Google DeepMind Launches AlphaGenome Atlas With Predictions for 9 Billion DNA Variants

Google DeepMind launched AlphaGenome Atlas on September 8, 2026, giving researchers searchable access to predicted molecular effects for approximately 9 billion possible single-letter changes in the human genome.

The one-petabyte database applies DeepMind’s AlphaGenome model across the human reference genome in advance. Instead of selecting variants individually, writing code and running the computationally demanding model for each analysis, researchers can retrieve previously calculated results through a web portal or programmatic interface.

The Atlas also introduces the AlphaGenome Variant Impact score, or AVI score. It combines AlphaGenome’s predictions of regulatory effects with AlphaMissense predictions for protein-altering variants, producing one score that can be used to rank changes in both coding and non-coding DNA.

This is a research-prioritization system, not a clinical diagnostic database. Its entries are model predictions rather than experimentally established findings, and Google explicitly prohibits using them for clinical decision-making.

1. DeepMind Has Precomputed Every Possible Single-Letter Substitution

The human genome contains roughly 3 billion base pairs. At each position, the reference DNA letter can theoretically be replaced by any of the other three letters, producing approximately 9 billion possible single-nucleotide substitutions.

AlphaGenome Atlas contains predictions for this entire search space relative to the human reference genome. The resulting dataset occupies about one petabyte, which DeepMind says is more than 30 times the size of the AlphaFold Database.

The scale does not mean the Atlas covers every possible form of human genetic variation. It focuses on single-letter substitutions. Insertions, deletions, structural variants and combinations of variants within an individual genome are outside the nine-billion-variant catalogue described in the launch.

Before the Atlas, researchers could submit selected sequences or variants to the AlphaGenome API. That approach works for targeted studies but requires users to decide what to test and then spend computational resources generating predictions. DeepMind says the ordinary model API is suited to analyses involving thousands of predictions and is unlikely to be appropriate for workloads exceeding one million predictions.

The Atlas moves that computation upstream. According to IEEE Spectrum’s reporting from the launch, DeepMind initially estimated that it needed an approximately 80-fold speed improvement to calculate the entire collection in a practical period. The team used model distillation, optimized GPU kernels and removed redundant calculations to reach the necessary scale.

The practical result is a shared lookup resource. Laboratories do not have to repeat the same model inference for a common variant, and researchers without access to newer accelerators can inspect the predictions through a browser.

2. What AlphaGenome Predicts

AlphaGenome was introduced in June 2025 and described in a peer-reviewed Nature paper published on January 28, 2026. It is a sequence-to-function model: it takes DNA sequence as input and predicts molecular measurements associated with gene regulation.

The model processes as many as one million DNA base pairs at a time. Its architecture uses convolutional layers to recognize local sequence patterns, transformers to exchange information across the input, and output layers specialized for different forms of genomic activity.

The published model simultaneously predicts 5,930 human or 1,128 mouse genomic tracks across 11 modalities. These include RNA expression, transcription initiation, chromatin accessibility, histone modifications, transcription-factor binding, three-dimensional chromatin contacts, splice sites, splice-site usage and splice-junction coordinates and strength.

For variant analysis, AlphaGenome compares predictions generated from the reference sequence with predictions from a sequence containing the alternate DNA letter. The difference indicates how the model expects the change to affect molecular processes in particular biological contexts.

In the Nature evaluation, AlphaGenome matched or exceeded the strongest external models in 25 of 26 variant-effect assessments and achieved state-of-the-art results in 22 of 24 genome-track tasks. These measurements evaluate individual model capabilities under defined benchmarks; they do not establish that every prediction in the new Atlas is correct.

The Atlas contains thousands of molecular-effect predictions for each variant. It also links variants to a collection of more than 2,500 recurring DNA sequence motifs, which can help researchers investigate whether a change disrupts a recognizable regulatory pattern.

3. The AVI Score Adds a Genome-Wide Ranking Layer

The number of outputs produced by AlphaGenome creates an interpretation problem of its own. A variant may affect predicted RNA abundance, chromatin accessibility or splicing differently across numerous tissues and cell types. Examining every output is impractical when a study begins with thousands or millions of candidates.

The AVI score condenses this information into a single impact value. For non-coding regions, it draws on AlphaGenome’s regulatory predictions. For protein-altering variants, it incorporates AlphaMissense, DeepMind’s specialized model for assessing missense changes.

This gives researchers a common ranking mechanism across the approximately 2% of the genome that codes for proteins and the much larger non-coding portion involved in controlling gene activity. A high score is intended to identify variants that deserve closer examination, not to declare that a variant causes disease.

DeepMind also provides feature attributions alongside the score. These decompose the result into biological categories, such as predicted changes to RNA splicing, gene expression, chromatin accessibility, sequence conservation or protein function. Researchers can therefore move from a ranked candidate to a proposed molecular mechanism.

That interpretability layer matters because a single composite score can hide important distinctions. An independent genomicist interviewed by IEEE Spectrum warned that the simplified number could be misinterpreted because genomic regulation involves many interacting processes. The detailed predictions and feature attributions remain necessary for understanding why a variant received its ranking.

4. Early Studies Show How the Atlas Can Narrow Searches

DeepMind reported that external research groups used the Atlas before launch in rare-disease, population-genetics and gene-regulation projects. These examples show its intended role as a filter for experimental work.

Researchers associated with the Broad Institute and the GREGoR Consortium used AVI scores to prioritize variants in unresolved rare-disease cases. One candidate affected the DNM1 gene, which is associated with developmental and epileptic encephalopathy.

AlphaGenome predicted that the change created an incorrect splice site and produced an abnormal extension of the resulting protein. DeepMind says experimental screens validated the predicted mechanism and identified nearby variants with similar effects. The experiment, rather than the original model score, supplied the validating evidence.

In a separate analysis, University of Exeter researcher Gareth Hawkes applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. DeepMind reports that grouping rare non-coding variants by their predicted molecular effects produced 22% more detectable genetic associations than the comparison approach.

One analysis reportedly reduced a region containing 526 initial candidates to four prioritized variants. The work identified regulatory candidates associated with circulating levels of proteins including PLA2G7 and EGLN1.

For a body-mass-index analysis, Hawkes restricted the search to the 1% of non-coding variants that the Atlas ranked as most impactful. That process identified 19 genomic regions for further investigation. These associations establish research targets; they do not demonstrate that the highlighted variants individually cause changes in body mass.

Researchers at the Stowers Institute for Medical Research also used the motif resource to study regulatory sequences involved in chromatin accessibility and gene activation. This illustrates a second use for the Atlas beyond disease-variant ranking: querying predicted effects at scale to investigate the basic rules controlling gene activity.

5. Access Conditions and Scientific Limits

AlphaGenome Atlas is available through a free website for academic and other non-commercial research. DeepMind also lists access through the AlphaGenome API and an AlphaGenome Atlas skill for Google Antigravity. Commercial Atlas access through Google Cloud was described at launch as coming soon, while the underlying AlphaGenome model already has a paid, allowlisted Google Cloud deployment route.

The public API’s terms restrict AlphaGenome outputs and Atlas information to non-commercial use except where expressly permitted. Google also says the outputs must not be used to train other machine-learning models.

Most importantly, AlphaGenome is neither intended nor cleared for clinical diagnostic use. Predictions cannot substitute for patient evidence, laboratory validation, clinical interpretation or medical advice.

The model’s one-million-base input is large compared with earlier sequence models, but regulatory elements can affect genes across long distances. DeepMind’s Nature paper found that accurately capturing influences more than 100,000 base pairs away remains challenging. Tissue-specific and condition-specific effects also remain difficult to reproduce consistently.

AlphaGenome was not designed or benchmarked for predicting the consequences of a person’s complete genome. Complex traits and diseases can depend on multiple variants, development, gene function and environmental factors that are outside a direct sequence-to-molecular-function model.

The Atlas therefore changes the cost and speed of hypothesis selection rather than the evidentiary standard. It can rank billions of candidate substitutions and expose predicted mechanisms immediately, but consequential findings still require population evidence, independent computational checks and experimental confirmation.

Frequently Asked Questions

What does AlphaGenome Atlas contain?

It contains precomputed molecular-effect predictions and AVI scores for approximately 9 billion possible single-nucleotide substitutions in the human reference genome.

Why are there 9 billion possible changes?

The genome has roughly 3 billion positions, and the reference letter at each position can be replaced by three alternative DNA letters.

Does the Atlas cover every type of genetic mutation?

No. Its nine-billion-variant catalogue covers single-letter substitutions, not every insertion, deletion, structural variant or combination of changes found in personal genomes.

Can doctors use an AVI score to diagnose a patient?

No. Google states that AlphaGenome predictions are for theoretical modelling and research and must not be used for clinical decision-making.

How can researchers access the Atlas?

Non-commercial researchers can use the browser-based portal or AlphaGenome API. DeepMind also offers an Atlas skill for Google Antigravity and plans commercial availability through Google Cloud.

Sources

• Original Google DeepMind announcement on X
• Google DeepMind: AlphaGenome Atlas—A Predictive Map of Every Possible DNA Letter Change in the Human Genome
• Google: Introducing AlphaGenome Atlas
• Google DeepMind: AlphaGenome
• Nature: Advancing Regulatory Variant Effect Prediction With AlphaGenome
• Google DeepMind AlphaGenome API Repository and Usage Terms
• IEEE Spectrum: Google DeepMind Maps 9 Billion Possible DNA Variants