AlphaGenome Atlas puts 9 billion predicted variant effects in a database

Google DeepMind has released AlphaGenome Atlas, a precomputed catalogue of predicted molecular effects for every possible single-nucleotide substitution in the human reference genome. By running AlphaGenome across the genome in advance, DeepMind has converted a sequence-to-function model into a searchable resource that can rank variants without requiring researchers to provision inference hardware.

Nine billion substitutions, one petabyte

The roughly 9 billion variants represent the three alternative DNA letters that could occur at positions across the approximately 3-billion-base human reference genome. Atlas links each substitution to an average of about 27,000 predictions covering processes such as gene expression, splicing, transcription and chromatin accessibility.

AVI compresses thousands of outputs into a ranking

The AlphaGenome Variant Impact score, or AVI, summarizes the predicted effect of each variant in one number. It combines AlphaGenome’s regulatory predictions across coding and noncoding sequence with AlphaMissense predictions for protein-altering variants.

Feature attributions accompany the score and identify which predicted processes contributed most, such as splicing, gene expression or chromatin accessibility. AVI provides a triage signal for sorting large variant sets; the underlying attributions and cell-type predictions supply hypotheses for experimental follow-up.

Precomputation removes the GPU step

About 98% of the human genome lies outside protein-coding regions. Variants in these regions are difficult to interpret because they can alter gene regulation, often in tissue-specific ways, without directly changing an amino-acid sequence. Atlas places coding and noncoding predictions in the same ranking workflow and returns stored results without running AlphaGenome for each query.

Three routes into the Atlas

DeepMind provides free web access for noncommercial use and plans commercial availability through Google Cloud. Teams preparing production integrations will need to confirm the required genome build, API quotas, bulk-access options, update cadence and score-versioning policy.

An intronic DNM1 variant rises to the top

Researchers at the Broad Institute used Atlas to rank variants from 814 families in the GREGoR Consortium rare-disease cohort. In one child’s unresolved case, the highest-ranked candidate was a single-nucleotide change in intron 10 of DNM1, a region removed from RNA before the cell produces a protein.

The model predicted that the change creates a cryptic splice site, giving the cell’s splicing machinery a new joining point. Use of that site would insert sequence encoding 13 amino acids into a transcript segment absent from the standard reference annotation. Atlas also predicted low expression of the affected segment in blood, which helps explain why a conventional blood-based analysis could miss it.

Experimental assays across five cell lines found that 12 tested variants behaved as predicted. Evidence from the DNM1 analysis supported reclassification of the child’s variant and gave the clinician additional information for the case review.

Effect-based grouping finds more UK Biobank signals

Dr. Gareth Hawkes applied Atlas to data from more than 54,000 UK Biobank participants. Grouping rare variants by their predicted molecular effects increased statistical power by pooling changes expected to disrupt the same biological process. The approach yielded 22% more noncoding associations than the baseline analysis used in the case study.

An analysis restricted to the top 1% of predicted-impact variants identified 19 genomic regions associated with body mass index. The same work found regulatory variants associated with circulating levels of proteins including PLA2G7, which has links to aging, and EGLN1, a sensor involved in cellular responses to oxygen.

Strong on splicing, weaker on enhancers

DeepMind reports that AlphaGenome performs best on variants affecting splicing and gene promoters. Researcher Žiga Avsec cautions that the model can miss effects in other regulatory elements, particularly enhancers, and remains less reliable overall than AlphaFold’s protein-structure predictions.

Launch materials also describe roughly 2,500 recurring DNA motifs associated with regulatory activity across cell types. These patterns can guide studies of how repeated sequence features influence gene control.

A practical analysis path

1. Align the inputs. Match variant coordinates and reference alleles to the genome build required by Atlas.
2. Choose an interface. Use the portal for individual lookups or the API for batch analysis and pipeline integration.
3. Rank with AVI. Sort candidate variants by score while retaining the original call-set metadata.
4. Inspect the mechanism. Review feature attributions and cell-type outputs to identify predicted effects on splicing, expression or chromatin.
5. Group related variants. For population studies, aggregate variants expected to perturb the same gene or molecular process.
6. Test the hypothesis. Validate shortlisted effects in an appropriate tissue, cell line or molecular assay before drawing biological or clinical conclusions.
7. Preserve provenance. Record the genome build, allele representation, Atlas version, AVI score and supporting outputs so results remain reproducible after database updates.

A database strategy borrowed from AlphaFold

The product strategy resembles the AlphaFold Database: DeepMind performs expensive model inference centrally, then distributes cached predictions as a reference resource. Downstream teams can spend their compute budgets on filtering, statistical analysis and experimental design while using Atlas for rapid variant lookup.

Adoption as a shared reference layer will depend on stable identifiers, versioned scores, reliable programmatic access and clear commercial terms. The current release supplies broad noncommercial access and a route to future Google Cloud deployment, while laboratory evidence remains the standard for confirming predicted molecular effects.