Artificial Intelligence
oi-Gaurav Sharma
Google DeepMind has launched AlphaGenome Atlas – a giant knowledge base to help scientists find out how certain changes in the DNA sequence affect the genes and physiology.
AlphaGenome tries to solve a key problem of current genetics: how to understand which of the many genetic variations discovered actually affect health. After all, even a small change in the DNA sequence can have serious consequences: a single nucleotide polymorphism can increase the risk of developing a disease.
Google DeepMind launched AlphaGenome Atlas, an AI knowledge base predicting the functional impact of billions of human DNA variations, including non-coding regions, to help scientists prioritize genomic research for disease understanding using the AVI score.
The task of the new AI is to prioritize the study of genomic variations and indicate which ones have the greatest impact on biology and potential for disease.
AI performed an analysis of billions of possible changes
The human genome consists of about three billion base pairs, with four nucleotides in each pair – A, T, G and C. Each of the nucleotides can change to another, resulting in a number of possible variations that affect the function of the gene.
DeepMind says that its Atlas has predicted the effect of all nine billion single nucleotide variations across the human genome, including those that occur in non-coding regions.
This is significant for research, since most of the genetic material in humans is non-coding DNA, the function of which is not yet fully understood. However, these regions can contain enhancer regions that control the activity of genes, and mutations in these areas can be the reason for various diseases.
The key to studying diseases is the score
To understand which variation has a greater impact, researchers need to assess the variants based on specific scores that take into account their effect on mRNA or protein-coding sequences.
DeepMind developed a new algorithm called AlphaGenome Variant Impact score (AVI), which analyzes and prioritizes variants based on their potential effect.
This score allows researchers to find candidate variants for further experimental study.
Researchers note that the system does not claim to replace the full analysis of such variations, as it only provides preliminary data that should be confirmed by experiments.
From AlphaFold to AlphaGenome
AlphaGenome continues the series of DeepMind projects that solve genetic puzzles using artificial intelligence. Previously, the company developed AlphaFold, a program that predicted protein folds – the three-dimensional shapes of proteins. This area of research has long been a focus of attention for biotechnology companies: in 2024, Demis Hassabis and John Jumper were awarded the Nobel Prize in Chemistry for their work on protein structure prediction.
AlphaFold results have already been used in practice: For instance, using the AlphaFold database, scientists were able to determine the structure of a protein associated with lung adhesion and take a step towards understanding the mechanisms of lung disease.
Meanwhile, AlphaGenome analyzes variations in the human genome. Some of the problems it solves are similar to those faced by AlphaFold, namely, analyzing the functional impact of amino acid substitutions on protein activity.
Early results are encouraging
As stated in the article, AlphaGenome has already shown its worth in practice. The tool has already been used to analyze the rarest and hardest-to-study diseases, such as epileptic encephalopathy. Using the AVI score, researchers were able to find a mutation in the DNM1 gene associated with the disorder and confirm experimentally that this change in the DNA affects the processing of genetic information.
In addition, the authors note that when analyzing the genome data of 54 thousand patients from the UK Biobank, scientists were able to increase the number of discovered non-coding connections by 22 percent using the clustering method.
An invaluable dataset, but not a complete solution
It is worth noting that DeepMind has indeed achieved a significant result by creating the largest database of genetic variations to date. This data can serve as raw material for years of research into the causes of various diseases and the development of new drugs.
At the same time, AlphaGenome Atlas is still a tool for exploring the effects of a single nucleotide exchange. Even if such a change is significant, it is not possible to predict on the basis of this whether a person will develop a disease without a long process of verification.
Researchers point out that to confirm the impact of a particular change in the DNA sequence, it is necessary not only to analyze the results of the experiments but also to conduct a large-scale analysis of patients’ health and genetic information.
Therefore, the tool can greatly speed up and facilitate the search for the causes of diseases in the future, but it is not sufficient for a complete analysis of heredity
