De novo protein design using geometric vector field networks

In this work, we introduce the Vector Field Network (VFN), a novel protein design encoder that leverages learnable vector computations to model protein residue backbones effectively, where traditional atom-wise features are not applicable. Unlike existing methods, VFN utilizes frame-anchored virtual atoms for dynamic residue representation, enabling superior frame modeling without relying on direct atomic details. Our results demonstrate VFN’s remarkable performance in protein design and inverse folding tasks, outperforming existing models such as IPA and PiFold in designability, diversity, and sequence recovery rate. Additionally, integrating VFN with the ESM model significantly enhances its capability, setting a new state-of-the-art benchmark. This advancement paves the way for more sophisticated de novo protein design, offering promising applications in life sciences.