biostudies-literatureZhu SNational Institute of Allergy and Infectious DiseasesNational Institute of General Medical SciencesNIGMS NIH HHSNational Science Foundatione70340https://www.ebi.ac.uk/biostudies/studies/S-EPMC12617271biostudies-literatureUnknown34(12)Computational design of functional proteins is of both fundamental and applied interest. This study introduces a generative framework for co-designing protein sequence and structure in a unified process by modeling their joint distribution, with the goal of enabling cross-modality interactions toward coherent and functional designs. Each residue is represented by three distinct modalities (type, position, and orientation) and modeled using dedicated diffusion processes: multinomial for types, Cartesian for positions, and special orthogonal group SO(3) for orientations. To couple these modalities, we propose a unified architecture, ReverseNet, which employs a shared graph attention encoder to integrate multimodal information and separate projectors to predict each modality. We benchmark ourProtein science : a publication of the Protein SocietyMultimodal diffusion for joint design of protein sequence and structure.PMC12617271R35GM1249521943008R01AI16543322037152514387Liu YSun QKotnis SShen YGulati SZhu SfalseMultimodal diffusion for joint design of protein sequence and structure.Computational design of functional proteins is of both fundamental and applied interest. This study introduces a generative framework for co-designing protein sequence and structure in a unified process by modeling their joint distribution, with the goal of enabling cross-modality interactions toward coherent and functional designs. Each residue is represented by three distinct modalities (type, position, and orientation) and modeled using dedicated diffusion processes: multinomial for types, Cartesian for positions, and special orthogonal group SO(3) for orientations. To couple these modalities, we propose a unified architecture, ReverseNet, which employs a shared graph attention encoder to integrate multimodal information and separate projectors to predict each modality. We benchmark our2025-01-01T00:00:00Z2025 Dec2026-06-05T14:36:38.84Z2026-05-18T03:07:58.789ZS-EPMC126172714123638610.1002/pro.70340