Google DeepMind’s launch artwork for AlphaGenome 2 (illustrative). Image: Google DeepMind
Google DeepMind has built an AI that reads DNA the way individual cells do. AlphaGenome 2, unveiled on Wednesday, predicts how a single-letter change in someone’s genome plays out in a specific type of human cell, from a gut immune cell to a brain astrocyte, and in DeepMind’s tests it found nearly twice as many gene-changing variants as the original model.
One cell type at a time #
Every cell in the body carries the same DNA, but a heart muscle cell and an immune cell fighting an infection switch on different genes. The first AlphaGenome, launched last year, took a million-letter stretch of DNA and predicted thousands of readouts across broad tissues all at once. DeepMind says that blurred the differences between closely related cells.
AlphaGenome 2 is instead a “conditional” model. Alongside the million letters of DNA, it takes one target condition (a tissue, cell type or cell state) and uses it at every stage of the network, so it learns how the same stretch of DNA is read differently in different cells. It was trained on bulk and single-cell data from humans and mice, including more than 107 million individual cell profiles across 68 tissues, and adds DNA methylation, a chemical tag linked to genes being switched off, to readouts such as gene expression and chromatin accessibility.
Nearly twice as many variants found #
The headline result comes from TenK10K, a large single-cell map of 28 immune cell types. There, DeepMind says AlphaGenome 2 identified 30% of causal variants, against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, and called the direction of a variant’s effect on its gene 43% of the time, against 23%. The technical paper describes this as “doubling recall of causal single-cell eQTLs at 90% precision”: variants that change how strongly a gene is expressed. These are DeepMind’s own benchmarks.
It also reproduced a known discovery. A 2024 Nature study found that a variant in a distant regulatory switch drives inflammatory bowel disease (IBD) by overactivating the ETS2 gene in gut macrophages, a kind of immune cell. Working across gut immune cells, AlphaGenome 2 singled out the same variant and showed it boosts ETS2 in macrophages while leaving other cell types and genes largely unaffected.
New cells without retraining #
Scientists can upload a single-cell RNA snapshot from just a few dozen cells. The model compresses it into a description of the cell’s state and makes predictions for it without being retrained, and DeepMind says it matches or beats models trained directly on those datasets. Given unseen data from a COVID-19 mRNA vaccine study, it told vaccine-activated T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells, an effect that would be missed in resting ones.
Žiga Avsec, DeepMind’s genomics lead, told The Verge why that detail matters: When you develop drugs, you’re targeting specific proteins and specific cell types.
Žiga Avsec, genomics lead, Google DeepMind, to The Verge
He also said the finer predictions are “more expensive” to compute, so AlphaGenome 2 will sit alongside the original rather than replace it. Longer term, Avsec said, models like it could feed into “virtual cells”, digital simulations of biology and the aim of the $1.8 billion project Google joined with Biohub, Meta and the US government in October.
Who can use it #
AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, with Colab tutorials, and in private preview for commercial research. DeepMind has also precomputed scores for trait-linked variants across hundreds of human cell and tissue types, which the paper puts at 7 million common variants, covering traits from height to type 2 diabetes.
“AlphaGenome 2 Skills” for Google Antigravity, its agentic development platform, are coming soon, and DeepMind is working with the Wellcome Sanger Institute to profile more cell types. Like the original, it is for research use only and has not been validated or approved for clinical or diagnostic use.
It joins a crowded field. Hugging Face released Carbon-X, a free model that lets you chat with DNA, last month, and OpenAI had thousands of GPT-Rosalind agents work through a mouse brain atlas in October.
Why it matters #
Most variants linked to disease sit outside genes, in switches whose effects depend on the cell, which is why genetics studies so often stop short of a drug target. If AlphaGenome 2’s cell-level predictions hold up in the lab, researchers can narrow down which variant, gene and cell type to chase before spending money on experiments. It also pushes Google DeepMind further into a biology race in which Anthropic is now developing its own medicines with Claude.
Sources: Google DeepMind; AlphaGenome 2 technical paper; The Verge (Avsec interview); Nature (ETS2 study, 2024).