A lookup table for an impossibly large search space

Google DeepMind has released AlphaGenome Atlas, a research resource that stores predicted molecular effects for every possible single-letter substitution in a reference human genome. The scale behind the name is simple: a reference genome has roughly three billion positions, and each position can be swapped for three alternative DNA letters. That produces about nine billion possible single-nucleotide variants to examine.

Testing all of those changes one at a time in a lab is not realistic. DeepMind's approach was to run predictions in advance and make the results searchable, rather than asking each research team to run a model across a large genome. The company says the resulting resource is about one petabyte and is available to academic users through a portal and API.

The release is therefore not a new diagnosis engine. It is closer to a large reference map. It gives researchers a fast way to ask where a particular genetic change might be worth looking more closely.

What the score is designed to do

The Atlas includes an AlphaGenome Variant Impact, or AVI, score. DeepMind says it combines predictions from AlphaGenome, which looks at molecular and regulatory effects, with AlphaMissense, a model for protein-altering variants. The aim is to give researchers one starting point for ranking variants across both the protein-coding part of the genome and the much larger non-coding part.

A ranking score can be helpful because a human genome contains many harmless differences. Research teams often need to decide which changes deserve scarce experimental time. DeepMind also provides feature attributions, intended to show which predicted biological processes contributed to a score, such as gene expression or RNA splicing.

That is useful context, but a score is still a model output. It compresses a complicated biological question into a prioritisation aid. A high score can guide a hypothesis. It cannot, by itself, establish cause, predict a person's health outcome or settle the biology of a disease.

Why precomputing the map could matter

The practical advantage is access. A research group without its own large compute cluster can inspect a possible variant and see a set of model predictions immediately. DeepMind says collaborators have used the Atlas in rare-disease research and in analyses of UK Biobank data, where the resource helped focus attention on non-coding variants that can be difficult to interpret.

Those examples are promising, but they should be read correctly. They describe research uses and, in some cases, follow-up experimental validation. They do not show that the Atlas will produce the same result for every disease, population, laboratory or clinical setting.

The strongest version of this tool is probably as a bridge between data and experiments. It can reduce the number of unpromising paths before a researcher starts the slower work of replication, functional testing and clinical interpretation.

The clinical boundary is not a footnote

DeepMind's own announcement says the information in AlphaGenome Atlas is not a substitute for medical advice, diagnosis or treatment. It also says AlphaGenome has not been validated for, and is not approved for, clinical use. That distinction deserves to be in the headline, not buried at the end.

Genetic findings are rarely self-explanatory. A variant's meaning can depend on the evidence behind it, a person's symptoms, family history, ancestry, the quality of the underlying sequencing and the clinical question being asked. A predictive model can add evidence to that picture. It should not silently become the decision-maker.

That is especially important when a single number is easy to share. The AVI score may make a vast resource easier to navigate. It does not make the uncertainty disappear, and it does not turn a research result into a result a patient should act on alone.

What is confirmed, what DeepMind says, and what remains open

Confirmed: Google DeepMind announced AlphaGenome Atlas on 8 September 2026 and published an accompanying technical report. The company describes a one-petabyte resource containing predictions for nine billion single-nucleotide variants, alongside an AVI score and feature-level explanations. It provides academic access through a portal and API.

DeepMind's claims: the Atlas is the most comprehensive catalogue of this kind, its score performs strongly on the benchmarks reported in the technical paper, and the resource can accelerate genetic discovery. Those claims and the presented case studies are important evidence from the project team, not a blanket independent validation for every use.

Open questions: how well the score transfers across different populations and rare disease settings, how researchers will calibrate it against other evidence, where predictions fail systematically, how access develops for commercial users and how the community will measure real research benefit over time. The Atlas is a powerful starting point. It is not the end of the investigation.

Sources

  1. Google DeepMind — AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genomePrimary announcement, 8 September 2026. Source for the release, access terms, Atlas scope, AVI description, project case studies and the explicit non-clinical-use warning.
  2. Google DeepMind — AlphaGenomePrimary project page. Source for the Atlas portal, API and the project's description of predicted coding and non-coding variant effects.
  3. Google DeepMind — AlphaGenome Atlas technical reportPrimary technical report accompanying the release. Used for the scope of the precomputed predictions and the reported AVI evaluation and attribution methodology.