DeepMind has created an AI map of the human genome: 9 billion DNA variations can be looked up directly.

动察 Beating AI Flash News: Google DeepMind has released AlphaGenome Atlas, pre-computing all approximately 9 billion possible single-letter DNA changes in the human genome into a directly searchable database. The entire dataset is about 1PB, over 30 times larger than the AlphaFold Database.

The human genome contains roughly 3 billion DNA letters, and each position can be swapped for one of the other 3 options, resulting in about 9 billion single-letter changes in total. Previously, researchers had to use AlphaGenome to analyze variants individually; now they can directly look up results in the Atlas without rerunning the model each time.

The Atlas also provides an AVI score for each variant, combining predictions from AlphaGenome and AlphaMissense, helping researchers first identify variants most likely to have an impact. It covers both the approximately 2% protein-coding regions and the harder-to-study remaining 98% non-coding regions.

Research teams have already used it to uncover rare disease-related variants previously missed, and confirmed through experiments that one such variant indeed causes abnormal RNA splicing.