AlphaGenome 2: DeepMind's DNA model now reads single cell types, and the evidence is still its own

Google DeepMind has released AlphaGenome 2, a model that predicts how changes in DNA affect gene activity in individual cell types, not just whole tissues. The company reports roughly double the hit rate of its first model at spotting which genetic variants actually cause an effect. The claim is plausible and the tool is available to researchers now, but the numbers are DeepMind's own and the model is not a diagnostic.

What changed from the first AlphaGenome

The original AlphaGenome predicted many tissues in parallel from a long DNA sequence. AlphaGenome 2, published November 11, 2026 by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, flips the design. It is a conditional model: you tell it which tissue, cell type or cell state you care about, and it predicts the effect of the DNA in that one condition.

Three features follow from that. A cell encoder lets researchers build a condition from their own single-cell experiment. Cross-modality imputation lets one measurement, such as gene activity in single cells, predict readouts nobody measured, such as chromatin accessibility. And DNA methylation joins gene expression, histone modifications and chromatin accessibility among the outputs. The model takes a one million letter DNA sequence as input and was trained on more than 100 million cell profiles from human and mouse samples across 68 tissues. The blog gives "more than 100 million" in its introduction and "over 107 million" in the body, so treat the exact count as approximate.

The numbers DeepMind reports

DeepMind's headline test uses TenK10K, a dataset covering 28 primary immune cell types. There, the company says AlphaGenome 2 identified 30% of causal variants, against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools. For the direction of a variant's effect on a gene, it reports 43% against 23%. DeepMind also says the model covers roughly twice as many variants affecting gene expression as the previous state of the art.

Three case studies back this up. In inflammatory bowel disease, the model reproduced a published finding that a variant raises activity of the ETS2 gene in gut macrophages, and singled out the causal variant. In blood stem cell maturation, its predictions tracked observed expression more closely than the fine-tuned original. And in a COVID-19 mRNA vaccine study, given just a few dozen cells, it separated activated T cells from resting ones and showed that a variant linked to severe COVID-19 has a stronger effect in activated cells.

What an outside expert says

SiliconANGLE reported the launch the same day and quoted Sarah Teichmann, co-founder of the Human Cell Atlas and a professor at the University of Cambridge, who said the strength of the model is "how it adapts directly to new single-cell data without retraining." DeepMind's own post also quotes Peter Donnelly, chief executive of Genomics Limited, calling it "an impressive advance in the field." Both are praise from people with ties to the work or the field. Neither is a replication.

SiliconANGLE's report gave no benchmark figures of its own and described predicting how DNA variants drive gene activation as a major remaining challenge. The benchmark numbers above come from DeepMind's blog.

How to read the benchmark

A benchmark in genetics is only as good as its ground truth. TenK10K links genetic variants to gene expression across immune cells, so "causal" here means a variant that the dataset's statistical methods flag, which is itself an inference. When a model is scored against that, a higher match rate shows the model agrees better with the reference, which is useful, but it is not the same as proving biology at the bench. DeepMind is also comparing against the original model fine-tuned on single-cell data, a baseline of its own design, and "other specialized tools" that the post groups together. The comparison is fair as far as it goes. It simply shows how far the lab's own two generations have moved, with outside tools as a footnote. An independent group running the same test on held-out data would tell us far more.

Why a frontier lab is doing this

Genomics is where the labs say general AI will first pay for itself in the physical world. The same week, DeepMind also released Gemini Robotics 2.5, which we covered in our report on the robot model, and used AI to write formal proofs for software verification. The pattern is a lab arguing that capability gains show up as scientific tools, not only as chatbots. That is a fair argument, and a hard one to audit from outside, because the evaluation datasets and the models come from the same shop.

Access and limits

AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with interactive Colab tutorials. Commercial research is in private preview. DeepMind says "AlphaGenome 2 Skills" will come to its Antigravity coding environment soon, and that applications for use-case spotlights close November 30, with selected projects eligible for extra quota and credits.

The limits are stated plainly. The model is for research only and is not validated or approved for clinical or diagnostic use. Its predictions guide experiments but do not replace them. A causal-variant hit rate of 30% also means that most causal variants in the test were still missed, which is worth remembering when headlines say "doubles."

Our take

This is a real step for a hard problem, and the cell-type design is the right direction. The honest reading is that DeepMind has shown improvement on a benchmark it helped select, backed by a few case studies that reproduce known biology. The test that counts is whether independent labs, using their own data, find new causal variants and confirm them at the bench. We would watch for exactly that, and for whether the commercial preview prices out smaller groups.