Google DeepMind has unveiled AlphaGenome Atlas, a database that predicts at the molecular level how about 9 billion single DNA-letter variants that can occur in the human genome may affect biology. It is expected to accelerate genomics research and drug development by helping researchers quickly screen disease-related variants and analyse their biological impact amid vast genetic variation.

On Sept. 8 local time, foreign media outlets including The Verge reported that AlphaGenome Atlas is a prediction map that precomputes the effects of all possible single nucleotide variants (SNVs) in the human genome. Human DNA consists of four bases, A, C, G and T, and the genome has about 3 billion base pairs. Considering all possible substitutions at each position, the number of possible single-letter variants reaches about 9 billion.

Some of these variants have little effect, but others create differences between individuals or contribute to disease. The problem is that experimentally checking billions of variants one by one is virtually impossible. AlphaGenome Atlas uses the AlphaGenome AI model to predict in advance how each variant affects molecular-level biological processes such as gene regulation and protein production.

A key feature is that it expands predictions beyond the protein-coding regions, which account for about 2 percent of the human genome, to the non-coding regions that make up the remaining about 98 percent. Non-coding regions do not directly make proteins, but regulate when, where and how strongly genes are activated. For each variant, AlphaGenome Atlas predicts thousands of molecular effects including gene expression, RNA splicing, chromatin accessibility and evolutionary conservation, and also provides information on hundreds of types of human and mouse cells and tissues.

Google DeepMind also released an "AlphaGenome Variant Impact" score (AVI) that expresses a variant's potential impact as a single number. AVI combines results from AlphaGenome, which predicts gene-regulatory effects in non-coding regions, and AlphaMissense, which predicts protein changes. Researchers can use AVI to quickly prioritise variants expected to have a large impact among billions of variants and check which biological processes those variants affect.

AlphaGenome Atlas also includes information on more than 2,500 short DNA sequences that repeatedly appear in the genome, along with their locations. This allows analysis of regulatory functions of the genome, such as which cells a specific DNA sequence activates or suppresses gene expression in.

It has already been used in research. Broad Institute researchers used AVI to narrow candidate variants in rare-disease cases whose causes had not been found, and identified a DNM1 gene variant strongly associated with epileptic encephalopathy. AlphaGenome predicted the variant would create an incorrect RNA splicing site and produce an abnormally long protein, and experiments confirmed the prediction.

Gareth Hawkes (가레스 호크스), a Medical Research Council researcher at the University of Exeter in Britain, applied AlphaGenome Atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effects, he found 22 percent more associations between non-coding regions and traits than before. He also analysed links between hundreds of millions of non-coding variants and body mass index (BMI), then narrowed the scope to the top 1 percent that AlphaGenome predicted would have the highest impact and identified 19 genetic regions.

AlphaGenome, the AI model underpinning AlphaGenome Atlas, was released last year and predicts how DNA variants affect biological processes. Ziga Avsec (지가 아브섹), head of genomics at DeepMind, said the base model itself was already public but building it into a genome-wide database required time to precompute and analyse predictions for billions of variants.

The AlphaGenome Atlas dataset built this way is about 1 petabyte, more than 30 times larger than the AlphaFold database. Researchers can search variants on a web portal without separate coding work, and can also use it through the AlphaGenome API for large-scale analysis and via AlphaGenome features in Google's agent-based development platform, Antigravity.

Google DeepMind will provide AlphaGenome Atlas from Tuesday through its website for non-commercial research purposes. Commercial use will be offered later via Google Cloud. The information in AlphaGenome Atlas is research material based on AI predictions and has not been validated or approved for clinical use. It therefore cannot replace clinical judgement such as diagnosis or treatment.