ESM-2

Meta AI's transformer-based protein language model for high-resolution structure prediction and sequence analysis.

Meta AI (FAIR) engineered ESM-2 to interpret biological data through the lens of large language models. Trained on 138 million sequences from the UniRef database, the architecture scales up to 15 billion parameters. It serves as the backbone for ESMFold: a folding engine that generates atomic-level protein structures up to 60x faster than AlphaFold2. This speed allows researchers to map the metagenomic world (billions of proteins) with precision (predicting functional sites and mutation effects) using standard GPU hardware.

https://github.com/facebookresearch/esm