DeepMind unveils AlphaGenome 2 to trace how DNA changes affect specific cells

A disease clue found in 30 minutes

Google DeepMind unveiled AlphaGenome 2 on November 11, 2026, after training the artificial intelligence model on more than 107 million individual cell profiles across 68 tissues. Its central task is to predict which DNA changes affect which cells, helping researchers connect genetic clues to the biological processes behind disease.

In one demonstration, the model reproduced a key finding from a 2024 study of inflammatory bowel disease in about 30 minutes. Researchers had spent a year and a half conducting experiments to work out how a DNA variant increases activity of the ETS2 gene in immune cells involved in intestinal inflammation.

The variant sits in a stretch of DNA far from the gene it controls. AlphaGenome 2 selected it from seven candidates and predicted an effect on ETS2, about 290,000 DNA letters away. DeepMind highlighted its ability to identify increased ETS2 activity specifically in gut macrophages, immune cells that help defend the body and clear cellular debris.

James Lee, an immunologist and geneticist at the Francis Crick Institute in London who led the earlier study, described the computational result:

They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments.

The practical opportunity is to help laboratories decide which experiments deserve attention first. AlphaGenome 2 generates predictions about molecular activity; researchers still need laboratory tests to confirm its findings. DeepMind says the AlphaGenome models are intended for research and have not been validated or approved for clinical or diagnostic purposes.

Why the same DNA can behave differently

About 2% of the human genome contains instructions for making proteins. The remaining roughly 98% is called noncoding DNA. Some of it contains regulatory sequences that control when genes become active, where they operate and how much of their product a cell makes.

A useful comparison is a building with many rooms sharing the same electrical plans. Different switches are used in different rooms. Similarly, a heart muscle cell and an immune cell carry largely the same DNA but use different sets of genes to perform their jobs.

A DNA variant is a change in the sequence of DNA letters. A variant in a regulatory switch may increase a gene's activity in one cell type while having little effect in another. Its effect can also change when a cell becomes activated during infection or responds to vaccination.

DeepMind's AlphaGenome 2 announcement describes the new model as moving beyond broad tissue averages to individual cell types and states. That matters because a tissue contains a mixture of cells. Averaging their measurements can make an effect confined to a small group of cells harder to detect.

The inflammatory bowel disease example shows why location alone can be misleading. The variant lies in a gene desert, a region with few or no sequences that directly code for proteins. It forms part of an enhancer, a regulatory sequence that can influence a distant gene. Following that connection requires finding both the target gene and the cells where the switch operates.

A model built around the cell being studied

The original AlphaGenome, introduced in 2025, accepted a DNA sequence containing one million letters and predicted molecular properties across many tissues and experiments at once. Its outputs covered a fixed collection of tissue and cell conditions, making subtle differences between closely related cells difficult to capture.

AlphaGenome 2 instead takes both the DNA sequence and a biological condition as inputs. That condition can be a tissue, a cell type or a cell state. The model uses this information throughout its prediction process, focusing on one condition at a time.

Researchers call this a conditional model. In plain language, the question changes from what a DNA sequence generally does to what it does in a particular kind of cell under particular circumstances.

The training combined measurements from mixed tissue samples with data collected from individual human and mouse cells. It covered 637 human cell types and 311 mouse cell types. The more than 107 million profiles across 68 tissues give the model examples of how gene activity varies among specialized cells.

DeepMind's Korean launch statement describes the same redesign and the addition of DNA methylation predictions. DNA methylation involves chemical marks on DNA that can help regulate gene activity. Other predicted measurements include gene expression, chemical changes to proteins associated with DNA, and chromatin accessibility, which describes how exposed DNA is to the machinery that reads it.

What the performance figures mean

DeepMind evaluated AlphaGenome 2 using TenK10K, a large dataset covering 28 types of primary human immune cells. The company reported that the model identified 30% of causal variants in the evaluation, compared with 16% for the earlier AlphaGenome adapted to single cell data.

That is an improvement of 14 percentage points, or about 88% relative to the earlier figure. It explains DeepMind's description of the model as identifying nearly twice as many variants that change gene expression in specific cells.

The ability to identify the direction of a variant's effect also improved. The reported figure rose from 23% to 43%, a gain of 20 percentage points. Direction matters because a researcher needs to know whether a DNA change raises or lowers gene activity before deciding how to investigate its role in disease.

Another evaluation examined regulatory switches located between 100,000 and 2.5 million DNA letters from a gene. Against results from CRISPR interference screens, which reduce activity at selected DNA sites and measure the response, AlphaGenome 2 scored 0.62 on a measure combining precision and recall, compared with 0.48 for its predecessor.

Precision concerns how many selected candidates are correct; recall concerns how many relevant candidates the model finds. At a demanding confidence threshold, the research paper says the model still misses most variants thought to be causal. Scientists therefore face a tradeoff between obtaining a shorter list of stronger predictions and capturing a broader range of possible leads.

Learning from new cell states without retraining

AlphaGenome 2 can also use measurements from cells outside its original training collection. Researchers can upload single cell RNA sequencing data, which records gene activity in individual cells, from a sample containing just a few dozen cells.

The model compresses those measurements into a representation of the cell's state and uses it to guide predictions. This avoids retraining the entire model whenever a laboratory profiles a new biological condition.

Sarah Teichmann, cofounder of the Human Cell Atlas consortium and professor of stem cell medicine at the University of Cambridge, highlighted that feature in DeepMind's announcement:

What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells

In a demonstration using previously unseen data from a COVID-19 mRNA vaccine study, AlphaGenome 2 distinguished activated T cells from resting T cells. T cells are immune cells whose gene activity changes as they respond to a threat.

The model also predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in activated cells. The example makes cell state part of the genetic explanation: a variant's effect can depend on what the cell is doing when researchers examine it.

Researchers can also use one kind of measurement to predict another that is missing. For example, gene activity data can help the model estimate chromatin accessibility. This process, called cross modality imputation, could help a laboratory choose which additional measurements would be most useful.

From a genetic signal to a possible drug target

Genome wide association studies, usually shortened to GWAS, search for DNA differences associated with diseases or traits across many people. They often identify a region containing several possible variants. Researchers then have to determine which variant matters, which gene it affects and where in the body that effect occurs.

DeepMind has calculated AlphaGenome 2 predictions in advance for variants associated with traits across hundreds of human cell and tissue conditions. The aim is to make these connections easier to examine without requiring every laboratory to run large computing jobs itself.

The predictions also connect enhancers with target genes in primary human cells, including brain astrocytes, which support and protect neurons. Such connections can be difficult to map experimentally because many CRISPR screens use stable cells grown in laboratories rather than the specialized cells found in the body.

The inflammatory bowel disease demonstration brings these tasks together. The model connects a distant regulatory variant with ETS2 and identifies the relevant immune cells. The earlier experimental research had also identified potential treatment targets, giving scientists a biological finding against which to check the prediction.

Peter Donnelly, chief executive of Genomics Limited, explained the difficulty of turning genetic associations into drug research in DeepMind's announcement:

GWAS is central to that but it is difficult to move from an association signal to what a drug programme needs.

Donnelly said the company looked forward to incorporating AlphaGenome 2 into Mystra, its platform for evaluating drug targets using human genetics. No integration date or results from a drugmaker using the new model were announced.

How the nine billion variant Atlas fits in

AlphaGenome 2 follows another release in the same research programme: AlphaGenome Atlas, a catalogue containing predictions for all nine billion possible changes to a single DNA letter in the human genome. DeepMind released the Atlas about two months before announcing the new model.

The company's AlphaGenome research overview describes the Atlas as a dataset of one petabyte. It includes an AlphaGenome Variant Impact score, or AVI score, that helps scientists rank variants by their predicted molecular effects.

The two tools serve different computing needs. The original model provides broad coverage for a huge catalogue. AlphaGenome 2 focuses on more detailed questions about particular cells and states, with greater computing costs for its more specific predictions.

Ziga Avsec, the DeepMind researcher leading the AlphaGenome team, has described using the original model for extensive genome coverage and AlphaGenome 2 for closer investigation of disease mechanisms and possible drug targets. For researchers, this suggests a workflow that begins with a broad search and then examines selected candidates in the relevant cells.

The development has followed a short sequence:

• 2024: Lee's team established the ETS2 mechanism in inflammatory bowel disease through laboratory experiments.
• 2025: DeepMind introduced the original AlphaGenome model.
• About September 2026: AlphaGenome Atlas was released with predictions for nine billion possible changes to a single DNA letter.
• November 11, 2026: DeepMind announced AlphaGenome 2 and research access.
• November 30, 2026: The deadline for researchers to submit projects for possible promotion and additional computing quota or credits.

What researchers have already done with AlphaGenome

DeepMind's research overview gives examples of scientists using the wider AlphaGenome tools. At the Broad Institute, Laura Covill used AVI scores to identify a DNM1 variant associated with an error in gene processing. Laboratory experiments validated the prediction and supported reclassification of the variant.

At the University of Exeter, Gareth Hawkes used AlphaGenome in research involving more than 54,000 UK Biobank participants. The work examined rare noncoding variants and traits including plasma protein levels. In one region, the candidate list fell from 526 variants to four, illustrating how predictions can concentrate experimental attention.

Julia Zeitlinger at the Stowers Institute for Medical Research used AlphaGenome to investigate short regulatory DNA sequences and how their effects differ among cell types. Her work addresses the rules through which cells activate or reduce gene activity.

At the University of Copenhagen, Cesar Cunha and Marc Pielies Avelli, working under Tuomas Kilpelainen, used AlphaGenome to trace Alzheimer's genetic risk signals across tissues. Their analysis identified predicted regulatory effects in immune and metabolic tissues as well as the brain.

These examples show several practical routes from prediction to research: ranking rare disease candidates, reducing a long list of variants, examining regulatory sequences and tracing genetic signals across the body. AlphaGenome 2 adds a more detailed focus on the particular cells and states involved.

Access, costs and the next experiments

AlphaGenome 2 is available free for noncommercial research through the AlphaGenome application programming interface, or API, which lets researchers send requests to the model through software. Commercial researchers can seek access through a private preview. DeepMind has also provided interactive Colab tutorials.

DeepMind says it is working with researchers including teams at the Wellcome Sanger Institute to profile more human cell types, developmental stages and cells altered by different biological conditions. Those measurements can extend the range of states researchers are able to investigate.

The company also plans to release AlphaGenome 2 Skills in Google Antigravity, its platform for automated development and scientific workflows. It gave no release date. Researchers submitting use cases by November 30 may be selected for additional quota and credits.

For a laboratory, the immediate test is concrete: can the predictions help confirm useful candidates with fewer experiments, or reach a research decision sooner? The ETS2 demonstration offers an encouraging example of recovering an established mechanism. The wider value will depend on laboratories using the tool to investigate new candidates and checking them experimentally.

Key Points

• DeepMind announced AlphaGenome 2 on November 11, 2026, using more than 107 million individual cell profiles across 68 tissues.
• The model predicts DNA effects in specific cell types and states, including newly profiled states supplied by researchers.
• In the TenK10K evaluation, causal variant identification rose from 16% to 30%; identification of directional gene effects rose from 23% to 43%.
• A demonstration reproduced a key ETS2 finding in inflammatory bowel disease in about 30 minutes.
• AlphaGenome Atlas provides broad predictions for nine billion possible changes to a single DNA letter; AlphaGenome 2 supports more detailed cellular analysis.
• Access is free for noncommercial research, with commercial access through a private preview. Laboratory validation remains necessary, and the models are not approved for diagnosis.