TorsionBackDiff: a torsional diffusion model for generalized, transferable protein backmapping

Coarse-grained (CG) models have been an important tool in the study of protein structures, their thermodynamic characteristics, and conformational dynamics. However, the simplification inherent in CG models necessitates a subsequent conversion back to detailed all-atom structures for comprehensive analysis in fields like protein design and drug discovery. Despite advancements in data-driven methodologies for this backmapping process, a universal solution applicable across various proteins and CG model types remains elusive. Our study introduces TorsionBackDiff, an innovative generative framework engineered to address this challenge with a focus on generalizability and reliability in protein backmapping tasks. TorsionBackDiff utilizes the conditional score-based diffusion model with geometric representations. Developed upon its previous model that learns the atomic coordinates distribution, TorisionBackDiff considers torsion angles as its target of study, which greatly reduces the dimensionality. The new TorsionBackDiff yields improveed backmapping transferability and accuracy, allows more efficient sampling, while preserving versatility in handling various CG models. A pretrained TorsionBackDiff model can offer a convenient yet reliable plug-and-play solution for protein researchers.