AlphaGenome Atlas: Complete Guide

AlphaGenome Atlas is a free database that precomputes the predicted molecular impact of all ~9 billion possible single-letter DNA changes in the human genome — a 1-petabyte dataset searchable through a zero-code web portal, ranked by a new single-number AVI score. DeepMind announced it on September 8, 2026, alongside a preprint and case studies from the Broad Institute and the University of Exeter.

This guide follows the launch announcement, the DeepMind blog post, the 83-page technical preprint (Cheng et al.), Nature’s coverage, and the AlphaGenome API repository. It was checked on September 8, 2026, the day of the announcement.

What DeepMind launched

AlphaGenome Atlas is the second act of the AlphaGenome story. The base model — released June 2025 and published in Nature in January 2026 — takes up to 1 million DNA letters as input and predicts thousands of molecular properties at single-base resolution: gene expression, RNA splicing, chromatin accessibility, transcription factor binding, and more. It matched or beat the best external models on 24 of 26 variant-effect evaluations.

The Atlas scales that capability from “one variant per API call” to “every variant, precomputed”. Using in silico saturation mutagenesis (ISM), DeepMind ran the model across all three possible letter-changes at every position in the hg38 reference genome — roughly 9 billion single-nucleotide variants (SNVs) — plus more than 100 million short insertions/deletions (indels) observed in real populations through gnomAD, UK Biobank, and All of Us. The result is a 1-petabyte dataset, more than 30 times larger than the AlphaFold Database.

Four connected resources ship inside the Atlas:

• Molecular effect predictions — thousands of regulatory-readout predictions per variant, across hundreds of human and mouse cell types and tissues.
• AVI score — one PHRED-scaled number per variant summarizing predicted impact (details below).
• AVI feature attributions — a SHAP decomposition of every score into 18 interpretable components (splicing, chromatin, conservation, protein impact, and so on).
• DNA motif compendium — over 2,500 recurrent regulatory “words” of the genome and their locations, derived from the same contribution scores.

The announcement

DeepMind announced the Atlas on X with a 64-second explainer video and one sentence: an “AI-powered searchable database mapping the predicted impact of all 9 billion possible single-letter DNA changes.”

The launch echoes the AlphaFold playbook: take a model that researchers were already calling one at a time, precompute its outputs at planetary scale, and hand the result to a no-code portal. Around 9,000 researchers had used the AlphaGenome API since its release — a number DeepMind’s product team cited when explaining why the Atlas matters: the API required writing code, which blocked many biologists and clinical researchers entirely.

Why 98% of DNA was the problem

The human genome has about 3 billion base pairs. Only ~2% code for proteins; the other 98% is non-coding DNA that regulates when, where, and how much genes are expressed. Most trait-associated variants from genome-wide association studies live in that non-coding region, and interpreting them has been the field’s bottleneck.

Existing tools each covered part of the gap. Variant Effect Predictor (VEP) and ENCODE’s cCRE catalogue define broad regulatory regions but lack base-pair resolution. Evolutionary conservation models like GPN-Star prioritize variants across species but offer little molecular interpretability. CADD integrates features but trails on non-coding categories. AlphaMissense, DeepMind’s 2023 model, covers protein-altering variants — the coding 2%. Running AlphaGenome itself over a whole genome was computationally out of reach for most labs.

The Atlas’s pitch is simple: the computation is now done once, centrally, and served interactively — so the strongest in silico tool for non-coding interpretation became accessible to anyone with a browser.

Inside the Atlas

Three layers make the Atlas more than a dump of predictions:

1. Precomputed effects. For every SNV: the difference between reference and alternate allele predictions across thousands of molecular tracks — RNA-seq, splice sites and junctions, chromatin accessibility, histone marks, transcription factor binding, and contact maps. For indels: scores from a shift-augmentation strategy across the same modalities.
2. The AVI score and its decomposition. Each variant’s score decomposes additively into 18 SHAP features, so a researcher can see why a variant scores high — e.g. a splicing disruption versus a conservation signal.
3. The motif compendium. Over 2,500 recurrent DNA sequence motifs mapped genome-wide and annotated by function (activation, repression, accessibility change) per cell type — built from the same ISM contribution scores.

The AVI score, explained

The AlphaGenome Variant Impact (AVI) score is a supervised model trained to separate proxy-neutral from proxy-impactful human variants: observed variants with filtering allele frequency above versus below 0.1%. Its inputs:

Raw outputs convert to PHRED scaling: AVI 10 means top 10% of predictions, 20 means top 1%, and so on. That calibration matters for triage workflows — a clinical team can set a threshold and know roughly how many variants per genome will clear it.

Benchmark evidence

The preprint’s headline comparison is against ClinVar-asserted pathogenic versus benign variants, measured by AUPRC (area under the precision-recall curve — higher is better, and recall matters more than precision when hunting one causal variant in thousands):

Three caveats to read with the table. First, these are prediction benchmarks against clinical assertions, not proof of causation — the preprint itself pairs every benchmark with experimental validation stories. Second, AVI is trained on allele-frequency labels, which correlate with (but are not identical to) pathogenicity; the ClinVar evaluation is partially out-of-distribution, which is the honest way to test it. Third, the margins on protein-altering variants are small because AlphaMissense and conservation already carried most of that signal — the new value is concentrated in the non-coding rows.

Case studies: from a solved epilepsy case to BMI regions

The launch’s credibility rests on three worked examples, all with external collaborators:

• Rare disease (Broad Institute / GREGoR). Laura Covill and Anne O’Donnell-Luria’s team used the AVI score to prioritize variants in an unsolved epileptic encephalopathy case. The Atlas highlighted a variant in the DNM1 gene and the underlying AlphaGenome predictions showed exactly how it worked: the variant created an incorrect splice site, producing an abnormally extended protein. Experimental screens then validated the mechanism. The case was solved with a browser, not a compute cluster.
• Complex traits (University of Exeter / UK Biobank). Gareth Hawkes applied Atlas groupings to whole-genome data from 54,000+ UK Biobank participants. Grouping rare variants by predicted molecular effect uncovered 22% more non-coding associations than standard statistical approaches, and pinpointed regulatory variants driving abundance of circulating proteins including PLA2G7 (linked to aging) and EGLN1 (a cellular oxygen sensor). Scanning non-coding variants against body mass index surfaced 19 candidate genetic regions in the top 1% of predicted impact.
• Regulatory grammar (Stowers Institute). Julia Zeitlinger and Melanie Weilert used the motif compendium to classify transcription factors: which ones only change DNA accessibility versus which ones actively turn genes on or off — the kind of functional annotation the field had been assembling piecemeal for decades.

How to access it

The Antigravity integration is the quietly notable one for this site’s audience: DeepMind ships an Atlas-backed science skill for its agentic IDE, meaning an agent can plan a variant-prioritization workflow, query the Atlas, and interpret results in one session.

API quickstart

The Atlas portal needs no code, but the API remains the path for custom intervals and batch scoring. Smallest working example from the official repository (after pip install ./alphagenome and obtaining a free API key):

from alphagenome.data import genome
from alphagenome.models import dna_client

model = dna_client.create("YOUR_API_KEY")

interval = genome.Interval(chromosome="chr22", start=35677410, end=36725986)
variant = genome.Variant(
    chromosome="chr22",
    position=36201698,
    reference_bases="A",
    alternate_bases="C",
)

outputs = model.predict_variant(
    interval=interval,
    variant=variant,
    ontology_terms=["UBERON:0001157"],
    requested_outputs=[dna_client.OutputType.RNA_SEQ],
)
# outputs.reference.rna_seq vs outputs.alternate.rna_seq

Rate limits scale with demand; the docs position the API for analyses needing thousands of predictions, and explicitly not for million-variant batch jobs — that is precisely the workload the Atlas now covers precomputed.

Atlas vs AlphaFold DB: the same playbook

DeepMind says the comparison is intentional — Avsec’s team cites AlphaFold’s database as the inspiration. The telling detail from Nature’s coverage: AlphaFold’s database “has been accessed by millions of users” against AlphaGenome API’s 9,000. The Atlas is the attempt to close that access gap.

Limitations and caveats

• Not clinical. The disclaimer is explicit: predictions are research-use only, not validated or approved for clinical use, and not a substitute for professional medical advice. The DNM1 case needed experimental validation to count.
• Predictions, not truth. AVI is a model output trained on allele-frequency proxies. Martin Kircher (Max Delbrück Centre) told Nature it will not replace experiments or individual case assessment.
• Non-coding is the win area. On 5′UTR variants, GPN-Star-V still edges AVI out. The tool is strongest where the field was weakest.
• Complex traits remain complex. The Exeter result improved association detection by 22% — meaningful, but trait biology still involves development and environment beyond any sequence model’s scope.
• Reference genome bound. Precomputation covers hg38 SNVs and observed indels; rare structural variants and population-specific rearrangements are not in scope.

Use it if

Use the Atlas if you interpret variants — rare-disease pipelines, GWAS follow-up, clinical research triage — and want base-pair-resolution non-coding evidence without writing code or renting GPUs. The AVI score is now the sensible first-pass ranker for candidate variants, with the SHAP decomposition as the explanation layer.

Use the API instead if your variants of interest are structural, your intervals are custom, or you need predictions the precomputation does not cover — and accept the rate limits.

Skip it if you need clinically actionable conclusions today: no prediction database, this one included, replaces segregation analysis, functional assays, or clinical judgment. It is a triage and hypothesis-generation tool — a very good one.

FAQ

Sources

Sources checked September 8, 2026:

• @GoogleDeepMind launch post (announcement + explainer video)
• AlphaGenome Atlas: A predictive map of every possible DNA letter change — Google DeepMind blog
• Introducing AlphaGenome Atlas — Google blog (Kohli & Avsec)
• AlphaGenome Atlas preprint (Cheng et al., 2026, PDF)
• Nature news: DeepMind’s new genome ‘atlas’ (independent coverage)
• AlphaGenome Atlas portal
• AlphaGenome API — GitHub (quickstart code)
• AlphaGenome research repository (model code)
• AlphaGenome: AI for better understanding the genome — the 2025 base-model post
• Google Antigravity science use cases (Atlas skill)

Related reading: Karpathy’s LLM Wiki guide and Gemini 3.8 Flash guide.