DeepMind unveils the AlphaGenome Atlas, a pre-computed database that maps noncoding DNA and regulatory regions to functional outcomes, extending beyond protein-focused work like AlphaFold by covering the full genome.

The Atlas builds on AlphaGenome and expands beyond AlphaMissense by analyzing noncoding DNA to predict its impact on molecular function, including regulatory roles.

Researchers are introduced to the Variant Impact Score (AVI), a single metric designed to rapidly rank variants by predicted molecular impact and prioritize them for study.

AVI consolidates AlphaGenome and AlphaMissense predictions with conservation and loss-of-function features, and provides attributions for processes like RNA splicing and gene expression.

The AlphaGenome Variant Impact (AVI) score is intended to be user-friendly, indicating whether a variant is likely disease-causing and aiding prioritization in disease research.

Access to the Atlas will be available to non-commercial researchers via a web portal, with commercial use via Google Cloud planned in the near future.

As of September 8, 2026, the Atlas and AVI are available for non-commercial research, with a permissively licensed static AVI download for both commercial and non-commercial use; commercial access on Google Cloud is forthcoming.

Experts emphasize the Atlas’ usefulness for multi-variant questions and for guiding initial investigations when specific cell types or processes are not yet clear.

The authors caution that Atlas and AVI are research tools for predicting molecular effects and should not be used as standalone clinical evidence, noting data gaps and the need for future model improvements.

Early results show the Atlas can identify disease-causing variants missed by traditional methods and can rank known causal variants more effectively, aiding rare-disease analyses and linking noncoding variants to protein levels in large datasets.

The project processed about a petabyte of data, covering all three possible nucleotide changes at every genomic position and incorporating over 100 million observed insertions and deletions.

The Atlas comprises roughly a petabyte of data, about 30 times larger than AlphaFold’s database, and is intended as a research starting point rather than a final answer due to varying accuracy.