File size: 5,458 Bytes
b1e9bf7 9181540 b1e9bf7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 | ---
license: other
license_name: alphagenome-services-additional-terms-of-service
license_link: https://deepmind.google.com/science/alphagenome/terms
pretty_name: AlphaGenome AVI GRCh38 — lossless indexed VCF mirror
tags:
- genomics
- alphagenome
- variant-annotation
- grch38
- vcf
- biology
size_categories:
- n>1B
---
# AlphaGenome AVI GRCh38 — prepared VCF mirror
An unofficial, losslessly reformatted mirror of Google DeepMind's **AVI SNV
scores**, prepared for [GUIDE-IEI](https://github.com/yimingluo-md/guide-iei), and can be used directly for Ensembl VEP custom annotation.
Google DeepMind produced the predictions; GUIDE-IEI performed only the format
conversion described below. This mirror is not affiliated with or endorsed by
Google DeepMind.
## Source and terms
- [Official AlphaGenome downloads](https://deepmind.google.com/science/alphagenome/downloads)
- [Original AVI SNV ZIP](https://deepmind.google.com/science/alphagenome/_/download/atlas/avi_scores_snvs_tabix.zip)
- [AlphaGenome Services Additional Terms of Service](https://deepmind.google.com/science/alphagenome/terms), last modified September 8, 2026
- [AVI methodology](https://deepmind.google.com/science/alphagenome/learning#the-avi-model-and-scores)
The official downloads page lists **AVI SNV scores** in its permissive-use
category for commercial and non-commercial use. This is subject to the
applicable AlphaGenome terms, not a public-domain or Creative Commons release.
Review those terms directly for the permissions, exceptions and restrictions
that apply to your use. This mirror grants no additional rights and does not
change the upstream terms. It contains only AVI raw and Phred scores—not the
separately governed AVI feature breakdown, other Atlas predictions, model
weights, or API credentials.
Scores are computational predictions, not clinical pathogenicity
classifications. No clinical-use authorization or diagnostic validation is
claimed by this mirror. Cite and attribute the original AlphaGenome work as
specified by the upstream project; also identify this format conversion when
needed for reproducibility.
## Contents
Release: `atlas-2026-09-08-vcf-v1`.
| File | Size (bytes) | Purpose |
|---|---:|---|
| `avi.grch38.vcf.gz` | 75,764,802,676 | BGZF VCF containing all scores |
| `avi.grch38.vcf.gz.tbi` | 2,798,486 | Tabix VCF index |
| `manifest.json` | Small JSON | Source identity, counts, transformation and SHA-256 checksums |
| `SOURCE_AND_TERMS.txt` | Small text | Attribution and upstream links |
There are **8,812,917,339 scored SNV alleles**, grouped into **2,937,639,113 VCF
records**. Coverage is GRCh38 chromosomes **1–22, X and Y**. Mitochondrial and
alternate contigs and indels are not included. These are reference predictions
for possible alleles; the VCF contains **no patient samples or genotypes**.
## Exact transformation
The original BGZF TSV columns are `#CHROM`, `POS`, `REF`, `ALT`, `raw_score`
and `PHRED`. For every position, the three possible alternate nucleotides are
grouped into one VCF record. ALT order is A,C,G,T excluding REF. Both INFO
fields are declared **Number=A, Type=Float**, so each value corresponds to the
ALT in the same slot:
```vcf
1 10001 . T A,C,G . . raw=-0.03868,-0.032,-0.0372;phred=1.06466,1.3114,1.11839
```
The chromosome `chr` prefix was removed. Coordinates remain **1-based GRCh38**.
Score decimal text is unchanged: no rounding, imputation, thresholding,
filtering or liftover was performed. Negative raw values and zero are retained.
No gene or transcript assignment was added.
`raw` is the upstream AVI raw score; `phred` is the upstream Phred-scaled score.
A high impact score is not itself a clinical P/LP classification. Missing
coverage is not evidence of benignity.
## Using the files
Download the VCF and its `.tbi` together, preserving their names. Do not
recompress the VCF with ordinary gzip or use the index with a different file.
Standard Tabix queries work:
```bash
tabix avi.grch38.vcf.gz 1:10001-10001
```
For Ensembl VEP, use an exact-allele custom VCF annotation:
```text
--custom file=/path/avi.grch38.vcf.gz,short_name=AlphaGenomeAVI,format=vcf,type=exact,coords=0,fields=raw%phred
```
This produces `AlphaGenomeAVI_raw` and `AlphaGenomeAVI_phred`, including on
intergenic consequences. Match chromosome, position, **REF and ALT**, not
position alone. Select the matching ALT's score; do not take a maximum over
the three alternate alleles or interpret the list as three independent models.
## Integrity and provenance
The preparation checked the source schema and index, BGZF decompression CRCs,
sorted positions, finite scores, nonnegative Phred values, three distinct
alternate nucleotides at each position, and chromosome/total row counts. The
completed VCF was indexed with Tabix. VEP integration tests verified reversed
ALT ordering, intergenic annotations and rejection of a mismatched REF allele.
SHA-256:
```text
a34a9b48b6dbb3c769d86014e7f28062ef65420c5add7e88f30dfa50d5756b75 avi.grch38.vcf.gz
26d20b2855a478ec5b44bba8969a893453292bd432e6f969b01dd7a82334bbae avi.grch38.vcf.gz.tbi
```
The original source ZIP's SHA-256 is
`a237c198bc1c033da127fe0257129fdf246e7aa311f8fb584bb053ffa0c7966b`.
The manifest records the converter source SHA-256 and per-chromosome allele
counts. Original modification timestamps in the manifest are provenance, not
expected filesystem timestamps after downloading from this mirror.
|