AlphaGenome Atlas organizes predictions for nine billion possible DNA changes

Google DeepMind introduced AlphaGenome Atlas on September 8, a resource containing predictions for the effects of nine billion single-nucleotide variants in human DNA. The central distinction in this September 13 report is that the number describes possible single-letter changes, rather than people or clinical experiments.

Why the number is so large

Google's overview describes a human genome of approximately three billion base pairs. At each position, replacing the existing letter with another permits several alternatives. The scale comes from the space of possible changes, rather than performing that many laboratory experiments.

Atlas is a database of precomputed predictions. Google describes a dataset of approximately one petabyte and a portal through which researchers can query it. The idea is to retrieve an available prediction instead of recalculating every question from the beginning.

The role of AVI

DeepMind introduces the AlphaGenome Variant Impact, or AVI, score to rank variants and explain the molecular processes predicted to be affected. In the team's description, this can help researchers decide where to focus further attention.

Prediction, ranking and experimental verification remain different stages. Reading a scientific result requires knowing which stage is being described. A value appearing in a database does not automatically become a direct experimental result for that variant.

A simple research question illustrates the distinction: are we selecting candidates for investigation, or presenting the result of verification already performed? Those questions require different evidence. That is an explanation of how to read the resource, not an analysis we ran on its data or an independent assessment of model accuracy.

The use boundary

The AlphaGenome team states that the tool has not been validated or approved for clinical use. This report therefore concerns research infrastructure, rather than a way to diagnose an individual's condition or select treatment.

In our assessment, the technical value lies in organizing a very large prediction space so it can be explored. Evaluating an individual result still requires keeping the variant type, prediction measure and verification evidence visible.

Access to information and the strength of evidence it represents also need to be distinguished. An interface can make a result faster to retrieve without settling the scientific question of what it measures and how it was checked.

We have not analyzed Atlas data for this report. We present the announced resource and explain the distinction between prediction and evidence, so the database's breadth is not mistaken for complete understanding of every genetic change or medical validation of every result it contains.

Sources

• Google DeepMind — AlphaGenome Atlas, September 8
• Google — Atlas overview, September 8