Google DeepMind Releases AlphaGenome Atlas for 9 Billion DNA Variants

Google DeepMind has released the AlphaGenome Atlas, a comprehensive catalog that provides precomputed predictions for the molecular effects of approximately 9 billion single-nucleotide variants within the human genome. This extensive resource aims to accelerate genetic research by offering predictions for every possible single-letter change in human DNA. The release also introduces the AlphaGenome Variant Impact (AVI) score, a novel metric designed to rank genetic variants based on their predicted impact on molecular processes. In addition to the AVI score, the Atlas provides per-variant feature attributions and a genome-wide motif collection, offering deeper insights into the functional consequences of genetic alterations. The AlphaGenome Atlas is accessible as a free web portal for academic use and through the AlphaGenome API. Commercial access is planned for Google Cloud in the near future. The underlying AlphaGenome model, which predicts how DNA variants influence molecular processes like gene expression and RNA splicing, is already available for academic use on GitHub and for commercial deployment via Model Garden on Google Cloud. Previously, AlphaGenome was used to analyze variants one at a time or in specific genomic regions. The AlphaGenome Atlas represents a significant shift by processing all 9 billion single-nucleotide variants and storing the resulting predictions, creating a massive 1-petabyte dataset. This scale is substantially larger than the AlphaFold Database, which contains over 200 million protein structure predictions. The Atlas addresses the impracticality of laboratory testing for billions of mutations and the computational slowness of running large models on demand for genome-scale studies. By providing a lookup table with integrated interpretations, it removes these significant bottlenecks for researchers. The Atlas exposes four interconnected resources: thousands of molecular effect predictions per variant, covering various aspects of gene regulation across hundreds of human and mouse cell types and tissues; the AVI score, a single numerical value for each variant that synthesizes AlphaGenome's regulatory predictions with insights from AlphaMissense, DeepMind's model for predicting the pathogenicity of missense variants; per-variant feature attributions that highlight which specific molecular features contribute to a variant's predicted impact; and a genome-wide motif collection that identifies conserved DNA sequences associated with regulatory elements. The AlphaGenome model itself was initially released in June 2025, and its ability to predict molecular consequences of DNA variations has been widely adopted. The Atlas expands this capability from individual variant analysis to a genome-wide resource, making it a powerful tool for understanding the functional impact of genetic diversity.