Google DeepMind has launched AlphaGenome 2, a new artificial intelligence model that predicts how changes in DNA affect gene activity in specific human cell types, rather than in a tissue as a whole. The company announced the model on 11 November 2026, describing it as a step towards tracing disease back to the individual cells where a genetic variant does its damage.

Most DNA variants linked to disease sit outside the genes themselves, in regions that control when and where genes switch on. The same variant can be harmless in one cell type and disruptive in another, which means a prediction averaged across a whole tissue can miss the effect entirely. AlphaGenome 2 is designed to close that gap by conditioning each prediction on a chosen tissue, cell type or cell state.

What AlphaGenome 2 does differently

The first AlphaGenome model, previewed in June 2025 and published in Nature in January 2026, predicted molecular activity across broad tissues and cell lines. AlphaGenome 2 takes a DNA sequence of up to a million letters together with a biological condition, such as a heart muscle cell or an activated immune cell, and predicts gene expression, chromatin accessibility, histone marks and DNA methylation for that condition. Researchers can also condition the model on their own single-cell experiment, using as few as a few dozen cells, without retraining it.

DeepMind says the model was trained on more than 100 million individual cell profiles from human and mouse data, covering 68 tissues. On TenK10K, a single-cell map of 28 primary immune cell types, the company reports that AlphaGenome 2 recovered 30 per cent of putatively causal expression-linked variants at 90 per cent precision, compared with 16 per cent for the first AlphaGenome model fine-tuned on single-cell data. DeepMind also says the model identifies nearly twice as many variants that change gene expression in cells as its predecessor. Those figures are DeepMind's own, reported from its own benchmarks.

Who can use it, and what it is not for

AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, with commercial research access in private preview. DeepMind also plans to release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform. The company is clear about the limits: both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes, so the model cannot be used to interpret a person's own genome for medical decisions.

Early users include Genomics Limited, whose chief executive, Peter Donnelly, said the company looked forward to incorporating AlphaGenome 2 into its Mystra platform. Nature reported that researchers welcomed the release, with immunologist James Lee of the Francis Crick Institute in London among those commenting on the work. The technical paper that accompanies the launch cautions that even at high precision, the model recovers only a minority of the variants that statistical analyses nominate as likely causal, so its predictions are leads for laboratory follow-up rather than conclusions on their own.

Why cell-level prediction matters

Genome-wide association studies have linked thousands of variants to disease, but most of those variants sit in non-coding DNA, and working out which cell they act in has been slow, expensive work. A model that can screen a variant across cell types could help researchers prioritise which cells to study in the laboratory, and which genes to investigate as possible drug targets. DeepMind's announcement was written by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Ziga Avsec.

AlphaGenome 2 does not replace the first model. DeepMind says predictions from the new model cost more computing power, which makes a genome-wide atlas of the kind produced for the original AlphaGenome impractical, so the earlier model will stay in service alongside it. The launch comes in the same week as several other model releases, including Alibaba's Qwen4-Max preview and Runway's open-weight Praxis-1 robot model, as AI laboratories compete across genomics, robotics and general-purpose models at the same time.

Sources: Google DeepMind and Nature.