Diffusion in a quantized vector space generates non-idealized protein structures and predicts conformational distributions
The power of diffusion probabilistic models (DDPMs) in protein design was recently demonstrated by methods that performs three-dimensional protein backbone denoising. However, these DDPMs tend to generate protein backbones of idealized secondary structures and short loops, lacking diverse, non-idealized local structural elements which are essential for the rich conformational dynamics of natural proteins. Moreover, the sampling power of DDPMs have not yet been utilized for predicting the conformational distributions of natural proteins of dynamic structures. Aiming at these two needs, we developed a model named PVQD (protein vector quantization and diffusion), which used an auto-encoder with vector quantization and a generative diffusion model in the latent space to jointly performing the challenging task of modeling complicated protein structures within an end-to-end framework. Our study demonstrated that in design PVQD generated designable protein structures containing non-idealized elements, while in single sequence-based structure prediction PVQD reproduced experimentally observed conformational variations for a set of natural proteins of dynamic structures.