AlphaGenome 2 Is Live: DeepMind's DNA Model Reads One Cell at a Time

Every cell in your body carries the same DNA. A heart muscle cell and an immune cell fighting an infection read that identical blueprint in completely different ways, switching different genes on and off. Predicting which single-letter change in DNA matters, and in which cell, has been one of the hardest problems in genetics, because most models could only predict effects averaged across whole tissues. On November 11, 2026, Google DeepMind announced AlphaGenome 2, a model built to answer the question cell by cell. It takes a stretch of DNA up to a million letters long plus one chosen cell type, and predicts how that DNA behaves in that cell. DeepMind says it finds nearly twice as many variants that change gene expression as its predecessor, and it is free for non-commercial research through an API. The limits matter as much as the headline.

Contents: What AlphaGenome 2 does | The numbers DeepMind claims | The 30-minute test | AlphaGenome 1 vs 2 | Cost and access | What it cannot do | Who should try it | FAQ | Sources

What AlphaGenome 2 actually does

AlphaGenome 2 is a conditional model. Instead of predicting molecular activity across many tissues at once, it predicts the effect of a DNA sequence for one condition at a time, a tissue, a cell type or a cell state. The condition can come from a reference catalog, or from a scientist's own single-cell experiment through a component DeepMind calls a cell encoder, which turns a snapshot of gene activity from just a few dozen cells into a usable condition with no retraining. DeepMind says the model can then match or outperform models trained or fine-tuned directly on those new datasets, including for cell states it has never seen, such as cells changed by disease or vaccination.

The model predicts the same core readouts as the first AlphaGenome, gene expression, histone modifications, chromatin accessibility and RNA splicing, and adds DNA methylation, a widely requested readout. It also does cross-modality imputation: give it one measurement, such as single-cell gene activity, and it can predict unmeasured properties, such as chromatin accessibility, which helps researchers decide which experiments to run next. The trade-off is compute. Every condition needs its own prediction run, so predictions across hundreds of cell types mean hundreds of runs. That is why the original AlphaGenome stays in service for broad scans, and why a genome-wide atlas built with AlphaGenome 2 would be impractical.

DeepMind trained the model on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. Nature's report adds that the training covered 637 types of human cell and 311 types of mouse cell. The first AlphaGenome, previewed in June 2025 and published in Nature in January 2026, was trained on bulk tissue samples, so it blurred specialized cells into tissue averages.

The numbers DeepMind is claiming

All performance figures below are DeepMind's own, from its announcement and the technical paper published alongside it. The paper is a preprint, not yet peer-reviewed, and no independent lab had replicated the results at launch.

TenK10K is a large single-cell map of 28 primary immune cell types. On it, AlphaGenome 2 identified nearly twice as many causal variants and got the direction of gene effects right nearly twice as often as the original model fine-tuned on single-cell data and other specialized tools. DeepMind also says the model tracks continuous changes in gene expression, such as blood stem cells maturing into red blood cells and immune lineages.

The 30% figure needs its other half stated plainly. At that precision setting, about seven in ten likely causal variants in the immune benchmark were still missed. DeepMind's own paper cautions that even at high precision, the model recovers only a minority of the expression-linked and disease-associated variants that statistical analyses nominate as likely causal. Doubling recall from 16% to 30% is real progress, and the gap between 30% and 100% is the honest measure of where the technology stands.

DeepMind also precomputed two resources for researchers who do not want to run large-scale comparisons themselves. One maps enhancer-gene links across 640 primary single-cell reference conditions and 39,323 protein-coding and long non-coding RNA genes. The other scores more than 6.8 million unique common and trait-associated single-letter variants across 510 human conditions. The same precomputed scores cover traits from height to type 2 diabetes, helping researchers connect genome-wide association study signals to the specific variants and cell types most likely involved.

The 30-minute test that caught Nature's attention

The most quoted result from the launch is not a benchmark table. It is a reproduction. A 2024 study published in Nature had traced part of inflammatory bowel disease risk to a variant in a distant regulatory switch that overactivates the ETS2 gene in gut macrophages. Establishing that mechanism took the research team about a year and a half of lab work. DeepMind says AlphaGenome 2 reproduced the finding computationally: it flagged the variant rs2836882 from a seven-variant credible set, linked it to ETS2 roughly 290 kilobases away, and localized the predicted effect to gut monocytes and macrophages, leaving other cell types and genes largely unaffected.

James Lee, the immunologist who led the 2024 study, told Nature: "They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments." Read that quote carefully. The model reproduced a mechanism that was already established experimentally. It did not discover a new one, and reproducing a known result computationally is not the same as producing a new wet-lab result. What it does show is that the model can single out the right variant, the right gene and the right cell type from sequence alone, which is the step researchers need before they commit months to a bench experiment.

DeepMind reports a second example. Given single-cell data from a COVID-19 mRNA vaccine study it had never seen, AlphaGenome 2 distinguished vaccine-activated immune T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect that would be missed in unactivated cells.

AlphaGenome 1 vs AlphaGenome 2

AlphaGenome 2 is a separate model release, not an extension of the AlphaGenome Atlas. The Atlas, released in September 2026, contains precomputed predictions for all 9 billion possible single-letter changes in the human genome. The new model is for asking your own questions about specific cells, not for looking up a precomputed answer.

This is not the only DNA model aimed at researchers this year. In October we covered Hugging Face's Carbon-X, which also predicts variant effects from DNA sequence, with its own limits around wet-lab validation. Carbon-X is an open model card on Hugging Face; AlphaGenome 2 is an API-only release from a single lab. They are different tools with different access terms.

What it costs and how to get access

AlphaGenome 2 has no price tag in the usual sense. The AlphaGenome API documentation states that the API is offered free of charge for non-commercial use, with query rates that vary with demand. Atlas lookups allow higher rates than live model predictions, which suit analyses of thousands of predictions rather than millions. There is no published per-query price and no subscription tier, so there is no USD, GBP or CAD figure to convert.

Commercial research access is in private preview. DeepMind has published no price, no timeline for general availability and no terms for the preview, so any company planning around it is working from an application, not a rate card. Interactive Colab tutorials for expression conditioning are available now alongside the API, so a researcher can try the conditioning workflow without setting up infrastructure.

Two things are coming but undated. DeepMind says it will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, and the lab is collecting community use cases through a form open until November 30, 2026, with selected projects eligible for additional quota and credits. There is no download: access runs through the AlphaGenome API, and the announcement mentions no release of model weights, so local or offline use is not an option at launch. DeepMind is working with the Wellcome Sanger Institute to profile more cell types and states, and Genomics Limited chief executive Peter Donnelly said the company looks forward to incorporating AlphaGenome 2 into its Mystra platform.

What it cannot do

DeepMind states the central limit twice in its own announcement, and it is worth repeating in full: AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model predicts molecular effects. It does not diagnose anyone, and it cannot tell you what will happen to you. Anyone presenting it as a medical test is misrepresenting it.

The paper's own cautions go further. The model can underestimate effects outside accessible chromatin, at very distant regulatory elements, or through mechanisms that act after transcription. Transient, stimulated and disease states remain underrepresented in training, and fine-grained neuronal subpopulations remain difficult to distinguish. The condition representations come from observational, steady-state profiles, so they capture correlations rather than causal relationships, and it remains uncertain whether the model can predict responses to gene knockouts, overexpression or dynamic cell-state transitions. Every hypothesis it generates still needs experimental follow-up.

The benchmark limits matter too. The headline test is an immune-cell dataset, so performance elsewhere is less established. The preprint has not been peer-reviewed, and the comparisons are DeepMind's own tests against its earlier model and tools it selected. Until independent groups rerun them, the 30% and 43% figures are launch-day results from the lab that built the model.

DeepMind has had a busy November. Days after this announcement, the lab also published AlphaProof Code results on Signal's cryptography, a different kind of AI verification story from the same lab.

Who should try it, and who should wait

Try it this week if you work in genomics or drug discovery and have a variant list that needs narrowing. A genome-wide association study hands you a region and a list of candidate variants; what you need next is which variant, which gene and which cell. AlphaGenome 2 is built to narrow exactly that search, and the free non-commercial API plus the Colab tutorials make it a cheap experiment to run.

Wait if you are a clinician, a patient, or anyone looking for a diagnosis or a health risk score. Wait as well if you need commercial terms, a service-level agreement or local deployment, because none of those exist yet. If you work outside immunology, treat the headline numbers as a starting point: the flagship benchmark is an immune dataset, and the model's own authors say transient and disease states are underrepresented in training.

Frequently asked questions

Is AlphaGenome 2 free to use?

Yes, for non-commercial research. The AlphaGenome API documentation states the API is offered free of charge for non-commercial use, with query rates that vary with demand. Commercial research access is in private preview, and DeepMind has published no price for it.

Can AlphaGenome 2 diagnose disease or predict my health risks?

No. DeepMind states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model predicts how DNA variants change gene activity in cells. It is not a medical test.

How is AlphaGenome 2 different from the first AlphaGenome?

The first AlphaGenome predicted molecular activity across broad tissues and cell lines at once from a million-letter DNA sequence. AlphaGenome 2 is conditional: it takes the DNA sequence plus one chosen tissue, cell type or state, adds DNA methylation as a readout, and accepts a condition built from a researcher's own single-cell data, with no retraining.

Can I download and run AlphaGenome 2 locally?

No. Access runs through the AlphaGenome API. DeepMind's announcement mentions no release of model weights, and the model is not available as a download at launch.

What is the TenK10K test?

TenK10K is a large single-cell map of 28 primary immune cell types that DeepMind used as its headline benchmark. On it, AlphaGenome 2 identified 30% of putatively causal expression-linked variants at 90% precision, against 16% for the original AlphaGenome fine-tuned on single-cell data. The figures are DeepMind's own, from its preprint.

Is AlphaGenome 2 the same thing as the AlphaGenome Atlas?

No. The Atlas, released in September 2026, is a precomputed resource with predictions for all 9 billion possible single-letter changes in the human genome. AlphaGenome 2 is a separate model release for asking your own questions about specific cell types and states.

AlphaGenome 2 is a sharper lens for researchers, not a finished answer. It finds nearly twice as many causal variants as its predecessor on DeepMind's own benchmark, works from a researcher's own single-cell data without retraining, and is free for non-commercial research today. Its limits are stated in its own announcement: research only, no clinical validation, most causal variants still missed, and no independent replication yet. If you have single-cell data and a variant you care about, the free API makes it worth an afternoon. Start with DeepMind's announcement and the API documentation linked below.

Sources

• Google DeepMind, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease," November 11, 2026. deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
• Ewen Callaway, "DeepMind's ultra-powerful AI for genome mapping wows scientists," Nature, November 11, 2026. nature.com/articles/d41586-026-03673-0
• AlphaGenome documentation, API access and terms. alphagenomedocs.com
• Unite.AI, "Google DeepMind's AlphaGenome 2 Predicts Cell-State Variant Effects," November 11, 2026. unite.ai/google-deepminds-alphagenome-2-predicts-cell-state-variant-effects/
• MLQ, "Google DeepMind releases AlphaGenome 2 to predict how DNA changes affect specific cell types and disease," November 11, 2026. mlq.ai/news/google-deepmind-releases-alphagenome-2-to-predict-how-dna-changes-affect-specific-cell-types-and-disease/
• Paul Lovelace, "AlphaGenome 2 Predicts How DNA Variants Act in Specific Cells," DEV Community, November 11, 2026. dev.to/paullovelace/alphagenome-2-predicts-how-dna-variants-act-in-specific-cells-twice-the-recall-of-its-predecessor-29p9

This is general information, not medical or legal advice.