Google DeepMind released AlphaGenome Atlas today, a database containing predictions for the effects of 9 billion single-nucleotide variants across the entire human genome. The resource is freely available for academic research and represents the most comprehensive map of how genetic mutations affect molecular biology ever assembled.
The scale is difficult to overstate. The human genome contains about 3 billion base pairs, and at each position, three different letter substitutions are possible. DeepMind precomputed predictions for all of them, generating a 1-petabyte dataset more than 30 times larger than the AlphaFold protein structure database the company released in 2022.
Why the Non-Coding Genome Matters
Scientists understand roughly 2% of the human genome well. That's the fraction that codes directly for proteins. The remaining 98% regulates gene activity, controlling when genes switch on, how much protein they produce, and in which tissues they operate. Most disease-associated genetic variants identified in population studies reside in these regulatory regions, yet interpreting what any given mutation actually does has remained slow and uncertain.
AlphaGenome, the AI model underlying the Atlas, analyzes DNA sequences up to one million base pairs long and predicts how changes might affect gene expression, RNA splicing, chromatin accessibility, and other molecular processes across hundreds of human cell types. The Atlas removes the computational barrier that previously required researchers to run the model one variant at a time. Now, they can simply look up the answers.
A Clinical Test Case
Researchers at the Broad Institute, working with the GREGoR Consortium on unsolved rare diseases, have already applied the resource. Laura Covill and Anne O'Donnell-Luria used the Atlas's new AlphaGenome Variant Impact (AVI) score to sift through candidate mutations and flagged a deep intronic variant in the DNM1 gene, which is strongly linked to epileptic encephalopathy. The underlying predictions showed the variant created a brain-specific cryptic splice site that produced an abnormally long protein. Laboratory experiments confirmed the prediction, and the variant was reclassified as likely pathogenic.
At the University of Exeter, Gareth Hawkes applied the Atlas to data from more than 54,000 UK Biobank participants. By grouping variants based on predicted molecular effects, he identified 22% more non-coding genetic associations. Focusing on the top 1% of impactful variants, his team found 19 genetic regions linked to body mass index.
Where CRISPR and Programmable DNA Fit In
Gene therapy and CRISPR-based treatments have advanced rapidly in recent years. The FDA has approved gene therapies for sickle cell disease, certain inherited blindness conditions, and other monogenic disorders. But the challenge has always been knowing which mutations to target and predicting what happens when you change them.
A resource like AlphaGenome Atlas addresses part of that problem by providing computational predictions that can guide experimental validation. Researchers designing guide RNAs for CRISPR interventions can now consult the Atlas to understand the predicted downstream effects of their edits before committing to wet lab work.
This matters especially for tools like CRISPRware, software developed at UC Santa Cruz to design guide RNAs for precision genome editing. CRISPRware enables researchers to target any region of the genome, including the poorly annotated non-coding portions. Paired with AlphaGenome Atlas, these tools could streamline the development of therapeutic interventions for genetic conditions including certain metabolic and blood disorders.
The broader vision is sometimes called programmable DNA: the ability to design genetic sequences with predictable regulatory properties. For synthetic promoters, enhancers, and therapeutic antisense oligonucleotides, having a reliable map of variant effects transforms what was previously trial-and-error into something closer to rational design.
Limitations Worth Noting
DeepMind is clear about what AlphaGenome Atlas is not. The company states explicitly that the resource has not been validated or approved for clinical use. The technical paper notes that Atlas predictions can serve only as part of the evidence chain leading to clinical diagnoses. They are not sufficient evidence on their own.
The Atlas also covers only single-letter substitutions. Many diseases involve insertions, deletions, or structural rearrangements that fall outside the current model's scope. There are gaps in training data, including missing cell types and certain RNA categories. DeepMind describes this release as a baseline rather than an endpoint.
Still, the practical utility is immediate. Around 9,000 researchers have already accessed AlphaGenome predictions through DeepMind's API since the model's release last year. Now, those predictions are packaged into a searchable platform requiring no coding skills. As Martin Kircher, a bioinformatician at the Max Delbrück Centre, told Nature: "This is a useful and generous way to scale up access to a strong model."
The Atlas is available through a web portal, an API, and as a skill within Google's Antigravity agentic platform. Commercial access through Google Cloud is coming soon.
