DeepMind's new genomics model doubles the ability to find disease-causing variants

Google DeepMind has released AlphaGenome 2, a model that predicts how DNA variants, such as single-letter differences in genetic code, affect the body. What is new is that it can make predictions for one specified cell type and state at a time. According to the company, it finds about twice as many variants that change gene activity as the best previous method. That could speed up the work of narrowing down candidates in research aimed at finding variants that cause disease. The model is available via an API for non-profit research, but its weights (the trained parameters that make up the model) have not been made public.

For Japanese companies: Most companies will not be affected. The ones it matters to are research divisions at pharmaceutical, genetic testing and diagnostics companies. If you want to use it for in-house research, first check DeepMind's terms of use with your legal team to see whether your research qualifies under the API's restriction to non-profit research.

What it can do now

Liver cells and brain cells carry the same DNA, but different genes are active in each. Even in the same cell, genes behave differently when the cell is inflamed than when it is at rest. That means whether a given variant leads to disease can depend on which cell, and which state, you look at.

AlphaGenome 2 predicts the effects of a variant for one specified cell type and state at a time. According to DeepMind, it finds about twice as many variants that change gene activity as the best previous method. That gives researchers more clues for narrowing down which of the many variants found in a patient's DNA are actually involved in disease.

Who can use it and who can't

Non-profit researchers can use it through an API. Because the model weights have not been released, it cannot be installed and run on your own computers. Hospitals and companies that cannot send data outside their organizations will need to think about how to use it. Whether highly sensitive data, such as patients' genetic information, can be sent to the API will be up to each organization's own rules.

What we could not confirm

We were not able to read DeepMind's blog post or technical paper directly for this story. This article does not report the announcement date, the names of the methods it was compared against, the evaluation used to measure the "about twice" figure, the length of DNA the model can handle, how it differs from the previous version of AlphaGenome, or the terms for commercial use. The "about twice" figure was reported by DeepMind itself and has not yet been independently verified by outside researchers.