Google DeepMind Maps 9 Billion DNA Variants With AI

Google DeepMind has unveiled AlphaGenome Atlas, an artificial intelligence-powered resource that predicts the biological effects of all 9 billion possible single-letter changes in the human genome.

The database could help researchers tackle one of genetics’ biggest challenges: identifying which DNA variants are actually important from among an enormous number of possibilities. DeepMind says the Atlas pre-computes predictions for every possible single-nucleotide variant and makes them available through an online research platform.

The resource is designed to help scientists prioritise potentially important genetic changes, investigate rare diseases and study how mutations affect gene regulation. It is available for non-commercial research, while commercial access is planned through Google Cloud.

Atlas Covers 9 Billion Changes

The human genome contains roughly 3 billion DNA bases. At each position, a DNA letter can theoretically be replaced by one of the other three bases, producing about 9 billion possible single-letter substitutions.

DeepMind used its AlphaGenome model to calculate molecular-effect predictions for these variants in advance. The resulting dataset is about 1 petabyte in size, according to the company.

Instead of researchers having to run the computationally demanding model separately for individual variants, the Atlas provides an accessible catalogue of pre-computed predictions.

Focus Beyond Protein-Coding DNA

A key feature of the project is its focus on the vast majority of the genome that does not directly code for proteins.

Only about 2% of the human genome consists of protein-coding regions, while much of the remaining 98% is involved in regulating when, where and how genes operate. Changes in these non-coding regions can therefore influence biological processes even when they do not alter a protein's amino-acid sequence.

AlphaGenome Atlas provides predictions across both coding and non-coding regions, allowing researchers to investigate regulatory effects alongside protein-related changes.

New Score Ranks Variants

DeepMind has also introduced the AlphaGenome Variant Impact (AVI) score to make the large dataset easier to use.

The score combines information from AlphaGenome with predictions from AlphaMissense, DeepMind's model for protein-altering variants. Researchers can use it to rank genetic variants according to their predicted biological impact and then examine the molecular processes behind those scores.

The Atlas includes predictions covering hundreds of human and mouse cell types and tissues, along with information on processes such as gene expression, RNA splicing and other regulatory activity.

Rare Disease Case Shows Promise

Early research suggests the tool could help solve some difficult rare-disease cases.

DeepMind says researchers working with the Broad Institute used the AVI score to prioritise variants in an unsolved case involving DNM1, a gene associated with epileptic encephalopathy. AlphaGenome predicted that a variant created an incorrect splice site, and subsequent laboratory experiments supported the prediction.

In a retrospective analysis of previously solved rare-disease cases, the known causal variant appeared among the top 50 candidates in 29.5% of cases using AVI, compared with 12.5% for CADD, an established variant-ranking method.

UK Biobank Data Adds Evidence

The Atlas has also been tested against large population datasets.

In research involving whole-genome data from more than 54,000 UK Biobank participants, investigators used AlphaGenome predictions to study rare non-coding variants associated with protein levels. DeepMind reported 22% more associations after incorporating the predicted molecular effects.

In one example, the approach reduced a region containing 526 candidate variants to just four candidates for further investigation.

These results suggest AI could help researchers reduce the enormous search space involved in genomic studies.

AI Predictions Need Testing

Despite its potential, AlphaGenome Atlas is not a diagnostic system or a replacement for laboratory research.

DeepMind itself describes the Atlas as a resource for prioritising and interpreting variants. Predictions indicate what a genetic change may do biologically, but experimental evidence is still needed to establish whether a variant actually causes a disease or produces a particular biological effect.

Independent experts have similarly cautioned that computational predictions cannot replace experiments or consideration of individual patient circumstances.

The model can also perform differently across types of variants and biological processes, meaning researchers will still need to validate important findings experimentally.

A New Genomic Research Tool

The AlphaGenome Atlas represents a major expansion of AI-assisted genomic research by turning billions of individual variant predictions into a searchable resource.

Its significance lies not simply in the number of DNA changes mapped, but in making predictions about their potential molecular effects easier for researchers to access and compare. By helping scientists move from millions of possible variants towards a smaller group of promising candidates, the technology could accelerate studies of rare diseases, gene regulation and potential therapeutic targets.

However, the Atlas should be viewed as a research and prioritisation tool rather than proof that a particular DNA mutation causes disease. Its ultimate value will depend on how researchers combine its predictions with genomic data, clinical evidence and laboratory experiments.