IndustryAugust 2023

ESM-2 Published in Science: Protein Language Models Reach Evolutionary Scale

Meta AI's team publishes ESM-2 in Science with 15 billion parameters, trained on 250 million protein sequences. The model predicts structure and function with unprecedented accuracy, opening the door to AI-directed protein engineering without experimental data.

JFInnova Perspective

ESM-2 is the central engine of JFInnova's directed evolution pipeline. We use it to generate and score CDR3 variants for sequence naturalness (pseudo-perplexity), achieving candidates with 18% better PPL than the wild-type.

References

ESM-2 protein language model Science Meta AI