Launchedby Google DeepMind on September 8, 2026, AlphaGenome Atlas provides predictions on the molecular effects of nine billion possible single-letter substitutions in the human genome. These are not nine billion mutations observed in patients, nor is it a catalog of diagnoses, but rather a map calculated from the reference genome. The challenge is therefore less about celebrating its sheer volume than about understanding what this infrastructure allows us to prioritize, what it does not demonstrate, and under what conditions its results can be used responsibly.
What Google DeepMind Actually Published
AlphaGenome Atlas is a database of pre-calculated predictions based on AlphaGenome, Google DeepMind’s genomics model. For each position in the human reference genome, the team simulated the three possible substitutions of one DNA base with another. This theoretical space yields approximately nine billion single-nucleotide variants, often abbreviated as SNVs. The term “atlas” therefore refers to a systematic exploration of possibilities, not a census of variants actually present in the population.[1][2]
The resource comprises approximately one petabyte of data. It is accessible to researchers through a free web portal for academic use, as well as through a dedicated programming interface. Google DeepMind has also announced integration with Google Antigravity and plans to offer commercial access via Google Cloud, though no detailed timeline was available at the time of this review.[1][4]
The Atlas associates variants with several predictions of molecular effects and a composite score called AlphaGenome Variant Impact (AVI). This score combines information from AlphaGenome and AlphaMissense to classify variants in coding and non-coding regions. A high score indicates a priority for further evaluation, but does not, on its own, constitute proof of causality or a diagnosis.[1][2]
Clinical caution is explicitly stated in Google DeepMind’s documentation. AlphaGenome and the Atlas are neither validated nor approved for clinical use. They should not replace medical advice, regulated genetic testing, or experimental validation.[1][4]
What's Really New
The AlphaGenome model did not first appear with the Atlas. It had already been described in a Nature article published in January 2026. This article presents a model capable of processing up to one million base pairs and predicting thousands of functional signals, including gene expression, RNA splicing, chromatin accessibility, transcription factor binding, and certain interactions between regions of the genome.[3]
The main innovation lies in the change in scale and access. Instead of running a prediction for each variant studied, researchers can query a database where the effects of every possible substitution have already been calculated. This pre-calculation reduces the initial cost of exploration and makes it easier to compare variants across the entire genome.
Another contribution is the goal of linking coding regions—which are directly involved in protein production—to noncoding regions, which play a role in regulation, among other functions. These noncoding regions account for a large portion of human variation and remain difficult to interpret. The AVI proposes a common metric for prioritizing them, but its scope must be evaluated separately from that of the underlying model, as the study on the Atlas was still a preprint as of October 6, 2026.[2][3]
This distinction is essential. The article in *Nature* provides a peer-reviewed evaluation of the AlphaGenome model. It does not constitute an independent validation of all the Atlas’s features, the AVI score, or every prediction contained in the database.
Executive MBA in AI & Business Transformation
The MBA Redesigned for the Age of AI. For experienced executives who want to lead the transformation of their organizations. Paris, Nice, and Dubai.
How AlphaGenome Atlas Works
AlphaGenome belongs to the family of “sequence-to-function” models. It takes a DNA sequence as input and predicts biological measures that could be observed in laboratory experiments in certain tissues or cell types. To estimate the effect of a variant, the system compares the predictions obtained with the reference sequence and with the modified sequence.
The model analyzes a context of up to one million base pairs. This length is designed to account for relationships that are not immediately adjacent to the variant. Its architecture combines convolutional layers—suited for local patterns—and transformer-like mechanisms, used to represent more distant dependencies. The outputs cover 5,930 human genomic tracks and 1,128 mouse tracks, distributed across eleven signal types.[3]
The Atlas applies this mechanism on a very large scale. It stores the predicted differences between the reference sequence and each possible substitution, then provides tools to search for a variant, compare its effects across tissues, and identify recurring regulatory motifs. Google DeepMind reports having identified more than 2,500 sequence motifs in this analysis.[2]
AVI adds a layer of classification. It aggregates signals from AlphaGenome and, for variants that affect proteins, from AlphaMissense. This aggregation can speed up the sorting of very long lists. It can also mask discrepancies between the signals if the score is interpreted without examining the detailed predictions.
Technology Framework
| Capacity or Constraint | What You Need to Know |
|---|---|
| Scope | Approximately nine billion possible single-base substitutions, calculated based on the reference human genome. These are not nine billion observed mutations. |
| Background of the Model | Up to one million base pairs as input to represent local effects and certain more distant regulatory relationships. |
| Organic Outings | 5,930 human tracks, 1,128 mouse tracks, and eleven types of signals, including expression, splicing, chromatin, and genomic contacts. |
| AVI Score | A prioritization score combining AlphaGenome and AlphaMissense. It ranks variants without proving their causality or pathogenicity. |
| Volume | Approximately one petabyte of pre-computed predictions—more than thirty times the reported size of the AlphaFold database. |
| Access | Free academic portal, API for non-commercial use, and integration with Antigravity announced. Commercial access planned on Google Cloud. |
| Scientific Status | The AlphaGenome model was published in *Nature*. The study specifically on the Atlas and the AVI score was a preprint as of October 6, 2026. |
| Clinical Use | No clinical validation or approval has been announced. The predictions should remain hypotheses to be tested and validated. |
Google DeepMind describes certain performance metrics as "state-of-the-art." This claim remains an assertion by the authors until independent teams have replicated the comparisons using the same data, the same partitions, and predefined criteria. Nor is the size of the dataset an indicator of accuracy.
An independent analysis reported by Live Science highlights this distinction between accessibility and scientific breakthroughs. Several experts view the interface as a useful tool, while noting that sequence-to-function models already existed and that laboratory validation remains essential. The Atlas’s current value therefore lies in its scale, unification, and ease of exploration, rather than in demonstrating a complete understanding of the genome.[5][6][7]
MSc in Data Management
Lead the strategic and operational management of data: acquisition, analysis, and governance. RNCP Level 7 degree (equivalent to a 5-year post-secondary degree).
Why This Infrastructure Is Important for Research
The immediate benefit of the Atlas is prioritization. Sequencing can reveal a large number of differences compared to the reference genome. Biologists must then identify which of these warrant further experimentation, a family study, or additional clinical analysis. A pre-calculated database can narrow this scope of investigation by highlighting the variants whose predicted molecular effects are most consistent with the hypothesis under study.
This approach is particularly useful for the non-coding genome. A variation located outside a gene can alter promoter activation, transcription factor binding, RNA splicing, or the accessibility of a chromatin region. The Atlas brings together these potential effects within a single interface, where multiple specialized tools were previously required.
In rare diseases, the potential value lies in the speed of triage. Google DeepMind presents a case related to the DNM1 gene in which Atlas helped identify a non-coding variant that creates an incorrect splicing site. The molecular effect was subsequently confirmed experimentally. This case demonstrates a potential utility within a research workflow, but a single successful observation does not measure sensitivity, specificity, or the error rate across the entire spectrum of rare diseases.[1][2]
At the cohort level, the team reports an analysis of more than 54,000 genomes from the UK Biobank. Grouping rare variants based on their predicted molecular effects reportedly identified an additional 22% of noncoding associations. Another analysis focusing on body mass index identified 19 genetic regions by limiting the search to the top 1% of non-coding variants. These results suggest an increase in statistical power and prioritization, not a demonstrated clinical link for each variant. [1][2]
For scientific organizations, data governance is also a key challenge. A petabyte of predictions can speed up access to information, but requires stable versions, traceable identifiers, metadata on the data sets, citation rules, and mechanisms for reproducing an analysis when the model or dataset changes.
What the results actually allow us to conclude
The available evidence is of different types. The Nature article evaluates AlphaGenome on technical tasks. The preprint on the Atlas evaluates the AVI score and presents several applications. The examples of rare diseases and cohorts illustrate possible uses. They do not all address the same question and should not be combined as general evidence of clinical validity.
| Evaluation | Published Results | Read Carefully |
|---|---|---|
| Genomic clues | AlphaGenome achieved the best reported results in 22 of the 24 genomic track prediction tasks. | An evaluation published in *Nature* based on test datasets defined by the authors. A good molecular prediction does not guarantee a correct clinical interpretation. |
| Effects of Variants | The model matches or outperforms the best external models in 25 of the 26 evaluations presented. | Comparison methods vary depending on the task. Specialized models may still be preferable in certain contexts, and the results do not cover all human variants. |
| DNM1 Case | A prioritized non-coding variant is associated with abnormal splicing, which is then confirmed by experimental analysis. | An informative case study, but insufficient to estimate the success rate for other diseases, genes, or populations. |
| UK Biobank | More than 54,000 participants and an additional 22% of non-coding associations, according to the analysis published by the team. | A statistical gain reported in a preprint. It demonstrates neither general biological causality nor clinical benefit. |
| Body Mass Index | Nineteen genetic regions are identified after selecting the top 1% of non-coding variants. | A result of prioritization involving a complex trait, which is also influenced by numerous biological and environmental mechanisms. |
| Independent Analysis | An independent preprint describes a common underestimation of the magnitude of effects and difficulties in linking certain distant elements to their target genes. | This research paper has not yet been peer-reviewed. Nevertheless, it confirms a point already acknowledged in the Nature article: long-range regulation remains challenging. |
Google DeepMind describes certain performance metrics as "state-of-the-art." This claim remains an assertion by the authors until independent teams have replicated the comparisons using the same data, the same partitions, and predefined criteria. Nor is the size of the dataset an indicator of accuracy.
An independent analysis reported by Live Science highlights this distinction between accessibility and scientific breakthroughs. Several experts view the interface as a useful tool, while noting that sequence-to-function models already existed and that laboratory validation remains essential. The Atlas’s current value therefore lies in its scale, unification, and ease of exploration, rather than in demonstrating a complete understanding of the genome.[5][6][7]
MSc in Data Management
Lead the strategic and operational management of data: acquisition, analysis, and governance. RNCP Level 7 degree (equivalent to a 5-year post-secondary degree).
What the Atlas Does Not Yet Allow Us to Conclude
AlphaGenome predicts molecular consequences based on DNA. It does not directly predict whether a person will develop a disease. Between an effect on gene expression and a clinical phenotype lie development, gene interactions, the environment, treatments, age, and many other factors that are still poorly understood.[3]
The model remains limited for certain long-range regulatory interactions, particularly those beyond 100,000 base pairs. The authors also note difficulties in reproducing effects specific to certain tissues or biological states, lower coverage of non-coding genes, and the lack of an evaluation of predictions for individual genomes. Its species coverage is limited to humans and mice.[3]
The Atlas focuses on single-base substitutions and also includes many small observed variants. It does not encompass the full range of genetic diversity. Large insertions, deletions, rearrangements, copy-number variations, and combinations of multiple variants can produce effects that cannot be inferred from a single mutation.
Finally, performance depends on the experimental data used to train the model on biological signals. The best-documented tissues, cell types, and populations are likely to be better represented. Without a subgroup analysis, an overall score may mask significant differences in reliability.
Genetic Data, Bias, and Accountability
The Public Atlas is a database of predictions calculated based on a reference genome, not a database of medical records. The risk changes when a laboratory submits variants from participants or patients to the Atlas. In the European Union, genetic data that can identify an individual falls under the special categories of personal data protected by Article 9 of the GDPR. Processing such data requires an appropriate legal basis, a defined purpose, and the safeguards provided for research or healthcare.[8][9]
Pseudonymization does not automatically make a genome anonymous. A genetic sequence is persistent, highly distinctive, and can indirectly reveal information about family members. Organizations must therefore limit the data they send, separate identifiers, document access, regulate data transfers, and verify the storage conditions of the portal or API.
Representation biases pose a second problem. If the reference, training, or validation data better represent certain ancestries and tissues, variants from other populations may be classified with less reliability. Unequal prioritization can influence research budgets and, ultimately, reinforce existing disparities in biomedical research.
The AVI score must also be safeguarded against genetic determinism. It represents an estimate of molecular impact, not a certainty about a person’s medical future. A simple interface can create the illusion of a definitive answer. Teams must maintain access to detailed indicators, highlight uncertainty, cross-reference multiple sources, and reserve all clinical interpretation for qualified professionals.
Finally, the physical scale of the project warrants documentation. The pre-processing and storage of one petabyte of data require significant computing resources. As of October 6, 2026, no sufficiently detailed public data on energy, water, or emissions associated with the creation and operation of the Atlas had been identified. It would therefore be premature to quantify its environmental footprint, but it is reasonable to expect comparable and verifiable indicators.
What to Watch for Now
The first step will be the peer-reviewed publication of the study on the Atlas and the AVI score. The most useful evaluations should distinguish between the ability to identify a previously known variant and the ability to prioritize new variants, using independent datasets defined prior to the analysis.
Next, robustness will need to be assessed across different ancestries, tissues, diseases, and variant categories. The results will need to be compared with those from specialized models, against reference scores, and, where possible, with functional experiments. An average measure will not be sufficient to determine whether to use the model in a sensitive context.
Version management will be just as critical. A variant may receive a different score after an update to the model or the data. Researchers will need to be able to retrieve the version that was queried, the parameters, the reference genome, the detailed output, and the rules for calculating the AVI.
Finally, the announced commercial launch will need to specify the terms regarding confidentiality, ownership of results, availability, and portability. AlphaGenome Atlas has the potential to become an important infrastructure for generating hypotheses more quickly. Its actual impact, however, will depend on the quality of independent validations, equity across populations, and the ability to maintain a clear distinction between algorithmic prioritization, biological evidence, and clinical decision-making.
Learn more
To understand AlphaGenome Atlas in the context of the evolution of genomics, augmented scientific research, and human validation in healthcare, be sure to check out these analyses from the aivancity blog.
Sources
[1] Google DeepMind, September 8, 2026. AlphaGenome Atlas: Molecular Predictions for 9 Billion Human DNA Variants. View source
[2] Cheng, J. et al., September 2026. AlphaGenome Atlas: In Silico Mutagenesis of the Entire Human Genome Improves Prioritization and Interpretation of Non-Coding Variants. medRxiv preprint. View the preprint
[3] Avsec, Ž. et al., Nature, 2026. Advancing Regulatory Variant Effect Prediction with AlphaGenome. View the study
[4] Google DeepMind, AlphaGenome and AlphaGenome Atlas documentation, accessed October 6, 2026. View the documentation
[5] Callaway, E., *Nature*, September 9, 2026, corrected September 22, 2026. DeepMind’s New Genome Atlas Charts Effects of All Nine Billion Human Gene Mutations. Read the article
[6] Kreimer, A. et al., September 2026. " Causal Variant Underestimation Is a Major Overlooked Driver of Poor Expression Prediction by Sequence-to-Function Models." bioRxiv preprint. View the preprint
[7] Live Science, September 23, 2026. Google DeepMind’s Latest AI Tool Promises a New Era for Genetics Research, but Experts Warn to Take Its Predictions with Caution. Read the article
[8] CNIL. The GDPR as Applied to the Healthcare Sector, accessed October 6, 2026. Visit the CNIL website
[9] European Union. Regulation (EU) 2016/679 on the protection of personal data, Article 9 and provisions relating to scientific research. View the regulation