AlphaGenome 2: DeepMind's DNA Model Now Works Cell by Cell

DeepMind's new model, trained on 107 million cell profiles, finds nearly twice as many causal variants as AlphaGenome, giving disease research sharper leads. Google DeepMind on Wednesday, November 11, 2026, announced AlphaGenome 2, an AI model that predicts how changes in DNA affect specific human cell types, such as a heart muscle cell or an immune cell fighting an infection.

It is free today for non-commercial research through the AlphaGenome API, while companies can apply for a private preview. The headline number: on a large immune cell benchmark, the new AlphaGenome model identified nearly twice as many likely causal variants as the original version, 30% versus 16%.

For scientists hunting the genetic roots of disease, that means more candidates worth testing in the lab. What Is AlphaGenome 2 and How Does It Work? Every cell in your body carries the same DNA, but each cell type switches different genes on and off.

A genetic variant, a single-letter difference in the DNA code, can therefore be harmless in one cell and harmful in another. The first AlphaGenome, released last year, predicted these effects across broad tissues. DeepMind then built AlphaGenome Atlas, a catalog of predictions for all 9 billion possible single-letter changes in the human genome.

According to the DeepMind announcement , AlphaGenome 2 changes the design. Instead of predicting thousands of experiments in parallel, it is a conditional model. It reads a one-million-letter stretch of DNA together with one target condition, meaning a tissue, a cell type or a cell state, and makes its prediction for that condition only.

As a result, it can learn how the same DNA is read differently by closely related cells. DeepMind trained it on human and mouse data, including more than 107 million individual cell profiles across 68 tissues. Besides gene expression, chromatin accessibility (how open a DNA region is to the cell's machinery) and histone modifications, it now also predicts DNA methylation, a frequently requested readout.

How Much Better Is It Than the First AlphaGenome? DeepMind's main test used TenK10K, a single-cell map of 28 primary immune cell types. The comparison was against the original AlphaGenome fine-tuned on single-cell data.

Metric (TenK10K) AlphaGenome 2 AlphaGenome (fine-tuned) Causal variants identified 30% 16% Correct gene effect direction 43% 23% Training cell profiles 107M+ Broad tissue data DNA methylation Yes No In a real-world check, the model also reproduced a published finding about inflammatory bowel disease (IBD).

It singled out a variant in a distant regulatory switch that boosts the ETS2 gene in gut immune cells called macrophages, while leaving other cells largely alone. Can It Handle Cells It Has Never Seen? Yes, and this is the part researchers seem most excited about.

Scientists can upload single-cell RNA data from just a few dozen cells, and AlphaGenome 2 adapts to that new cell state without retraining. With data from a COVID-19 mRNA vaccine study, it told vaccine-activated T cells apart from resting ones. It also showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells.

"What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining," said Sarah Teichmann, a co-founder of the Human Cell Atlas consortium at the University of Cambridge. Meanwhile, Peter Donnelly, CEO of Genomics, said the company plans to add the model to its Mystra drug discovery platform.

Who Can Use AlphaGenome 2, and What Are the Limits? Access is research-only. DeepMind says neither AlphaGenome version is validated or approved for clinical or diagnostic use, so this is not a tool that will read your personal genome.

Non-commercial users get free API access and Colab tutorials, and DeepMind also published precomputed scores for disease-linked variants across hundreds of cell conditions, covering traits from height to type 2 diabetes. Next, AlphaGenome 2 Skills are coming to Google Antigravity, the agentic development platform that runs on models like Gemini 4 .

There are trade-offs, though. DeepMind researcher Ziga Avsec told The Verge that the new model is "more expensive" to run, so it will not replace the original AlphaGenome or its genome-wide Atlas soon. The accompanying paper also notes it still recovers only a minority of the variants that statistics flag as likely causal, and it learns correlations rather than proven cause and effect.

Why AlphaGenome 2 Matters In our view, this follows the same playbook DeepMind used with AlphaProof : a specialist model that narrows a huge search space so human experts can focus on the best leads. We saw a similar approach in its RSA-1024 research as well. AlphaGenome 2 will not cure anything by itself, but it can shorten the path from "this gene region is linked to a disease" to "this variant, in this cell type, is the likely culprit." If a relative asks what changed, the simple answer is this: AlphaGenome used to tell scientists what a DNA change does in an organ, and AlphaGenome 2 tells them what it does in a specific cell inside that organ.