Google DeepMind has unveiled the AlphaGenome Atlas, a map of the human genome that calculates the potential consequences of all 9 billion possible single-point mutations. These are mutations where one “letter” of the genetic code is replaced by another. Such changes represent one of the most common types of genetic variation; while most are harmless, some influence disease risk, and rare variants can directly cause disease.
At the heart of the atlas is the AlphaGenome model, which predicts thousands of potential effects for each mutation. These include changes in the activity of nearby genes across different tissues and alterations to chromatin (the structure in which DNA is packaged inside the cell). DeepMind calculated the effects for each of the three possible substitutions for every letter in the human genome, generating a dataset of approximately one petabyte. The atlas also incorporates over 100 million short DNA insertions and fragments found in human genomes.
For each mutation, DeepMind has included a unified metric called the AlphaGenome Variant Impact (AVI) score, designed to help quickly identify which variants warrant further study. In testing, the AVI score successfully distinguished between disease-causing mutations and harmless changes. The atlas’s predictions have already helped a team at the Broad Institute identify a variant in a non-coding region of DNA as a potential cause of severe epilepsy. However, the atlas is intended primarily for identifying and prioritizing variants rather than for making clinical diagnoses.
The ability to analyze non-coding DNA is particularly significant; 98% of the human genome does not contain instructions for protein production, and the functions of many of these regions remain unclear. Using AlphaGenome predictions, the DeepMind team mapped thousands of short sequence motifs and hypothesized their roles across different cell types: some may activate genes, while others might suppress them or alter DNA accessibility for transcription. AlphaGenome Atlas aims to lower the computational barrier for geneticists: previously, obtaining such predictions via AlphaGenome required programmatic interaction with the model, whereas the results are now compiled in a ready-made database available free of charge for non-commercial use.
However, the system analyzes mutations individually against a reference genome and does not directly account for combinations of multiple variants or individual genetic differences. The authors and independent experts also emphasize that these predictions do not yet replace experiments and should not be used in isolation to make clinical decisions.
About the Author
