De novo protein prediction is a grand challenge in understanding the protein structure and generation. In this direction, we introduce the Vector Field Network (VFN), a novel approach to de novo protein design. VFN stands out by leveraging learnable vector computations, enabling enhanced modeling of protein structures. This research focuses on VFN's application in protein diffusion models, where it demonstrates significant advantages in designability and diversity over existing methods like IPA. Additionally, we'll touch on its success in inverse folding, showcasing VFN's superior sequence recovery rates. VFN's integration with the latest ESM model marks a significant advancement in protein design, offering promising future applications in life sciences.
W. Mao, M. Zhu, Z. Sun, Lin Yuanbo Wu, H. Chen, C. Shen. De novo Protein Design Using Geometric Vector Field Networks. International Conference on Learning Representations (ICLR), Spotlight Paper (5% acceptance), Austria 2024.