Google DeepMind’s AlphaGenome 2 predicts how genetic changes may affect specific cell types, and noncommercial researchers can now use it through an API.
Why cell type matters
Many disease-associated DNA variants are outside genes that code for proteins, in regions that help control when and where genes switch on. Studies can point to a stretch of DNA linked to a disease but leave researchers unsure which change matters, which gene it affects, and in which cell. The same DNA can be read differently in an immune cell and a heart cell.
AlphaGenome followed the original AlphaGenome, announced in 2025 and later described in Nature. The earlier model analyzed DNA sequences up to one million letters long and predicted several kinds of molecular activity, but made predictions across a fixed set of tissues and cell lines. With AlphaGenome 2, researchers specify a cell type or state along with the DNA sequence. DeepMind says it trained the model on more than 100 million cell profiles across 68 tissues. Researchers can also use their own single-cell data to describe a cell state the model has not seen before.
What the predictions could show
A variant that affects only a small group of cells can be hard to detect in measurements averaged across many cell types. Cell-specific predictions could help researchers narrow a genetic association to a possible mechanism and choose variants and cells to test in the lab. DeepMind says the model can also predict some molecular measurements researchers have not collected, using other measurements as clues. These are research predictions, not demonstrations that the model can diagnose disease or identify a treatment.
In the TenK10K dataset, which covers 28 immune-cell types, DeepMind reports that at a 90% precision threshold AlphaGenome 2 recovered 30% of likely causal variants affecting gene expression. Its predecessor, fine-tuned on single-cell data, recovered 16%. This measures how many relevant variants the model found in that test, not whether it is right 30% of the time in general. The company’s evaluation on one dataset does not establish how well the model will perform across diseases, tissues, or populations.
A November 2026 bioRxiv preprint, which has not been peer reviewed, tested the earlier AlphaGenome by simulating disruptions to 2,111 promoters. Researchers found that some predicted signals of lost promoter activity did not consistently carry through to predictions of RNA levels in downstream parts of genes. The study tested AlphaGenome 1, not the new model, so it does not show that AlphaGenome 2 has the same weakness. It does point to the need for outside tests and experiments.
What comes next
Noncommercial researchers can try AlphaGenome 2 through DeepMind’s API; commercial research access is in private preview. The model could help labs prioritize experiments, but its predictions still need to be checked in biological samples. Researchers outside DeepMind will need to test it on other datasets and see whether its predictions hold up in the lab. Medical advances would require further steps, including showing that a predicted mechanism matters to patients.