On September 8, 2026, Google DeepMind released AlphaGenome Atlas β€” a database of precomputed predictions about the molecular effects of all 9 billion possible single-nucleotide substitutions in the human genome. Along with the atlas, AVI β€” a single numerical score of variant impact β€” was presented, and the resource is open for academic research for free.

image
image

What happened

The atlas occupies about 1 petabyte, which is more than 30 times larger than the AlphaFold Database. The predictions cover 2% coding and 98% non-coding regions of the genome, including regulatory elements that affect gene activity and RNA splicing. Each variant is associated with thousands of molecular predictions: gene activity, splicing variants, and more than 2,500 repetitive DNA motifs in hundreds of human and mouse cell types. For academic research, the resource is available for free via a no-code web portal, API, and a skill for Google Antigravity. A practical example has already been obtained: a Broad Institute team used AVI to find a variant in the DNM1 gene that creates an incorrect splicing site and solve an unsolved rare disease case.

Context

AlphaGenome is a Google DeepMind model that predicts the molecular consequences of DNA changes. The release continues the line of AlphaFold, AlphaMissense, and AlphaGenome, and AVI itself combines AlphaGenome predictions with AlphaMissense scores into a single score. The AlphaFold Database has already shown that a precomputed database can become a standard reference layer: for protein structures, querying it has become the norm. For a long time, running models individually for each change was expensive, so in-depth analysis remained focused mainly on the coding part of the genome. A precomputed reference for all possible substitutions removes this limitation.

Why this matters for the industry

The atlas turns individual queries to the AlphaGenome model into a ready-made genomic reference: researchers rank variants by AVI and get thousands of molecular predictions without running the model for each variant. This sharply lowers the computational barrier and accelerates variant prioritization in rare disease research pipelines: on UK Biobank data with more than 54,000 participants, a 22% increase in associations in non-coding regions is reported. The β€œrun a mutation through a model” layer becomes widely available, and the starting point of capabilities shifts down the chain β€” to analysis and interpretation tools. If experimental validation confirms the quality of AVI, precomputed genomic references may become the de facto first filter in variomics β€” following the example of the AlphaFold Database for structures.

Why this matters for users

For the reader, this is a clear example of how precomputation at a scale familiar from language models is applied beyond text: 1 petabyte of predictions for the entire genome. Any researcher can freely open the alphagenome.google/atlas portal and in seconds get an AVI score and molecular predictions for a single mutation that previously had to be run through the model individually. For those following biotech AI, the release is a continuation of the AlphaFold, AlphaMissense, and AlphaGenome line and a benchmark for where applied AI modeling of biology is heading.

What is still unknown / limitations

The claimed best-in-class results of AVI on pathogenicity and rare disease benchmarks and the 22% increase in associations in non-coding regions on UK Biobank are the developers' own claims: independent replication, evaluation protocol, and metrics are absent in the available materials. DeepMind explicitly states that predictions require experimental validation and do not replace clinical interpretation: the atlas is a prioritization tool, not a diagnostic. Finally, the release does not announce a new architecture or new predictive capabilities: the model remains the existing AlphaGenome, and the atlas is its precomputed results.

Sources

Author

Look at AI, editorial team