DeepMind's AlphaGenome 2 predicts how DNA variants act in specific cell types

Google DeepMind has introduced AlphaGenome 2, a model that predicts how genetic variants behave in particular cell types and states, Robert Hart reports for The Verge. It follows last year's AlphaGenome.

Why cell type matters

Almost every cell in your body carries the same DNA, but a heart cell, a skin cell and an immune cell fighting an infection turn different genes on and off, and that changes with what the cell is doing. So one variant can be harmless in one cell and important in another. AlphaGenome 2 is built to see that difference, which broader models miss.

What it can do

• Training data: over 100 million single-cell profiles drawn from 68 tissues. DeepMind says it turns static cell atlases into a predictive engine for human genetics.
• Results: nearly twice as many variants that change gene expression as DeepMind's previous state-of-the-art model.
• Generalizing: it can make predictions for cell types and states it wasn't trained on, without retraining, and can point out gaps in experimental data worth filling in the lab.

Genomics lead Žiga Avsec told The Verge this specificity matters for treatment, since drugs aim at specific proteins in specific cell types. Research scientist Adam Gayoso called the big step tying human genetics to cell biology.

The cost

It doesn't replace AlphaGenome. Its predictions cost more compute, which would have made it impractical for the genome-wide Atlas of every possible single-letter DNA change that DeepMind released in September.

AlphaGenome 2 is available through the AlphaGenome API for non-commercial research now; commercial researchers can get in through a small private preview. DeepMind is collecting more data, including for poorly covered cell states, and Avsec mentioned multi-species training and, longer term, "virtual cells."

Scoreboard update: AI +1. An AI model maps how DNA changes act inside individual cell types.

Stay undead.