Google DeepMind announces AlphaGenome 2 DNA model for cell-level predictions

In short: Google DeepMind has introduced AlphaGenome 2, an AI model that predicts how specific DNA changes could affect different kinds of human cells.

What happened

Google DeepMind announced AlphaGenome 2, a new version of its AlphaGenome model released last year. The goal is to better understand how small changes in DNA, often called variants or mutations (like a single letter change in a long instruction manual), can lead to disease.

DeepMind says the key improvement is detail. Even though almost every cell in your body has the same DNA, different cells use different parts of it. A change in DNA might matter a lot in one cell type, like an immune cell, and barely matter in another, like a skin cell.

According to DeepMind, AlphaGenome 2 was trained on more than 100 million individual cell profiles across 68 tissues. The company says it can find nearly twice as many DNA variants that may change gene expression, meaning whether a gene is turned up or down in a cell (like a dimmer switch). DeepMind also says the model can make predictions about cell types or cell “states” it has not explicitly seen before, and it can point out where lab data is missing.

DeepMind is making AlphaGenome 2 available for non-commercial research through the AlphaGenome API, which is a way for researchers to send questions to the system and get results back. It will also offer a limited private preview for commercial research.

Why it matters

This kind of tool could help scientists narrow down which DNA changes are most likely to cause problems in specific cells, which is important for understanding disease and for designing treatments that target the right cells and proteins. DeepMind noted that AlphaGenome 2 is more costly to run than the older model, so it is not replacing the original AlphaGenome for every use.

Source: The Verge AI