Google DeepMind releases AlphaGenome 2 to study disease

Google DeepMind on Wednesday, November 11, released AlphaGenome 2, a model that predicts how DNA sequences and their variants affect specific human cell types and states. The company describes it as a tool to help scientists find the cellular drivers of disease. It was trained on more than 100 million single-cell profiles from 68 tissues. According to the company, it identifies nearly twice as many expression-changing variants as the original AlphaGenome, which launched last year.deepmind.google+1

Researchers can use the model now for non-commercial work through the AlphaGenome API. Commercial research access is in private preview. DeepMind says the models are meant for research only and have not been validated or approved for clinical or diagnostic use.deepmind.google

A conditional redesign

The first AlphaGenome made predictions across thousands of experiments in parallel. AlphaGenome 2 is a conditional model instead. Along with a million-letter DNA sequence, it takes one tissue, cell type or state as input at a time, so it can learn how the same stretch of DNA is read differently in different cells. It was trained on human and mouse data, including more than 107 million individual cell profiles. It also adds DNA methylation to the readouts it predicts.deepmind.google

DeepMind tested it on the TenK10K dataset, which covers 28 immune cell types. There, it identified causal variants 30% of the time, compared with 16% for the previous model, and directional gene effects 43% of the time, compared with 23%. The model can also work with new single-cell data from just a few dozen cells without being retrained. In one test, it showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in vaccine-activated T cells.europapress.es+1

Linking genetics to drug targets

DeepMind genomics lead Žiga Avsec told The Verge that this level of detail matters for drug development. "When you develop drugs, you're targeting specific proteins and specific cell types," he said. He added that AlphaGenome 2 will not replace its predecessor soon, because its predictions are "more expensive" to compute. That cost would have made it impractical to build the genome-wide Atlas DeepMind released in September.theverge.com

Sarah Teichmann, co-founder of the Human Cell Atlas, said what makes the model exciting is that it "adapts directly to new single-cell data without retraining".businessinsider.es

What comes next

DeepMind has precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions. It plans to release AlphaGenome 2 Skills in its Google Antigravity agentic development platform. Peter Donnelly, CEO of Genomics Limited, called the model "an impressive advance in the field" and said his company plans to add it to its Mystra platform. Avsec said models like this could eventually feed into efforts to build "virtual cells," which are digital simulations of biological systems.theverge.com+1