Google DeepMind has a new model that looks at how our DNA changes affect individual cells. The company says AlphaGenome 2 gives scientists a far sharper view of a problem that has frustrated geneticists for years. The work could help explain where some diseases begin, and which treatments are worth testing.
AlphaGenome 2 is the follow-up to DeepMind’s genome-predicting model AlphaGenome, which launched last year. Almost every cell in the human body carries the same DNA. But heart muscle cells, skin cells and immune cells switch different genes on and off. Those patterns also shift as a cell changes what it is doing.
That means a single genetic variant can behave very differently depending on the cell type and its state. Broad models tend to blur those differences together. AlphaGenome 2 tries to keep them apart, and it predicts variant effects cell by cell.
Trained on 100 million cell profiles
DeepMind trained the model on more than 100 million individual cell profiles drawn from 68 tissues. The company says this turns static cell atlases into what it calls a predictive engine for human genetics. Cell atlases are large repositories that map and catalogue individual cells in fine detail.
According to Google, AlphaGenome 2 finds nearly twice as many variants that can change gene expression inside cells. That is roughly double the tally of its previous best model. The team also found the model can handle new cell types and biological states without being retrained from scratch.
It can flag gaps in experimental data too. That could help researchers decide which experiments to run next, instead of guessing.
Why cell-level detail matters for drugs
Adam Gayoso, a research scientist at Google DeepMind, said the big breakthrough is tying human genetics to cell biology. The model lets researchers take variants flagged in large population studies and ask what they do at the cellular level. From there, scientists can connect those effects to disease.
Ziga Avsec, DeepMind’s genomics lead, said the specificity matters when the goal is treating a particular illness. Drugs target specific proteins and specific cell types, he told The Verge. Having granular predictions makes that targeting far easier to plan.
Cost is the trade-off. Avsec said AlphaGenome 2 is more expensive to run than its predecessor. Those demands would have made it impractical to build the genome-wide Atlas that DeepMind released in September, which covers every possible DNA letter change in the human genome. He still expects the original AlphaGenome to stay in use for broader sweeps.
Available now for non-commercial research
Google is releasing AlphaGenome 2 for non-commercial research through the AlphaGenome API starting today. Commercial researchers get access through a limited private preview. Avsec said future versions could be trained on data across multiple species, and that models like this may eventually feed into efforts to build virtual cells.
The release lands in a busy stretch for AI in science. Anthropic has been experimenting with drug design and gave Claude a wet lab, while OpenAI has been posting results on long-standing math problems. Google has also pushed hard on weather and materials research, including the open-sourcing of its WeatherNext cyclone model. For healthcare watchers, the wider trend is tracked in our look at AI medical diagnosis breakthroughs in 2026.
The near-term test is whether labs can turn cell-level predictions into experiments that hold up. AlphaGenome 2 gives them a much more detailed map. It does not tell them which road to take.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
