Zefeng's work on the geometric and probabilistic modeling of protein backbones has been published in J. Chem. Inf. Model (link). This study proposed using a concise protein backbone representation and demonstrated peptide units' 3D rotations that unify the angular parts of internal coordinates on a single rotation manifold, composing the primary degrees of freedom of the protein backbone. This representation enables efficient differentiable coordinate conversions, effective conformation optimization ability free from the lever-arm effect, and straightforward geometric reasoning capability within deep neural networks. By integrating it into protein backbone generative models, we achieved enhanced performance in protein design regarding in silico stereochemical quality, diversity, novelty, and length generalizability. We anticipate its future application in advancing biomolecular structure prediction and design. Congratulations!
