TL;DR — Key Takeaways
- Google DeepMind released AlphaGenome Atlas with precomputed predictions for roughly 9 billion possible single-nucleotide variants.
- The searchable resource adds browser access and a new AVI score that helps researchers rank variants for further investigation.
- AlphaGenome is not validated for clinical use, and researchers say its predictions still require experimental and case-specific follow-up.
Google DeepMind has turned its AlphaGenome AI model into a searchable atlas containing predictions for the molecular effects of every possible single-letter change in the human genome, giving researchers a way to screen billions of genetic variants using precomputed predictions.
AlphaGenome Atlas, released today, contains predictions for roughly 9 billion single-nucleotide variants, representing the three possible substitutions at each of the genome’s approximately 3 billion DNA positions. DeepMind said the precomputed predictions total about one petabyte of data, more than 30 times the size of its AlphaFold Database.
The Atlas is available free for noncommercial research through a web portal, the existing AlphaGenome API and as a skill in Google Antigravity. DeepMind said it plans to offer commercial access through Google Cloud.
The release builds on the AlphaGenome AI model, which was described in a Nature paper published in January. AlphaGenome takes up to one million DNA letters at a time and predicts thousands of molecular measurements, including gene expression, RNA splicing and chromatin accessibility. In evaluations reported in the Nature paper, the model matched or outperformed leading external models on 25 of 26 variant effect prediction benchmarks.
The Atlas turns AlphaGenome’s predictions into a genome-wide resource, combining much greater scale with easier access to the results. Researchers previously had to submit variants through an API, which requires coding. DeepMind told Nature that about 9,000 researchers had used the API since AlphaGenome’s release. The Atlas makes those outputs available in advance and adds a browser interface for searching them directly.
DeepMind is also introducing the AlphaGenome Variant Impact, or AVI, score, which combines AlphaGenome predictions with AlphaMissense, its model for protein-altering variants. The score condenses a variant’s predicted biological effects into a single number so researchers can prioritize which variants to investigate. The work behind the Atlas and AVI score is described in a preprint here.
Early collaborators have tested the approach on disease and population genetics questions. Broad Institute researchers used the AVI score to help prioritize a noncoding DNM1 variant in a previously unsolved rare disease case. A University of Exeter researcher used Atlas predictions to identify additional associations between rare noncoding variants and traits in UK Biobank data.
DeepMind says AlphaGenome has not been validated or approved for clinical use. An outside researcher told Nature that its predictions will not replace laboratory experiments or the details of individual cases when diagnosing disease.
