AI for Science
AlphaGenome Atlas Precomputes 9 Billion Single-Nucleotide Variants, Turning Genetic Effect Inference Into a Data Lookup
Google DeepMind has precomputed AlphaGenome predictions for every possible single-nucleotide substitution across the human genome, creating a 1 PB atlas that uses AVI scores to rank coding and noncoding variants on a unified scale. It sharply reduces the cost of large-scale screening, but the data remains model-generated and cannot replace functional experiments or clinical diagnosis.
Google DeepMind has released AlphaGenome Atlas, transforming genetic variant analysis from a process that requires repeated model calls into a precomputed data layer that can be queried directly. The team generated predictions for approximately 9 billion single-nucleotide variants—the three possible alternative bases at each of roughly 3 billion positions in the human genome—producing a dataset totaling 1 PB. Access options include a no-code website, the AlphaGenome API, and a skill available in the Google Antigravity scientific workbench.
The Atlas stores AlphaGenome’s predictions for molecular phenotypes such as gene expression, RNA splicing, and chromatin accessibility, while also introducing the AlphaGenome Variant Impact (AVI) score. AVI combines AlphaGenome’s predictions of regulatory effects with AlphaMissense’s predictions of protein-level effects, allowing coding and noncoding variants to be ranked on the same scale. The system also provides feature attribution, indicating whether a high score may be driven by signals such as splicing, expression level, or evolutionary conservation. The main technical change is not a new backbone model, but the conversion of expensive inference into low-latency lookups, making genome-wide cohorts and large population analyses more practical.
Early case studies show that a Broad Institute team used AVI to identify a DNM1 variant associated with aberrant splicing, with follow-up experiments confirming the effect. Another analysis of more than 54,000 UK Biobank participants reported a 22% increase in noncoding association signals. However, these cases do not directly demonstrate generalizability across diseases or populations. External researchers quoted by Nature cautioned that the Atlas should be used to prioritize research hypotheses, not as a substitute for individual context, clinical evidence, or wet-lab experiments.
Engineering and research teams should next examine API quotas, batch export options, AVI calibration across ancestries, and whether a single score obscures conflicting molecular signals. The core atlas of 9 billion predictions primarily covers single-nucleotide variants; compound variants, long-range regulation, and large structural variants still cannot be fully addressed through this lookup layer.