AlphaGenome 2 Nearly Doubles Variant Finds

Google DeepMind on November 11 released AlphaGenome 2, a DNA sequence model that predicts how genetic variants change gene activity inside specific human cell types rather than across whole tissues. The company says the model finds nearly twice as many expression-altering variants as its predecessor and is available now for non-commercial research through the AlphaGenome API.

Key takeaways

• On the TenK10K single-cell dataset covering 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants versus 16% for the fine-tuned original model, and 43% of directional gene effects versus 23%.
• The model was trained on more than 107 million individual cell profiles from human and mouse samples, and adds DNA methylation as a new prediction target.
• Access is free for non-commercial research today, with commercial use in private preview; DeepMind says the model is not validated for clinical or diagnostic use.

How AlphaGenome 2 predicts variant effects by cell type

The first AlphaGenome, introduced in June 2025 and published in Nature in January 2026, read up to 1 million DNA letters and predicted thousands of experimental readouts across many tissues at once. According to DeepMind's announcement, that parallel design spread the model's capacity thin and blurred differences between closely related cells.

AlphaGenome 2 is instead a conditional model. It takes the same million-letter sequence plus one target condition, such as a tissue, cell type or cell state, and feeds that condition into every stage of the network. That lets it learn how one stretch of DNA is read differently in, for example, a heart muscle cell and a blood immune cell.

The conditional design also allows what DeepMind calls cross-modality imputation. If a lab supplies only single-cell gene activity data, the model can predict unmeasured properties such as chromatin accessibility, which could help researchers decide which experiment to run next.

Benchmark results and the ETS2 test

The headline figures come from TenK10K, a large single-cell map of immune cells. There, DeepMind reports that AlphaGenome 2 roughly doubled recall of causal variants and of the direction in which they push a gene, compared with the original model fine-tuned on single-cell data and other specialized tools.

DeepMind also checked the model against a known result. A 2024 Nature study showed that a variant in a distant regulatory region drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. Predicting across gut immune cell types, AlphaGenome 2 singled out the same causal variant and showed it boosting ETS2 in macrophages while leaving other cells and genes largely unaffected.

New cell states without retraining

Researchers can upload a single-cell RNA-seq snapshot from just a few dozen cells, which the model compresses into a cell-state representation and uses as its condition. DeepMind says this matches or beats models trained directly on those datasets. In data from a COVID-19 mRNA vaccine study the model had not seen, it separated vaccine-activated T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state.

"What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining," said Sarah Teichmann, co-founder of the Human Cell Atlas consortium and a professor at the University of Cambridge.

Why it matters for genetics research

Genome-wide association studies flag broad stretches of DNA tied to traits, but rarely pin down the exact variant, gene and cell involved. DeepMind has precomputed AlphaGenome 2 scores for trait-associated variants across hundreds of cell and tissue conditions, covering traits from height to type 2 diabetes, and extended enhancer-to-gene mapping to primary cells such as brain astrocytes.

The release follows September's AlphaGenome Atlas, a lookup table of predictions for all 9 billion possible single-letter changes in the human genome, and continues DeepMind's push to turn research models into tools, as it did with WeatherNext's cyclone forecasts. Peter Donnelly, CEO of Genomics Limited, called the model "an impressive advance in the field" and said his company plans to incorporate it into its Mystra platform.

What's next

DeepMind says AlphaGenome 2 Skills will "soon" arrive in Google Antigravity, its agentic development platform. It is collecting use cases from researchers until November 30, with selected projects eligible for extra quota and credits, and is working with the Wellcome Sanger Institute to profile more cell types and states.

FAQ

Who can use AlphaGenome 2?
Academic and other non-commercial researchers can use it now through the AlphaGenome API. Commercial research access is in private preview.

Can it diagnose genetic disease?
No. DeepMind states that AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic purposes.

How is it different from the first AlphaGenome?
The original predicted effects across bulk tissues in parallel. Version 2 predicts one cell type or state at a time, adds DNA methylation, and can adapt to new single-cell data without retraining.

References

• AlphaGenome 2: Deciphering how cells read DNA to better understand disease — Google DeepMind (Nov. 11, 2026)
• AlphaGenome: AI for better understanding the genome — Google DeepMind (June 25, 2025)
• A disease-associated gene desert directs macrophage inflammation through ETS2 — Nature (June 5, 2024)
• Everything That Happened in AI Today Thursday, November 12 — The Neuron (Nov. 12, 2026)