Google DeepMind unveils AlphaGenome 2 for cell-level DNA analysis

A model built around individual cell types

AlphaGenome 2 is the successor to Google DeepMind’s AlphaGenome, which predicted how DNA variants affect molecular activity across broad tissue types. The new model is designed to examine how the same DNA sequence is interpreted differently by particular tissues, cell types and biological states.

It processes a DNA sequence of up to one million letters together with a biological condition. That condition can come from a reference catalog of cell types or from a scientist’s own single-cell experiment, using the model’s cell encoder. Google DeepMind says this conditional architecture helps the system focus its capacity on specific cell types instead of predicting many tissue outputs in parallel.

The model can also use one type of measurement to predict another. For example, a researcher can provide single-cell gene activity and use AlphaGenome 2 to estimate unmeasured properties such as chromatin accessibility. The model retains AlphaGenome’s predictions for gene expression, histone modifications and chromatin accessibility, while adding DNA methylation.

Training and reported performance

Google DeepMind says AlphaGenome 2 was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. The company says the model can resolve activity in specialized cells, including distinct heart muscle and blood immune cells.

On the TenK10K dataset, which maps 28 primary immune cell types, AlphaGenome 2 identified nearly twice as many causal variants as the original AlphaGenome fine-tuned on single-cell data and other specialized tools. Its reported results were 30% versus 16% for causal variants and 43% versus 23% for directional gene effects.

Google DeepMind also says AlphaGenome 2 can analyze new single-cell experiments without retraining. Scientists can upload gene activity from a few dozen cells, allowing the model to create a representation of an unseen cell state. In an example involving a COVID-19 mRNA vaccine study, it distinguished activated immune T cells from resting cells and found that a variant linked to severe COVID-19 had a stronger effect in the activated state.

Connecting genetic variants to disease biology

AlphaGenome 2 is intended to help researchers connect variants identified in genome-wide association studies with the cell types where disease processes occur. Google DeepMind says it has precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions.

The system can extend enhancer-gene mapping to primary human cell types, including brain astrocytes. These maps may help researchers connect distant regulatory variants with the genes they affect in specific cells, an area where experimental CRISPR screens often provide limited information because they mainly use stable, lab-grown cell lines.

In a disease example, AlphaGenome 2 reproduced findings that a distant regulatory variant linked to inflammatory bowel disease increases activity of the ETS2 gene in gut macrophages. According to Google DeepMind, the model identified the variant’s selective effect in those cells while leaving other cell types and genes largely unaffected.

AlphaGenome 2 is available for non-commercial research through the AlphaGenome API and is in private preview for commercial research. The release gives scientists a tool for exploring variant effects at the level of individual human cell types and states, rather than relying only on broad tissue averages.