AlphaGenome Atlas: data model, scorers, AVI, and access

Reviewed 2026-09-30 against alphagenome 0.9.0 (released 2026-09-08, the first version with alphagenome.atlas), the Atlas technical report (Cheng, Taylor, Nicolaisen, Pan, Bycroft, Perino, Ward, et al., "AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants", 2026 preprint), and DeepMind's own Atlas agent skills in google-deepmind/science-skills. Source and local SDK checks cover the client contracts below. No authenticated Atlas query was run in this review; server-side catalogues and portal rendering remain unverified. The report now has medRxiv DOI 10.64898/2026.09.16.26363192.

What is in the Atlas

Index key semantics

Each record is keyed by chromosome:position:alt on GRCh38. The REF allele is implied by the reference and is not part of the key, so:

• a variant written with REF and ALT swapped (for example minor-allele-first from a GWAS table) can produce a lookup miss;
• non-reference to non-reference substitutions do not exist;
• positions where the reference is N were never scored;
• GRCh37 coordinates return wrong-but-plausible answers. Lift over first.

The SDK sends the full variant (including REF) and returns the server variant in obs["variant"]; these helpers reject a response with a different variant identity. Do not assume the service independently validates REF or the assembly. Check REF against the reference FASTA before trusting any Atlas result (genomic-coordinates has normalize_variant.py and check_contigs.py).

The scorers (Table S1 of the report)

requested_scorers= takes these names. Track counts are for human; the _ACTIVE names match the SDK recommended configuration keys, but the served Atlas catalogue was not checked with an authenticated scorer_metadata() call.

"Signed" raw scorers can go negative (ALT lowers the signal). Their calibrated quantiles are on [-1, 1] (CDF p mapped to 2*p-1); unsigned scorers use [0, 1]. Use abs(quantile_score) for the CLI magnitude filter. The AVI CDF described below is unsigned and has its own background. These semantics follow the official scoring FAQ; authenticated Atlas quantile values were not checked in this review. For unsigned effect scorers the direction is not meaningful; _ACTIVE ranks activity rather than an allelic change.

Table S1 totals 9,440 scorer-track entries, including repeated tracks for active and difference scorers; it is not 9,440 distinct model outputs. python scripts/atlas_query.py tracks --scorer RNA_SEQ --query liver shows the metadata for any of them.

AlphaGenome Variant Impact (AVI)

AVI is a small neural network trained on proxy labels: gnomAD v4.1 variants with a group-maximum filtering allele frequency above 0.001 are "proxy benign" (about 2e7), below 0.001 "proxy impactful" (about 1e8). It therefore learns to rank variants by how strongly negative selection appears to act on them, using 18 features:

The 18 SHAP values (expected-gradients approximation, reference = all features zero) sum to the raw AVI score, so the attribution tells you which modality drives a high score: splicing, a TF-binding change, conservation alone, or a coding consequence via AlphaMissense.

Raw, quantile, Phred

The Atlas returns the raw score in X and a cumulative quantile against all genome-wide SNVs in layers['quantiles']. Convert as DeepMind's tools do:

tail  = 1 - cdf_quantile            # fraction of SNVs scoring at least this high
phred = -10 * log10(tail)           # 10 = top 10 %, 20 = top 1 %, 30 = top 0.1 %, 40 = top 0.01 %
top % = 10 ** (-phred / 10) * 100

Indel Phred scores are placed on the SNV quantile curve. The report's guidance: AVI is applicable to coding and non-coding variants, but use region- or application-aware thresholds and prefer ranking over hard cut-offs. Pathogenic regulatory variants land in lower AVI bins than pathogenic protein-truncating or splice-motif variants, so one genome-wide cut-off under-calls regulatory hits.

Benchmarks and case studies (report, for context only)

State-of-the-art on ClinVar non-coding, saturation genome editing, and TraitGym style benchmarks against CADD, GPN-Star, and AlphaMissense after removing training-set overlap; resolution of an epileptic-encephalopathy case in the GREGoR cohort via a DNM1 variant missed by prior tools; a 22 % increase in detectable rare non-coding associations for circulating proteins in UK Biobank.

Client API (alphagenome.atlas.atlas)

from alphagenome.atlas import atlas
from alphagenome.data import genome

client = atlas.create(api_key, timeout=30)          # gRPC to gdmscience.googleapis.com:443

client.scorer_metadata()   # -> {name: ScorerMetadata(name, is_signed, track_metadata: DataFrame)}

client.query_variant(
    genome.Variant.from_str("chr22:36201698:A>C"),  # 1-based position
    requested_scorers=["AVI_SCORE", "AVI_SCORE_FEATURE_IMPORTANCE", "RNA_SEQ"],
    ontology_terms=None,      # e.g. ["UBERON:0001157", "CL:0000084"]; strings or OntologyTerm
    gene_ids=None,            # Ensembl IDs, gene-centric scorers only
    gene_names=None,          # symbols, gene-centric scorers only
)                              # -> {scorer_name: AnnData}

client.query_variants([...], requested_scorers=[...], progress_bar=True, max_workers=10)
client.query_interval(genome.Interval("chr11", 5225726, 5226575), requested_scorers=[...])

• genome.Interval is 0-based half-open; genome.Variant.position is 1-based. Variant.from_str also accepts gnomAD (22-1024-A-C), GTEx (chr22_1024_A_C_b38), and Open Targets spellings when the matching variant_format=genome.VariantFormat... is explicitly passed; the SDK does not autodetect formats (the helper scripts do).
• query_interval splits the window into 32 bp chunks, walks the pagination, and returns stored SNVs (up to 3 per non-N base). A 1 kb window can yield 3,000 variants; latency depends on scorers, filters and service load. DeepMind's tool refuses windows above 1,000 bp by default, and so does scripts/atlas_query.py (--max-window).
• query_variants re-raises a failed future; other submitted calls may already have run. Results arrive in completion order, not input order; scripts/atlas_query.py makes separate variant requests so a single miss becomes an error cell.
• Each AnnData: X = raw scores (obs x tracks, float32), obs['variant'] = genome.Variant, plus gene_id, gene_name, strand, junction_Start, junction_End for gene-centric scorers; var = track metadata (name, strand, ontology_curie, biosample_name, biosample_type, assay title, gtex_tissue, transcription_factor, histone_mark where applicable); layers['quantiles'] when the server sent calibrated quantiles.
• AVI_SCORE: X shape (n, 1). AVI_SCORE_FEATURE_IMPORTANCE: X shape (n, 18), var['name'] = the feature keys above.
• An ontology_terms filter is ignored for scorers whose tracks have no ontology metadata (the filter string is built that way), so SPLICE_SITES and the AVI scorers still come back.

Transport and pagination

atlas.create uses TLS gRPC, service google.gdm.gdmscience.alphagenome.atlas.v1main.AtlasService, with x-goog-api-key authentication metadata. This SDK has no REST URL or JSON request body to construct. query_variant maps to GetDenseVariantScores (full variant, human organism, filter string), query_interval maps to ListDenseVariantScores (interval, human organism, filter, page token), and scorer_metadata maps to ListVariantScoresMetadata. The list response is variant_scores plus next_page_token; the SDK follows tokens until empty within each 32 bp chunk. Do not infer completion from a short page. Gene ID/name filters are ORed together, then ANDed with scorer/ontology filters; they can remove non-gene scorer results. Return order from parallel queries is unspecified.

create(timeout=30) is only the wait for channel readiness. The Atlas service config separately sets a 60 s deadline and up to five attempts, with 1 s initial backoff, multiplier 2, maximum backoff 60 s. It retries RESOURCE_EXHAUSTED, UNAVAILABLE, and DEADLINE_EXCEEDED; this does not guarantee the whole batch completes in 30 or 60 seconds.

Errors

atlas.handle_rpc_error maps gRPC status to Python exceptions:

Invalid-key errors can be ValueError (INVALID_ARGUMENT) or PermissionError (UNAUTHENTICATED); inspect the status/message. No current authenticated or invalid-key probe was run in this review.

Access tiers and terms

From the report's Data Availability section and the AlphaGenome terms:

These API outputs are for research and theoretical modelling, may not be used to train other machine-learning models, and are not for diagnostic procedures or medical decision-making. The website is free for non-commercial use. Query rates are demand-dependent and unpublished; the docs say the Atlas "will typically have a larger query rate" than on-demand model predictions.

The portal

https://deepmind.google.com/science/alphagenome/atlas (also reachable as https://alphagenome.google/atlas). Query parameters, from DeepMind's link builder:

Because RNA-seq and DNase tracks carry no TF code, a filter made only of TF predicates hides them; add SCORER_MODALITY:RNA-seq,SCORER_MODALITY:DNase alongside (scripts/atlas_link.py does this automatically).

Bulk files and annotations

• GENCODE v46 GTF as Feather, the annotation behind both Atlas and model gene scores: https://storage.googleapis.com/alphagenome/reference/gencode/hg38/gencode.v46.annotation.gtf.gz.feather (about 318 MB, roughly 4 GB RAM in pandas; column names are capitalised: Feature, Start, End, Strand). Load with pd.read_feather and use alphagenome.data.gene_annotation helpers (filter_to_mane_select_transcript, get_gene_interval, extract_tss) so gene models match the scores.
• AVI Tabix downloads: https://alphagenome.google/downloads (public page resolves to a sign-in shell; individual files and licence acceptance were not exercised).

Citing

Cite both the Atlas report (above) and the model paper: Avsec Ž. et al., "Advancing regulatory variant effect prediction with AlphaGenome", Nature 649, 1206-1218 (2026), doi:10.1038/s41586-025-10014-0. Software: https://github.com/google-deepmind/alphagenome (Apache-2.0 client; the model weights and outputs carry separate terms).

Official contract sources

• Released client and Atlas SDK.
• Atlas protobuf schema, filter builder, gRPC retry policy.
• Official portal link builder.
• Scoring/terms FAQ and Atlas release announcement.