Meta AI Computer Model Rapidly Predicts Hundreds of Millions of Metagenomic Proteins

Using a large language model for predicting protein structures, researchers from Meta AI have built a database containing the predicted atomic-level structures for hundreds of millions of metagenomic proteins. According to the researchers, the model — called ESMFold — can make protein predictions up to 60 times faster than other predictors including DeepMind's AlphaFold2, but with nearly the same accuracy, enabling large-scale structural characterization of vast collections of metagenomic proteins. As reported in Science this week, the team has used ESMFold to predict structures for more than 617 million metagenomic protein sequences, including more than 225 million that are predicted with high confidence, to create the open-access ESM Metagenomic Atlas. "Structure prediction at the scale of evolution can open a deep view into the natural diversity of proteins and accelerate the discovery of protein structures and functions," the authors write.