AlphaGenome 2: DeepMind’s DNA Model Now Reads Cell by Cell

Google DeepMind's updated genetics model predicts what a DNA change does inside a single cell type, but its headline results are the company's own.

Jach Hoggen
16 Min Read
A rendering of the DNA double helix, shown for illustration. Image: PantheraLeo1359531 / Wikimedia Commons, CC0
Highlights
  • Google DeepMind released AlphaGenome 2 on November 11, 2026, a model that predicts how DNA variants change gene activity in specific cell types.
  • DeepMind says it was trained on more than 100 million individual cell profiles from 68 human and mouse tissues.
  • On an immune-cell benchmark, the company reports it found causal variants 30% of the time, up from 16% for earlier tools (DeepMind, November 2026).
  • Non-commercial researchers can use it through the AlphaGenome API now; commercial research access is a private preview.
  • DeepMind states the model is for research only and has not been validated or approved for clinical or diagnostic use.

Google DeepMind released AlphaGenome 2 on November 11, 2026. It is an artificial intelligence model that predicts how a change in DNA alters gene activity inside a specific cell type, not just across a whole tissue. Non-commercial researchers can use it through an API today. However, the company says it is not approved for diagnosing or treating anyone.

The shift matters because nearly every cell in the body carries the same DNA but uses it differently. For example, a variant that does nothing in a liver cell can disrupt an immune cell. A tissue-wide average can hide that effect.

The performance figures below come from DeepMind’s own testing, and no outside group has reproduced them yet. Here is what the model does, how it differs from the 2025 original, and what it means for researchers and patients in Canada and the US.

What Did Google DeepMind Announce?

AlphaGenome 2 is a next-generation DNA sequence model. DeepMind says it predicts how DNA sequences and variants affect specific human cell types, across hundreds of them. Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec wrote the announcement. Source: Google DeepMind.

The first AlphaGenome arrived on June 25, 2025. It predicted molecular activity across broad tissues and cell lines. DeepMind later added AlphaGenome Atlas, a set of precomputed predictions for all 9 billion possible single-letter changes in the human genome.

The new version keeps the readouts scientists already used, including gene expression, histone modifications and chromatin accessibility. In addition, it predicts DNA methylation, which DeepMind calls a widely requested readout. Source: SiliconANGLE.

DeepMind links to a paper describing the work. However, the post does not name a journal, and it is not clear whether the paper has been peer reviewed.

How Does AlphaGenome 2 Work?

The model takes a DNA sequence 1 million letters long, plus a description of the condition a scientist cares about. That condition can be a tissue, a cell type or a cell state. For instance, it could be a resting immune cell or an activated one.

DeepMind calls this a conditional design. In contrast, the first model produced a fixed set of parallel outputs for bulk tissues and cell lines. The company now says that approach spread the model’s capacity thin.

AlphaGenome 2 predicts for one condition at a time and feeds that condition into every stage of the network. As a result, three features follow, according to the announcement:

  • A cell encoder: researchers can generate a condition from their own single-cell experiment instead of relying on a preset list.
  • No retraining for new cell states: DeepMind says the model can work from a single-cell RNA snapshot of a few dozen cells.
  • Cross-modality imputation: from one measurement, such as gene activity, the model predicts properties that were not measured, such as chromatin accessibility.

The Numbers Behind the Launch

DeepMind’s post gives two figures for the training data. The introduction says more than 100 million individual cell profiles across 68 tissues. The body, however, says over 107 million. Both describe a mix of bulk and single-cell data from human and mouse samples.

The headline benchmark uses TenK10K, a dataset covering 28 primary immune cell types. DeepMind reports that AlphaGenome 2 identified causal variants 30% of the time, against 16% for the comparison tools. Those tools included the first AlphaGenome fine-tuned on single-cell data. Source: Google DeepMind.

100M+

Individual cell profiles used in training, per DeepMind

68

Human and mouse tissues covered

30% vs 16%

Causal variants found on the TenK10K immune-cell benchmark

1 million

DNA letters read per prediction

The model also predicted the direction of a variant’s effect on a gene, up or down. It was right 43% of the time versus 23% for the comparison tools, according to Google DeepMind. Across its tests, the company says it finds nearly twice as many expression-changing variants as the previous best model.

These numbers also show the ceiling. Specifically, a 30% hit rate means the model still misses most causal variants in that benchmark. It narrows a search more than it settles one.

How Does It Compare With the First AlphaGenome?

The original model was peer reviewed. Nature published it on January 28, 2026. The paper reported that it takes 1 megabase of DNA as input and matched or exceeded the strongest external models in 25 of 26 variant-effect evaluations. Source: Nature, via PubMed Central.

DeepMind itself listed that model’s weak spots in 2025. In particular, capturing cell- and tissue-specific patterns needed work. Regulatory elements more than 100,000 letters from a gene were also hard to predict. Source: Google DeepMind.

AlphaGenome 2 targets the first of those gaps. The table below sets out the reported differences.

FeatureAlphaGenome (2025)AlphaGenome 2 (2026)
AnnouncedJune 25, 2025November 11, 2026
DNA inputUp to 1 million letters1 million letters plus a target condition
Output designFixed parallel outputs for bulk tissues and cell linesOne tissue, cell type or cell state at a time
DNA methylationNot predictedPredicted
New cell statesLimited to the conditions it was trained onAdapts without retraining, per DeepMind
Peer reviewPublished in Nature, January 2026Paper linked; review status not confirmed
Clinical useNot validatedNot validated
Sources: Google DeepMind announcements of June 25, 2025 and November 11, 2026; Nature.

Why Do Cell Types Matter for Disease Research?

Only about 2% of the human genome codes for proteins. The other 98% is non-coding DNA that controls when and where genes switch on, and it holds many disease-linked variants. Source: Google DeepMind.

Those control regions behave differently from one cell type to the next. DeepMind’s inflammatory bowel disease example shows why. Earlier research found that a variant in a distant regulatory switch overactivates the ETS2 gene in gut immune cells. The company says AlphaGenome 2 reproduced that finding by computer and singled out the causal variant.

In a second test, the model received single-cell data it had never seen from an mRNA vaccine study. DeepMind reports it told activated T cells from resting ones. It also showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells.

Drug developers are the obvious audience. For example, Genomics Limited, a company led by chief executive Peter Donnelly, plans to build the model into its Mystra platform. DeepMind also says it is collaborating with researchers at the Wellcome Sanger Institute.

Human genetics is one of the most powerful tools in drug discovery.

Peter Donnelly, CEO of Genomics Limited, quoted in Google DeepMind's announcement

What Does It Mean for Canada and the US?

For patients, nothing changes today. DeepMind states that AlphaGenome and AlphaGenome 2 are intended for research use only. They have not been validated or approved for clinical or diagnostic purposes. Therefore, no clinic in either country should be using the model to read a patient’s genome.

For university and hospital research labs, access is immediate. The model is available for non-commercial research through the AlphaGenome API, with interactive Colab tutorials. The post lists no country restrictions. That said, researchers should read the API terms before sending data.

Companies face a slower route. Commercial research access is a private preview, and DeepMind’s post gives no pricing. Meanwhile, the model’s weights have not been released, so users cannot run it on their own servers. Source: Startup Fortune.

The research need is real in Canada. Health Canada has said one in twelve people in the country, many of them children, has a rare disease. In March 2023, Ottawa committed up to $1.5 billion over three years to rare-disease drugs. Tools that sort harmless variants from harmful ones feed the earliest stage of that work. Source: Health Canada.

A Fair Counterpoint: The Results Are Self-Reported

A fair counterpoint is that every benchmark above comes from the company that built the model. Startup Fortune noted on launch day that it had seen no third-party replication. Similarly, no reaction from independent scientists had appeared in the coverage reviewed for this article in the first hours after release.

The reaction to the first AlphaGenome is a useful guide. When Nature published that paper, Ben Lehner of the Wellcome Sanger Institute said the model was “far from perfect”. Aldo Faisal of Imperial College London advised treating the results as promising until independent groups reproduce them. Source: Science Media Centre.

Xianghua Li of King’s College London said in the same roundup that such predictions “are not yet ready for use in clinics.” By DeepMind’s own statement, that caution applies equally to the new version.

Other known limits may persist. IEEE Spectrum reported in February 2026 that the original model leaned toward false negatives, meaning it could miss variants that matter. It also struggled with regulatory regions far from their target genes. Christina Leslie of Memorial Sloan Kettering Cancer Center called generalization to new cell types “a huge limitation.” Source: IEEE Spectrum.

AlphaGenome 2 is DeepMind’s answer to Leslie’s point. Nevertheless, whether it holds up outside the company’s benchmarks is unproven. The same question hangs over other recent AI launches built on vendor-run tests, including OpenAI’s GPT-6 Cyber and Anthropic’s latest Claude models.

There is a commercial question as well. Startup Fortune pointed out that DeepMind’s sister company, Isomorphic Labs, works in drug discovery. DeepMind has not said how it will handle that overlap for firms building on its API.

What Happens Next and What Readers Can Do

Independent testing is the next milestone. Academic groups can now run the model on their own data. Their results will show whether the reported gains hold on cell types and diseases DeepMind did not test.

DeepMind says AlphaGenome 2 Skills for its Google Antigravity tool will be released soon, without giving a date. In addition, it is asking researchers to submit use cases by November 30. Selected projects may receive additional quota and credits.

Practical steps depend on who you are:

  • Patients and families: do not expect this model in a clinic or a consumer DNA test. Questions about a genetic result belong with a doctor or genetic counsellor.
  • Academic researchers: the API and Colab tutorials are open for non-commercial work now, and the use-case call closes November 30.
  • Biotech companies: commercial access requires the private preview, and pricing has not been disclosed.
  • Everyone else: treat the 30% and 43% figures as company-reported until outside groups publish their own tests.

Frequently Asked Questions

What is AlphaGenome 2?

AlphaGenome 2 is a Google DeepMind AI model released on November 11, 2026. It reads a DNA sequence 1 million letters long and predicts how genetic variants change gene activity in a chosen tissue, cell type or cell state. It updates the first AlphaGenome from June 2025, which worked mainly at the level of bulk tissues.

Can doctors use AlphaGenome 2 to diagnose genetic diseases?

No. DeepMind says both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. Its predictions are meant to help scientists decide which variants to test in the lab.

How accurate is AlphaGenome 2?

DeepMind reports that on the TenK10K immune-cell benchmark it identified causal variants 30% of the time, compared with 16% for earlier tools. It predicted the direction of gene effects 43% of the time versus 23%. However, these are company-reported results that outside groups have not yet reproduced.

Who can access AlphaGenome 2, and what does it cost?

Non-commercial researchers can use it now through the AlphaGenome API. Commercial research access is a private preview. DeepMind’s announcement does not give pricing, and the model’s weights have not been released.

How is it different from the first AlphaGenome?

The first model gave fixed outputs for bulk tissues and cell lines. In contrast, AlphaGenome 2 predicts for one condition at a time and adds DNA methylation as a readout. DeepMind also says it adapts to new single-cell data without retraining.

Sources and Further Reading

This article was researched with AI assistance and checked against the sources listed above. Facts are current as of the publication date; follow the source links for updates.

Share This Article
Leave a Comment

Leave a Reply