---
license: other
pretty_name: Ensembl variation (split parquet)
tags:
- genomics
- variants
- GRCh38
configs:
- config_name: core
  data_files: data/*.parquet
  default: true
- config_name: consequence
  data_files: groups/consequence/*.parquet
- config_name: clinical
  data_files: groups/clinical/*.parquet
- config_name: predictors
  data_files: groups/predictors/*.parquet
- config_name: cadd
  data_files: groups/cadd/*.parquet
- config_name: frequency
  data_files: groups/frequency/*.parquet
- config_name: frequency_populations
  data_files: groups/frequency_populations/*.parquet
- config_name: transcripts
  data_files: groups/transcripts/*.parquet
---

# Ensembl variation, split parquet

dbSNP variants for GRCh38 with Ensembl VEP annotations, converted from the Ensembl variation VCF
(`variation.vcf.gz`, release directory `2026_07_13`) into
parquet tables that DuckDB joins on demand. Release id: `ensembl-variation-2026_07_13-vep116.0`.

Built by [prepare-annotations](https://github.com/dna-seq/prepare-annotations) for
[just-dna-lite](https://github.com/dna-seq/just-dna-lite).

## Layout

- `data/homo_sapiens-chr{N}.parquet`: the core, one row per ALT allele: `chrom, start, ref, alt, id, row_idx`.
  Positions are 1-based VCF POS; REF and ALT are spelled as in the VCF. Same file names and column
  names as the earlier `just-dna-seq/ensembl_variations` cache, which pipe-joined multi-allelic ALTs
  into one row; here each ALT has its own row.
- `groups/{group}/homo_sapiens-chr{N}.parquet`: sidecar columns keyed by `(chrom, row_idx)`.
  Sidecars are sparse: an allele with no value in a group has no row there.
- `manifest.json`: release id, sources, and every file with its row count, size and SHA-256.

`row_idx` is valid only within one release. Never join files whose `manifest.json` `release_id` differs.

| Group | Default | Rows | Size | Content | Columns |
|---|---|---:|---:|---|---|
| `core` | yes | 1,391,478,689 | 14.54 GB | Alleles and rsIDs | `chrom`, `start`, `ref`, `alt`, `id`, `row_idx` |
| `consequence` | yes | 1,391,478,685 | 4.14 GB | Most severe transcript consequence per allele (by IMPACT, protein-coding first) | `variant_class`, `consequence`, `impact`, `gene_symbol`, `gene_id`, `transcript_id`, `biotype`, `exon`, `intron`, `hgvsc`, `hgvsp`, `amino_acids`, `protein_position`, `sift`, `polyphen` |
| `clinical` | yes | 269,631,208 | 0.93 GB | ClinVar significance, PubMed ids and phenotypes Ensembl attaches to the variant | `clin_sig`, `somatic`, `pheno`, `pubmed`, `phenotypes`, `var_synonyms` |
| `predictors` | yes | 486,820,688 | 1.63 GB | REVEL, SpliceAI, AlphaMissense, ClinPred and MaveDB (maximum over transcripts) | `revel`, `spliceai_ds_ag`, `spliceai_ds_al`, `spliceai_ds_dg`, `spliceai_ds_dl`, `alphamissense_score`, `alphamissense_class`, `clinpred`, `mavedb_score`, `mavedb_urn` |
| `cadd` | optional | 1,323,076,840 | 10.41 GB | CADD v1.7 scores | `cadd_phred`, `cadd_raw` |
| `frequency` | yes | 1,359,614,704 | 7.36 GB | Global allele frequencies: gnomAD v4.1 exomes and genomes, 1000 Genomes phase 3 | `af_gnomad_exomes`, `af_gnomad_genomes`, `af_1kg`, `ancestral_allele` |
| `frequency_populations` | optional | 755,450,474 | 8.16 GB | Per-population allele frequencies: gnomAD v4.1 exomes and genomes, 1000 Genomes | `af_gnomad_exomes_afr`, `af_gnomad_exomes_amr`, `af_gnomad_exomes_asj`, `af_gnomad_exomes_eas`, `af_gnomad_exomes_fin`, `af_gnomad_exomes_mid`, `af_gnomad_exomes_nfe`, `af_gnomad_exomes_remaining`, `af_gnomad_exomes_sas`, `af_gnomad_genomes_ami`, `af_gnomad_genomes_afr`, `af_gnomad_genomes_amr`, `af_gnomad_genomes_asj`, `af_gnomad_genomes_eas`, `af_gnomad_genomes_fin`, `af_gnomad_genomes_mid`, `af_gnomad_genomes_nfe`, `af_gnomad_genomes_remaining`, `af_gnomad_genomes_sas`, `af_1kg_afr`, `af_1kg_amr`, `af_1kg_eas`, `af_1kg_eur`, `af_1kg_sas` |
| `transcripts` | optional | 2,757,116,851 | 4.32 GB | Every GENCODE primary transcript consequence of each allele | `consequence`, `impact`, `gene_symbol`, `gene_id`, `transcript_id`, `biotype`, `exon`, `intron`, `hgvsc`, `hgvsp`, `amino_acids`, `protein_position`, `sift`, `polyphen`, `distance`, `strand` |

## Querying

Match a sample to the core first, then join only the sidecars you need on `(chrom, row_idx)`:

```sql
WITH m AS (
  SELECT v.*, c.row_idx, c.id
  FROM read_parquet('sample.parquet') v
  LEFT JOIN read_parquet('data/*.parquet') c USING (chrom, start, ref, alt)
)
SELECT m.*, k.* EXCLUDE (chrom, row_idx)
FROM m LEFT JOIN read_parquet('groups/consequence/*.parquet') k USING (chrom, row_idx);
```

## Sources and licences

- VEP: `##VEP="v116.0" API="v116" time="2026-06-15 16:58:35" cache="/nfs/production/flicek/ensembl/variation/new_website/run_setup/outputs/tmp/homo_sapiens/116_GRCh38" ensembl=116.c0cf13d ensembl-compara=116.46f80af ensembl-funcgen=116.90049ea ensembl-io=116.6afb5dc ensembl-variation=116.2fb834b 1000genomes="phase3" COSMIC="102" ClinVar="202509" HGMD-PUBLIC="20204" assembly="GRCh38.p14" dbSNP="156" gencode="GENCODE 50" genebuild="GENCODE50" gnomADe="v4.1" gnomADg="v4.1" polyphen="2.2.3" regbuild="1.0" sift="6.2.1"`
- CADD: free for non-commercial use only (https://cadd.gs.washington.edu/)
- Ensembl data: no restrictions (https://www.ensembl.org/info/about/legal/)

The `cadd` group carries CADD scores, which are free for non-commercial use only; it is a separate
group so that commercial users can leave it out.