Google DeepMind launches AlphaGenome Atlas, an AI map of 9 billion possible DNA variants

Google DeepMind has released AlphaGenome Atlas, an AI platform that predicts how every possible single-letter change in human DNA could affect the body at a molecular level. According to The Verge, the company unveiled the tool on Tuesday alongside a web portal that lets scientists explore the predictions at no cost for noncommercial research, with commercial access through Google Cloud promised soon.

A predictive map of 9 billion possible changes

Human DNA is written in four chemical letters, A, C, G and T, and the genome contains roughly 3 billion letter pairs carrying the instructions for how, when and where genes are switched on and off, and to what degree. Because each position can in principle switch to any of the other three letters, there are about 9 billion possible single-letter substitutions, The Verge notes. A central problem in genetics is working out which of those changes are harmless, which account for ordinary differences between people, and which play a role in disease.

Atlas contains a prediction for each of those variants, covering effects such as changing how much of a particular protein is produced. DeepMind's researchers describe the result as the most extensive catalogue to date of how genetic mutations influence molecular biology.

Coverage beyond protein-coding DNA

The project builds on AlphaGenome, the model DeepMind introduced last year to help identify genetic drivers of disease, and on the earlier AlphaMissense tool, which focused on predicting which small mutations might alter proteins. As The Verge reports, Atlas goes considerably further by extending predictions across the entire genome, including the vast majority of DNA that does not directly code for proteins but instead controls how genes behave.

A petabyte-scale precomputation

The underlying AlphaGenome model was trained on public databases of human and mouse genomes, allowing it to learn patterns connecting DNA changes to biological processes. Turning that model loose on billions of hypothetical variants was itself a substantial engineering effort. At a press briefing, DeepMind genomics lead Ziga Avsec acknowledged the model had already existed, but said the team needed considerable time to precompute and analyze so many variants because the space is so large. The resulting dataset is roughly 1 petabyte in size, Google says.

To help researchers sift through the billions of entries, Google is also releasing what it calls a Variant Impact Score (AVI), which draws on its other models for predicting the effects of DNA changes. The company says this lets scientists rapidly rank variants and interpret their molecular effects in the same workflow.

Access and the bigger picture

Scientists can query the catalogue through a web portal, as a skill inside Google's agentic development platform Antigravity, and through the AlphaGenome interface. Noncommercial access is open now via the website, while commercial availability on Google Cloud is described as coming soon.

The launch is the latest in a series of Google efforts to apply AI to core problems in science and medicine. It arrives as DeepMind cofounder Demis Hassabis steps back from running the AI lab to focus on research, including leading the drug-discovery spinoff Isomorphic Labs. Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold, the protein-structure prediction model, and the company has also built AI systems for weather forecasting, for finding new solutions in computing and mathematics, and an agentic co-scientist assistant for researchers.

Why it matters

Deciding which of billions of possible DNA changes actually matter is one of the slowest steps in genetics, and one that often requires laborious experimental work. If Atlas's predictions prove reliable, researchers gain a searchable shortlist of variants worth investigating, which could accelerate the search for the genetic roots of diseases and, ultimately, the development of treatments. The inclusion of non-coding regulatory regions matters too, since these make up most of the genome and have been difficult to interpret with earlier tools focused on protein-altering mutations.

There is an important caveat: the catalogue consists of model predictions, not laboratory measurements, so promising entries will still need experimental validation before they inform clinical work. Even so, giving the research community free access to a genome-wide, petabyte-scale reference, plus a scoring system for ranking variants, marks a significant milestone in the application of large-scale AI to biology, and a natural continuation of the AlphaFold line of work.