Graph-Based Diffusion Enables Interface-Aware Sequence-Structure Co-design of Protein Complexes

The rational design of protein–protein complexes remains a fundamental challenge in synthetic biology and therapeutic development. Current generative methods often fall short in performing sequence–structure co-design, particularly in treating the functionally critical protein-protein interface as a first-class target. To bridge this gap, we present InterfaceDiff, a graph-based diffusion framework for interface-aware co-design of protein complexes. The complex is encoded by intra-chain graphs coupled through an explicit bipartite interface graph, concentrating modeling capacity on physically interacting residues. InterfaceDiff learns a joint distribution over discrete amino acid sequences and continuous local rigid frames (rotations and translations) by a simultaneous denoising process. To achieve this efficiently, we develop a novel graph neural network denoiser inspired by Invariant Point Attention, which performs message passing on the sparse graph representation while avoiding the computational overhead of fully SE(3)-equivariant networks. We evaluate InterfaceDiff across multiple design tasks, demonstrating its ability to generate diverse, high-quality, and physically plausible all-atom complexes. Our method achieves strong performance on key biophysical and geometric metrics, offering a scalable and geometrically efficient approach for controllable protein complex engineering. This work establishes a foundation for generative co-design of novel molecular interactions.