just-dna-seq/alphagenome_avi

AlphaGenome AVI scores, re-encoded

AlphaGenome's Variant Impact (AVI) scores for 8,812,917,339 SNVs on GRCh38, re-encoded from the 88.5 GB published tabix TSV into ~34 GB of parquet by just-dna-enricher.

This is a re-encoding, not a re-analysis. No score is changed, recomputed or filtered.

What is in it

data/alphagenome_avi-<contig>.parquet	chrom, pos (1-based VCF), ref, alt, raw_score_e5
avi_knots.parquet	the PHRED reconstruction curve — not optional, see below
release.json	row counts, the source digest, and the artifact's own timestamp
LICENSE.txt	the Output Terms' "Use restrictions" section, verbatim

raw_score is stored as Int32 at a scale of 10^5. The published file prints at most five decimals, so the integer is exactly lossless where a Float32 would be both larger and lossy. Compare in the integer domain — score >= 0.1 is raw_score_e5 >= 10_000. Dividing back (raw_score_e5 / 1e5) disagrees with the printed value on ~53% of rows, because the division rounds a second time.

PHRED is not stored. Measured over every row it is an exact within-corpus rank — PHRED >= p keeps 10^(-p/10) of the corpus — so it is a function of raw_score, and storing it costs 24.7 GB. avi_knots.parquet carries the curve in 466 KB instead, as an interval per printed score rather than a point, because the source prints raw_score to four significant digits and PHRED to six.

That interval makes threshold safety decidable in advance: a threshold is unsafe iff it falls inside some knot's span. Genome-wide, exactly one knot straddles any integer threshold from 1 to 50 — 0.00076, 676,356 rows, spanning PHRED 2.99961 to 3.00027. Every other integer threshold is decided.

Absence is row-absence. AVI covers ~95% of the assembly and includes 672,931 genuine zeros, so a position with no row is unscored and a row holding zero is scored zero. They are not the same.

Terms — read these

Governed by the AlphaGenome Services Additional Terms of Service and the AlphaGenome Output Terms of Use. The AVI SNV scores are classified by the AlphaGenome download page as a Permissive Use Downloadable Artifact, for commercial and non-commercial use; the merged-splicing and feature-importance artifacts are not and are absent here deliberately.

LICENSE.txt in this repository carries the Output Terms' "Use restrictions" section as an enforceable provision, which restriction 3b requires of anyone attaching their own terms. By using this data you agree to the AlphaGenome Output Terms of Use at http://deepmind.google.com/science/alphagenome/output-terms.

Two bars worth stating plainly because they are easy to miss:

• No training of machine-learning models for predicting genetic variant effects (prohibition 4).
• Google may request deletion of Output and Derivatives in your possession on breach — not only on termination.

The applicable version of the terms is the one effective on the date the Output was generated, which is why release.json records the source artifact's own timestamp.

Citation

Avsec et al., Advancing regulatory variant effect prediction with AlphaGenome, Nature 649(8099):1206-1218, 2026. doi:10.1038/s41586-025-10014-0

Provenance

Built by just-dna-enricher alphagenome build --input <the artifact> from the publisher's own download; the source digest is in release.json. Nothing in this repository was fetched by a tool — the source is behind a sign-in whose eligibility clause bars classes of holder, so acquisition is the operator's act under their own acceptance of the terms.