Relation-Aware Graph–Hypergraph Networks for Antibody Sequence–Structure Co-Design

The design of antibody complementarity-determining regions (CDRs) is pivotal for therapeutic development, yet remains challenged by the complexity of the antibody–antigen interface. Existing methods, which largely rely on pairwise relationships, struggle to capture intrinsic multibody interactions and global structural constraints. To address this limitation, we propose the Relation-Aware Graph–Hypergraph Network (RA-GHN), a novel framework that systematically unifies graph-based pairwise modeling with biophysical hypergraphs to capture high-order co-operative effects. Specifically, we construct a heterogeneous hypergraph representation featuring three complementary hyperedge types: epitope–paratope interaction patches, spatial structural clusters, and high-degree hub structures, enabling a comprehensive encoding of interface physics. Architecturally, RA-GHN adopts a dual-pathway design that synergistically integrates a Relation-Aware Equivariant Graph Network (RA-EGN) with an Equivariant Hypergraph Encoder , strictly preserving SE (3)-equivariance throughout the sequence–structure co-design process. Furthermore, to mitigate folding hallucination and sequence overfitting commonly observed in generative models, we introduce the Structure–Sequence Decoupling Index (SSDI), which employs a harmonic mean to non-linearly balance structural physical reliability with sequence novelty. Extensive experiments demonstrate that RA-GHN significantly outperforms state-of-the-art baselines across standard benchmarks and affinity optimization tasks, validating that the fusion of high-order geometric modeling with rigorous evaluation criteria enables the generation of physically feasible and functionally innovative antibodies.