AlphaGenome Atlas maps every possible human DNA letter change

Overview

Updated 16 hours ago

Google DeepMind opened AlphaGenome Atlas to researchers on September 8. The searchable database predicts the molecular consequences of all 9 billion possible single-letter changes in the human genome — the result of testing all three alternative DNA bases at every position across the 3-billion-letter genome.

Each variant carries an AlphaGenome Variant Impact (AVI) score, a single number ranking its biological impact. The resource turns what previously required running AI models and writing code into a simple lookup. Early applications have already found a rare disease variant that earlier methods missed, and researchers found 22% more noncoding genetic associations in UK Biobank data than standard approaches.

Why it matters

Any researcher can now look up the predicted effect of any single-letter DNA change in seconds, shortening the search for disease-causing variants and linking noncoding DNA to common diseases.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Play on this story Voices Debate Predict Higher or Lower Connections

Key Indicators

9 billion

Single-nucleotide variants mapped

All three possible single-letter changes at every position in the human genome.

1 petabyte

Prediction data volume

Total size of precomputed molecular effect predictions stored in the atlas.

22%

Increase in noncoding genetic associations found

University of Exeter analysis of 54,000 UK Biobank genomes found more signals using atlas predictions.

9,000

Researchers using AlphaGenome API before atlas launch

DeepMind product manager Dhavi Hariharan reported the API user count ahead of the September release.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Play

Exploring all sides of a story is often best achieved with Play.

Most of these play right now — no account needed. Sign up to save scores, keep a streak, and unlock Debate and Predict. Log in Sign Up

Predict 3 ways this could play out. Back the one you believe — contrarian picks score more when a scenario has a resolution date. Log in to play

People Involved

Dhavi Hariharan

Product manager, Google DeepMind

Overseeing AlphaGenome Atlas release

Anne O'Donnell-Luria

Researcher, Broad Institute

Leading rare disease collaboration with GREGoR Consortium

Gareth Hawkes

Medical Research Council fellow, University of Exeter

Applied atlas to UK Biobank whole-genome data

Martin Kircher

Bioinformatician

Independent commentator

Julia Zeitlinger

Molecular biologist

Preprint co-author; spoke at the AlphaGenome Atlas press briefing

Organizations Involved

Google DeepMind

AI research laboratory

Developer and publisher of AlphaGenome Atlas

AI research lab that released the AlphaGenome model in 2025 and the AlphaGenome Atlas database in September 2026.

GREGoR Consortium

Rare disease research collaboration

Beta tester applying AVI scores to unsolved cases

Research consortium that applied AlphaGenome AVI scores to prioritize causal variants in unsolved rare disease cases.

Timeline

2025 September 2026

4 events Latest: 3 days ago

Tap a bar to jump to that date

September 2026
Atlas sparks broad research coverage
Latest Response

Nature, Scientific American, and Ars Technica covered the release; GREGoR and Exeter findings demonstrated early applications in rare disease and common trait genetics.
3 days ago
AlphaGenome Atlas launched
Product Launch

DeepMind unveiled the Atlas: predictions for all 9 billion single-nucleotide variants, a 1-petabyte searchable database with AVI impact scores.
4 days ago
2025
AlphaGenome model released
Model Release

DeepMind released AlphaGenome, an AI model analyzing noncoding DNA and predicting variant effects on gene regulation.
2025
API use grows among researchers
Adoption

Around 9,000 researchers accessed AlphaGenome predictions through the programming interface after the model's 2025 release, though it required writing code.
2025

Scenarios

AlphaGenome Atlas accelerates rare disease diagnosis

Likely Resolves by Sep 8, 2027

Discussed by: GREGoR Consortium, Broad Institute, University of Exeter researchers

The AVI score shortens variant prioritization from months to minutes. The DNM1 finding validated the approach end to end, from prediction to experimental confirmation. Widespread clinical adoption hinges on replication studies and integration into diagnostic pipelines at medical genetics centers.

Experimental validation exposes prediction gaps

Possible Resolves by Sep 8, 2027

Discussed by: Martin Kircher, Max Delbrück Centre for Molecular Medicine

Kircher cautioned that predictions won't replace laboratory experiments or case-specific clinical detail. Some AlphaGenome predictions, especially in regulatory regions, may not survive experimental testing. Even a modest failure rate could slow clinical adoption and shift researchers toward hybrid prediction-plus-validation workflows.

DeepMind expands atlas beyond single-letter variants

Possible Resolves by Sep 8, 2027

Discussed by: DeepMind, genomics research community

The current atlas covers single-nucleotide variants and 100 million short insertions or deletions. DeepMind could extend coverage to structural variants, additional populations beyond the reference genome, or other species. The company said the base model is available commercially via Google Cloud, which could fund further expansion.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

1990-2003

Human Genome Project (1990-2003)

An international consortium sequenced the human genome's 3 billion base pairs, completing the reference genome in 2003. The project cost roughly $3 billion and took 13 years.

Then

Researchers gained a reference genome but no systematic map of how variations affect function.

Now

Enabled a decade of genome-wide association studies linking variants to diseases, though most hits fell in noncoding DNA that researchers couldn't interpret.

Why this matters now

AlphaGenome Atlas extends the Human Genome Project's goal from reading the genome to interpreting its variations, covering the noncoding majority that GWAS could not explain.

2003-present

ENCODE project (2003-present)

The National Human Genome Research Institute launched ENCODE to catalog functional elements in the human genome — promoters, enhancers, and regulatory regions. Phase 3 mapped these across hundreds of cell types.

Then

Produced the first systematic functional annotation of the noncoding genome.

Now

Gave researchers tissue-specific regulatory maps but no direct predictions of how individual DNA variants perturb those elements.

Why this matters now

AlphaGenome Atlas builds on ENCODE-style annotations by predicting variant effects across hundreds of human and mouse cell types, adding the impact dimension ENCODE lacked.

2020-2021

AlphaFold protein structure prediction (2020-2021)

DeepMind's AlphaFold solved the protein folding problem, predicting 3D structures for hundreds of millions of proteins. CASP14 judges scored it near experimental accuracy in 2020.

Then

Structural biologists gained instant access to predicted protein shapes that previously required years of lab work.

Now

AlphaFold became a standard resource cited in tens of thousands of papers, establishing DeepMind's pattern of releasing large prediction databases to the research community.

Why this matters now

AlphaGenome Atlas follows the same playbook: a massive precomputed prediction resource, free for academics, built on a model DeepMind released earlier.