Google DeepMind’s AlphaGenome Atlas: A 1PB Map of 9 Billion DNA Variants

Google DeepMind’s AlphaGenome Atlas is a searchable database of predicted molecular effects for every possible single-nucleotide variant in the human genome—about 9 billion possible one-letter DNA substitutions. Released on September 8, 2026, it packages those predictions into a roughly 1-petabyte research resource for academic users. 67

What AlphaGenome Atlas is

A single-nucleotide variant is a change to one DNA letter. Such changes can occur in protein-coding sequences, which directly specify proteins, or in non-coding DNA that helps control when and where genes are active.

AlphaGenome Atlas aims to make the potential consequences of these changes easier to investigate. DeepMind used its AlphaGenome AI model to precompute predicted regulatory effects across all possible single-letter changes in the human genome, rather than requiring scientists to request or run a new prediction for each candidate variant. 67

That distinction matters: AlphaGenome is the predictive model; AlphaGenome Atlas is the large-scale, searchable reference built from its predictions. AlphaGenome was introduced as a model for predicting how DNA variants may affect biological processes involved in gene regulation. 21

What the 1-petabyte database contains

The Atlas contains predicted effects for approximately 9 billion possible single-nucleotide variants, along with an AlphaGenome Variant Impact score for each variant. 16

The resource is designed to help researchers move from a long list of observed DNA differences to a shorter list of candidates worth studying further. Rather than presenting a clinical verdict, it provides model-based signals about which changes may be more biologically consequential and what molecular mechanisms may be involved. 17

How researchers can access AlphaGenome Atlas

DeepMind says the Atlas is available for academic research through an intuitive, free-to-use website portal. The portal is intended to let researchers query the dataset without needing to run the model or write code. 7

This makes the Atlas especially useful as a first-pass research layer: a scientist can examine a candidate variant, review its predicted effects, and use that information to decide which experiments or genetic analyses should come next.

What the AVI score does

The AlphaGenome Variant Impact (AVI) score is a single score intended to help rank variants by predicted biological impact. Its key feature is that it combines information across both major categories of genomic variation:

• Protein-coding variants, which can alter a protein’s sequence or processing
• Non-coding variants, which can affect gene regulation and related molecular activity

By offering a common prioritisation framework, AVI is meant to make coding and non-coding candidates more directly comparable during genome-wide analysis. 16

The provided sources describe AVI as a practical ranking score but do not establish its full mathematical formulation. It should therefore be understood as a summary signal for prioritisation, not a self-explanatory measure of disease causality.

Early research findings reported by collaborators

DeepMind and reporting on the launch highlighted two early applications.

In rare-disease work, a team at the Broad Institute reportedly used AVI to surface a variant affecting DNM1 that creates an incorrect splice event. DNM1 has been associated with epileptic encephalopathy. 10 This is a potentially useful example of a model helping researchers identify a candidate that might otherwise be missed, but the available material does not provide enough detail to judge replication, clinical interpretation, or the case’s broader significance independently.

In population genetics, researchers analyzing whole genomes from more than 54,000 UK Biobank participants reportedly found 22% more non-coding associations using the Atlas. 310 More detected associations do not by themselves establish causality or clinical actionability; the supplied sources do not specify the traits, statistical thresholds, or downstream validation for this result.

How Atlas compares with the AlphaFold strategy

AlphaGenome Atlas follows a familiar DeepMind pattern: pair a predictive AI system with a widely accessible scientific database. The AlphaFold Protein Structure Database, developed with EMBL-EBI, made predicted protein structures broadly available to researchers; AlphaGenome Atlas applies a related access model to predicted consequences of DNA variants. 3134

The scientific targets are different:

• AlphaFold predicts three-dimensional protein structures.
• AlphaGenome Atlas predicts molecular and regulatory effects associated with single-letter DNA changes. 731

By data volume, the Atlas is reported to be more than 30 times larger than the AlphaFold Database. 215

The essential limitation: predictions are research leads

Atlas predictions are not experimental measurements, and an elevated AVI score does not demonstrate that a DNA variant causes a disease. The resource is best treated as a research baseline for prioritising variants, generating mechanistic hypotheses, and planning functional or genetic follow-up studies. 16

That caution is particularly important for non-coding DNA, where interpretation remains difficult. A useful workflow is to use the Atlas to narrow a large candidate set, then evaluate the leading variants with independent genetic evidence, functional studies, and appropriate clinical assessment where relevant.

AlphaGenome Atlas’s real contribution is scale and accessibility: it gives researchers a ready-to-query map of AI-generated hypotheses across the full set of possible single-letter human DNA changes. Its value will depend on how effectively those hypotheses are tested in the lab and validated in human studies.