Google DeepMind Releases ‘AlphaGenome Atlas,’ With Predicted Effects of 9 Billion Human Genome Variants

Google DeepMind has released the AlphaGenome Atlas, a resource containing precomputed predictions of the effects of about 9 billion possible single-nucleotide variants in the human genome.

According to Nature News on the 9th (local time), DeepMind expanded an existing AI model into a large-scale prediction data resource to make it easier to interpret the function of noncoding variants and identify candidates associated with disease.

Only about 2% of the roughly 3 billion bases in the human genome encode proteins. It remains difficult to interpret how variants in the rest of the noncoding regions affect gene regulation and disease.

The atlas generated predictions for about 9 billion possible single-nucleotide variants by calculating the effects of replacing each base in the human genome with each of the other three bases. It also includes more than 100 million short insertions and deletions observed in the human genome.

The resulting dataset is about 1 petabyte and is available free of charge for noncommercial use. For each variant, it contains thousands of predictions about a range of molecular effects, including tissue-specific gene expression and chromatin structure.

AlphaGenome is an AI model that predicts variant effects from DNA sequences; using it through an API required coding and computational resources. About 9,000 researchers accessed the API after the model was released, but it posed a barrier to entry for some biology researchers.

By running large-scale predictions in advance and making the results available, the atlas broadens access from “running the model directly” to “finding and using the predictions needed.” The approach is intended to build a shared research resource, similar to AlphaFold’s large-scale protein structure prediction database.

To reduce the burden of interpreting numerous predictions, the researchers also developed the AlphaGenome Variant Impact (AVI) score. The metric summarizes the potential effects of a variant on biological function and helps identify candidates for further analysis.

A Broad Institute research team used AVI scores and the atlas’s predictions to prioritize noncoding variants as candidates that may be responsible for severe epilepsy in patients. A researcher at the University of Exeter in the United Kingdom used the resource to interpret rare noncoding variants associated with blood protein levels, drawing on whole-genome sequences and clinical information from about 54,000 UK Biobank participants.

However, the current atlas changes one variant at a time in a reference genome, limiting its ability to directly account for individual genetic backgrounds and the combined effects of multiple variants. Experts emphasized that clinical decisions should not be made based solely on predictions from genomic AI models.

Researchers say a framework is needed that uses AI to prioritize promising variants, then verifies their actual function and disease relevance through follow-up experiments and patient information. A key challenge going forward is to obtain experimental data to validate predictions and improve the model as the predictions accumulate.