Finding a DNA change linked to disease is a bit like discovering a suspicious wire behind your wall. Useful information. Still doesn’t tell you which light it controls or why the kitchen goes dark when you turn on the kettle.
For researchers, the missing connection can be the cell where that genetic change actually matters.
Google DeepMind’s AlphaGenome 2, announced November 11, predicts how DNA variations affect specific cell types and states. Its central idea: understanding a genetic change requires understanding the cellular context around it.
That could help scientists answer a more useful question than “Which mutation looks suspicious?” Namely: “What should we test next?”
We covered the original AlphaGenome earlier this year. The broader challenge is interpreting genetic variation outside protein-coding regions, including regulatory DNA that helps control gene activity.
The preceding AlphaGenome Atlas project generated predictions for every possible single-letter human DNA variant. Its accompanying preprint describes using those predictions to prioritize potentially consequential changes and investigate their effects.
But a comprehensive catalog still leaves a practical problem: researchers need to know which predictions deserve attention in the particular biology they’re studying.
An average can hide the thing you’re looking for. Imagine measuring a building’s electricity consumption when the problem is one overloaded socket.
AlphaGenome 2 takes a million-letter DNA sequence plus a specified biological condition, rather than producing fixed parallel tissue outputs. DeepMind trained it on more than 107 million individual cell profiles.
The reported benchmark improvement is substantial. On the TenK10K immune-cell dataset, DeepMind reports causal-variant recall of 30%, versus 16% for the original model fine-tuned on single-cell data.
That result needs its denominator attached. It describes performance on a particular evaluation, not a percentage of diseases solved or patients diagnosed. Even within that test, there is considerable room left to improve.
One example makes the potential easier to understand.
In a 2024 Nature study, researchers investigated a stretch of DNA associated with inflammatory diseases, including inflammatory bowel disease. The region was a “gene desert,” meaning it contained no protein-coding genes.
Yet it wasn’t doing nothing.
The team connected a regulatory region there to ETS2, a gene roughly 290,000 DNA letters away. Experiments showed that increasing ETS2 activity could drive an inflammatory program in macrophages, a type of immune cell. Disrupting the gene or its regulatory region reduced inflammatory activity.
That is the detective work genetic associations demand. A disease-linked location is a starting point; researchers still need to establish the mechanism.
DeepMind says AlphaGenome 2 computationally reproduced the variant’s selective effect on ETS2 in gut macrophages.
Reconstructing an established finding is a useful check. The more consequential test will be whether predictions guide scientists toward previously unknown mechanisms that hold up in experiments.
For a research team, the potential value is straightforward: a better shortlist.
Suppose a study implicates a region containing several plausible variants. A useful model could help researchers prioritize which change to investigate, which gene to measure, and which cells to use. Each decision makes the next experiment more specific.
That is also where the commercial opportunity becomes interesting. For biotech companies, a model’s value could come from reducing the number of plausible but unproductive investigations—not simply producing more predictions.
It fits the larger question running through our coverage of AI drug discovery: how does computational progress translate into biological evidence?
Our read is that specialized models will earn their place in research workflows by improving those decisions. A system that helps choose a revealing experiment can be valuable even when it cannot explain an entire disease.
AlphaGenome 2 is available through an API for noncommercial research, with commercial access in private preview. It remains research-only and has not been validated or approved for diagnosis.
The next evidence to watch is whether outside researchers can use it to uncover mechanisms they would otherwise have missed—and then confirm them.
The promise is a shorter route from a suspicious stretch of DNA to an experiment worth running. Biology will still grade the answer.
Related reading
- Our original AlphaGenome coverage — Background on the challenge of predicting how genetic changes affect gene regulation.
- Isomorphic Labs’ AI drug ambitions face the human-trial test — A look at the distance between promising computational results and evidence that a medicine works.
