Google DeepMind launched AlphaGenome Atlas on September 8, 2026, making precomputed molecular-effect predictions for roughly 9 billion possible single-nucleotide variants searchable. Researchers can look up candidate variants instead of computing every prediction from scratch. It is an AI-generated map for narrowing scientific searches, not a tool that diagnoses a disease from DNA.
What do the 9 billion predictions cover?
Think of DNA as a long text written with four letters. Replacing one letter can affect a protein or how a gene is regulated in different cells. Atlas focuses on those single-letter substitutions. It should not be described as a complete simulation of all genetic changes, large deletions or every difference between people.
The new release is not AlphaGenome’s first appearance. It turns large-scale predictions into a roughly 1 PB dataset and an accessible platform. About 2% of the human genome directly encodes proteins; interpreting the rest, including regulatory regions, is harder. Atlas offers an entry point for exploring both coding and non-coding variation.
Helping researchers prioritize
Google also introduced the AlphaGenome Variant Impact, or AVI, score to rank candidates. It combines AlphaGenome and AlphaMissense predictions, while feature attributions connect a score to predicted effects such as gene expression or RNA splicing. The aim is to pair prioritization with a biological explanation that researchers can investigate.
One collaboration described by Google used AVI to identify a previously overlooked DNM1 variant. The predicted abnormal splice site was then investigated experimentally. The important sequence is a more specific hypothesis followed by testing, not accepting an AI response as the conclusion. A successful case does not establish that every disease can be solved the same way.
Available now, with important boundaries
The non-commercial research website is available, and Google lists API and Antigravity Skill access. A visual portal reduces the coding barrier, not the expertise needed to interpret results. Google says commercial Atlas access on Cloud is coming later. The base AlphaGenome model’s commercial availability should not be confused with an already launched commercial Atlas.
Google explicitly states that AlphaGenome has not been validated or approved for clinical use. A high AVI score is not a probability that someone has a disease and cannot by itself determine diagnosis or treatment. Research prediction, experimental evidence and clinical decisions are different stages.
What this changes for AI research tools
For AI readers, Atlas illustrates a concrete deployment pattern: perform expensive computation in advance, then expose search, ranking and explanations to specialists. Its value depends on whether it helps identify useful candidates and whether later experiments support those leads. The advance is a more inspectable starting point for research, not a claim that AI understands everyone’s health.
