Key takeaways
- Google DeepMind said on Nov. 11 that AlphaGenome 2 takes a tissue, cell type or cell state as input alongside the DNA sequence and predicts for that one condition.
- On the TenK10K immune cell dataset, DeepMind reported 30% of causal variants identified versus 16% for the original model, and 43% versus 23% for directional gene effects.
- The model is available now for non-commercial research through the AlphaGenome API, is in private preview for commercial use and is not approved for clinical or diagnostic purposes.
In this story 3 sections
Scientists who study disease genetics can now ask a Google DeepMind model a narrower question than before: not just what a DNA change does in general, but what it does inside one particular kind of cell. AlphaGenome 2, which DeepMind described in a blog post on Nov. 11, takes a tissue, cell type or cell state as an input alongside the DNA sequence, and the lab said it is open today to non-commercial researchers through the AlphaGenome API.
DeepMind is part of Alphabet (NASDAQ: GOOGL). The release follows the first AlphaGenome, which the lab launched last year, and it lands in the same week as other Google AI releases such as Gemini 4 Hydrogen.
From fixed outputs to one condition at a time
The original model read a stretch of 1 million DNA letters and produced a fixed set of predictions for thousands of lab experiments at once, covering bulk tissues and cell lines. DeepMind wrote that this approach spread the model's capacity thin.
AlphaGenome 2 works differently. It is a conditional model: the chosen biological condition is fed in at every stage together with the sequence, and the model makes its prediction for that one condition. According to the post, the changes include:
- Training data: more than 107 million individual cell profiles from human and mouse, drawn from both bulk and single cell experiments.
- Signals covered: gene expression, histone modifications and chromatin accessibility, as before, plus DNA methylation as a new output.
- Cross signal prediction: a researcher can supply one kind of single cell measurement, such as gene activity, and ask the model to estimate another, such as chromatin accessibility.
- New cell data without retraining: RNA sequencing from a few dozen cells can be used to condition the model, the lab said.
The benchmark DeepMind highlighted
The headline result comes from TenK10K, a single cell dataset covering 28 primary immune cell types. On that set, DeepMind said AlphaGenome 2 identified causal variants 30% of the time, against 16% for the original AlphaGenome fine tuned on single cell data and other specialized tools. For directional gene effects, meaning whether a variant raises or lowers a gene's activity, the figures were 43% against 23%.
By our calculation, those numbers work out to about 1.9 times the earlier rate in both cases (30 divided by 16, and 43 divided by 23), which matches the lab's description of nearly twice as many hits. These are DeepMind's own results on a dataset it selected; the post does not report independent testing.
The post also walks through qualitative examples, including a case involving the ETS2 gene and inflammatory bowel disease, where the model singled out a causal variant that earlier lab work had identified, and a test separating activated from resting T cells in data from people who received COVID-19 mRNA vaccines.
Who can use it, and what it is not for
Access is split in two. Non-commercial research goes through the AlphaGenome API, while commercial research is in private preview. The post does not mention pricing or open model weights. DeepMind said the model is meant for research only and has not been validated or approved for clinical or diagnostic use.
Researchers who want to test it can submit a use case by Nov. 30, and selected projects may receive extra quota and credits, DeepMind said. It also said AlphaGenome 2 Skills are coming soon to Google Antigravity.
Sarah Teichmann of the University of Cambridge, a co-founder of the Human Cell Atlas consortium, said in the post that the appeal is how the model "adapts directly to new single-cell data without retraining." Genomics Limited, led by CEO Peter Donnelly, plans to build AlphaGenome 2 into its Mystra platform, according to DeepMind.
DeepMind has published a technical paper with the full method alongside the post. The model joins a growing list of DeepMind science tools, after the math focused AlphaProof Code. More coverage of models like this is on our AI page.
Sources
Share this story
Spotted an error? Tell us


