AlphaGenome Atlas: Mapping Every Possible DNA Variant

Google DeepMind has released AlphaGenome Atlas, a free, searchable database that predicts the regulatory effects of all nine billion possible possible single-nucleotide variants across the human genome.

For researchers hunting disease-linked mutations in a near-infinite sea of genetic possibilities, the tool offers the first practical, whole-genome map of mutational impact, reaching far beyond the two per cent of coding DNA.

Image

What Does Google DeepMind’s AlphaGenome Atlas Actually Map?

The human genome contains around three billion DNA base pairs, most of which still remains poorly understood.

Scientists have mapped the two per cent that codes for proteins in reasonable detail, but the remaining 98%, the non-coding genome that regulates when and how genes switch on, has stayed largely opaque to researchers and drug developers alike.

DeepMind’s earlier AlphaGenome model already demonstrated that single-nucleotide variations in these non-coding regions can disrupt processes such as protein production and gene splicing.

AlphaGenome Atlas takes that capability to genome scale, pre-calculating the regulatory impact of all nine billion single-nucleotide variants and compressing the results into a one-petabyte dataset, more than 30 times larger than the AlphaFold Database that DeepMind released for protein structures in 2022.

That earlier AlphaFold database offers a useful comparison. When DeepMind expanded it, coverage grew from around 190,000 experimental protein structures to more than 200 million predictions, and it became one of the most widely used resources in structural biology within a few years.

AlphaGenome Atlas is designed to do the same for genetic variants, giving researchers a browsable starting point instead of a blank, three-billion-letter sequence.

The resource is available now through a free website portal that requires no coding experience, alongside an AlphaGenome API and dedicated Google Antigravity, DeepMind's agentic research platform.

Access is open for non-commercial academic research today, with commercial availability planned via Google Cloud.

The scale of AlphaGenome Atlas explains why this kind of resource has not existed before.

Testing nine billion individual variants in a laboratory, one experiment at a time, would take longer than any research programme could reasonably fund or complete.

By precomputing every prediction in advance and packaging them into a queryable atlas, DeepMind has effectively converted a problem of experimental capacity into a problem of computational storage and search, which is far easier to solve at scale.

View interview How Should Pharma Companies Audit AI Vendors?

How Should Pharma Companies Audit AI Vendors?

“This is one of the riskiest industries there is because you're dealing with human lives, you're dealing with experimental protocols, and you're dealing with regulatory bodies where you might not get another shot at that clinical trial."

Image

Inside the AlphaGenome Variant Impact Score

Scanning nine billion data points variant-by-variant is not realistic for any single research team, however well resourced.

To solve that problem, DeepMind introduced the AlphaGenome Variant Impact (AVI) score, a single figure that merges predictions from AlphaGenome, which covers non-coding regulatory effects, with AlphaMissense, DeepMind’s model for protein-altering variants, all into one ranking.

The AVI score works across both coding and non-coding DNA in a single pass, and DeepMind reports that it delivers best-in-class performance across several published variant pathogenicity and rare disease benchmarks.

Each score is also paired with AVI feature attributions, which decompose the result into the specific biological mechanism responsible, whether that is splicing disruption, chromatin accessibility, or evolutionary conservation.

Image

AlphaGenome Atlas is built from four linked resources that researchers can query together:

• Molecular effect predictions spanning hundreds of human and mouse cell types and tissues
• The AVI score, ranking every variant's likely impact on a single scale
• AVI feature attributions, linking each score back to its underlying biological mechanism
• A catalogue of more than 2,500 recurring DNA sequence motifs, described by DeepMind as the regulatory “words” of the genome.

Together, these layers let a researcher move from a single letter change to a ranked, mechanistically explained hypothesis in one query, rather than running a separate experiment for every candidate variant.

Early Evidence: From Rare Disease Diagnosis to Population Genetics

AlphaGenome Atlas has already supported real diagnostic and research findings through DeepMind’s academic collaborators, several of which point directly at problems pharmaceutical R&D teams face daily.

• Rare disease diagnosis. Working with the GREGoR Consortium, researchers at the Broad Institute used the AVI score to prioritise variants that earlier research had overlooked in unsolved rare disease cases. The finding flagged a variant in the DNM1 gene, predicting it created an incorrect splice site linked to epileptic encephalopathy, later confirmed through laboratory validation.
• Population-scale genetics. A Medical Research Council fellow at the University of Exeter applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping variants by their predicted molecular effect uncovered 22% more non-coding genetic associations than standard statistical methods found alone, including regulatory variants tied to two circulating proteins, PLA2G7 and EGLN1.
• Trait mapping. The same analysis narrowed hundreds of millions of non-coding variants down to the top one percent by predicted impact, pinpointing 19 genetic regions linked to body mass index for further investigation.
• Regulatory motif discovery. Separately, researchers at the Stowers Institute for Medical Research used the Atlas’s motif catalogue to categorise which transcription factors only open up DNA for access, and which ones also switch genes on or off directly.

Image

How Pharma R&D Can Best Utilise the AlphaGenome Atlas

The practical value of AlphaGenome Atlas sits squarely in the 98% of the genome that most drug discovery programmes still struggle to interpret.

Non-coding regulatory variants drive much of the biology behind complex, chronic, and rare diseases, yet validating them experimentally, one variant at a time, has never been commercially viable at genome scale.

A free, precomputed map changes the economics of that early triage work. Target identification teams can use AVI scores to shortlist promising non-coding candidates before committing wet-lab resources, while biomarker teams can lean on the same scores to strengthen patient stratification strategies in genomics-led clinical trials.

The feature attributions add a further layer of value here, showing whether a candidate variant disrupts splicing, chromatin accessibility, or another mechanism entirely, which shapes how a therapeutic hypothesis actually gets built.

The rare disease angle carries particular weight for pharma. Orphan drug programmes routinely stall because a single causal variant cannot be pinned down among thousands of candidates in a patient’s genome, and every month spent narrowing that list delays a diagnosis or a trial enrolment.

A prioritisation layer that returns a ranked, mechanistically explained shortlist in minutes rather than months could meaningfully shorten that search, even before any wet-lab work begins.

None of this replaces experimental validation, and DeepMind is explicit on that point. Indeed, AlphaGenome has not been validated for, and is not approved for, any clinical use.

The Atlas functions as a hypothesis-generation and prioritisation layer, not a diagnostic tool.

Decision-makers evaluating genomics partnerships, internal bioinformatics roadmaps, or vendor selection should treat it accordingly, while keeping an eye on the promised Google Cloud commercial release, which would extend access well beyond pure academic research.

View interview AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech

AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech

Kate O’Reilly, President & Chair of the Healthcare Businesswomen’s Association (HBA) Dublin-Ireland Chapter & Healthcare Transformation Partner at Roche, discusses AI in pharma, patient engagement, and the future of healthcare innovation.

Image

AlphaGenome Offers a Chance to Solve ‘Biological Mysteries’

AlphaGenome Atlas turns a previously impossible task, interpreting nine billion individual DNA changes one by one, into a searchable, computational starting point.

Early results from the Broad Institute and the University of Exeter already show it surfacing disease-linked variants that conventional analysis missed, and the AVI score gives non-specialist teams a fast, ranked way into a dataset that would otherwise be unusable.

The AlphaGenome Atlas allows faster, better-informed target and biomarker triage across the non-coding genome, provided every lead is still validated experimentally before it shapes a therapeutic programme.

Watching how quickly the commercial Google Cloud release matures will be a useful signal for how soon this kind of precomputed genomic intelligence becomes a standard part of the drug discovery stack.

At Pharmatica, we track the platforms, models, and datasets reshaping how pharmaceutical organisations discover and validate new therapeutic targets. Explore our HealthTech and AI Insights for more analysis on how genomic AI tools are moving from research labs into pharma pipelines, and see our related feature on AI-assisted surgical innovation for another example of AI reaching clinical practice.

Pharmatica: Insight. Connection. Impact.