AlphaGenome Atlas: 9 Billion DNA Variants, Precomputed

AlphaGenome Atlas: 9 Billion DNA Variants, Precomputed

Google DeepMind has done something unusual with a genomics model: instead of asking researchers to run the model every time they want to investigate a variant, it ran the model across the human genome first.

On September 8, 2026, DeepMind introduced AlphaGenome Atlas, a database containing precomputed predictions for 9 billion possible single-nucleotide variants — essentially every possible single-letter change in the human genome.

The resulting dataset is about 1 petabyte in size. DeepMind says that makes it more than 30 times larger than the AlphaFold Database. The Atlas is available for academic research through a web interface, the AlphaGenome API, and an AlphaGenome skill for Google Antigravity.

The size is impressive, but that is not really the important part.

The interesting change is what happens when a prediction model becomes a dataset.

From running a model to querying a map

The human genome contains roughly 3 billion base pairs. At each position, there are three possible alternative bases, giving roughly 9 billion possible single-letter substitutions.

Researchers obviously cannot test all of those changes experimentally.

Before AlphaGenome Atlas, there was also a computational bottleneck. A researcher with a candidate variant could ask AlphaGenome to compare the reference and altered sequences and predict how the change might affect gene regulation. That works well when you have a manageable list of variants.

It becomes a different problem when the question is:

Which variants across the entire genome are worth looking at?

DeepMind's answer is to precompute the expensive part — precomputation as a pattern that is increasingly how scientific computing absorbs AI models.

AlphaGenome can process up to 1 million base pairs of DNA and produce thousands of functional genomic predictions. These include gene expression, transcription initiation, chromatin accessibility, transcription-factor binding, chromatin contacts and splicing-related signals.

The model was introduced in 2025 and subsequently published in Nature in January 2026. Its published evaluation covered 26 variant-effect prediction tasks, where AlphaGenome matched or exceeded the strongest external models in 25 of them.

AlphaGenome Atlas takes that model and applies it at genome scale.

Instead of:

variant
   ↓
run AlphaGenome
   ↓
get prediction

the workflow becomes:

9 billion possible variants
        ↓
  AlphaGenome runs
        ↓
precomputed predictions
        ↓
      Atlas
        ↓
researcher queries what matters

That is a very different way of using an AI model.

The expensive computation happens once. Researchers can then start from the biological question instead of starting from a model invocation.

The difficult part is not the number 9 billion

A catalogue containing 9 billion variants would not be very useful if every entry simply returned a giant pile of model outputs.

Each variant can have thousands of predicted molecular effects across different tissues and cell types.

So DeepMind added another layer: the AlphaGenome Variant Impact (AVI) score.

AVI reduces the information into a single score that researchers can use to rank variants. It combines AlphaGenome's predictions for regulatory effects with AlphaMissense's predictions for protein-altering variants.

That distinction matters.

Only around 2% of the human genome directly codes for proteins. The remaining 98% is largely non-coding DNA involved in regulating when, where and how genes are expressed.

A tool that only focuses on protein-changing mutations therefore misses a huge part of the genome's regulatory machinery.

AVI is designed to cover both sides.

More importantly, the score is not intended to be a black-box number. The Atlas also provides AVI feature attributions, which show which biological signals contributed to a variant's score.

For example, a high-impact prediction can be associated with changes involving processes such as RNA splicing, gene expression or chromatin accessibility.

That makes the system more useful for research because the question is not simply:

"Which variant has the highest score?"

It is also:

"Why did this variant rank highly?"

What the underlying model is actually doing

The basic idea behind variant scoring is a comparison between two versions of the same DNA sequence.

One is the reference sequence.

The other contains the alternative base.

AlphaGenome predicts molecular signals for both and looks at the difference.

A simplified example of an expression-related comparison looks like this:

import numpy as np

def expression_score(ref_cov, alt_cov, eps=0.001):
    """Simplified log-fold change between REF and ALT predictions."""
    return (
        np.log(np.mean(alt_cov) + eps)
        - np.log(np.mean(ref_cov) + eps)
    )

ref_cov = np.array([3.1, 2.9, 3.3, 3.0])
alt_cov = np.array([7.8, 8.2, 7.9, 7.5])

score = expression_score(ref_cov, alt_cov)

print(f"log-fold expression change: {score:+.3f}")

This is only a simplified illustration, not an implementation of the Atlas scoring pipeline.

The real AlphaGenome system operates across many genomic modalities and much larger sequence contexts. The published model processes sequences up to 1 Mb and predicts thousands of functional genomic tracks at high resolution.

The important idea is the comparison: what does the model predict for REF, and what changes when one base is replaced?

At Atlas scale, those comparisons have already been computed for the genome.

The first results are more interesting than the dataset size

DeepMind is not presenting the Atlas as a giant database for its own sake.

Its collaborators have already used it on real research problems.

Rare disease

At the Broad Institute, researchers working with the GREGoR Consortium used AVI to prioritize variants in an unsolved rare-disease case.

The system highlighted a variant affecting DNM1, a gene associated with epileptic encephalopathy.

The important part was not simply that the variant received a high score. The underlying AlphaGenome predictions suggested that the variant created an incorrect splice site, resulting in an abnormal extension of the resulting protein.

Experimental work subsequently supported the prediction.

That is a useful example of where this kind of system fits into research.

AI did not replace the experiment.

It helped narrow down where researchers should spend their experimental effort — AI narrowing the search before experimental proof, a pattern now repeating across the sciences.

Population genetics

Another case comes from Gareth Hawkes at the University of Exeter.

Using whole-genome data from more than 54,000 UK Biobank participants, Hawkes used AlphaGenome Atlas predictions to group rare non-coding variants according to their predicted molecular effects.

According to DeepMind, this approach uncovered 22% more non-coding genetic associations than the analysis would otherwise have detected.

The work also identified regulatory variants associated with proteins including PLA2G7 and EGLN1.

In another analysis, focusing on the top 1% of predicted impactful non-coding variants led to 19 genetic regions associated with body mass index.

Again, the useful part is the filtering.

The genome contains an enormous number of possible variants. The problem is not a lack of candidates. The problem is deciding which candidates deserve attention.

There is also a map of the genome's "words"

The Atlas is not limited to variant scores.

DeepMind says it includes more than 2,500 recurrent DNA sequence motifs and their locations.

These short sequences can act like regulatory building blocks. They can help researchers investigate where transcription factors bind and how particular sequences influence gene activity.

Researchers at the Stowers Institute for Medical Research used these resources to examine which transcription factors affect DNA accessibility and which can also influence gene activation.

This is where the Atlas starts looking less like a variant database and more like a reference layer for genome analysis.

You can move from:

variant
  ↓
predicted effect
  ↓
biological feature
  ↓
DNA motif
  ↓
regulatory mechanism

That chain is much more useful than a score on its own.

The API is not the same thing as the Atlas

There is an important distinction here.

The AlphaGenome model and API are for generating predictions. The Atlas is the massive precomputed resource built from those predictions.

The official AlphaGenome API repository says the API supports predictions across gene expression, splicing, chromatin features and other genomic outputs. It can analyse sequences up to 1 million base pairs and is currently offered free for non-commercial use, subject to its terms.

The API is therefore useful when you want to run your own analysis on specific regions or variants.

The Atlas is useful when you want to search the precomputed genome-wide results.

That difference sounds small, but it changes the economics and workflow of genomic analysis.

You do not need to spend compute generating the same prediction that DeepMind has already generated.

And this is where the infrastructure gets interesting

A 1-petabyte dataset sounds like an infrastructure problem.

It is.

But it is also an infrastructure strategy.

DeepMind is effectively turning a model into a reusable scientific data layer.

The traditional workflow is:

research question
      ↓
prepare sequence
      ↓
run model
      ↓
analyse prediction

The Atlas introduces another option:

research question
      ↓
query precomputed predictions
      ↓
rank candidates
      ↓
inspect biological context
      ↓
run experiments

That can make a major difference when the search space is enormous.

It is similar in spirit to what happened with AlphaFold: once large amounts of expensive prediction work were turned into an accessible database, researchers no longer had to begin every project by generating the same basic structural information.

The interesting question now is what happens when the same idea is applied to genome-wide variant effects.

But this is not a diagnostic system

This distinction matters.

AlphaGenome Atlas is a research resource. DeepMind explicitly says that AlphaGenome has not been validated for, and is not approved for, clinical use.

An AVI score is therefore not a diagnosis and should not be treated as one — the same reason unvalidated model outputs need guardrails before deployment in any high-stakes setting.

A high score means that the model predicts a potentially important molecular effect. It does not prove that the variant causes a disease.

That is why the DNM1 example is important. The computational prediction helped researchers prioritize a candidate, but experimental validation was still required.

There is also a broader limitation worth keeping in mind: model predictions are still predictions.

Precomputing them for every possible variant does not magically turn them into ground truth.

It simply makes the predictions much easier to search.

The bigger idea is the database, not the model

There will be better genomics models.

AlphaGenome will improve. New training data will become available. Prediction accuracy will change. More biological modalities will eventually be modelled.

The Atlas architecture can benefit from those improvements without changing the basic idea.

Generate better predictions.

Rebuild the map.

Make the new map searchable.

That is why DeepMind describes AlphaGenome Atlas as a baseline rather than an endpoint.

From an AI infrastructure perspective, this is the part I find more interesting than the 1-petabyte headline.

We have spent years thinking about AI models as things you call when you need an answer.

AlphaGenome Atlas is another pattern:

run the model once, store the knowledge, and let other systems build on top of it.

For genomics, that means researchers can start with billions of possible mutations and progressively narrow the search to the variants, biological processes and regulatory regions that are actually worth investigating.

Nine billion variants is a huge number.

The real achievement is making that number searchable.

Sources

• Google DeepMind — AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
• Google Blog — AlphaGenome Atlas: a high-resolution map of human DNA
• Google DeepMind — AlphaGenome
• Nature — Advancing regulatory variant effect prediction with AlphaGenome
• PubMed — Advancing regulatory variant effect prediction with AlphaGenome
• Google DeepMind — AlphaGenome API