Google’s AlphaGenome Atlas Makes Nine Billion Possible DNA Substitutions Searchable

Google DeepMind has released AlphaGenome Atlas, a one-petabyte collection of precomputed predictions for roughly nine billion possible one-letter substitutions across the human reference genome. Academic and other non-commercial researchers can search it through a browser or API.

The release changes the first step of variant research. Instead of selecting a DNA change, running a large model and waiting for its outputs, a researcher can look up predictions that DeepMind has already calculated. Independent scientists told IEEE Spectrum and Nature that this can remove computational and coding barriers.

AlphaGenome Atlas is separate from the AlphaGenome model introduced in 2025, which is the system that takes a long stretch of DNA and predicts how sequence changes could alter molecular activity. The Atlas is the stored output of running that model at genome scale, supplemented with tools for ranking and inspecting variants. 

Why Three Billion Positions Produce Nine Billion Substitutions

The arithmetic comes from the four-letter DNA alphabet. At each position in a reference genome, the recorded nucleotide is A, C, G or T. A single-nucleotide substitution replaces that reference letter with one of the other three. Applying three alternatives to roughly three billion reference positions produces roughly nine billion possible substitutions.

An insertion adds nucleotides and a deletion removes them. Multi-base substitutions and larger structural variants also change more than one position. DeepMind’s exhaustive nine-billion count covers reference-to-alternate single-nucleotide substitutions. The Atlas separately includes more than 100 million short insertions and deletions observed in large datasets, not every insertion or deletion that could exist.

The one-petabyte resource stores thousands of predicted molecular effects for each substitution. AlphaGenome estimates changes in signals such as gene expression, RNA splicing and chromatin activity across hundreds of human and mouse tissues and cell types. These are computational forecasts tied to the model’s training data and sequence context.

AVI Turns Predictions Into a Research Queue

The AlphaGenome Variant Impact score, or AVI, helps researchers sort that volume. At a high level, it joins AlphaGenome’s predictions about regulatory effects with AlphaMissense’s protein-impact predictions. The accompanying preprint describes a broader 18-feature score that also uses conservation, predicted protein-termination effects and indicators for insertions or deletions.

AVI ranks candidates for follow-up. A high score can move a variant toward the front of a research queue, while feature attributions show which modeled inputs pushed the score up or down.

Researchers can also inspect a catalogue of 2,601 sequence motifs derived from patterns the model learned. A motif is a recurring DNA pattern that may help explain why the model predicts an effect in a particular molecular readout or cell type. Some patterns match known regulatory biology and some have experimental support, while the paper labels others as broad, unresolved or unknown. The catalogue gives researchers testable patterns whose certainty ranges from known biology to unresolved hypotheses.

Access Depends on the Research Workflow

The A browser portal is the simplest entry point for exploring individual variants without code. The AlphaGenome API supports programmatic and higher-volume access to AlphaGenome Atlas scores, AVI and feature importances. Both are available for academic or other non-commercial research under Google’s terms. Google also exposes the Atlas as a specialized Antigravity science skill for agent-assisted research workflows.

Commercial AlphaGenome Atlas access is planned for the future. DeepMind says it will come to Google Cloud “soon”, but has not disclosed a launch date, price or licensing terms. Google Cloud already offers the base AlphaGenome model through its Model Garden.

Two Applications Show What Prioritization Can Do

Google highlights how one collaborator case followed an initial prediction done by the system advanced protein related findings. Researchers in the GREGoR Consortium and at the Broad Institute used AVI to rank a deep intronic variant in DNM1, a gene involved in a severe developmental disorder. AlphaGenome predicted that the variant would create a brain-specific splice site and alter the resulting protein. A minigene assay then tested that local splicing mechanism across five cell lines and reproduced the predicted effect for the specific variant, alongside related changes.

A separate analysis tested whether Atlas-derived filters could help find statistical associations in UK Biobank data. Researchers examined 2,028 circulating proteins in 54,189 participants, focusing on rare variants and comparing several association methods. Filtering with Atlas features found 728 significant non-coding rare-variant associations with protein levels that remained statistically distinct after overlapping signals were separated, up from 595 without those aggregated features. The reported 22 percent increase counts additional association signals under this analysis; it measures neither predictive accuracy nor causation.

DeepMind also reports that a body-mass-index analysis identified 19 associated regions after focusing on the top one percent of non-coding variants predicted to have the greatest effect. The result is statistical association and prioritization in the UK Biobank cohort. Causal explanations or clinical performance would require different evidence.

Faster Triage Does Not Settle Biology

DeepMind describes its AlphaGenome Variant Impact scoring as best in class across a broad evaluation. AVI led many benchmark tests, while other methods narrowly led a ClinVar task involving variants in a gene’s untranslated region and a TraitGym task involving non-coding variants tied to inherited disorders. Those results show that performance depends on the variant context and test dataset.

The model has structural limits. It can miss effects from enhancers that act beyond the sequence window it analyzes. Indirect regulatory effects, absent cell types, uneven experimental coverage and gaps in training data can all weaken a prediction. Scientific American characterized AlphaGenome’s predictions as less accurate overall than AlphaFold’s protein-structure predictions, although the systems predict different objects and do not share a single comparable accuracy measure.

AlphaGenome Atlas is a research resource and so far has no clinical validation or approval. Its practical advance is narrower and still substantial: researchers can scan, rank and inspect a vast set of possible DNA substitutions before deciding which ones deserve scarce laboratory time. The release removes repeated AlphaGenome inference from that initial triage. Experiments still have to establish which predictions describe biology in the real world.