CZ Biohub releases “ESM Atlas” with 1.1 billion ESMFold2-predicted protein structures, claiming to surpass AlphaFold

Summary

A claim has emerged that a new protein structure prediction artificial intelligence (AI) model surpasses the performance of Google DeepMind’s AlphaFold, which was developed by DeepMind and even led to the 2024 Nobel Prize in Chemistry. Experts are divided between seeing it as an important new resource for protein structure prediction and expressing doubts about its performance. According to a report in the international journal Nature on the 27th (local time), the U.S. biomedical research institute Chan Zuckerberg Biohub (CZ Biohub) released a database called “ESM Atlas,” which compiles the results of its protein structure prediction AI model “ESMFold2.” CZ Biohub is a research institute founded by physician Priscilla Chan, the wife of Meta (formerly Facebook) founder Mark Zuckerberg. According to CZ Biohub, ESMFold2 outperforms other models, including AlphaFold3, the latest protein structure prediction model from Google DeepMind. The report states that the number of protein structure data predicted by ESMFold2 exceeds that of the “AlphaFold database,” which contains AlphaFold’s predictions, by 800 million. Alex Reeves, chief scientific officer at CZ Biohub and lead author of the ESM Atlas study, said, “ESM Atlas reveals even the most unknown regions of the protein universe,” adding that “it will provide a powerful foundation for new biological discoveries.” ESMFold2 is based on a “protein language model” that the team released in 2024. A protein language model learns amino acid sequences of proteins as if they were a language. The researchers trained the model on billions of protein data obtained from living organisms. This included “metagenomes,” which are the total genomes of microorganisms living in various environments such as soil and oceans. Metagenomic data are not included in the AlphaFold database. According to the team, ESMFold2 predicts protein complex structures, such as how antibody molecules bind to their target antigens, more accurately than existing models. ESMFold2 also designed new antibodies and proteins that bind strongly to proteins involved in cancer and immune diseases. The team produced these antibodies and proteins in the laboratory and confirmed that they functioned as predicted by the model. Using ESMFold2, the researchers built an atlas containing 1.1 billion predicted protein structures and 6.8 billion protein sequences. Reeves…