AI-assistant de novo protein design aims to efficiently explore the vast protein structure space and generate rationally designable protein. Despite the impressive performance of generative models like denoising diffusion models, customizing them specifically to de novo protein design remains challenging, including (a) ensuring simple and effective modeling of structures; (b) enabling reasonable expression of protein sequences; and (c) capturing complex relationships among sequences. To address these challenges, we propose a novel de novo Protein backbone design model with SE(3)-Equivariant Diffusion, dubbed ProSEED, to enable efficient and diverse protein structure design. Specifically, ProSEED contains three submodules: (1) a residue information interaction network; (2) SE(3)-equivariant neural network; and (3) an improved denoising diffusion probabilistic model, to precisely learn and capture the distribution of various protein backbone structures. By sampling from Gaussian noise, ProSEED could generate new protein backbones through reverse diffusion on residue backbones. Extensive experiments show that ProSEED achieves high protein structure designability, diversity, and novelty while requiring less data. These results highlight ProSEED’s exceptional ability to comprehend protein structures and enable efficient information transfer and interaction between its modules, paving the way for successful and innovative protein design with a high success rate.