Latest InsightsIssuesImpressionsStories 

DeepMind DNA variant prediction

DeepMind’s AlphaGenome 2 predicts how DNA variants act in specific cell types, with early benchmark gains

The company says the model finds nearly twice as many expression-altering variants as the previous best, though the results come from a preprint and are for research use only.

Google DeepMind has released AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect specific human cell types and states. The company announced it on November 11, and Nature’s news section ran a feature on it in its November 19 issue. Unlike the original AlphaGenome, which predicted many tissues in parallel, the new model takes one condition (a tissue, cell type or cell state) alongside up to one million DNA bases and analyses them together, according to DeepMind and TechRepublic.

DeepMind says AlphaGenome 2 was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles from 68 tissues. It adds DNA methylation to the model’s outputs, alongside gene expression, histone modifications and chromatin accessibility, and can predict an unmeasured readout from a measured one.

The company reports that the model identifies nearly twice as many variants that change gene expression as the previous state of the art. On the TenK10K dataset, which covers 28 immune cell types, it found causal variants in 30% of cases against 16% for comparison models, and the direction of gene effects in 43% against 23%. DeepMind says it reproduced a computational finding linking a distant regulatory variant to overactivation of the ETS2 gene in gut macrophages in inflammatory bowel disease, and that it separated activated from resting T cells in a COVID-19 vaccine example using a few dozen cells and no retraining. The benchmark comparisons come from the authors’ preprint on bioRxiv and were not independently verified in the reports reviewed.

Sarah Teichmann of the University of Cambridge, a co-founder of the Human Cell Atlas, said the model “adapts directly to new single-cell data without retraining”, according to TechRepublic. Žiga Avsec, DeepMind’s genomics lead, told The Verge that its detailed predictions are more expensive to produce, so the original AlphaGenome will stay in use for genome-wide scans.

The model is free for non-commercial research through the AlphaGenome API, and commercial organisations can apply for a private preview through Google Cloud. DeepMind says it is intended for research use only and has not been validated or approved for clinical or diagnostic use, and that its predictions need experimental validation. The reports do not say how the model performs on cell types that are rare in its training data, or how its results compare when tested by groups outside DeepMind. This article is for information only and is not medical advice.

Tags:

No responses yet

Leave a Reply