Google DeepMind has built an enormous new map of the human genome that attempts to predict what could happen when any single DNA letter changes.

Called AlphaGenome Atlas, the new resource contains predictions for all roughly 9 billion possible single-nucleotide variants in the human genome. The resulting dataset is about 1 petabyte in size, giving researchers a large-scale way to explore how genetic changes could affect molecular processes across the body.

The project builds on AlphaGenome, DeepMind’s AI model introduced in 2025 to predict how DNA variants can affect processes involved in gene regulation. Rather than asking researchers to run the model individually on specific variants, however, AlphaGenome Atlas precomputes predictions across the entire human genome and makes those results available through an interactive research platform.

That distinction is important. There are billions of possible single-letter changes in human DNA, making it impractical to experimentally test every possibility. AlphaGenome Atlas is designed to give researchers a starting point for narrowing that enormous search space.

A map of the genome’s 9 billion possible changes

The human genome contains roughly 3 billion DNA base pairs, but understanding what those sequences actually do remains one of biology’s major challenges.

Only around 2% of the genome directly codes for proteins. The remaining 98% is largely non-coding DNA involved in regulating when, where and how genes are activated. Those regions contain many genetic variants associated with traits and diseases, but their effects can be difficult to interpret.

AlphaGenome was designed to tackle that problem by examining long stretches of DNA and predicting thousands of molecular properties, including gene expression, RNA splicing and other aspects of gene regulation. The model can then compare an unchanged DNA sequence with one containing a particular variant to estimate how that change could alter biological activity.

AlphaGenome Atlas takes that capability and scales it up dramatically.

DeepMind says it has used AlphaGenome to calculate the molecular effects of every possible single-letter change in the human genome — approximately 9 billion variants. Those predictions make up the new 1-petabyte Atlas dataset. Google says that makes it more than 30 times larger than the AlphaFold Database.

The goal isn’t to claim that every prediction is a confirmed biological fact. Instead, the Atlas gives researchers computational predictions that can help identify which variants deserve closer investigation and laboratory testing.

The AlphaGenome Variant Impact score

One of the main ways researchers can navigate the Atlas is through a new metric called the AlphaGenome Variant Impact, or AVI, score.

Rather than requiring researchers to examine thousands of individual molecular predictions for every variant, the AVI score condenses information into a single value indicating the predicted impact of a genetic variant. It combines AlphaGenome’s predictions with AlphaMissense, Google’s AI model for assessing the effects of protein-altering variants.

The score can be used across both coding and non-coding regions of the genome. That is particularly significant because most of the genome is non-coding, where understanding the functional consequences of genetic variation has historically been more difficult.

Researchers can also inspect the factors contributing to an AVI score. DeepMind says AlphaGenome Atlas provides feature attributions showing which predicted biological processes are driving the result, including effects related to gene expression, RNA splicing and chromatin accessibility.

The Atlas also includes a collection of more than 2,500 recurrent DNA sequence motifs and their locations, giving researchers another way to connect predicted variant effects with the underlying regulatory sequences of the genome.

Researchers are already using it to investigate disease

DeepMind says AlphaGenome Atlas has already been used in research involving rare diseases and complex human traits.

In one collaboration involving the Broad Institute and the GREGoR Consortium, researchers used the AVI score to prioritize genetic variants in an unsolved rare-disease investigation. The system highlighted a variant in the DNM1 gene that was predicted to create an incorrect splice site, ultimately providing a lead that was experimentally validated.

Another project led by researchers at the University of Exeter used AlphaGenome Atlas alongside data from more than 54,000 UK Biobank participants. The researchers were investigating rare non-coding variants associated with protein levels and other traits. According to Google, using AlphaGenome’s predictions helped narrow large groups of candidate variants and identify additional genetic associations.

These examples illustrate how DeepMind expects the Atlas to be used: not as a replacement for laboratory research, but as a way to identify promising candidates before researchers invest significant time and resources in experimental validation.

A 1-petabyte dataset that researchers can actually explore

The size of AlphaGenome Atlas is impressive on its own, but DeepMind is also putting considerable emphasis on accessibility.

The Atlas is available through a web-based interface designed for researchers who don’t necessarily need to write code to explore the data. Users can investigate variants, examine predicted molecular effects and move between the variant information and relevant genomic sequences.

DeepMind is also making AlphaGenome Atlas available through the AlphaGenome API, while an AlphaGenome Atlas skill can connect the dataset with Google’s Antigravity scientific workbench. That integration is intended to let researchers automate tasks such as ranking variants, explaining predicted biological effects and generating visualizations.

Google says the Atlas is available for non-commercial research through its website, with commercial access planned through Google Cloud. The underlying AlphaGenome model is already available for academic use through GitHub and the AlphaGenome API, as well as commercially through Google Cloud’s Model Garden.

The approach is reminiscent of what made the AlphaFold Database so useful to researchers. Instead of making scientists repeatedly perform expensive computational work themselves, DeepMind is taking a huge amount of computation and packaging the resulting information into a resource that can be searched and explored.

It’s a research tool, not a diagnostic system

There is an important limitation to keep in mind.

AlphaGenome Atlas provides AI-generated predictions about the molecular effects of genetic variants. Those predictions can help researchers decide which variants to investigate, but they don’t establish that a particular variant causes a disease in a person.

DeepMind explicitly says AlphaGenome has not been validated for clinical use and is not approved for clinical use. The information provided by the Atlas is therefore intended as a research resource rather than a substitute for medical diagnosis or professional medical advice.

That distinction will remain important as increasingly capable AI systems move deeper into genomics. Predicting what a DNA change might do is one problem; proving what it actually does in living cells, tissues and people is another.

For now, AlphaGenome Atlas represents a major expansion of the first part of that process. By precomputing predictions for roughly 9 billion possible single-letter DNA changes, Google DeepMind has effectively created a searchable computational map of possible genetic variation.

And as AlphaGenome itself improves, DeepMind says the Atlas can be updated with increasingly comprehensive and precise predictions. The company’s longer-term ambition is to make these resources part of broader AI-assisted scientific workflows, where models can help researchers move from genomic data to hypotheses and eventually to experiments.