Google DeepMind's AlphaGenome 2 predicts how DNA variants affect individual cell types

A sharper view of the genome

Google DeepMind has released AlphaGenome 2, an AI model designed to predict how changes in DNA play out across different kinds of cells, a capability the company says could surface disease clues that coarser models miss. According to The Verge, the model succeeds AlphaGenome, DeepMind's earlier genome-prediction system released last year.

The biological problem it targets is subtle. Nearly every cell in the body carries the same DNA, yet a heart muscle cell, a skin cell and an immune cell behave very differently because each switches distinct sets of genes on and off, and those patterns shift depending on what a cell is doing. A single genetic variant can therefore have sharply different consequences depending on the cell involved. AlphaGenome 2 is built to model that variation directly, predicting a variant's effects on individual cell types and cell states rather than averaging across tissues.

Training at single-cell scale

To get that resolution, DeepMind trained the model on more than 100 million individual cell profiles spanning 68 tissues, according to The Verge. The company frames the result as turning static cell atlases — vast reference datasets that catalogue individual cells — into something closer to a predictive tool for human genetics. DeepMind claims the new model identifies nearly twice as many variants capable of altering gene expression as its previous state-of-the-art system.

Two further capabilities stand out. The researchers report that AlphaGenome 2 can generalize to cell types and biological states it never saw during training, without being retrained. It can also flag gaps in existing experimental data, which could help scientists decide what to study next at the bench.

A tool aimed at drug development

DeepMind staff involved in the project emphasized the medical payoff. Adam Gayoso, a research scientist at DeepMind, told The Verge that the core advance is linking human genetics to cell biology: researchers can take variants flagged in large population studies, examine what they actually do inside specific cells, and connect those effects to disease. Žiga Avsec, who leads genomics at DeepMind, noted that drug development targets particular proteins in particular cell types, which makes this level of detail important when the goal is treating a specific illness.

Building such a model only became possible recently, Avsec said, because single-cell technologies can now capture cells and their states at high resolution and generate the large datasets needed for training.

Cost and availability

AlphaGenome 2 does not retire its predecessor. Avsec told The Verge that the finer-grained predictions are computationally more expensive, which is why DeepMind relied on the earlier model to build the genome-wide atlas it released in September, covering every possible single-letter change in human DNA. Google is making AlphaGenome 2 available for non-commercial research through the AlphaGenome API as of the announcement, with a limited private preview for commercial research.

Why it matters

Most genetic variants linked to disease in population studies arrive with little explanation of mechanism — which tissue, which cell, which moment in a cell's life the variant does its damage. A model that predicts variant effects at the level of individual cell types gives researchers a way to prioritize leads, design sharper experiments and, eventually, justify drug targets with more evidence. The longer-term ambitions are larger still: Avsec suggested future models could be trained across multiple species for a broader view of biology, and that systems like AlphaGenome 2 could feed into digital simulations of living cells. The release also fits a wider race among AI labs to attack scientific problems, with The Verge pointing to Anthropic's drug-design experiments and wet lab, OpenAI's results in mathematics, and DeepMind's planned automated materials science lab in the UK.