AlphaGenome Atlas Predicts Effects of Billions of DNA Changes

Google DeepMind has released AlphaGenome Atlas, a one-petabyte computational resource containing artificial-intelligence-generated predictions for the molecular effects of more than 9 billion possible single-letter changes across the human genome. The atlas contains thousands of molecular-effect predictions for each variant across hundreds of human cell types and tissues. It is designed to let researchers search predicted effects across the genome, rank variants and examine the biological processes they may disrupt.

The resource is available for non-commercial research through Google DeepMind’s website, where scientists can explore it through a web browser without writing code. A preprint describing the work is available on bioRxiv. The atlas is a computational research tool, not a collection of laboratory measurements: its predictions have not been validated or approved for clinical use.

The scale of the project addresses a central challenge in human genetics. The human genome contains approximately 3 billion DNA letters, creating more than 9 billion possible single-letter changes. Testing every possible change in a laboratory would be practically impossible.

AlphaGenome Atlas instead provides a searchable collection of predicted molecular effects intended to help researchers identify variants and regulatory features that warrant further investigation. Researchers can use the resource to examine predicted effects on gene expression, RNA splicing and protein function, as well as the activity of DNA sequences that help regulate genes. The development is a computational modeling effort rather than a clinical study, and the release does not establish that any individual prediction represents a confirmed biological effect in people.

Stowers Institute for Medical Research researchers contributed biological expertise, feedback on the model’s interpretation and practical testing of the resource. Institute Investigator Julia Zeitlinger, who also leads the Stowers AI Initiative, partnered with the Google DeepMind team led by Vice President of Science and Chief Scientist Žiga Avsec to connect predictions with biological processes that give cells their identities and allow them to function. Stowers Bioinformatics Scientist and Zeitlinger Lab member Melanie Weilert served as a lead author on the AlphaGenome project.

Researchers from the Broad Institute, the University of Exeter, Memorial Sloan Kettering Cancer Center and Stanford University also contributed scientific input and explored applications. Their work covered questions in gene regulation, rare disease, protein biology and population-scale genomics, providing examples of how a genome-scale prediction resource can be evaluated across distinct research settings. One major use of the atlas is to investigate how cells read short regulatory DNA sequences, or motifs.

These sequences help control which genes are active, when they are activated and how strongly they operate. Their effects depend on cellular context: every cell in the human body contains essentially the same DNA, yet different cells use that information in different ways. Zeitlinger and her team used AlphaGenome Atlas to analyze regulatory motifs across the genome and determine what they reveal about transcription factors, the proteins that control gene activity.

The researchers categorized regulatory signals by function, distinguishing transcription factors associated with changes in DNA accessibility from those that also activate or repress genes. By making genome-wide predictions available in one searchable resource, the team could look for general patterns across many cell types and begin defining rules that may be shared rather than specific to a single cellular environment. Stowers scientists also used the resource to examine recurring DNA motifs, including locations where transcription factors are predicted to bind.

Zeitlinger described the atlas as a searchable dictionary for non-coding DNA with motifs mapped at base-pair resolution across the genome and in many cell types. Conducting the equivalent analysis experimentally across thousands of sites and many cell types would require substantial time and resources, according to the research team. The computational resource therefore helped them identify broad patterns in gene regulation and formulate focused questions.

This interpretation illustrates why the collaboration combined biological expertise with AI development: predictions become most informative when they can be connected to recognizable molecular processes and tested against meaningful biological questions. The atlas does not replace those tests, but it can help determine which variants, regulatory elements and mechanisms should receive experimental attention first. AlphaGenome Atlas also supports the AlphaGenome Variant Impact score, abbreviated AVI.

The score combines predictions from AlphaGenome, the related AlphaMissense system and evolutionary conservation data. It provides a single measure for ranking variants by potential impact across protein-coding and non-coding regions. Researchers can then inspect which molecular processes are predicted to be affected, including gene expression, RNA splicing and protein function.

The score is intended as a prioritization tool for navigating a very large set of possible variants. The source material presents no clinical validation of AVI, and the release explicitly states that the atlas predictions are for research rather than clinical use. Collaborators examined several exploratory applications of the resource and its ranking system.

At the Broad Institute, scientists used the AVI score to prioritize a previously overlooked non-coding variant associated with an unsolved rare-disease case. The result shows how genome-wide predictions can direct attention to a regulatory variant outside protein-coding DNA. At the University of Exeter, researchers applied the atlas to genomic data from more than 54,000 UK Biobank participants and uncovered additional associations between rare non-coding variants and protein levels.

That population analysis demonstrates an association-based application of the resource; it does not establish that the AI model captures every biological mechanism contributing to protein-level variation. The project builds on earlier work in computational biology. In 2019, an international collaboration that included Avsec, Zeitlinger and her Stowers team developed BPNet, an AI framework now widely used to extract and interpret DNA sequences associated with genome-wide biological data.

More recently, the Zeitlinger laboratory unveiled PISA, an AI method that generates high-resolution visualizations of what models have learned from DNA. These earlier efforts supplied relevant experience in model interpretation and regulatory genomics. For AlphaGenome Atlas, Zeitlinger’s group applied its knowledge of gene regulation and computational biology to relate molecular-effect predictions to recognizable cellular processes, while Stowers scientists investigated applications involving impactful human genetic variation.

The release also reflects a broader, years-long effort: Google DeepMind developed the underlying technology, and the scientific collaborators helped guide the biological questions and potential uses addressed during development. The collection includes molecular-effect predictions, recurring motifs and the AVI ranking score, creating a single environment in which researchers can move from a broad search to a smaller set of candidate biological questions. The release is best understood as foundational computational infrastructure with potential applications across gene regulation, rare disease research and human genetics.

By placing billions of predicted variant effects and genome-wide regulatory patterns in one searchable resource, AlphaGenome Atlas could help researchers focus experiments and investigate biological mechanisms that are difficult to survey at the same scale through conventional methods. The participation of Stowers, Broad Institute, University of Exeter, Memorial Sloan Kettering Cancer Center and Stanford University provides several settings in which the resource has been explored, from functional genomics to analysis of UK Biobank data. As those links are tested, the atlas may help move the study of DNA variation from sequence-scale cataloguing toward focused investigation of the rules that control gene activity.

Subject of Research: AI-predicted molecular effects of human DNA variants and regulatory sequences

Article Title: Stowers Institute partners with Google DeepMind and leading research institutions to help reveal the regulatory language of the human genome

Article References: Stowers Institute partners with Google DeepMind and leading research institutions to help reveal the regulatory language of the human genome. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: AlphaGenome Atlas, computational genomics, DNA variants, gene regulation, regulatory motifs, transcription factors, rare disease, UK Biobank, protein levels, artificial intelligence, genome biology, Stowers