DeepMind Put Nine Billion DNA What-Ifs Into a Searchable Atlas

Choose a position in the human reference genome. If the DNA letter there is A, it could be replaced by C, G or T. A position containing C also has three alternatives.

Now repeat that exercise across roughly three billion positions.

The result is more than nine billion possible single-nucleotide substitutions relative to the GRCh38 human reference genome. Google DeepMind has used its AlphaGenome model to calculate predicted molecular effects for essentially that entire set and placed the results in a searchable resource called AlphaGenome Atlas.

DeepMind released the Atlas on September 8, 2026, describing the complete resource as approximately one petabyte, more than 30 times the size of the AlphaFold Database.

That can sound as if somebody has experimentally decoded every possible mutation in human DNA. Nobody has. What DeepMind has built is subtler: a catalogue of more than nine billion computational hypotheses.

A lookup table for DNA counterfactuals

Sequencing reveals where a person’s DNA differs from a reference. It does not automatically explain which differences matter.

That interpretation problem becomes especially difficult outside protein-coding sequences. Roughly 2% of the genome directly encodes proteins. The rest includes introns, regulatory elements, repetitive sequences and large regions whose functions remain uncertain.

A change inside a protein-coding sequence can sometimes provide a recognizable clue: it changes an amino acid or introduces a premature stop. A non-coding change may instead weaken a transcription-factor binding site, alter when a gene is activated or cause RNA to be spliced incorrectly. Its effect might appear only in a particular tissue or developmental stage.

AlphaGenome was designed to predict those kinds of molecular consequences. According to the model’s peer-reviewed paper in Nature, it accepts DNA sequences as long as one million base pairs and predicts thousands of genomic signals. These include gene expression, transcription initiation, chromatin accessibility, transcription-factor binding, histone modifications, chromatin contacts and several aspects of RNA splicing.

For a variant, the model can make predictions from the reference sequence, repeat the calculation after inserting the alternative letter and compare what changes.

The underlying AlphaGenome model was peer reviewed. The newer Atlas manuscript, including its AlphaGenome Variant Impact score and application studies, was still in review when the Atlas launched.

The Atlas changes how researchers reach those predictions. Instead of selecting each variant and running the model themselves, they can retrieve effects that DeepMind has already calculated. The resource therefore shifts part of the work from computation to search.

That may sound like a small distinction. At the scale of a genome, it is the point.

One score, followed by thousands of details

A model capable of producing thousands of outputs per variant creates another problem: where should a researcher begin?

The Atlas introduces the AlphaGenome Variant Impact score, or AVI, as a first-pass ranking tool. The Atlas manuscript says AVI combines AlphaGenome’s regulatory predictions with other information, including AlphaMissense protein-effect predictions, evolutionary conservation and protein-disruption annotations.

A high score does not mean that a variant has a known probability of causing disease. AVI was not trained as a universal clinical pathogenicity classifier. Its training uses population rarity as a proxy for evolutionary constraint, and rarity is not the same thing as harm.

To make the ranking more interpretable, the Atlas supplies feature attributions showing which inputs pushed a score higher or lower. A researcher might learn that a variant ranked highly because of a predicted splicing disruption, an expression change or conservation across species.

This combination, ranking plus a proposed molecular mechanism, is more useful than a leaderboard alone. The score suggests where to look; the underlying predictions suggest what experiment to perform.

The variant inside DNM1

The Atlas manuscript describes an early example involving the GREGoR rare-disease consortium.

In an unresolved case of epileptic encephalopathy, AVI prioritized a deep-intronic variant in DNM1. Introns are copied into preliminary RNA but normally removed when the cell assembles the finished message used to make a protein.

AlphaGenome predicted that the variant would create a cryptic splice site: a convincing but incorrect signal telling the cell where to cut the RNA. The proposed result was an aberrant splice product affecting a neurologically important form of DNM1.

The collaborators did not treat that output as a diagnosis. They built a minigene assay to test the proposed mechanism. According to the preprint, the experiment examined 796 substitutions across 265 nucleotides in five cell lines and supported the predicted splicing effect of the candidate variant.

The study authors then combined the functional result with genetic, clinical and published evidence and recommended classifying the variant as likely pathogenic.

That sequence matters. The AI ranked a candidate and proposed a mechanism. An experiment tested the mechanism. A broader body of evidence informed classification.

It is a strong example of what the Atlas could help researchers do. It is not evidence that every high AVI score identifies a disease-causing variant, or that the system has demonstrated prospective diagnostic accuracy across unresolved cases.

More signals, and a warning about what that means

The Atlas manuscript also reports a population-scale test led by researchers at the University of Exeter. The team analyzed 2,028 circulating proteins in as many as 54,189 UK Biobank participants, looking for collective effects from rare non-coding variants.

Rare variants pose a statistical problem: few people carry any particular one. Researchers can group variants together, but a group filled with irrelevant changes may bury a real signal. The Exeter analysis used Atlas predictions to create narrower groups enriched for variants that the model considered consequential.

The preprint reports that this approach produced 22% more conditionally independent rare non-coding aggregate associations than the comparison analyses. That is an increase in statistical discoveries under the study’s method, not a 22% increase in confirmed causal variants.

The manuscript highlights a region near PLA2G7. An unfiltered set containing 526 rare variants produced a minimum P = 1.10 × 10^-8 with the ACAT-O test. Filtering for the top 1% of AVI-ranked variants reduced the set to four; that set produced P = 1.74 × 10^-13 in a burden test.

The smaller P value is striking, but it is not a clean head-to-head comparison: both the set of variants and the statistical test changed. The result shows how prediction-guided filtering behaved in this analysis, not that the four variants have been proven causal.

The preprint’s replication results add another reason for restraint. Of 31 additional associations, 25 could be evaluated in the independent All of Us cohort. Four were nominally significant, and none survived Bonferroni correction for multiple testing.

Rare-variant signals can be difficult to replicate across cohorts, so that outcome does not establish that the method fails. It does mean the reported 22% increase should be treated as promising preprint evidence rather than a settled measure of general performance.

What nine billion leaves out

The Atlas’s scale is real, but its boundary is easy to miss.

It covers modeled single-nucleotide substitutions relative to GRCh38. It does not enumerate every insertion, deletion, duplication, large rearrangement or combination of variants. The project also analyzed more than 100 million observed short insertions and deletions, but those are not part of the exhaustive nine-billion figure.

Nor is a reference genome the genome of an individual person. A real variant exists alongside millions of other inherited differences. Those neighboring variants can alter molecular effects. Age, environment, development and cell state add further context that a reference-sequence model cannot fully represent.

AlphaGenome’s one-million-base window is large for a sequence model, but some regulatory relationships span greater distances. Prediction quality also depends on the cell types, tissues and experimental measurements represented in its training data.

DeepMind’s official API documentation is explicit: the predictions are for theoretical modeling and research, and they must not be used for clinical decision-making. DeepMind also says AlphaGenome has not been validated or approved for clinical use.

That limitation matches how clinical genetics already handles computational tools. ClinGen’s variant-classification guidance places predictive evidence within a larger framework that can include population data, family segregation, patient phenotype, functional experiments and established gene-disease relationships.

The Atlas does not replace those forms of evidence.

What it may replace is some of the wandering.

A researcher facing a large set of plausible variants can rank them, inspect the molecular disturbance predicted for each and choose a test suited to that disturbance. A splicing prediction can lead to a splicing assay. A predicted regulatory effect in a particular cell type can suggest where to investigate next.

DeepMind has not experimentally tested nine billion DNA changes. It has made more than nine billion model-generated predictions available for other scientists to question, compare and test.

That is less magical than decoding every possible genetic future. It may be more useful: a map whose purpose is not to announce the answer, but to show researchers where to start digging.