Google DeepMind's New Atlas Maps 9 Billion Human DNA Variants for Disease Research

News Summary

Google DeepMind on Tuesday, September 8, 2026 (Eastern Time), released the AlphaGenome Atlas, a free online resource that provides precomputed predictions of the molecular effects of every one of the roughly 9 billion possible single-letter DNA changes across the human genome. Built on top of AlphaGenome, DeepMind's sequence-to-function AI model, the Atlas is designed to let scientists instantly look up how a given genetic variant might disrupt gene regulation, rather than running new computations themselves — a shift researchers say could meaningfully speed up the search for the genetic roots of disease.

What the Atlas contains

The Atlas is built from roughly 1 petabyte of predicted data, which DeepMind says is more than 30 times the size of the AlphaFold Protein Structure Database the company expanded in 2022. For each of the 9 billion possible variants, the resource offers thousands of molecular effect predictions spanning multiple aspects of gene regulation across hundreds of human and mouse cell types and tissues. It also includes more than 2,500 identified DNA sequence motifs and direct links between variants and the functional sequences they affect.

A central feature is the AlphaGenome Variant Impact (AVI) score, a single unified number that combines outputs from AlphaGenome with those of AlphaMissense, DeepMind's earlier model for predicting the effects of protein-coding mutations. The AVI score lets researchers quickly rank and prioritize variants across both the roughly 2 percent of the genome that codes for proteins and the 98 percent that does not — the "noncoding" regions where most disease- and trait-associated variants are now known to sit, but which have historically been far harder to interpret.

How it works

AlphaGenome, the underlying model, was first introduced by DeepMind last year as a sequence-to-function system: it reads a stretch of DNA — DeepMind says up to roughly 1 million base pairs surrounding a variant — and predicts how a single-letter change would alter gene regulation, such as by creating an incorrect splice site or disrupting a regulatory element. Rather than requiring researchers to query the model one variant at a time, the Atlas precomputes and stores these predictions for every possible substitution in the genome, turning what was previously a heavy computational task into a simple lookup.

Pushmeet Kohli, DeepMind's vice president of science, said the release means "any researcher in the world can access a comprehensive map of the human genome and its variations by simply opening a browser." Žiga Avsec, the genomics lead on the project, described the Atlas as "a starting point" that is especially useful for research involving multiple variants at once, with the AVI score helping scientists "prioritize variants and try to find that needle in the haystack."

Early scientific applications

DeepMind and outside groups have already tested the Atlas on real research questions. Researchers at the Broad Institute used AVI scores to help identify a variant in the DNM1 gene linked to a severe form of epileptic encephalopathy; the Atlas indicated the variant created an incorrect splice site, a prediction the team says was later supported by lab experiments. At the University of Exeter, an analysis of roughly 54,000 UK Biobank participants using the new scores identified 22 percent more noncoding genetic associations with protein levels than previous methods, including variants tied to aging (near the gene PLA2G7) and to cellular oxygen sensing (EGLN1), as well as 19 genetic regions potentially linked to body mass index. At the Stowers Institute, scientists used the resource to sort transcription factors — proteins that switch genes on and off — by their specific regulatory roles.

Jonathan Sebat, a psychiatric geneticist at UC San Diego, said the tool could streamline everyday lab work: "Our own workflows in the lab can be streamlined quite a bit because we don't actually have to compute anything. We literally can just look up everything."

Independent expert reaction

Outside scientists convened by the UK's Science Media Centre offered a generally positive but measured assessment. Kristian Helin, chief executive of the Institute of Cancer Research in London, called AlphaGenome "a major advance in computational genomics" with long-term significance he compared to AlphaFold's impact on structural biology. Robert Goldstone, head of genomics at the Francis Crick Institute, said the model marks a shift in noncoding DNA prediction "from theoretical to practical utility," particularly for splice-site prediction, but cautioned that "AlphaGenome is not a magic bullet for all biological questions."

Other researchers were more cautious. Xianghua Li, a lecturer in medical and molecular genetics at King's College London, said that "when we look at each prediction, this AI performs as well as the best existing tools, but not better," and stressed the predictions are "not yet ready for use in clinics." Ben Lehner of the Wellcome Sanger Institute, who said his team ran more than half a million experiments to check the model's performance, agreed it was strong but noted "AlphaGenome is far from perfect and there is still a lot of work to do," pointing to data quality and standardization as ongoing challenges. Aldo Faisal of Imperial College London recommended treating the results as promising pending independent reproduction.

Limitations DeepMind acknowledges

DeepMind is explicit that the Atlas is a research tool, not a diagnostic one. Its documentation states that "Atlas and AVI are research tools that predict molecular effects and can serve only as part of the evidence chain leading to clinical diagnoses," and that AlphaGenome "has not been validated for, and is not approved for, any clinical use." The company also acknowledges gaps in the underlying model, including missing cell types in its training data and a lack of coverage for non-polyadenylated RNA, describing the current release as "a baseline that will grow more precise as the underlying models improve." Because the model's field of view is limited to roughly 1 million base pairs around a variant, it may also miss the effects of distant regulatory elements — some of which are known to influence genes from more than 100,000 DNA letters away.

Access and availability

The AlphaGenome Atlas website opened on September 8, 2026 for free, non-commercial academic research use, alongside the AlphaGenome API on GitHub and a new skill for Google Antigravity, DeepMind's agentic development platform. DeepMind says commercial access through Google Cloud is coming soon, and a static download of AVI scores is being released under a license permissive enough for both commercial and non-commercial use. The dataset builds on existing public resources, incorporating variant information from gnomAD, the UK Biobank and the All of Us research program rather than replacing them.

Why it matters

The Atlas extends a pattern DeepMind established with AlphaFold, whose underlying method won the 2024 Nobel Prize in Chemistry: take a previously slow, expertise-heavy scientific task and turn it into a searchable public database. Where AlphaFold mapped the shapes of more than 200 million proteins, the Atlas targets the much larger and less understood noncoding genome, where scientists believe many of the genetic contributors to complex diseases — from cancer to inherited disorders such as Tay-Sachs disease and sickle cell anemia — remain to be found. Researchers caution the tool will not replace laboratory experiments or clinical judgment, but many see it as a way to narrow enormous search spaces down to the variants most worth investigating first.