Google DeepMind released AlphaGenome 2 on November 11, a model that predicts how a DNA variant changes gene activity in a named cell type instead of a tissue average. On a map of 28 immune cell types, DeepMind says it catches roughly double the causal variants its predecessor did. Access is split: non-commercial researchers can use it now, while commercial users get a private preview.
The gain comes from an architecture change. The first AlphaGenome took a million-letter stretch of DNA and returned predictions for thousands of experiments in one pass, across tissues and cell lines. Version 2 drops that fixed menu. It is a conditional model, meaning it takes the DNA plus one target condition (a tissue, cell type, or cell state) and predicts for that condition alone. DeepMind’s account is that the old parallel design stretched the model too thin to separate closely related cell types.
The headline numbers come from TenK10K, a single-cell dataset covering 28 primary immune cell types. On it, AlphaGenome 2 correctly identified causal variants 30 percent of the time against 16 percent, and directional gene effects 43 percent of the time against 23 percent. The comparison set was a version of the first model retrained on single-cell data, plus other specialized tools. Those figures are DeepMind’s own, and the announcement cites no independent replication.
One worked example is a retrospective. An earlier published study had traced inflammatory bowel disease to a variant sitting in a far-off regulatory switch, which overactivates ETS2 in macrophages of the gut. Given gut immune cells, the model singled out that variant and showed it boosting ETS2 in macrophages, with little effect on other cell types or other genes. That is a rediscovery of a known result, which makes it a sanity check rather than a new finding.
The more practical feature is adaptation without retraining. A scientist can upload single-cell RNA-seq data from as few as a few dozen cells, and a cell encoder compresses it into a state representation that conditions the model. DeepMind says the result can match or beat models trained or fine-tuned on those same datasets. On data from a study of mRNA vaccination against COVID-19, the model told activated T cells apart from dormant ones. It also flagged that a variant tied to severe disease from the virus does more damage to gene activity in the activated state.
The model also fills gaps. Feed it one measurement, such as single-cell gene activity, and it can predict others, such as chromatin accessibility, which helps decide which experiment to run next. It adds DNA methylation as a readout and was trained on more than 100 million cell profiles from human and mouse samples. DeepMind also ships precomputed scores for variants linked to traits, covering hundreds of conditions, with traits ranging from body height to diabetes.
The limits are stated plainly. DeepMind says AlphaGenome and AlphaGenome 2 are for research only and have not been validated for clinical or diagnostic use. Access runs through the AlphaGenome API, and the company plans to release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform. Teams with projects to showcase have until November 30 to apply for extra quota and credits.
For biotech and pharma groups, the open item is the commercial track. The announcement describes a private preview but no pricing or licensing terms, so any company wanting to run the model on proprietary single-cell data for target discovery should ask for those terms before building a pipeline around it.
Reported by Google DeepMind on November 11, 2026.
