Open-Source AI Model Predicts Structures of 1 Billion Proteins

Launch of ESMFold2

Researchers have released ESMFold2, an open-source artificial intelligence model capable of predicting the 3D shapes of proteins. The tool has been used to generate a comprehensive atlas containing more than one billion predicted protein structures.

According to Biohub, ESMFold2 is a "world model of protein biology" that translates evolutionary patterns encoded in ESMC into precise, atomic-resolution 3D models. These models cover not only individual proteins but also their interactions.

Performance and Benchmarks

Biohub states that ESMFold2 leads across standard protein folding benchmarks, specifically in the prediction of antibody-antigen and protein-protein interactions. The organization further claims that the performance of ESMFold2 surpasses that of other protein-structure prediction AIs, including AlphaFold3, the most recent version of the system developed by Google DeepMind.

The predictions generated by the model were detailed in a preprint released alongside the tool. This release expands the known protein universe by providing a massive scale of predicted structures that were previously unavailable.

Context of Protein Prediction

Protein structure prediction has long been a central challenge in biology, as the shape of a protein determines its function. While Google DeepMind's AlphaFold has been a dominant tool in the field, recent updates to that system have focused on expanding its capabilities to predict very large and complex protein structures and integrating experimental data into its workflow.