AlphaGenome Atlas predicts the effect of nine billion DNA changes
AlphaGenome Atlas predicts the molecular effects of nine billion changes in human DNA. Here is its usefulness, its access and its limitations.
The human genome contains about three billion base pairs. At almost every position, a letter of DNA can theoretically be replaced by one of the other three, representing some nine billion possible substitutions.
Individual laboratory testing would be unrealistic. Google DeepMind therefore used its AlphaGenome model to predict their molecular effects, then consolidated the results in AlphaGenome Atlas, launched on September 8, 2026.
This map could speed up research on rare diseases and gene functioning. However, it does not constitute a diagnostic test or evidence that each prediction is accurate.
A prediction map, not nine billion experiments
AlphaGenome compares a normal DNA sequence to a modified version. He then tries to predict how this difference could influence gene expression, RNA splicing, chromatin accessibility, or protein binding.
The model can examine up to one million base pairs around a variant and produce thousands of predictions covering different tissues and cell types. Its architecture and evaluations were published in a peer-reviewed study in Nature.
Previously, a researcher had to choose a variant, prepare his data and execute the model. AlphaGenome Atlas instead contains pre-calculated results for all single letter substitutions in the reference human genome.
According to the official presentation of Google DeepMind, the complete set occupies about one petabyte, or one million gigabytes. However, the user does not need to download everything: a portal allows you to view a specific variant.
A score to rank variants
The Atlas adds an indicator called AlphaGenome Variant Impact, or AVI. It combines the predictions of AlphaGenome with those of AlphaMissense, another model specialized in protein modification variants.
The result is a number designed to help scientists quickly rank thousands of opportunities. The Atlas also indicates which mechanisms — such as splicing or gene expression — have most influenced the score.
This simplification entails a risk. Carl de Boer, a researcher in genomics at the University of British Columbia, warns in IEEE Spectrum that only one number could be misinterpreted in the face of such a complex biological system.
A high score means that a molecular effect seems more likely according to the models. It does not automatically mean that the carrier will develop a disease.
First results to be confirmed
Google presents several uses made with university partners.
A team studying rare diseases used AVI to identify a previously neglected variant of the DNM1 gene. Experiments then confirmed that this variant created a bad splicing site and produced an abnormal protein.
In another analysis of more than 54,000 UK Biobank participants, a team from the University of Exeter reportedly discovered 22% more associations involving non-coding variants. By focusing on the 1% most likely to have an effect, she would also have identified 19 areas associated with the body mass index.
These results are described by Google DeepMind. They do not demonstrate a 22 per cent universal improvement and will need to be replicated by independent teams.
Why non-coding regions matter
About 2% of the genome code proteins directly. Some of the remaining 98% is involved in controlling the location, timing and intensity with which genes are activated.
A variant located away from a gene can therefore change its functioning. These relationships are difficult to study because they vary between cells, tissues and stages of development.
The Atlas could help researchers reduce a long list of variants before undertaking costly experiments. As [IEEE Spectrum] (https://spectrum.iee.org/alphagenome-atlas) points out, its main purpose is to prioritize the assumptions to be checked, not to replace the laboratory.
Prices and availability in Canada
The portal and API are available free of charge for non-commercial research. No exclusions for Canada are posted.
Google plans to offer Atlas to businesses in Google Cloud, but this commercial service is not yet available and no rates have been communicated.
Actual costs may vary depending on the use of cloud resources, storage, sequencing, validation experiments, staff and institutional conditions.
AlphaGenome is neither validated nor approved for clinical use. A patient or health care professional should not use his or her predictions as a diagnostic or medical recommendation.
Limits or Considerations
The Atlas covers single letter substitutions. It does not map the large deletions, duplications, inversions or other structural changes of the genome in the same way.
Its window of one million bases is vast, but some regulatory elements act over even greater distances. The model may therefore lack important interactions.
Common diseases often depend on many variations, the environment, age and lifestyle. An isolated prediction does not summarize these interactions.
Finally, the quality of the results depends on the tissues, populations and data represented during training. High average performance does not guarantee the same reliability for all genetic ancestry or diseases.
Conclusion
AlphaGenome Atlas transforms previously demanding calculations into a searchable resource covering some nine billion genetic substitutions.
His interest is not to predict a person's medical destiny. It is to show scientists which variants seem likely to alter a biological mechanism and which deserve experience.
Independent validations will now determine whether this huge map actually solves more rare diseases and better understands the still poorly known regions of the genome.
