AlphaGenome 2: DeepMind's model predicts DNA variant effects in single cell types

Google DeepMind announced AlphaGenome 2 on 11 November 2026, a DNA model that predicts how genetic variants change gene activity in specific human cell types rather than in broad tissue averages. DeepMind says it finds nearly twice as many expression-changing variants as its previous best model, and it is open to non-commercial researchers through the AlphaGenome API from today. For the AI-for-science field, it is a clear example of where genomics models are heading: conditioning on a cell's state instead of predicting everything at once.

Key points

• AlphaGenome 2 was announced on 11 November 2026 and is available for non-commercial research through the AlphaGenome API, with a private preview for commercial research.
• DeepMind says it was trained on more than 100 million individual cell profiles across 68 tissues (the post also cites "over 107 million", from human and mouse samples).
• On the TenK10K single-cell dataset of 28 immune cell types, DeepMind reports it identified causal variants at 30% against 16% for the original AlphaGenome fine-tuned on single-cell data and other tools, and directional gene effects at 43% against 23%.
• It is a conditional model: it takes a DNA sequence of one million letters plus a chosen cell type or state, instead of predicting many tissues in parallel.
• It does not replace the first AlphaGenome yet. DeepMind's genomics lead told The Verge it is more expensive to run, which ruled it out for the genome-wide Atlas.

What happened

DeepMind published a blog post and a paper on AlphaGenome 2, written by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec. The model predicts how a DNA sequence and its variations affect specific human cell types. Every cell in the body carries the same DNA, but a heart muscle cell and an immune cell switch different genes on and off, so the same variant can matter in one cell type and do nothing in another. Earlier models averaged over tissues and could miss those effects.

AlphaGenome 2 predicts the same core readouts as the first model: gene expression, histone modifications and chromatin accessibility. It adds DNA methylation, which DeepMind calls a widely requested readout.

DeepMind gives three worked examples. First, a published study linked a variant in a distant regulatory switch to inflammatory bowel disease (IBD) through overactivation of the ETS2 gene in gut macrophages; the model, DeepMind says, reproduced this computationally by singling out the causal variant and showing it boosted ETS2 in those macrophages while leaving other cell types largely unaffected. Second, given unseen single-cell data from a COVID-19 mRNA vaccine study, it distinguished vaccine-activated T cells from resting ones, and indicated that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. Third, DeepMind precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions, to help researchers map enhancer-gene links in primary human cells such as brain astrocytes and to narrow down causal variants from genome-wide association studies (GWAS) for traits ranging from height to type 2 diabetes.

Background

The first AlphaGenome was released last year, according to The Verge, as a model that took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues at once. DeepMind says that parallel approach spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types.

In September 2026 DeepMind followed with the AlphaGenome Atlas, a resource holding predictions for all 9 billion possible single-letter changes in the human genome, which was reported by News-Medical. The Atlas was computed with the first model. AlphaGenome 2 is a different design aimed at a different question: not "what does this variant do somewhere in the body" but "what does it do in this cell, in this state".

The enabling change is data. Single-cell technologies and efforts such as the Human Cell Atlas now map gene expression for millions of cells, and DeepMind's genomics lead said building such a model has only recently become possible because those datasets exist.

How it works

The key change is architectural. The original model produced fixed parallel outputs for bulk tissues and cell lines. AlphaGenome 2 is a conditional model that predicts DNA effects for one condition, defined as a tissue, cell type or state, at a time. The target condition is fed in "at every stage of its architecture" alongside the sequence, so the network can learn how the same stretch of DNA is read differently across cell types.

Two practical features follow. The condition can come from a reference catalog or be generated on the fly from a scientist's own single-cell experiment through a cell encoder. DeepMind says a researcher can upload single-cell RNA-seq data from as few as a few dozen cells, which is compressed into a cell state representation and used to condition the model on states it has never seen, with no retraining. DeepMind says this can match or outperform models trained or fine-tuned directly on those datasets. And because the cell's state is separated from the outputs, giving it one measurement such as gene activity lets it predict a missing one such as chromatin accessibility, which could help prioritise lab experiments.

Causal variants, AlphaGenome 116

Causal variants, AlphaGenome 230

Directional effects, AlphaGenome 123

Directional effects, AlphaGenome 243

Share identified on TenK10K (28 immune cell types), DeepMind's reported results (%)

	%
Causal variants, AlphaGenome 1	16
Causal variants, AlphaGenome 2	30
Directional effects, AlphaGenome 1	23
Directional effects, AlphaGenome 2	43

Source: Google DeepMind, AlphaGenome 2 announcement

These figures are DeepMind's own, from its post. The comparison model is the original AlphaGenome fine-tuned on single-cell data "and other specialized tools", so it is a strong baseline rather than the untuned first release.

What it means

For genetics researchers, the pitch is resolution. GWAS studies flag broad stretches of DNA, and the hard part is finding the causal variant, the target gene and the cell type where it acts. A model that scores variants per cell type, including cell states a lab has just profiled, narrows that search. Professor Sarah Teichmann of Cambridge, a Human Cell Atlas co-founder, is quoted by DeepMind as saying what makes it exciting is that it adapts to new single-cell data without retraining.

For drug discovery the argument is specificity. DeepMind's Avsec told The Verge that "When you develop drugs, you're targeting specific proteins and specific cell types." Peter Donnelly, CEO of Genomics Limited, is quoted by DeepMind saying his company looks forward to incorporating it in its Mystra platform. These are statements of intent, not results.

For the AI field, it is another case of a general lesson: when a fixed multi-output model spreads capacity too thin, conditioning on context can be better than adding outputs. The compute cost is the trade-off the Verge reports.

What we don't know yet

• How the results hold up outside DeepMind's own benchmarks. We have only the company's reported figures, and the paper's peer-review status was not clear from the pages we could open.
• The cost and speed of running it, beyond the statement that it is "more expensive" than AlphaGenome.
• The terms of the commercial private preview, and how many groups have access.
• Whether the precomputed GWAS scores will be published as a downloadable resource like the Atlas.
• How well predictions transfer to disease states with little training data. DeepMind says it is generating more data for poorly represented cell states.

What to watch next

Independent evaluations on held-out datasets, the paper's review, and any extension of the Atlas with cell-type-specific scores. DeepMind says it is working with groups including the Wellcome Sanger Institute, and Avsec mentioned future versions trained on multiple species and a longer-term link to "virtual cells". If you work with variant data, the AlphaGenome API is where to try it.

Sources

• Google DeepMind: AlphaGenome 2, Deciphering how cells read DNA
• The Verge: Google's AlphaGenome 2 zooms in on the cellular roots of disease
• Google DeepMind: AlphaGenome, AI for better understanding the genome (June 2025)
• News-Medical: AlphaGenome Atlas maps billions of genetic changes with AI