AlphaGenome Atlas: DeepMind's map of 9 billion DNA variants

In 60 seconds: Google DeepMind released AlphaGenome Atlas on September 8, 2026: an approximately 1-petabyte resource with precomputed predictions for about 9 billion possible single-letter changes in the human genome. Atlas is not a new model that diagnoses disease. It is an access layer over AlphaGenome that lets researchers look up variants, rank them with the AlphaGenome Variant Impact (AVI) score, and inspect the molecular signals behind that score. The portal, API, and a Google Antigravity skill are available for non-commercial research. As of September 30, 2026, commercial access to Atlas itself was still described as “coming soon” on Google Cloud; the base AlphaGenome model had a separate commercial offering.

The practical value is a smaller search space. The human genome has roughly 9 billion possible single-letter substitutions, and testing each one in a lab does not scale. Atlas lets a researcher start with variants whose predicted molecular effects look more relevant, then design experiments to test them.

That final step matters: a prediction prioritizes work. By itself, it does not prove that a variant causes disease, replace an assay, or constitute a diagnosis.

AlphaGenome Atlas is not the AlphaGenome model

AlphaGenome is the model that takes DNA sequences and predicts signals related to gene regulation. It can estimate effects on gene expression, splicing, chromatin accessibility, and other processes across different tissues and cell types.

AlphaGenome Atlas is the catalogue Google DeepMind published on September 8, 2026. DeepMind ran the model at scale and precomputed the effects of every possible single-nucleotide variant. The result links four layers:

• molecular-effect predictions for each variant;
• an AVI score for ranking variants;
• attributions showing which signals contribute to the score;
• more than 2,500 recurring DNA motifs and their locations.

The difference resembles consulting a map that has already been drawn instead of calculating a route from scratch. For an included variant, Atlas avoids a full inference run and provides a connected view of the results. The base model remains useful when an analysis needs custom queries over particular sequences or regions.

The dataset is approximately 1 petabyte. That number describes scale, not clinical accuracy.

What the AVI score summarizes

The AlphaGenome Variant Impact (AVI) score compresses signals from AlphaGenome and AlphaMissense into one number. AlphaGenome supplies regulatory predictions, which are especially useful outside protein-coding regions. AlphaMissense contributes a signal about the possible effects of protein-altering changes.

The score therefore covers both the roughly 2% of the genome that codes for proteins and the 98% that does not. That common measure helps rank candidates, but a number by itself does not explain a mechanism.

Atlas also publishes feature attributions for that reason. They break the score into interpretable contributions, including:

• chromatin accessibility;
• gene expression;
• RNA splicing;
• AlphaMissense’s predicted protein impact;
• evolutionary conservation.

A high score can mean “look here first.” Its attributions help form a hypothesis about the molecular process that may be disrupted. Neither is a probability of disease, proof of causality, or clinical recommendation.

Access: portal, API, Antigravity, and Cloud

Availability needs to separate Atlas from the base AlphaGenome model. This was the published position checked on September 30, 2026:

Route	What it provides	Published status and terms
Atlas web portal	No-code search and visualization	Available and free for non-commercial use
AlphaGenome API	Programmatic queries to the model and Atlas dataset	Free for non-commercial use, subject to terms and query limits
AlphaGenome Atlas skill in Google Antigravity	Access to the resource inside assisted scientific workflows	Available; users should check the terms that apply to their use case
AlphaGenome on Google Cloud	Commercial deployment of the base model for inference	Available through a paid subscription, sales contact, and self-hosted infrastructure
Commercial Atlas on Google Cloud	Commercial access to the precomputed catalogue	DeepMind said it was “coming soon”; the checked sources gave no public date or price

The API documentation says its outputs and Atlas information are non-commercial except where the terms expressly allow otherwise, and must not be used to train other machine-learning models. A research prototype does not automatically confer permission to turn the same workflow into a product.

The Cloud documentation, updated September 28, 2026, says the base model requires a paid subscription, contact with sales, and payment for hosting infrastructure. It lists no fixed price. It also does not say that commercial Atlas access is live. Before committing a roadmap, get written confirmation of which product the contract covers, available regions, quotas, costs, and rights over outputs.

Who should care

Atlas is designed for teams that already have a genomic question and need to decide where to investigate first:

Team	Reasonable use	What it still needs
Rare-disease research	Rank candidate variants and inspect predicted mechanisms	Phenotype, segregation, prior evidence, and experimental validation
Population genetics	Group rare variants by predicted molecular effects	Statistical design, bias controls, and replication
Regulatory biology	Explore motifs and signals in non-coding regions	Functional assays and the right cellular context
Genomics or biotech startups	Evaluate hypotheses and plan experiments	Commercial rights, specialist talent, data governance, and independent validation

It is less relevant to a general software startup looking only to “add AI.” Access to the map does not replace appropriate data, genetics expertise, or a concrete experimental question.

What DeepMind reported, and how to read it

In its September 8, 2026 announcement, DeepMind described results from external collaborators. These should be treated as claims from the provider and the named research teams, not as established clinical efficacy.

In work with Broad Institute researchers and the GREGoR Consortium, DeepMind says AVI helped prioritize a previously overlooked DNM1 variant. According to the announcement, the underlying prediction pointed to an incorrect splice site, and experimental screens validated that molecular mechanism. This is a case of prioritization followed by validation; it does not make Atlas an autonomous diagnostic test.

In a separate analysis, Gareth Hawkes at the University of Exeter applied Atlas to data from more than 54,000 UK Biobank participants. DeepMind reports that grouping variants by predicted molecular effects uncovered 22% more non-coding associations, and that focusing on the top 1% by predicted impact identified 19 regions linked to body mass index. Those are research results described by DeepMind and its collaborators. They need to be read alongside the study design, cohorts, and validation—not as individual predictions for patients.

The limit that must remain visible

AlphaGenome Atlas predicts molecular effects. DeepMind explicitly says it has not been validated or approved for clinical use and that its information is no substitute for professional advice, diagnosis, or treatment.

The commercial documentation for the base model also lists a 1-megabase context horizon: genomic interactions across longer distances cannot be captured in one query. Like any model, it also reflects the data, tissues, assays, and objectives used to build it.

A responsible workflow keeps three steps separate:

1. Prediction: Atlas highlights a variant and a possible mechanism.
2. Validation: the team tests that hypothesis against independent genetic, experimental, and statistical evidence.
3. Decision: a qualified person integrates the full evidence within the appropriate clinical, regulatory, or research framework.

Jumping from the first step to the third is exactly the use the official sources exclude.

Sources and freshness

We checked this guide on September 30, 2026 against official sources:

• Google DeepMind’s AlphaGenome Atlas announcement, September 8, 2026: scale, Atlas content, AVI, collaborator cases, access, and the clinical disclaimer.
• Google’s announcement, September 8, 2026: launch summary, portal, and reported results.
• Official AlphaGenome API repository: model and Atlas access, non-commercial terms, query limits, and the exclusion of clinical decision-making.
• Google Cloud AlphaGenome documentation, updated September 28, 2026: the base model’s commercial offering, deployment, contact-based pricing, and limitations.

Availability, terms, and pricing can change. Check the sources and applicable contract before using Atlas or AlphaGenome in a commercial product.