Open Source Model Predicts Shape of 1B Proteins (www.nature.com)

Researchers at the Chan Zuckerberg Initiative’s Biohub have launched a groundbreaking open-source AI tool called ESMFold2, which has generated an extensive atlas of over 1.1 billion predicted protein structures and 6.8 billion protein sequences. This new database, named the ESM Atlas, significantly expands our understanding of the protein universe, particularly in areas that remain largely unexplored. ESMFold2 claims to outperform existing models like AlphaFold3 in predicting protein structures, especially in complex interactions such as antibody-antigen binding. The team is optimistic that this resource will catalyze discoveries in protein biology and pave the way for new medical breakthroughs. The significance of the ESM Atlas lies in its potential to bridge gaps in protein knowledge, using a ‘protein language’ model trained on vast metagenomic data from diverse environments. Researchers have successfully employed ESMFold2 to design new antibodies targeting proteins linked to cancer and immune disorders, demonstrating the practical applications of this AI tool. By making this atlas freely accessible, the Biohub aims to foster collaboration across the scientific community, encouraging insights that could enhance our understanding of protein functions and lead to novel therapeutic strategies.