Abstract: Breakthroughs in de novo protein design methods are of great significance for advancing applications in biology and chemistry. In recent years, flow matching-based models have emerged as important tools for generating protein backbone structures. Building upon the framework of AlphaFold2, this study develops the Prodes model. The model employs an invariant point update mechanism to iteratively optimize coordinates, effectively improving three-dimensional protein structure prediction by progressively updating the rotation and translation parameters of each residue. Subsequently, a Transformer architecture is further introduced to enhance the modeling of long-range dependencies, and the three-dimensional structure of protein backbone is reconstructed through an improved backbone update module. Furthermore, this paper redesigns the temporal sampling function to better balance diversity and designability in backbone generation, enhancing the model's adaptability and robustness across various tasks. To comprehensively evaluate performance, we employ multi-dimensional metrics such as designability, diversity, and novelty for quantitative analysis. Additionally, we conduct in-depth exploration on the challenging task of generating protein backbones with lengths of up to 800 residues, a problem rarely addressed in previous studies. This work provides new perspectives for the field of protein design and expands its research scope and application potential. Experimental results show Prodes excels in de novo protein backbone design. It outperforms baseline models in generating standard-length backbones and demonstrates superior performance in long-sequence tasks, providing strong support for advanced de novo protein design methods.