alphagenome 2: DeepMind DNA Model Zooms Into Single Cells
The alphagenome 2 model from Google DeepMind, released Nov 11, predicts how DNA variants affect specific cell types and finds nearly twice as many expression-changing variants as its predecessor.
- The model was trained on more than 100 million single-cell profiles from human and mouse samples across 68 tissues.
- It can adapt to a new cell state from a single-cell snapshot of just a few dozen cells, without retraining.
- Access is non-commercial via API for now, with a private preview for commercial research, and it will not replace the original model.
in this block
The alphagenome 2 model from Google DeepMind arrived on November 11: a DNA model that predicts how genetic variants change gene activity in specific human cell types rather than in broad tissue averages. DeepMind says it finds nearly twice as many expression-changing variants as its previous best model, and it is already open to non-commercial researchers through the AlphaGenome API.
What actually happened
DeepMind published the announcement in a blog post by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, alongside a paper. The Verge and MarkTechPost covered it the same day.
The first AlphaGenome, launched last year, read a one-million-letter DNA window and predicted thousands of experimental readouts across tissues in parallel. That was efficient, but DeepMind says it spread the model's capacity thin and blurred differences between closely related cells. In September, the lab also released AlphaGenome Atlas, with predictions for all 9 billion possible single-letter changes in the human genome.
How AlphaGenome 2 is different
The headline change is architecture. AlphaGenome 2 is a conditional model: alongside the DNA sequence, it takes one target condition at a time, such as a tissue, a cell type or a cell state, and uses it at every stage of the network. The condition can come from a reference catalog or be generated on the fly from a scientist's own single-cell experiment through a cell encoder.
That design brings a few practical tricks:
- Cell-type resolution. It can separate, for example, distinct heart muscle cells or blood immune cells instead of averaging them.
- Filling gaps. Given one measurement, like single-cell gene activity, it can predict unmeasured properties such as chromatin accessibility, which helps decide which lab experiments to run next.
- A new readout. On top of gene expression, histone marks and chromatin accessibility, it now predicts DNA methylation.
MarkTechPost reports that training used 2,048 TPU v4 chips and that weights are promised "upon final publication," so for now it is hosted-only.
The numbers DeepMind shared
On TenK10K, a single-cell map of 28 primary immune cell types, DeepMind says the model identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data, and 43% versus 23% for directional gene effects. These are the lab's own benchmarks, not independent results.
It also reproduced a known finding: a variant in a distant regulatory switch that drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. In another test with unseen data from a COVID-19 mRNA vaccine study, it separated activated T cells from resting ones and showed a severe-COVID variant matters more in the activated state.
"When you develop drugs, you're targeting specific proteins and specific cell types," DeepMind genomics lead Žiga Avsec told The Verge.
The catch: cost and access
Avsec told The Verge the more detailed predictions are "more expensive" to compute, which is why the genome-wide Atlas stayed on the first model. That is also why the two will coexist. Commercial users are limited to a private preview, and there is no local download yet.
Avsec said future versions might train on data "across multiple species" and feed into longer-term "virtual cells" work. Those are goals, not products. DeepMind also says it is generating more data for cell states that are poorly covered today, which suggests the next gains will come from better datasets as much as from new architecture. Kohli, who leads the lab's AI-for-science work, has hinted that more scientific results are coming, according to The Verge.
What to do as a reader
If you follow AI-for-science, here is how to read the AlphaGenome 2 launch:
- Treat benchmark gains as claims until replicated. The TenK10K comparison is DeepMind's own setup.
- Note the license. Non-commercial API access is useful for academics, less so for startups.
- Watch for the weights. Open weights after publication would let outside labs test it properly.
- Ignore any token or "genome coin" using the name. DeepMind has not launched one.
For more Google AI coverage, read our take on gemini 4 carbon and on google schools ai. This article is informational, not medical or financial advice.
Not financial advice. DYOR, ser.
