DeepMind publishes AlphaGenome Atlas of nine billion DNA letter swaps

Google DeepMind released AlphaGenome Atlas on Tuesday, 8 September. The company described it as a one-petabyte dataset of predicted molecular effects for all roughly nine billion possible single-letter DNA changes in the human genome. Scientific American said the atlas could change how genetic disease is read, because it scores the consequence of any one-base substitution rather than only the variants already seen in patients.

A single-letter change is a substitution of A, C, G or T at one position. The human genome contains about three billion bases. Counting every possible substitution at every position produces a space on the order of nine billion. Most of those changes have never been measured in a wet lab. The atlas assigns each one a predicted effect on molecular processes such as transcription and splicing. Researchers can query a rare variant without waiting for a matching case report.

That is the information gain against existing catalogues. ClinVar and gnomAD list variants that have been observed and, in some cases, labelled pathogenic or benign. They are silent on the far larger set of possible changes. AlphaGenome Atlas fills that grid with model output. The output is a prediction, not a clinical diagnosis. A lab that treats the scores as ground truth will misclassify some alleles. A lab that uses the scores to rank which alleles to test will waste fewer assays.

DeepMind has published biological models before, from protein structures to gene regulation. This release is a lookup table at genome scale. One petabyte is large enough that most hospital IT shops will not download it whole. Access will run through cloud queries and slices. The practical question is whether clinical genetics groups can call a score for a patient variant in seconds, and whether those scores move a decision about a child with an unexplained syndrome.

No peer-reviewed accuracy table was attached to the public summaries on Tuesday. That paper, when it appears, will decide how far the atlas travels beyond a research curiosity. Until then the fact on the table is simpler. Every possible one-letter human DNA change now has a predicted molecular label in a single public set. That label is only as good as the model. It is still a new object in the field.