HFGuidedDesign framework achieves 75% success in de novo cyclic peptide design for protein targets.

Cyclic peptides are promising scaffolds for targeting protein surfaces and protein-protein interactions (PPIs) due to their unique structural and functional advantages, offering enhanced stability and target affinity compared to linear counterparts. However, their therapeutic potential is often hindered by the limited availability of experimentally determined cyclic peptide-protein complex structures. This scarcity severely restricts the rational design of novel, target-specific cyclic peptides, creating a significant bottleneck in drug discovery. Current design methods struggle to accurately predict binding poses and optimize sequences for desired properties, necessitating advanced computational approaches to bridge this structural knowledge gap and accelerate development.

Researchers developed HFGuidedDesign, a de novo cyclic peptide design framework integrating a discrete diffusion model with external structure guidance. The framework incorporates the high-accuracy complex structure predictor HighFold to perform real-time structural evaluation during reverse diffusion sampling. This dynamic steering mechanism guides sequence generation toward cyclic peptides with favorable structural plausibility and binding potential. The discrete diffusion model underwent a two-stage training strategy: initial pre-training on peptide monomers, followed by fine-tuning using known peptide-protein complex structures. The framework was then applied to design tasks targeting two distinct proteins, evaluating both head-to-tail and disulfide bond cyclization strategies.

The HFGuidedDesign framework demonstrated significant effectiveness and generalizability in designing cyclic peptide binders for two distinct protein targets. For the first target, the framework achieved a sequence design success rate of 75%, indicating its strong capability to generate viable peptide sequences. For the second target, the success rate was 66.7%, further demonstrating robust performance across different protein interfaces and cyclization strategies. The integration of HighFold for real-time structural evaluation during the reverse diffusion sampling process proved critical in dynamically steering sequence generation towards peptides exhibiting high structural plausibility and strong predicted binding potential. This iterative, structure-guided approach ensured that the generated sequences were not only novel but also conformationally viable and likely to interact effectively with their intended targets. The two-stage training, initially on peptide monomers and subsequently fine-tuned on peptide-protein complex structures, optimized the model's ability to learn complex sequence-structure relationships inherent to cyclic peptides.