Artificial Intelligence

Google DeepMind's AlphaGenome 2 Predicts How DNA Variants Act in Specific Cell Types

DeepMind says the research-only model, trained on more than 100 million individual cell profiles, finds nearly twice as many expression-changing variants as the original AlphaGenome; it is free for non-commercial research through the AlphaGenome API.

DailyMech Staff

Published

Last updated / Fact-checked on

Google DeepMind's AlphaGenome 2 Predicts How DNA Variants Act in Specific Cell Types
PublicDomainPictures / Pixabay — Pixabay Content License

A glass pipette with a droplet above laboratory test tubes holding green and purple liquid

Illustrative stock photo: a pipette above laboratory test tubes. Not AlphaGenome 2 or DeepMind research.

Google DeepMind on Wednesday announced AlphaGenome 2, which it says "predicts how DNA sequences and their variations impact specific human cell types." DeepMind says it identifies nearly twice as many variants that change gene expression as its previous model, a claim DailyMech has not independently verified.

DeepMind says the model was trained on more than 100 million individual cell profiles across 68 tissues. It is available now for non-commercial research through the AlphaGenome API and in private preview for commercial research. DeepMind says AlphaGenome and AlphaGenome 2 "are intended for research use only and have not been validated or approved for clinical or diagnostic purposes."

The announcement comes a day after Google opened Gemini 4 Carbon to paying developers.

At a glance

Item Detail Source/status
What it does Predicts how DNA variants affect gene regulation in a chosen tissue, cell type or state DeepMind blog (fully read)
Training data More than 100 million individual cell profiles across 68 tissues DeepMind
Headline claim Nearly twice as many expression-changing variants as the original AlphaGenome DeepMind's benchmark; not independently verified
New cell states Works from a few dozen cells' single-cell data without retraining DeepMind
Access Non-commercial research via AlphaGenome API; commercial research in private preview DeepMind
Limits Research use only; not validated or approved for clinical or diagnostic use DeepMind

What is different from the original AlphaGenome

DeepMind says the original AlphaGenome, launched last year, took a million-letter DNA sequence and produced fixed, parallel outputs for bulk tissues and cell lines.

AlphaGenome 2 is instead a conditional model that predicts for one condition at a time, defined as a tissue, cell type or state. DeepMind says the condition can come from a reference catalog or from a scientist's own single-cell experiment, through a component it calls a cell encoder. Scientists can upload single-cell RNA-seq data from just a few dozen cells without retraining, the company says.

DeepMind also describes cross-modality imputation: given one measurement, such as single-cell gene activity, the model can predict properties that were not measured, such as chromatin accessibility, which DeepMind says can help researchers prioritize experiments. Readouts include gene expression, histone modifications, chromatin accessibility and, new in this version, DNA methylation.

DeepMind says it trained AlphaGenome 2 on both bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles. Nature reports the model was trained on 637 human and 311 mouse cell types.

What DeepMind says it can find

On TenK10K, which DeepMind describes as a large single-cell map of 28 primary immune cell types, DeepMind says AlphaGenome 2 identified causal variants at 30 percent versus 16 percent and directional gene effects at 43 percent versus 23 percent, compared with the original AlphaGenome fine-tuned on single-cell data and other specialized tools. These are DeepMind's own figures.

DeepMind also says AlphaGenome 2 reproduced a published finding computationally: that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating ETS2 in gut macrophages. This is a reproduction of a known result, not a new discovery. James Lee of the Francis Crick Institute, who led that 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."

DeepMind says that on unseen data from a COVID-19 mRNA vaccine study, the model distinguished vaccine-activated T cells from resting ones and found that a severe-COVID-linked variant disrupts gene activity more strongly in the activated state.

DeepMind says it has also precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions, to help prioritize variants for traits ranging from height to type 2 diabetes.

Sarah Teichmann, co-founder of the Human Cell Atlas consortium, is quoted 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."

Access, cost and limits

DeepMind says interactive Colab tutorials for expression conditioning are available now. It has not published commercial preview terms or pricing, and it says AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, are coming soon.

Žiga Avsec told The Verge that AlphaGenome 2 is not replacing AlphaGenome any time soon. Its predictions are "more expensive" to produce, which The Verge says would have made it impractical for the genome-wide AlphaGenome Atlas. Adam Gayoso told The Verge the key advance is "tying human genetics to cell biology."

What to watch

Watch for outside researchers' tests and laboratory follow-up, review of DeepMind's technical paper, commercial terms and query limits, and the Antigravity Skills release.

How we reported this

DailyMech read DeepMind's Nov. 11 announcement and The Verge's report in full and Nature's news story in part (the rest did not load). All performance figures are DeepMind's and were not independently verified. DailyMech did not use the model and contacted no one.

Published and fact-checked November 11, 2026.

Sources

DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.

In Artificial Intelligence

From other sections

Sections