AlphaGenome Atlas, launched by Google DeepMind on September 8, 2026, turns AlphaGenome from a model queried variant by variant into a large precomputed catalogue of predicted molecular effects. The headline number — about 9 billion variants — does not mean 9 billion variants found in patients or 9 billion diagnoses. It essentially represents every possible single-letter DNA substitution in the human reference genome: at each position, the existing letter can be replaced by three alternatives.
The scientific value of the Atlas is that it makes this enormous space of possibilities searchable. Instead of testing every change in the laboratory or running the model from scratch for each variant, researchers can use precomputed predictions to prioritize hypotheses, investigate regulatory mechanisms and decide which changes deserve experimental validation. The key distinction is this: the Atlas predicts molecular effects; it does not determine on its own whether a person has, will develop or will not develop a disease.
What “9 billion variants” actually means
The human genome contains roughly 3 billion DNA positions. For a single-nucleotide variant, or SNV, each position can be changed to any of the other three letters. Applying that logic genome-wide produces about 9 billion possible substitutions. AlphaGenome Atlas precomputed predictions across that space, creating a dataset of roughly 1 petabyte.
That scale is different from three concepts that are often confused:
- Possible variants: all single-letter substitutions that could exist relative to the reference genome.
- Observed variants: changes that have actually been identified in people, populations or genomic databases.
- Causal variants: changes for which there is sufficient evidence that they contribute to a specific biological mechanism or phenotype.
The Atlas covers the first category and can help study the other two. It does not automatically turn a computational possibility into an observed, functional or disease-causing variant.
What kinds of predictions the Atlas provides
The Atlas brings together connected resources ranging from rapid ranking to mechanistic investigation. Its foundation is AlphaGenome, a model that takes DNA sequence context and predicts thousands of functional signals. The work published in Nature describes AlphaGenome processing up to 1 million base pairs and producing predictions across multiple modalities, many at single-base resolution.
| Resource | What it is for | How not to interpret it |
|---|---|---|
| Molecular effect predictions | Estimate how a change could alter signals such as gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription-factor binding and splicing. | They are not direct measurements from a patient's sample. |
| AVI score | Summarize in one number a variant's priority according to predicted molecular impact, combining information from AlphaGenome and, where relevant, AlphaMissense. | It is not a disease probability or an automatic clinical classification. |
| AVI feature attributions | Indicate which components — for example splicing, expression, chromatin or conservation — contributed most to the score. | They do not prove that the mechanism actually occurs in a living organism. |
| Sequence motifs | Connect variants with recurring DNA patterns that can help generate hypotheses about gene regulation. | They are not, by themselves, causal evidence of protein binding or a phenotypic effect. |
Why one score helps — and why it is not enough
A practical challenge in genomics is that a single genome can contain millions of differences from the reference. In rare-disease research, for example, the problem is often not finding variants but deciding which few deserve attention first. The AlphaGenome Variant Impact (AVI) score was created to assist with that triage.
The score allows coding and non-coding variants to be ranked on a common scale of predicted impact. That matters because most of the genome does not directly encode proteins but can influence when, where and how strongly a gene is activated, how RNA is processed and how chromatin is organized.
Ranking, however, is not diagnosis. A high AVI means the system predicts a relatively strong molecular alteration. By itself, it does not specify the penetrance of a condition, age of onset, clinical severity, interactions with other variants, environmental effects, or whether the predicted mechanism is active in the tissue and cell state relevant to a particular person.
What changes compared with the original AlphaGenome
AlphaGenome already allowed researchers to analyze specific sequences and variants through an API. The Atlas mainly changes scale and accessibility: instead of requiring each group to repeatedly run predictions, results for billions of SNVs were computed in advance and organized for access through a portal and API.
This removes an important computational barrier. The original AlphaGenome paper reported performance that matched or exceeded the strongest external models in 25 of 26 variant-effect evaluations. The same paper also describes limitations: some cell-type-specific effects remain difficult, and models of molecular effect are not equivalent to models of phenotypic consequences.
How researchers are already using the Atlas
The examples released with the launch illustrate the type of gain the Atlas is intended to provide: prioritization and hypothesis generation, followed by verification through other methods.
In a collaboration involving Broad Institute researchers and the GREGoR Consortium, AVI helped prioritize a variant involving the DNM1 gene in an unresolved rare-disease case. The predictions indicated that the variant created an incorrect splice site, and subsequent experiments validated the functional effect. The important point is not that the Atlas “made the diagnosis,” but that it helped surface and explain a candidate that was then experimentally tested.
In another use reported by DeepMind, Gareth Hawkes of the University of Exeter applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. By grouping rare variants according to predicted molecular effects, the analysis identified 22% more non-coding genetic associations than would otherwise have been detectable with that strategy. Again, this is a research tool for increasing analytic power, not a way to turn a score into an individual clinical conclusion.
What AlphaGenome Atlas does not diagnose
Google DeepMind is explicit: information from AlphaGenome Atlas is not a substitute for professional medical advice, diagnosis or treatment, and AlphaGenome has not been validated for and is not approved for any clinical use.
There are scientific reasons for that boundary. A molecular-effect prediction answers questions such as “could this base change alter splicing?” or “does the model expect a change in expression in this context?”. A diagnosis requires much more: the patient's phenotype, family history, quality and origin of the observed variant, population frequency, inheritance pattern, functional evidence, published literature, classification criteria and, where appropriate, laboratory confirmation.
A variant may also have a large predicted molecular effect but little relevance to a specific clinical outcome, or the reverse. Regulatory biology depends on tissue, cell state, developmental context, interactions among variants and environmental factors that are not fully represented by a single score.
A safe way to read an Atlas result
- Use AVI to prioritize, not to conclude. A high score signals that a variant deserves closer investigation.
- Open the attributions and detailed predictions. Check whether the signal is driven by splicing, expression, chromatin, protein impact or another category.
- Check the biological context. Ask whether the predicted tissue, gene and mechanism fit the research hypothesis.
- Compare with external evidence. Population databases, literature, experimental data and other tools remain necessary.
- Validate when the conclusion depends on the mechanism. A strong computational prediction can guide an experiment, but it does not replace one.
What the launch changes for genomic research
The main advance of AlphaGenome Atlas is making searchable a layer of prediction that would be too expensive for many groups to generate genome-wide on their own. It can accelerate screening, draw more attention to non-coding variants and connect variant analysis more directly with hypotheses about molecular mechanisms.
Its usefulness, however, depends on keeping the boundary between prediction, evidence and diagnosis clear. The system provides a map of computational hypotheses at unprecedented scale; the scientific work is still to determine which of those hypotheses match real biology and under which conditions.
Sources
- Google DeepMind, AlphaGenome Atlas announcement: https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/
- Avsec et al., Nature (2026), AlphaGenome paper: https://www.nature.com/articles/s41586-025-10014-0
- Nature, independent coverage of the Atlas launch: https://www.nature.com/articles/d41586-026-02835-4
- UCSC Genome Browser, AVI track description for hg38: https://genome.ucsc.edu/goldenPath/newsarch.html
