading anchor base in VCF convention (e.g. AG>A); the build applies the same anchor-trim and start + 1 shift that vepyr's runtime allele normalisation produces, so indel rows land on the key the engine actually probes. SNV rows pass through unshifted.

Frequency tiering

Each shard is sorted by (tier, start, allele_string). 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 — 25,038,237 rows, 0.28% 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 700 M-row shard. Per-shard warm/cold counts are in manifest.json.

Usage

At ~69 GB this is the largest of the vepyr plugin caches — pull single chromosomes unless you need the whole genome.

# a single chromosome
hf download biodatageeks/vepyr_116_GRCh38_plugin_cadd \
  chr21.parquet manifest.json --repo-type dataset --local-dir ~/vepyr_plugin_cache/plugin/cadd

# the whole cache (~76 GB)
hf download biodatageeks/vepyr_116_GRCh38_plugin_cadd \
  --repo-type dataset --local-dir ~/vepyr_plugin_cache/plugin/cadd

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

SELECT start, allele_string, cadd_raw, cadd_phred
FROM 'chr21.parquet'
WHERE start BETWEEN 33000000 AND 33100000
  AND CAST(cadd_phred AS DOUBLE) >= 20;

Licence and intended use

CADD is free for non-commercial use only. Commercial use requires a licence — see the CADD download terms and contact the University of Washington / Hudson-Alpha licensing contacts listed there. This cache is a format conversion of CADD v1.7 scores and is redistributed under those same terms; the scores are unmodified. By downloading it you accept CADD's upstream conditions.

The indel component additionally derives from gnomAD genomes r4.0, which is released under CC0.

CADD is a research tool. A high PHRED score means a variant ranks as more likely deleterious than most possible substitutions — it is not a pathogenicity assertion and not a clinical result.

Citation

Schubach, M., Maass, T., Nazaretyan, L., Röner, S. & Kircher, M. CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions. Nucleic Acids Research 52, D1143–D1154 (2024). doi:10.1093/nar/gkad989

Rentzsch, P., Witten, D., Cooper, G. M., Shendure, J. & Kircher, M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Research 47, D886–D894 (2019). doi:10.1093/nar/gky1016