What happened?
AlphaGenome Atlas is a database of predicted effects for every possible single letter change in the human genome, about 9 billion changes in total Google blog ↗. Google DeepMind released it on 8 September 2026 through a web portal that needs no coding skills Google blog ↗.
The human genome has about 3 billion DNA letters, and each one could be swapped for one of three others SiliconANGLE ↗. DeepMind used its AlphaGenome model, which predicts how DNA changes affect gene control, to calculate the likely effect of every one of these swaps ahead of time DeepMind blog ↗. The result is about one petabyte of data, which DeepMind says is more than 30 times the size of the AlphaFold Database DeepMind blog ↗.
Before the Atlas, researchers had to send their DNA changes to the AlphaGenome API, which needs programming SiliconANGLE ↗. DeepMind also added the AlphaGenome Variant Impact (AVI) score, a single number that combines AlphaGenome with AlphaMissense, its model for changes inside protein recipes, to help rank which variants to look at first DeepMind blog ↗. According to a Stowers Institute release, the full description of the work is a preprint, meaning it has not yet been peer reviewed News-Medical ↗.
What is inside?
Variant predictions
Thousands of predicted molecular effects for each change, across hundreds of human and mouse cell types and tissues DeepMind blog ↗.
The AVI score
One ranking number covering both protein coding DNA and the much larger non coding part DeepMind blog ↗.
A motif collection
Over 2,500 recurring short DNA patterns and where they appear, which proteins may bind to DeepMind blog ↗.
A researcher can now look up a DNA change in seconds instead of running the model themselves, which could make this kind of analysis available to many more labs.
Leapscope interpretation of the reported result.How did AI help?
Every entry in the Atlas is a model prediction, not a lab measurement News-Medical ↗. DeepMind researchers ran the model at huge scale and built the scoring and portal, while outside collaborators tried it on real problems DeepMind blog ↗. At the Broad Institute, the AVI score helped flag a change in a gene called DNM1, linked to a severe form of epilepsy, in an unsolved rare disease case, and lab screens then confirmed the predicted effect DeepMind blog ↗. At the University of Exeter, applying it to over 54,000 UK Biobank participants uncovered 22% more non coding genetic links DeepMind blog ↗.
All figures as reported by Google DeepMind Google blog ↗ DeepMind blog ↗.
The limits matter. DeepMind says AlphaGenome has not been validated or approved for any clinical use DeepMind blog ↗. Its genomics lead, Žiga Avsec, said the predictions point in the right direction but should not be treated as the universal truth, and that the model does well on some DNA switches but poorly on others called enhancers SiliconANGLE ↗. Claims of top benchmark performance come from DeepMind, with no figures given in the announcement DeepMind blog ↗.
Which fields could this affect?
Some uses are relevant now and others are possible future value; the connections below are our assessment.
Genetics research
Researchers can look up predicted effects without running the model or writing code Google blog ↗. One Exeter researcher described it as a way to shrink the haystack SiliconANGLE ↗.
Explore scienceRare disease research
The Broad Institute example shows the score helping pick a variant that lab tests then confirmed DeepMind blog ↗. This was one case, not a large trial.
Explore healthcareClinical genetic testing
Ranked predictions could one day support doctors reviewing a patient's DNA. DeepMind says the model is not approved for any clinical use today DeepMind blog ↗.
Explore healthcarePersonal health predictions
Reading one person's genome to predict their health is a different and harder task. This release demonstrates no such use.
What has been checked?
The evidence is a developer announcement and preprint, early results from named collaborators, and news coverage, built on a model with a peer reviewed Nature paper. Leapscope reviewed these sources; we did not repeat the experiments.
Shown so far
- The portal and dataset are publicly available for non commercial research DeepMind blog ↗ SiliconANGLE ↗.
- Lab screens confirmed the predicted effect of one DNM1 variant found with the AVI score DeepMind blog ↗.
- An analysis of over 54,000 UK Biobank participants found 22% more non coding links DeepMind blog ↗.
Still unknown
- How accurate the predictions are across all 9 billion changes, since only a few have been checked in the lab.
- How well it handles enhancers and indirect effects, which the team says are weaker areas SiliconANGLE ↗.
- What peer review of the Atlas preprint will conclude News-Medical ↗.
Evidence status: Research resource. Stage: Usable. The database of predictions is published for anyone to explore.
From a map to real answers
This is our suggested way to follow it, not a promised timetable.
Can I use it today?
Researchers can explore the Atlas for free on DeepMind's web portal for non commercial use, and DeepMind says commercial access through Google Cloud is coming DeepMind blog ↗ SiliconANGLE ↗. It is not a medical service and should not be used to interpret your own DNA.
A few things you might be wondering
Has someone tested all 9 billion DNA changes in a lab?
No. These are computer predictions News-Medical ↗. Only a small number have been checked by experiments so far DeepMind blog ↗.
Can I upload my DNA and find out my risks?
No. The Atlas is a research tool, and DeepMind says AlphaGenome is not validated or approved for any clinical use DeepMind blog ↗.
How is this different from AlphaGenome?
AlphaGenome is the model. The Atlas is its results, worked out ahead of time for every possible single letter change and put in a searchable portal Google blog ↗ DeepMind blog ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01The launch announcement describing the database, its scale and early collaborator results.
Technical blog with details on contents, the AVI score, access routes and the clinical use disclaimer.
News report with access details, researcher quotes and limits described by DeepMind's genomics lead.
Collaborator release explaining the scale, that results are predictions, and that the work is a preprint.