Large language model for atomic protein structure prediction developed by Meta, operating orders of magnitude faster.
Model weights and ESM Metagenomic Atlas predicting over 600 million proteins freely available
ESMFold is an atomic-level protein folding engine developed by Meta AI FAIR. Departing from legacy techniques requiring multiple sequence alignments, ESMFold leverages the evolutionary scale language model ESM-2 to predict atomic 3D structures directly from solitary primary amino acid sequences. Operating up to 60 times faster than conventional pipelines, it enabled the creation of the ESM Metagenomic Atlas mapping over 600 million candidate microbial protein structures.