Google's AI can read 1 million DNA letters at once
What's the story
Google's DeepMind has unveiled a revolutionary deep learning model, AlphaGenome, which can analyze long sequences of DNA with remarkable accuracy. The model was introduced in a study published in the journal Nature. It is hoped that AlphaGenome will be an invaluable tool for understanding how minor changes in human DNA impact health and biology, especially in the largely unexplored non-coding regions of our genome.
Genetic exploration
AlphaGenome: A breakthrough in understanding genetic 'dark matter'
Pushmeet Kohli, VP of research at Google DeepMind, expressed excitement over the launch of AlphaGenome as a solution to decode complex regulatory codes. The human genome is composed of coding and non-coding regions. While only about 2% of our genes code for proteins, the remaining 98% are made up of non-coding regions or "genetic dark matter." These areas were once thought to be useless junk DNA but are now known to contain sequences crucial for regulating protein-making genes.
Model prowess
AlphaGenomes's training and capabilities
DeepMind trained AlphaGenome on human and mouse genomes. It can analyze up to one megabase (about one million DNA letters) at a time, which is a significant increase in capacity compared to older models. From this sequence, the model predicts thousands of functional genomic tracks, including gene expression and interactions between coding/non-coding regions of DNA. In tests measuring its ability to predict genetic variant effects, AlphaGenome matched or outperformed other existing AI models in 25 out of 26 cases.
Future prospects
AlphaGenome's potential applications and limitations
AlphaGenome can simultaneously predict nearly 6,000 human genetic signals linked to specific functions. However, it is not a one-stop solution for all genetic code mysteries. Ben Lehner from the Wellcome Sanger Institute praised AlphaGenome's performance but also noted that there is still much work to be done. He cautioned that AI models are only as good as the data used to train them and most existing biological data isn't suitable for AI due to small size and lack of standardization.