AlphaGenome 2: Google's DNA AI predicts what gene variants do in each cell type

What's new

Google DeepMind unveiled AlphaGenome 2 on November 11, 2026. The AI model predicts how stretches of DNA and their variants affect individual human cell types. Variants are small differences in the genome, such as a single changed "letter." The original AlphaGenome from 2025 only looked at averages across whole tissues. The new version goes down to specific cell types and cell states, such as heart muscle cells or immune cells in the middle of fighting an infection.

AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, a programming interface that lets you query the model from your own software. According to the official documentation, this access is free, and how many requests you can make depends on demand. Commercial research gets a private preview with limited access. DeepMind is also offering interactive Colab tutorials you can work through in your browser. "AlphaGenome 2 Skills" for Antigravity, Google's agent development platform, are due soon. DeepMind is not releasing the model weights.

Why cell type matters so much

Almost every cell in your body carries the same DNA. But each cell type switches different genes on and off, and that pattern also shifts depending on what the cell is doing at the moment. That's why the same variant can have big consequences in one kind of cell and none at all in another. If you only look at the average across a tissue, an effect that shows up in a single rare cell type easily gets drowned out. It's a bit like trying to pick out one voice in the roar of a packed stadium.

Projects like the Human Cell Atlas map which genes are active in millions of individual cells (known as gene expression). Until now, though, it has been hard to predict how small DNA differences change that activity. That's the gap AlphaGenome 2 is meant to fill. DeepMind says the model turns static cell atlases into a "predictive engine for human genetics." Žiga Avsec, who leads genomics at DeepMind, told The Verge:

"When you develop drugs, you target specific proteins and specific cell types."

His colleague Adam Gayoso sees the real breakthrough in "connecting human genetics to cell biology."

How the model works

The first AlphaGenome read a DNA sequence of one million letters and predicted thousands of measurements across many tissues from it at once. That was efficient, but it spread the model's capacity thin. Subtle differences between closely related cell types got lost.

AlphaGenome 2 is instead a conditional model: it only computes for one condition at a time, meaning one tissue, one cell type or one cell state. It receives that condition as input alongside the DNA sequence at every stage of its architecture. That way it learns how the same DNA is "read" differently in different cells. As before, it predicts gene expression, RNA splicing (the trimming of gene transcripts), histone modifications (chemical marks on the proteins DNA is wrapped around) and chromatin accessibility (how open a stretch of DNA is to regulation). New is DNA methylation, a much-requested addition: a chemical mark that often silences genes.

Training covered tissue and single-cell data from humans and mice, including more than 107 million single-cell profiles from 68 tissues. The design enables two capabilities:

• New cell states without retraining: Researchers upload a single-cell RNA measurement. A few dozen cells are enough. A cell encoder condenses them into a description of the cell state, which the model then uses for its predictions. DeepMind says it matches or beats models trained specifically on that data.
• Filling in missing measurements: If you've only measured gene activity, the model can predict other properties such as chromatin accessibility. That helps you decide which experiment is worth running next.

The trade-off is that predictions cost more: every condition needs its own forward pass. Avsec says that's why AlphaGenome 2 doesn't replace its predecessor. The AlphaGenome Atlas released in September wouldn't have been feasible with the new model. According to HPCwire, it spans roughly one petabyte of predictions for all of the roughly 9 billion possible single-letter changes in the human genome.

Numbers from the paper

In its technical paper, DeepMind reports these figures, among others (company-reported):

There are also two case studies. For inflammatory bowel disease (IBD), the model picked variant rs2836882 out of seven candidates. It linked it to the ETS2 gene, about 290,000 bases away, and placed its effect in gut monocytes and macrophages, the immune system's scavenger cells. That matches a mechanism a major study had previously confirmed in the lab. In the second example, the model was given previously unseen data from a study on mRNA vaccination against Covid-19. It distinguished vaccine-activated T cells from resting ones and predicted that a variant linked to severe Covid has a strong effect mainly in the activated state. In resting cells, that effect would have gone unnoticed.

Precomputed data and early users

So that not every research group has to run big computations itself, DeepMind has precomputed results. According to Unite.ai, one is a map of enhancer-gene links across 640 single-cell reference conditions and 39,323 genes. Enhancers are regulatory switches in DNA that often sit far away from the gene they control. Until now, such maps have mostly come from CRISPR experiments on lab cell lines. Now they exist for real human cell types too, such as astrocytes in the brain. The other is a set of scores for more than 6.8 million common and trait-associated variants across 510 human conditions. These are meant to help trace hits from genome-wide association studies (GWAS, large comparisons of many people's genomes) back to a specific variant, target gene and cell type, for traits ranging from height to type 2 diabetes.

Peter Donnelly, CEO of Genomics PLC, plans to build AlphaGenome 2 into his company's Mystra platform. DeepMind is working with the Wellcome Sanger Institute, among others, to cover more cell types and cell states. Avsec said future versions could be trained on data from "multiple species." In the long run, models like this could feed into "virtual cells," digital simulations of biological systems. DeepMind says that if you submit interesting use cases by November 30, you can get extra quota and credits.

The bottom line

AlphaGenome 2 is a research tool, not diagnostic software. DeepMind stresses that neither version has been validated or approved for clinical or diagnostic use. The paper itself lists clear limitations. Even at high precision, the model only finds a minority of the statistically suspected causal variants. It may underestimate effects outside accessible DNA regions, from very distant switches, or after genes have been transcribed. Short-lived or disease-related cell states are underrepresented in the training data, and it struggles to tell apart fine subtypes of neurons. Because the cell states come from observational data, the model captures correlations, not causes. Whether it can predict the consequences of, say, knocking out a gene remains an open question.

So its value lies in triage: out of thousands of candidates, it can filter out the few that are worth testing in the lab. The results remain hypotheses that need experimental confirmation. So far, the benchmarks all come from DeepMind itself. AlphaGenome 2 is part of a race among the big AI labs over scientific applications. According to The Verge, Anthropic is experimenting with drug design and a wet lab for Claude, while OpenAI is working on open math problems. DeepMind has the experience of AlphaFold behind it, which earned Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry.