Google DeepMind Releases AlphaGenome 2 to Predict How DNA Variants Act in Specific Cells

The distinction matters for drug development. Almost every cell in the body carries the same genome, but heart, skin and immune cells switch different genes on and off, and the same cell behaves differently when it is resting or responding to an infection. A genetic variant that does nothing in one cell can be harmful in another. "When you develop drugs, you're targeting specific proteins and specific cell types," Žiga Avsec, DeepMind's genomics lead, told The Verge.

According to DeepMind's announcement, AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model. It is available now for non-commercial research through the AlphaGenome API, with a private preview for commercial research. DeepMind said the model is for research use only and has not been validated or approved for clinical or diagnostic purposes.

One Condition at a Time

The original AlphaGenome, released last year, read a sequence of up to one million DNA letters and predicted many tissues in parallel. In September DeepMind used it to publish the AlphaGenome Atlas, which gives predictions for all 9 billion possible single-letter changes in the human genome.

AlphaGenome 2 works differently. It is a conditional model that makes predictions for one tissue, cell type or cell state at a time. Scientists can pick a condition from a reference catalogue or feed in their own single-cell experiment through what DeepMind calls a cell encoder. The company said the model can handle new cell states without retraining, using a compressed snapshot of single-cell RNA data from as few as a few dozen cells. It can also infer readouts that were not measured, such as chromatin accessibility, from a single measurement of gene activity.

DeepMind said the model was trained on bulk and single-cell data from humans and mice, covering more than 100 million individual cell profiles across 68 tissues. It predicts gene expression, histone modifications, chromatin accessibility and, for the first time in the AlphaGenome family, DNA methylation.

The Immune Cell Test

On the TenK10K dataset, which covers 28 primary immune cell types, AlphaGenome 2 found the causal variant in 30% of cases, against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, according to the company. For predicting whether a variant raises or lowers a gene's activity, the figures were 43% and 23%.

DeepMind gave two disease examples. In inflammatory bowel disease, the model reproduced earlier laboratory findings that a variant in a distant regulatory switch overactivates the ETS2 gene in gut macrophages, and showed that the effect was specific to those cells. In data from an mRNA vaccine study it had not seen, it separated vaccine-activated T cells from resting ones and found that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells.

Avsec told The Verge that the model's finer-grained predictions are more expensive to compute than its predecessor's, which is why the original AlphaGenome remains in use and why the newer model would have been impractical for building the genome-wide atlas.

DeepMind said it is working with the Wellcome Sanger Institute to profile more human cell types and states. Peter Donnelly, chief executive of Genomics Limited, said his company plans to build the model into its Mystra platform, according to the announcement. DeepMind also said AlphaGenome 2 skills are coming to its Google Antigravity agent platform, and researchers can apply until November 30 for extra API quota and credits under a community programme.