Google DeepMind has launched AlphaGenome 2, a DNA model that predicts how a single-letter change in DNA affects one specific cell type at a time. Announced on 11 November 2026, it is the follow-up to the original AlphaGenome, and it answers a question the first model could not: not just what a genetic variant does, but in which cell it does it.

That distinction matters because every cell in the body carries the same DNA, yet a heart muscle cell and an immune cell read that DNA very differently. A variant that raises the risk of bowel disease may act only in gut immune cells, while leaving every other cell untouched. Tissue averages hide that detail. AlphaGenome 2 is built to find it.

On DeepMind’s headline test, the new model found nearly twice as many gene-changing variants as its predecessor: 30 percent against 16 percent. The figures are DeepMind’s own, from a technical paper that had not been peer reviewed at launch, so they are a strong signal rather than a settled fact. Even so, the early reaction has been warm. Nature reported that scientists were impressed, and the immunologist whose bowel disease finding the model reproduced said it reached the answer in 30 minutes, against a year and a half of laboratory experiments.

Key takeaways

  • AlphaGenome 2 predicts how DNA variants affect individual cell types and states, not tissue averages.
  • It was trained on more than 107 million individual cell profiles across 68 tissues, from human and mouse samples.
  • On the TenK10K immune benchmark it identified 30 percent of causal variants, against 16 percent for the original AlphaGenome, and called the direction of the effect 43 percent of the time, against 23 percent.
  • It reproduced a 2024 bowel disease finding computationally, pinpointing the variant that overactivates the ETS2 gene in gut macrophages.
  • It is available now for non-commercial research through the AlphaGenome API. It is for research only and has not been validated or approved for clinical or diagnostic use.

What AlphaGenome 2 does differently

The first AlphaGenome, previewed in June 2025 and published in Nature in January 2026, took a DNA sequence of up to a million letters and predicted its effects across thousands of tissue samples all at once. That parallel approach was efficient, but it spread the model’s capacity thin and made it hard to capture subtle differences between closely related cell types. A variant that matters in one rare cell type could disappear into the average.

AlphaGenome 2 redesigns that step. It is a conditional model: alongside the million-letter DNA sequence it takes a specific biological condition as input, one tissue, cell type or cell state at a time. The condition can be chosen from a reference catalogue of cell types, or generated on the fly from a scientist’s own single-cell experiment through a cell encoder. Because the model sees the condition at every stage of its architecture, it learns how the same stretch of DNA is read differently across cell types.

The redesign also adds a new readout. Like its predecessor, AlphaGenome 2 predicts gene expression, chromatin accessibility and histone modifications. It now also predicts DNA methylation, a widely requested readout that describes chemical marks on DNA that help switch genes on or off. And because the model separates a cell’s biological state from its outputs, a scientist who supplies just one measurement, such as single-cell gene activity, can get predictions for unmeasured properties, such as chromatin accessibility, helping decide which experiments are worth running next.

Nearly twice as many variants found

The headline numbers come from TenK10K, a large single-cell map of 28 primary immune cell types. On that dataset, AlphaGenome 2 reliably identified 30 percent of the variants believed to cause changes in gene expression, against 16 percent for the original AlphaGenome fine-tuned on single-cell data and other specialised tools. It also called the direction of a variant’s effect on its gene, whether it turns activity up or down, 43 percent of the time, against 23 percent before. DeepMind reports the variant recovery at 90 percent precision, meaning most of the variants it flags are genuine.

Those numbers deserve honest framing. Thirty percent is nearly twice sixteen, but it also means roughly seven in ten likely causal variants still went unnoticed, and the direction of the effect was resolved less than half the time. The test is also an immune-cell dataset, so performance in other parts of the body is less established.

AlphaGenome AlphaGenome 2
Cell resolution Broad tissue averages Individual cell types and states
Training data Bulk tissue data Bulk and single-cell data, over 107 million cell profiles across 68 tissues
Outputs Gene expression, chromatin accessibility, histone modifications Those three, plus DNA methylation
Causal variants found on TenK10K 16 percent 30 percent at 90 percent precision
Direction of effect called 23 percent 43 percent
Access AlphaGenome API AlphaGenome API for non-commercial research, commercial research in private preview

The bowel disease test: 30 minutes against 18 months

The most striking demonstration is a case where the answer was already known. In 2024, researchers showed that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut immune cells called macrophages. It took a year and a half of experiments to establish that chain.

DeepMind asked AlphaGenome 2 to make predictions across immune cell types in the gut. The model reproduced the finding computationally: it singled out the causal variant and showed that it selectively boosts the ETS2 gene in gut macrophages, leaving other cell types and genes largely unaffected. James Lee, the immunologist and geneticist at the Francis Crick Institute in London who led the 2024 study, told Nature that the model reached the answer in 30 minutes, against a year and a half of experiments.

That is a demonstration, not a discovery. The model confirmed something scientists had already proven in the laboratory. Its value is speed and scale: if it can reproduce a known result in half an hour, researchers can use it to prioritise which of thousands of candidate variants deserve expensive laboratory follow-up, and which can wait.

New cell types without retraining

AlphaGenome 2 can also generalise to single-cell experiments it has never seen, without being retrained. Scientists can upload a single-cell RNA sequencing snapshot of gene activity from just a few dozen cells. That activity is compressed into a cell state representation, which conditions the model on the unseen state. DeepMind reports that it can match or outperform models trained or fine-tuned directly on those datasets.

The example DeepMind gives comes from a COVID-19 mRNA vaccine study. Given unseen single-cell data, AlphaGenome 2 distinguished vaccine-activated immune T cells from resting ones. It also revealed that a variant linked to severe COVID-19 disrupts gene activity more strongly when cells are in the activated state, an effect that would be missed in unactivated cells. The same variant can behave differently depending on what the cell is doing at the time.

From gene maps to disease clues

One of the hardest problems in human genetics is connecting a variant to a disease. Genome-wide association studies flag broad stretches of DNA linked to disease, but they rarely identify the exact causal variant, the gene it acts on, or the cell type where it acts. Many disease-linked variants sit in distant regulatory switches called enhancers, far from the genes they control, which makes their effects hard to trace.

To help researchers make those connections without running large computations themselves, DeepMind precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions. Enhancer to gene mapping now extends to primary human cell types, such as brain astrocytes, the cells that support and protect neurons. Across complex traits ranging from height to type 2 diabetes, the precomputed scores help pinpoint the specific variants and cell types most likely to be involved, and when the model is confident it also predicts whether a variant turns a gene’s activity up or down. Static cell atlases describe what cells look like. AlphaGenome 2 turns those atlases into a predictive engine that asks what happens when the DNA changes.

How to get access, and what it is not

AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with interactive Colab tutorials for expression conditioning, and in private preview for commercial research. DeepMind also says it will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, pairing the specialised model with Gemini 4 for research workflows. The company is inviting researchers to share use cases through a form open until 30 November, with selected projects eligible for additional quota and credits, and it names the Wellcome Sanger Institute as an ongoing collaborator.

What it is not matters as much as what it is. DeepMind states plainly that both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model predicts molecular effects, not disease, and it cannot be downloaded and run locally. It also does not replace its predecessor: AlphaGenome 2 predictions cost more compute, so a genome-wide atlas of the kind built with the first model, covering all 9 billion possible single-letter changes in the human genome, is impractical with it.

DeepMind’s week has been busy on verification as well as prediction. Earlier this week the lab showed AI agents mathematically proving parts of Signal’s encryption code, a different kind of verification, but the same pattern: AI systems that check, rather than merely generate. Its Gemini Robotics robots have also begun factory work at Mercedes, sorting car parts on a production line.

Frequently asked questions

What is AlphaGenome 2?

AlphaGenome 2 is Google DeepMind’s next-generation DNA sequence model, announced on 11 November 2026. Given a DNA sequence of up to a million letters and a chosen cell type or state, it predicts how DNA variants affect gene activity in that specific cell, including gene expression, chromatin accessibility, histone modifications and DNA methylation.

How is it different from the first AlphaGenome?

The first AlphaGenome predicted variant effects averaged across broad tissues. AlphaGenome 2 is a conditional model that predicts effects for one cell type or state at a time, trained on more than 107 million individual cell profiles across 68 tissues. On the TenK10K immune benchmark it found nearly twice as many causal variants: 30 percent against 16 percent.

Can AlphaGenome 2 diagnose disease?

No. DeepMind states that both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model predicts molecular effects of DNA variants, which helps researchers prioritise laboratory work. It does not diagnose patients or predict who will develop a disease.

Who can use AlphaGenome 2?

Non-commercial researchers can use it now through the AlphaGenome API, with interactive Colab tutorials available. Commercial research access is in private preview. The model itself cannot be downloaded and run locally.

Sources and Further Reading