Atom-level Protein Representation Learning Improves Protein Structure Prediction
Researchers have developed TriProRep, a novel pretraining method for learning protein representations that incorporates atom-level and geometric data. This approach models three distinct views of protein residues: amino-acid identity, backbone geometry, and local full-atom geometry. The method aims to improve protein structure prediction by distinguishing plausible but incorrect cross-view augmentations from original protein data. AI
IMPACT Introduces a new method for learning protein representations that could advance biomolecular structure prediction and related fields.