DeepMind launches AlphaGenome for predicting genome regulatory activity

Google DeepMind released AlphaGenome, a deep-learning model that takes a DNA sequence of up to one million base pairs and predicts thousands of properties describing its regulatory activity — how strongly nearby genes are transcribed, where the DNA is accessible to other molecules, how it folds and contacts other DNA regions, and where RNA is spliced. The model outputs predictions at single-nucleotide resolution, meaning it can estimate the effect of changing one letter within a million-letter sequence, and was made available initially through an API for non-commercial academic use, alongside a preprint on bioRxiv.

Most of the human genome does not code for proteins; it consists of regulatory sequences that determine when, where and how strongly genes are switched on. Interpreting variants in this non-coding DNA — including many linked to disease risk in genome-wide association studies — has been harder than interpreting coding mutations, because the mechanism by which a given variant changes gene expression is often unclear. AlphaGenome was built to predict that mechanism directly from sequence, and DeepMind reported it performing competitively or better across a range of established genomics benchmarks compared with existing specialised models.

The release extended DeepMind’s earlier genomics and structural-biology work, following AlphaFold’s protein structure predictions, into the regulatory layer of the genome rather than the proteins themselves. DeepMind said it would open the model’s code and weights for non-commercial use, which it did the following January. Researchers cautioned that predicted regulatory effects still require experimental validation, and that the model’s usefulness for actual disease-variant interpretation would depend on independent testing outside DeepMind’s own benchmarks — the standard caveat attached to computational biology tools before they see clinical use.