HighMPNN: A Graph Neural Network Approach for Structure-Constrained Cyclic Peptide Sequence Design

Cyclic peptides become attractive therapeutic candidates due to their diverse biological activities. However, existing deep learning-based sequence design models, such as ProteinMPNN, are primarily optimized using cross-entropy loss and often overlook the unique topological constraints of cyclic peptides. This limits their ability to generate structurally accurate and foldable sequences. To address this challenge, we propose HighMPNN, a graph neural network model that builds upon ProteinMPNN by incorporating a structure prediction module and integrating cross-entropy loss with Frame Aligned Point Error (FAPE) loss. This dual-loss strategy enables the simultaneous optimization of sequence generation and structural fidelity, making HighMPNN better suited for cyclic peptide design. HighMPNN demonstrates superior performance in both sequence recovery and structural consistency compared to baseline models, particularly for short cyclic peptides and specific secondary structures. In summary, HighMPNN enables the design of sequences that closely resemble the native structures, thereby accelerating the discovery of high-quality cyclic peptides and advancing peptide-based drug development.