alphagenome

AlphaGenome and the AlphaGenome Atlas

AlphaGenome is DeepMind's sequence-to-function model: 1 Mb of DNA in, base-pair predictions for eleven assay types across thousands of human and mouse tracks out. The AlphaGenome Atlas (released 2026-09-08) is that model run once over every possible single-nucleotide change in GRCh38, about 9 billion variants, stored with a single ranking number, the AlphaGenome Variant Impact (AVI) score, its genome-wide percentile, and an 18-way attribution of what drives it. Both are reached through one pip install alphagenome and one API key.

Research and theoretical modelling only. Outputs must not be used to train other models, and are not for diagnostic procedures or medical decisions.

When to use which

You have	Use	Why
hg38 SNVs (a VCF, a credible set, a region up to ~1 kb)	Atlas via scripts/atlas_query.py	precomputed, higher quota, includes AVI and attributions
indels, mouse variants, a non-reference background, a custom scorer or window	model via scripts/score_variants.py or Python	the Atlas is SNV-only and hg38-only
a hypothesis to explain (which motif, which tissue, REF vs ALT tracks)	model predict_variant + plots, Atlas track scores, portal link	mechanism, not just rank
GRCh37 coordinates, rsIDs, unnormalised indels	genomic-coordinates first, then come back	wrong build or swapped REF gives a plausible wrong answer
ClinVar assertions, gene-disease validity, ACMG framing	folklore-variant-evidence, database-lookup	AlphaGenome is one evidence line, never the verdict
promoter/enhancer/expression predictions without a DeepMind key	genomic-intelligence	different provider, keyless demo tier

Setup

uv pip install alphagenome                     # PyPI; tested on Python 3.12 and 3.13, alphagenome 0.9.0
export ALPHAGENOME_API_KEY="..."               # https://deepmind.google.com/science/alphagenome
cd skills/alphagenome/scripts
python atlas_query.py scorers                  # proves key + network in one call

Never put the key on a command line or in a file you commit; the scripts only read it from the environment. An invalid key surfaces as ValueError: API key not valid, not as a permission error.

The coordinate contract

• A variant is 1-based chr:pos:ref>alt (chr22:36201698:A>C). gnomAD (22-36201698-A-C), GTEx (chr22_36201698_A_C_b38), and Open Targets spellings are accepted by the scripts and by genome.Variant.from_str.
• An interval on the command line is 1-based closed chr:start-end; the SDK's genome.Interval is 0-based half-open. The scripts convert.
• Human is GRCh38 only. The Atlas key is chr:pos:alt; REF is implied by the reference, so a variant with REF and ALT swapped, or on GRCh37, returns a wrong record silently. Check REF against the FASTA before trusting a lookup.
• rsIDs are not accepted by the API or the portal. Resolve them to coordinates.
• Use the chr prefix; MT becomes chrM.

Atlas workflow

1. Rank with AVI

python atlas_query.py avi --variant chr22:36201698:A>C chr9:128225994:G>A
python atlas_query.py avi --input candidates.vcf --min-phred 20 -o avi.tsv
python atlas_query.py avi --interval chr11:5225727-5226575 --top-k 25 -o hbb_window.tsv
python atlas_query.py avi --input credible_set.tsv --with-tracks -o avi_tracks.tsv

Output, one row per variant:

Column	Meaning
avi_raw	composite model output (the 18 attributions sum to it)
avi_cdf_quantile	cumulative quantile against all genome-wide SNVs, as served
avi_tail_quantile, avi_phred, avi_top_percent	tail = 1 - cdf, phred = -10 log10(tail); Phred 20 = top 1 %, 30 = top 0.1 %
top_feature, top_feature_value	largest absolute SHAP attribution and its value
fi_MERGED_SPLICING ... fi_IS_DELETION	all 18 attributions (keys in references/atlas.md)
top_track_* (with --with-tracks)	the strongest track behind the top feature: scorer, track, biosample, ontology CURIE, gene, raw score
atlas_url	deep link to the variant on the portal
error	per-variant lookup failure (indel, N base, wrong REF) instead of a crash

The Atlas report's advice: rank, do not threshold, and pick thresholds by region or application. Pathogenic regulatory variants sit in lower AVI bins than protein-truncating or splice-motif variants, so a single genome-wide cut-off under-calls exactly the variants this resource was built for.

Read the attribution before the number. MERGED_SPLICING or ALPHAMISSENSE on top means a splice or coding mechanism; MAX_ABS_DNASE, MAX_ABS_CHIP_TF, MAX_ABS_RNA_SEQ mean a regulatory mechanism you can resolve by track; CACTUS_241_WAY or PHASTCONS_470_WAY on top means conservation is carrying the score and the molecular mechanism is not resolved.

2. Resolve the mechanism by track

python atlas_query.py scorers                                   # what the server serves right now
python atlas_query.py tracks --scorer RNA_SEQ --query colon     # find ontology CURIEs
python atlas_query.py scores --variant chr22:36201698:A>C \
    --scorers RNA_SEQ DNASE SPLICE_SITE_USAGE --ontology UBERON:0001157 -o colon.tsv
python atlas_query.py scores --interval chr11:5225727-5226575 --scorers CHIP_TF --gene HBB -o hbb_tf.tsv

One row per variant x track (x gene for RNA_SEQ, POLYADENYLATION, SPLICE_*), with raw_score and, where served, quantile_score. Track-level scorer names: ATAC, DNASE, CHIP_TF, CHIP_HISTONE, CAGE, PROCAP, RNA_SEQ, POLYADENYLATION, SPLICE_SITES, SPLICE_SITE_USAGE, SPLICE_JUNCTIONS, CONTACT_MAPS, plus *_ACTIVE variants; scorers is the authority on the live list. Filter by the tissue the question is about, not by the genome-wide maximum: 9,440 tracks means something is always extreme somewhere.

3. Send the reader to the portal

python atlas_link.py variant chr22:36201698:A>C --biosample "colon" --modalities RNA_SEQ,DNASE,CHIP_TF
python atlas_link.py locus chr11:5225727-5226575 --tf GATA1
python atlas_link.py gene HBB --markdown

No key, no network. The site shows the AVI track, per-modality heatmaps over every biosample, REF-vs-ALT prediction tracks, and motif instances. Attach a link to every variant you report.

In Python

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

client = atlas.create(os.environ["ALPHAGENOME_API_KEY"], timeout=30)
scores = client.query_variant(
    genome.Variant.from_str("chr22:36201698:A>C"),
    requested_scorers=["AVI_SCORE", "AVI_SCORE_FEATURE_IMPORTANCE", "RNA_SEQ"],
    ontology_terms=["UBERON:0001157"],          # optional; ignored for scorers without ontology metadata
)
avi = scores["AVI_SCORE"]                        # AnnData: X (1,1) raw; layers['quantiles'] (1,1) cdf
fi = scores["AVI_SCORE_FEATURE_IMPORTANCE"]      # AnnData: X (1,18); var['name'] = feature keys
rna = scores["RNA_SEQ"]                          # AnnData: obs = variant x gene, var = tracks, X = log2 FC
client.query_interval(genome.Interval("chr11", 5225726, 5226575), requested_scorers=["AVI_SCORE"])

query_interval returns all 3 SNVs per base, in 32 bp chunks. Keep windows to about 1 kb (3,000 variants); atlas_query.py refuses more unless --max-window is raised. query_variants stops at the first failed lookup; the script queries one variant at a time so misses become error cells.

Model workflow

Score variants the Atlas does not hold

python score_variants.py --variant chr22:36201698:A>C -o scores.tsv                   # 12 recommended scorers, 1 Mb
python score_variants.py --input indels.vcf --scorers RNA_SEQ SPLICE_SITE_USAGE \
    --ontology UBERON:0001157 --min-abs-quantile 0.99 -o colon.tsv
python score_variants.py --organism mouse --variant chr7:45000000:A>G --sequence-length 500KB
python score_variants.py --list-scorers
python score_variants.py --list-tracks --output-type RNA_SEQ --query liver -o tracks.tsv

Output is the official tidy table from variant_scorers.tidy_scores: one row per variant x scorer x track (x gene) with raw_score and quantile_score, sorted by |raw|. Default scorers are the 12 recommended difference scorers; --include-active adds the seven *_ACTIVE activity scorers. At most 20 scorers per request.

from alphagenome.models import dna_client, variant_scorers
model = dna_client.create(os.environ["ALPHAGENOME_API_KEY"])
variant = genome.Variant.from_str("chr22:36201698:A>C")
interval = variant.reference_interval.resize(dna_client.SEQUENCE_LENGTH_1MB)
adatas = model.score_variant(interval, variant, variant_scorers=[variant_scorers.RECOMMENDED_VARIANT_SCORERS["RNA_SEQ"]])
df = variant_scorers.tidy_scores(adatas)         # filter df.ontology_curie afterwards; score_variant takes no ontology_terms

Predict tracks and mutagenise

vo = model.predict_variant(interval, variant,
                           requested_outputs=[dna_client.OutputType.RNA_SEQ, dna_client.OutputType.DNASE],
                           ontology_terms=["UBERON:0001157"])
vo.reference.rna_seq.values, vo.alternate.rna_seq.values      # (1048576, n_tracks)

window = genome.Interval("chr20", 3_753_000, 3_753_400).resize(dna_client.SEQUENCE_LENGTH_16KB)
ism = model.score_ism_variants(interval=window, ism_interval=window.resize(256),
                               variant_scorers=[variant_scorers.CenterMaskScorer(
                                   requested_output=dna_client.OutputType.DNASE, width=501,
                                   aggregation_type=variant_scorers.AggregationType.DIFF_MEAN)])

Supported windows: 16 kb, 100 kb, 500 kb, 1 Mb (2**14 to 2**20); 1 Mb is the default and is required for distal enhancers and contact maps. Ontology terms are CURIEs (UBERON:0002048 lung, CL:0000084 T cell); discover them with --list-tracks or model.output_metadata(...).concatenate(). Plotting, gene annotation (GENCODE v46 Feather on GCS), splicing and haplotype recipes: references/model-api.md.

Reading the numbers

Always report raw score and quantile or Phred, with the scorer, track, biosample CURIE, and gene. raw_score is the effect size on the scorer's scale (RNA_SEQ is log2 fold change: -1 is half); quantile_score is the rank against common variants and saturates near 0.99999. A quantile above 0.99 with |raw| < 0.1 is the standard artefact of a quiet region and means no effect. Unsigned scorers (SPLICE_*, POLYADENYLATION, CONTACT_MAPS, *_ACTIVE) have no direction. Most variants are benign; "AlphaGenome predicts no molecular effect" is a complete answer, and a variant inside a peak whose REF and ALT tracks are identical is not "disrupting" anything. Full rules, tissue matching, and the reporting checklist: references/interpretation.md.

What the model cannot see: trans effects, non-polyadenylated RNAs (snRNA genes such as RNU4-2), cell types absent from training, protein-level consequences (AlphaMissense is folded into AVI for that), RNA structure and miRNA biology, diploid dosage, developmental time, species other than human and mouse.

Limits, quota, terms

• Atlas: GRCh38 SNVs only for now; indels were scored for the paper and are promised later. Reference N bases were never scored.
• Quotas are per key and unpublished; the Atlas is documented as having a larger query rate than on-demand prediction. Transient RESOURCE_EXHAUSTED and UNAVAILABLE are retried by the client (5 attempts, back-off to 60 s).
• Access tiers (Atlas report): AVI scores are also a permissively licensed Tabix download at https://alphagenome.google/downloads; feature attributions and splicing scores are non-commercial downloads; all other raw track scores are API-only and non-commercial. Commercial API access is "coming soon" via Google Cloud Model Garden.
• The alphagenome client is Apache-2.0; model weights and outputs carry DeepMind's terms. Cite Avsec et al., Nature 649:1206 (2026) and the Atlas report (Cheng, Taylor, Nicolaisen, Pan, Bycroft, Perino, Ward et al., 2026).

References

• references/atlas.md - what the Atlas contains, the 19 scorer configurations with track counts, AVI training and the 18 features, quantile to Phred, the client API and AnnData layout, error mapping, access tiers, portal URL grammar, GTF and download locations.
• references/model-api.md - dna_client cheat sheet: coordinates, sequence lengths, output types and track counts, ontology metadata, predict and score calls, recommended scorer configurations, ISM, gene annotation, plotting.
• references/interpretation.md - raw versus quantile, AVI thresholds, tissue matching, negative results, model blind spots, coordinate hygiene, reporting checklist.
• Scripts: scripts/atlas_query.py (Atlas: avi, scores, scorers, tracks), scripts/score_variants.py (model scoring, --list-scorers, --list-tracks), scripts/atlas_link.py (portal deep links, offline).

Agent operating procedure

1. Check the environment. Confirm tool versions, the reference genome/annotation build and the input formats (FASTQ, BAM, VCF, h5ad).
2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
3. Run a small version first. Run the pipeline on a small subset (one sample, one chromosome, a few thousand cells) first.
4. Execute the full task using the instructions and references above.
5. Validate the result. Check QC metrics, sample identities, genome build consistency and batch effects before interpreting results.
6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.

If this happens	Do this
Genome builds or identifiers do not match between inputs	Stop and harmonize (liftover, ID mapping) before continuing.
A function, flag or endpoint in these instructions is missing in the installed version	Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous	Ask the user, or state the assumption explicitly before running.

Integrity rules

• Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
• Do not interpret biological significance beyond what the statistics support; report multiple-testing correction.
• Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
• Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

• shap: Explain and audit machine-learning predictions with SHAP.