AI atlas models the effects of nine billion DNA variants

A new AI-powered database predicts the effects of single-nucleotide changes in any of the three billion bases in human DNA.

The AlphaGenome Atlas, released by Google's DeepMind, contains precomputed predictions for all nine billion possible single-nucleotide variants, since each nucleotide has three possible replacements. The work also includes predictions for more than 100 million short insertions and deletions observed in human genomes. The Atlas uses predictions from DeepMind's AlphaGenome model, which predicts the molecular and regulatory effects of genetic variants in both coding and noncoding DNA (see BioNews 1325).

'Understanding our DNA is a grand challenge. Understanding this language of life can unlock so many things,' Pushmeet Kohli, vice president of research at Google DeepMind, told IEEE Spectrum.

Coding DNA encodes proteins, while certain sections of non-coding DNA can regulate the expression of nearby or far away genes. However, much of the function of noncoding DNA, which makes up around 98 percent of all DNA, is poorly understood. The Atlas can indicate which molecular processes, cells or tissues are most likely to be affected by a genetic variation in a non-coding region.

The AlphaGenome Atlas creators devised a metric, known as the AlphaGenome Variant Impact score, that helps grade the potential biological impact of each genetic variant. The score is established by combining predictions from AlphaGenome and AlphaMissense, a model that predicts the effects of protein-altering variants (see BioNews 1209). The score increases with the variant's predicted biological impact.

The Atlas dataset is about one petabyte (1000 terabytes) in size. To reduce the computational resources required to generate and access the predictions, the team at DeepMind implemented model distillation and GPU kernel optimisation, and eliminated redundant calculations. Unlike the previous AlphaGenome model, the Atlas does not require the user to write any specialised computer code.

The researchers note that the AlphaGenome uses a one-million-base-pair sequence window around each variant, meaning that regulatory DNA located outside this window is not included in its predictions.

Furthermore, many diseases involve variants in several genes; the Atlas analyses individual genetic variants rather than combinations of variants, so it does not capture effects that arise from interactions between multiple variants.

Researchers at the University of Exeter have already used the Atlas to analyse whole-genome data from more than 54,000 UK Biobank participants, identifying 22 percent more noncoding genetic associations by grouping rare variants according to their predicted molecular effects.

'It won't replace experiments or, in the case of diagnosing disease, accounting for differences specific to individuals,' Professor Martin Kircher, a researcher at the Max Delbrück Centre for Molecular Medicine in Berlin, Germany told Nature News. 'This is a useful and generous way to scale up access to a strong model.'

---

The impact of AI on genomics will be explored at this year's PET Annual Conference, AI and Automation in Fertility and Genomics: Pipeline? Or Pipe Dream?, taking place in central London on Wednesday 9 December 2026.

Find out more and register here.