AlphaGenome 2 predicts what a DNA change does inside one type of cell, but it is for research only

Every cell in your body carries the same DNA. Yet a heart cell and an immune cell switch on different genes. So one tiny DNA spelling change can matter greatly in one cell and do nothing in another.

On 11 November, Google DeepMind released AlphaGenome 2, an AI model built for that problem. It reads a stretch of 1 million DNA letters and predicts how a small change, called a variant, alters gene activity in one specific cell type or cell state. Non-commercial researchers can use it now. It is not a medical tool: DeepMind says it is for research only, not for clinical or diagnostic use.

Large genetic studies, called GWAS, link regions of DNA to traits from height to type 2 diabetes. They rarely say which exact letter causes the effect, or in which cell it acts. Drugs target specific proteins and cell types, DeepMind’s genomics lead Žiga Avsec told The Verge, so that missing detail matters.

How AlphaGenome 2 narrows the question to a single cell type

The first AlphaGenome, announced in June 2025, predicted thousands of results for broad tissues all at once. AlphaGenome 2 works differently. A scientist names the cell type or state they care about, and the model predicts for that one.

So what is AlphaGenome trained on? More than 100 million individual cell profiles from 68 tissues, DeepMind says; the body of its post puts the figure at over 107 million, from humans and mice. For a new cell state, such as cells after a vaccine, a snapshot of just a few dozen cells is enough, with no retraining.

What does AlphaGenome predict? Gene activity, how open the DNA is to be read, chemical marks on DNA and its packaging, and now DNA methylation. It can also fill in a missing measurement to help a lab pick its next experiment.

Excerpts from Google DeepMind’s announcement, as published on its blog on 11 November. Image: Google DeepMind

The model reproduced, on a computer, a known finding that a distant DNA switch boosts the ETS2 gene only in gut immune cells called macrophages, a link to inflammatory bowel disease. Using COVID-19 vaccine data, it told activated immune T cells from resting ones.

On DeepMind’s own test across 28 immune cell types, AlphaGenome 2 found 30% of the variants that cause changes, against 16% for a fine-tuned AlphaGenome 1 and other tools. It got the direction of the effect right 43% of the time, against 23%. Startup Fortune reported the same figures.

• Causal variants found: AlphaGenome 116%
• Causal variants found: AlphaGenome 230%
• Right direction of effect: AlphaGenome 123%
• Right direction of effect: AlphaGenome 243%

TenK10K test, 28 immune cell types. AlphaGenome 1 was fine-tuned on single-cell data. Not yet tested independently. Source: Google DeepMind, 11 November 2026

Strong results, but so far only DeepMind has checked them

All those numbers come from DeepMind’s own testing. The paper is a PDF on DeepMind’s site, and it does not say it has been peer reviewed. We found no public reaction yet from independent scientists.

The two outside voices in the post were chosen by DeepMind. Professor Sarah Teichmann, a co-founder of the Human Cell Atlas, praises how it adapts to new data; she also appears in the post’s thanks list. Peter Donnelly, CEO of Genomics Ltd, calls it “an impressive advance” and plans to add it to his company’s Mystra platform.

AlphaGenome 2 also costs more to run than the first model, according to The Verge and Startup Fortune. Both report that this is why the genome-wide Atlas still runs on AlphaGenome 1. An independent study of AlphaGenome 1 found real gains but “key limitations”.

If you study DNA, or just wonder about your own

For university and non-profit researchers, the AlphaGenome API is open now, with Colab tutorials and docs at alphagenomedocs.com. Commercial use is in private preview. No price has been published. DeepMind’s community spotlight form closes on 30 November, and chosen projects may get extra quota and credits.

For everyone else, this is not a DNA test. It cannot tell you your risk of a disease. Any benefit to patients is likely to come later, through better lab work.

People often confuse AlphaFold and AlphaGenome. AlphaFold predicts protein shapes. AlphaGenome predicts what DNA changes do to gene activity. AlphaGenome 2 Skills will also come “soon” to Google Antigravity, paired with the Gemini 4 models now reaching developers.

▶

Why cell-level detail matters more than the headline numbers

The good is real. Biology happens in specific cells, and this Google DeepMind DNA model finally asks questions at that level. In our reading, needing only a few dozen cells could help smaller labs.

The caution is real too. The scores are self-reported, peer review is not mentioned, and the price is unknown. On balance, our reading is that this is a meaningful step for lab scientists choosing their next experiment, not a breakthrough for patients yet.

The real test starts now, as independent labs run AlphaGenome 2 on their own cells and publish what they find.