AlphaGenome Atlas for variant research

The Atlas turns a difficult one-variant-at-a-time modeling problem into a searchable map. Google DeepMind precomputed AlphaGenome predictions for every possible single-nucleotide variant, producing a dataset of about one petabyte. Researchers can start with a specific location or a list of candidates, inspect predicted effects on gene regulation, and compare the resulting scores. The public website is intended to make this possible without writing code, while the AlphaGenome API serves computational workflows.

Its AlphaGenome Variant Impact (AVI) score condenses evidence from AlphaGenome and AlphaMissense into one prioritization signal across protein-coding and noncoding regions. The linked feature attributions matter as much as the headline number: they indicate which predicted mechanisms, such as RNA splicing or gene expression, drove a high score. The Atlas also links a catalogue of more than 2,500 DNA sequence motifs. This is useful when a team needs a plausible biological hypothesis to test, rather than merely a ranked list.

The most persuasive examples are research demonstrations, not blanket claims of clinical accuracy. Google describes a rare-disease case where an incorrect splice-site prediction helped support resolution of an unsolved case, and a complex-trait study using data from more than 54,000 UK Biobank participants. These examples show a route from model prediction to further analysis, but each condition, population, cell type, and variant still requires independent evidence. Model coverage cannot substitute for functional experiments, segregation evidence, or professional interpretation.