Google DeepMind introduces AlphaGenome 2 to predict the effects of single-letter DNA changes in specific cell types: the company reports that, with nine out of ten predictions correct, it identifies 30% of causal variants, compared with 16% for the previous version
Google DeepMind introduces AlphaGenome 2 to predict the effects of single-letter DNA changes in specific cell types: the company reports that, with nine out of ten predictions correct, it identifies 30% of causal variants, compared with 16% for the previous version
On November 11, DeepMind made the model available for noncommercial research through an application programming interface. The model takes a million letters of DNA and a “condition” specifying the tissue, cell type, or cell state, then predicts how much a single-letter change (a variant) will increase or decrease the activity of nearby genes in those cells.
Most disease-associated DNA variants lie within the 98% of the genome that does not encode proteins. Comparisons between people with and without a disease identify a chromosomal region containing several variants inherited together. Three questions remain: which variant causes the disease, which gene it affects, and in which cells it acts. One variant increases IL20RA gene activity in the lungs but decreases it in the lining of the esophagus. The model predicts that this variant decreases the gene’s activity in cells that line organs and increases it in cells that line blood vessels. The overall effect in a tissue depends on the relative abundance of these cells.
The first AlphaGenome, described in Nature, simultaneously predicted thousands of measures of DNA function averaged across whole tissues. The AlphaGenome Atlas of predictions for nine billion single-letter changes, released in September and built using the first version, did not identify the cells in which a variant acts. AlphaGenome 2 was trained on cell atlases containing measurements of gene activity in 107 million individual cells. The condition enters the neural network alongside the sequence and influences processing from the first layers onward. In the authors’ tests, supplying it only to the final layers almost entirely removed the improvement.
DeepMind tested the model on previously unseen data from the TenK10K project: DNA from 1925 people and gene activity in 5.4 million blood immune cells spanning 28 types. Statistical analysis of these participants identified variants highly likely to alter gene activity in a specific cell type, along with an equal number of control variants with no effect. According to the company’s calculations, when nine out of every ten variants classified as causal must be correct, the model identifies 30% of causal variants. The previous version, fine-tuned on single-cell data, identifies 16%.
In 2024, immunologist James Lee’s group at the Francis Crick Institute showed in Nature that a variant in a “gene desert” on chromosome 21 (a region without protein-coding genes) increases the risk of inflammatory bowel disease. It increases ETS2 gene activity in macrophages despite lying 290 thousand DNA letters away. Statistical analysis narrowed the search to seven candidates. The model evaluated all of them and predicted a strong effect only for the variant identified by Lee’s group, linked it to ETS2, and localized its effect to macrophages and their precursors. “They found the answer in 30 minutes, while it took us a year and a half of experiments,” Lee says.
For new data from a COVID-19 vaccine study, the model derived the condition solely from gene activity in blood cells from vaccinated participants, without further training. A variant associated with severe COVID-19 lies within FYCO1, but the model predicts that it reduces the activity of the neighboring gene CXCR6 more than that of FYCO1. The predicted effect is stronger in T-cells activated during the immune response than in resting T-cells. A 2021 analysis linked the risk of severe COVID-19 to lower CXCR6 activity in T-cells.
DeepMind precomputed the effects of 6.8 million variants across 510 conditions. For 17 of the 28 diseases and traits studied, the predictions identified cell types in which associated variants alter gene activity more often than common variants overall. For Alzheimer’s disease, these cells were microglia, the brain’s immune cells.
- https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
- https://doi.org/10.1038/s41586-025-10014-0
- https://t.me/UkhvatNews/3973
- https://doi.org/10.1038/s41586-026-10721-2
- https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf
- https://doi.org/10.1038/s41586-024-07501-1
- https://www.nature.com/articles/d41586-026-03673-0
- https://doi.org/10.1007/s00439-021-02305-z