Google DeepMind announced on Wednesday (11) AlphaGenome 2, the next generation of its artificial intelligence model for genomics. While the first AlphaGenome predicted the effects of DNA variants across broad tissues, the new model predicts how DNA sequences and variants affect specific types of human cells — such as a heart muscle cell, an immune cell fighting an infection or a brain astrocyte.
The advance comes from a redesigned architecture: AlphaGenome 2 is a conditional model. Alongside a one-million-letter stretch of DNA, it takes the target biological condition — tissue, cell type or cell state — and makes its prediction for that specific condition. Training used human and mouse samples, including more than 107 million individual cell profiles across 68 tissues.
Benchmark results
In benchmarks, the gain was striking. On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of likely causal variants, versus 16% for the original AlphaGenome fine-tuned on single-cell data, and got the direction of the effect on gene expression right in 43% of cases, versus 23%. The model also now predicts DNA methylation, a widely requested readout.
The model also generalizes to new experiments without retraining: just a few dozen cells from a single-cell RNA experiment are enough to condition it on unseen states, including disease or vaccination states. In one example cited by DeepMind, using data from a COVID-19 mRNA vaccine study, the model distinguished 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.
AlphaGenome 2 also computationally reproduced a discovery about inflammatory bowel disease: in about 30 minutes, according to Nature, the model reached the same answer that took a year and a half of experiments — a variant in a distant regulatory switch drives the disease by overactivating the ETS2 gene in gut macrophages. “They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments,” James Lee, an immunologist and geneticist at the Francis Crick Institute who led the 2024 study, told Nature.
To speed up research, DeepMind precomputed AlphaGenome 2 predictions for trait- and disease-associated variants across hundreds of human cell and tissue conditions, helping connect variants found in genome-wide association studies (GWAS) to the genes and cells where they act.
Availability
Availability: AlphaGenome 2 is available starting today for non-commercial research via the AlphaGenome API, with interactive Colab tutorials; commercial access is in private preview. DeepMind stresses the model is intended for research only and has not been validated or approved for clinical or diagnostic use.
Image credit: Google DeepMind — chart comparing the performance of AlphaGenome 1 and AlphaGenome 2.
Sources: Google DeepMind — AlphaGenome 2: Deciphering how cells read DNA to better understand disease · Nature — DeepMind's ultra-powerful AI for genome mapping wows scientists
Transparency: this article was produced with the assistance of artificial intelligence, drawing on public posts on X and sources from the internet. Please consult the original sources listed above to verify the details.
By Muse
