DeepMind's AlphaGenome 2 predicts DNA variant effects cell type by cell type

The successor model identifies nearly twice as many causal variants on an immune-cell benchmark, but it is costlier to run and will sit alongside the original rather than replace it.

AIPublished Updated
Image published by Google DeepMind with “AlphaGenome 2: Deciphering how cells read DNA to better understand disease”
Image: Google DeepMind · Source
What happened
Google DeepMind released AlphaGenome 2, a model predicting how DNA variants affect gene activity in specific human cell types and states.
What’s new
Instead of averaging across broad tissues, it conditions each prediction on one tissue, cell type or cell state, including states it has never seen.
Why it matters
Researchers studying disease genetics can trace variant effects to the specific cells involved, which DeepMind says broad-tissue models can miss.

Google DeepMind has released AlphaGenome 2, a successor to the DNA sequence model it launched last year. The new model predicts how genetic variants change gene regulation in particular human cell types and cell states, rather than in averaged tissues. It is available now through the AlphaGenome API for non-commercial research and in private preview for commercial users.

From parallel outputs to one condition at a time

According to DeepMind's announcement, the first AlphaGenome took a million-letter DNA sequence and produced predictions for thousands of tissue and cell-line experiments at once. The company says that approach spread the model's capacity thin and made it hard to separate closely related cell types.

AlphaGenome 2 is a conditional model. It still reads a million-letter sequence, but it also takes a single target condition as input, which can be a tissue, a cell type or a cell state. That condition can come from a reference catalogue or be generated from a scientist's own single-cell RNA-seq data covering as few as a few dozen cells, with no retraining required. The same design lets the model fill in missing measurements. For example, it can predict chromatin accessibility from gene activity data alone. DNA methylation has also been added to the readouts it covers.

DeepMind says the model was trained on bulk and single-cell data from humans and mice, including more than 100 million individual cell profiles across 68 tissues.

What the benchmarks show

The headline claim is that AlphaGenome 2 finds nearly twice as many variants that alter gene expression in cells. DeepMind's post grounds this in the TenK10K dataset, a single-cell map of 28 immune cell types. There the model identified 30% of causal variants against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools. It also captured 43% of directional gene effects against 23%.

DeepMind also describes two case studies. In one, the model computationally reproduced a published finding that a distant regulatory variant drives inflammatory bowel disease by boosting the ETS2 gene in gut macrophages. In the other, using unseen data from a COVID-19 mRNA vaccine study, it predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in activated T cells than in resting ones. These results are the company's own and come from its announcement.

The company has also precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions. It says these can help link genome-wide association study hits to target genes and cell types.

Costs and limits

AlphaGenome 2 will not replace its predecessor soon. DeepMind genomics lead Žiga Avsec told The Verge that its finer-grained predictions are more expensive to compute. He said that cost would have made it impractical to build the genome-wide Atlas of every possible single-letter change, which was produced with the original model.

DeepMind states that both AlphaGenome models are for research only and have not been validated or approved for clinical or diagnostic use.

Avsec argued the cell-level detail matters for therapeutics: “When you develop drugs, you’re targeting specific proteins and specific cell types.” He told The Verge that future versions could be trained on data from multiple species. DeepMind's post adds that it is working with the Wellcome Sanger Institute to profile more cell types and states. It also plans to offer AlphaGenome 2 skills in its Antigravity agentic platform.