Meet Meta AI's 'ESMFold,' An Artificial Intelligence-Based Model That Predicts Protein Structure 6x Faster Than AlphaFold2

Recent research has demonstrated that large language models can evolve with scale, moving beyond simple pattern matching to do higher-level reasoning and produce realistic visuals and text. There has been some research on language models trained on protein sequences, but when they are scaled up, little is known about what they learn about biology. Researchers at Meta AI have developed one of the most significant language models of protein to date, ESMFold, that can predict protein structure from a gene sequence. With an order-of-magnitude faster inference time, ESMFold, based on a 15B parameter Transformer model, delivers accuracy comparable to other state-of-the-art models. The paper describing the model and several tests carried out as part of this study have also been published on bioRxiv.