AlphaGenome Maps 9 Billion Variants—but It Cannot Diagnose Patients

In its September 8, 2026 launch announcement, Google DeepMind introduced AlphaGenome Atlas, a one-petabyte resource containing precomputed predictions for the molecular effects of 9 billion possible single-nucleotide variants in the human genome. Researchers can access the catalogue online through a web portal or retrieve its data programmatically.

The release turns AlphaGenome’s variant-effect forecasts into a searchable, genome-wide research resource rather than requiring scientists to request every prediction separately. Nature’s independent coverage of the Atlas describes the outputs as forecasts of what could happen when individual DNA letters are altered—not measurements from patients or laboratory experiments.

What the one-petabyte Atlas contains

The Atlas covers possible single-letter substitutions generated against the human reference genome. Its entries do not represent mutations observed in billions of people; they are model outputs calculated for alternative DNA letters at genomic positions.

For each variant, AlphaGenome can predict effects across molecular processes including gene expression, RNA splicing and chromatin features. This extends the catalogue across protein-coding regions and the much larger non-coding portion of the genome, where changes can affect the regulation of genes without altering a protein sequence directly.

The new AlphaGenome Variant Impact score, or AVI, combines AlphaGenome’s regulatory predictions with AlphaMissense predictions for protein-altering changes. A linked set of feature attributions indicates which predicted molecular processes contribute to the score, allowing researchers to rank candidates and investigate a proposed mechanism.

That ranking is not a finding of pathogenicity. A high AVI score means the models predict that a variant may have a consequential molecular effect; it does not establish that the variant causes a disease, accounts for a person’s symptoms or should change their care.

Precomputation changes the first stage of research

The practical advance is the removal of repeated computation from broad initial screens. Instead of submitting selected variants to AlphaGenome one at a time, a research group can search ready-made results, compare candidates and identify a smaller set for closer investigation.

This could accelerate hypothesis selection, particularly when researchers face many variants in non-coding regions whose functions are difficult to interpret. An entry might predict altered splicing or gene expression and point to the features behind that forecast. The resulting hypothesis still needs to be tested in a biological system suited to the gene, tissue and mechanism under study.

The Atlas also has narrower boundaries than its genome-wide scale might suggest. It maps single-nucleotide substitutions, not every structural rearrangement, combination of variants or interaction with genetic background and environment. Nor can a general impact score establish inheritance, population frequency, tissue relevance or a match with a patient’s phenotype.

Prediction, experiment and clinical evidence remain separate

The Atlas occupies the first level of an evidence ladder. Its value depends partly on keeping that level distinct from the evidence required to validate a biological mechanism or interpret an individual case.

• Model prediction: AlphaGenome estimates whether a DNA substitution could alter molecular outputs. AVI helps rank that estimated impact, while feature attributions suggest possible mechanisms.
• Experimental validation: A laboratory assay tests a selected prediction in an appropriate biological system. Confirmation of one result supports that finding; it does not validate every entry in the catalogue.
• Clinical evidence: Patient interpretation requires case-specific information such as validated measurements, phenotype, family history and other evidence assessed under applicable professional standards.

A computational prediction can be useful before its correctness is established because it helps researchers allocate limited experimental capacity. Clinical decisions carry a different evidentiary burden: prioritizing a candidate is not equivalent to demonstrating that it explains disease in a particular person.

Portal and API access come with firm use limits

The web portal is the direct route for searching and inspecting precomputed entries. The API supports repeatable, scripted retrieval of Atlas data and also provides access to on-demand AlphaGenome predictions. Retrieving an existing Atlas record and asking the model to calculate a new prediction are therefore related but distinct operations.

• Web portal: browse and interpret precomputed variant entries without running the model locally.
• AlphaGenome API: retrieve Atlas results programmatically or request focused model predictions after obtaining an API key.
• Commercial access: the announced research access is governed by non-commercial conditions, while separate commercial routes apply through Google Cloud.

The boundary for medicine is explicit. The official AlphaGenome API documentation and terms state that predictions are for theoretical modelling and research, must not be used for clinical decision-making, and cannot be relied upon as medical or professional advice.

Scale is established; broader validity is not

The launch establishes the Atlas’s scale, access routes and prediction-based design. Its longer-term scientific value will depend on how reliably prioritized variants survive independent experiments across different genes, tissues and populations, and whether researchers can reproduce useful results outside the collaborations involved in its development.

Clinical use would require evidence and authorization for a defined medical purpose beyond what the Atlas currently provides. For now, it is a large map of testable hypotheses: potentially valuable for deciding where genomic research should look next, but not evidence that can diagnose a patient or determine treatment.

Also read:

• Minnesota Criminalized Prediction Markets, but Its Ban Is Blocked
• One Pitch Cannot Do Every Job: Match Funnel Copy to the Next Decision
• AI Is Reaching Patients—but Its Proven Benefits Remain Task-Specific