APCyc: Property-Informed Design of Cyclic Peptides via Automated Cyclization

Technical Summary: APCyc – Property-Informed Design of Cyclic Peptides via Automated Cyclization

Problem Statement
Cyclic peptides offer significant therapeutic advantages over linear peptides, including improved stability, binding affinity, and pre-organized bioactive conformations. However, the de novo design of cyclic peptides remains a formidable challenge. Existing generative models are predominantly trained on linear peptide data and often treat cyclization as a post-hoc structural modification rather than an integrated design variable. This approach fails to capture the complex, target-specific topological requirements necessary for high-affinity binding. Furthermore, current methods typically rely on rigid, pre-defined cyclization heuristics (e.g., fixed linkage sites or types) and largely overlook the simultaneous optimization of critical drug-relevant properties such as membrane permeability, protease resistance, solubility, and immunogenicity. Consequently, cyclic peptide design is inherently a multi-objective optimization problem that requires balancing therapeutic constraints beyond simple binding activity.

Methodology
The authors propose APCyc, a target-aware, latent diffusion framework designed to automate the selection of cyclization topologies while jointly optimizing multiple physicochemical properties. The method operates through three core mechanisms:

1. Automated Cyclization Pair Injection:

   • Vocabulary Expansion: APCyc extends the standard amino acid residue vocabulary to explicitly distinguish between cyclized and non-cyclized residues (e.g., distinguishing a disulfide-linked cysteine from a standard one), even if they share the same amino acid identity.
   • Topology Modeling: The model formulates cyclization as a discrete, target-conditioned decision-making problem. It predicts both the cyclization linkage types (e.g., amide, disulfide, isopeptide) and the specific residue-level linkage sites directly from the receptor binding pocket context.
   • Injection Mechanism: A pairwise representation tensor is constructed to capture semantic compatibility and geometric proximity. This information is projected into task-specific subspaces to predict a linkage probability matrix and cyclization type distribution. These predictions are injected into the denoising network (an AM-EGNN) via edge-feature augmentation and message modulation, guiding the diffusion trajectory toward valid topological closures.

2. Geometry Latent Diffusion:

   • APCyc utilizes a dyMEAN-style autoencoder to map peptide graphs into residue-level latent variables (invariant and equivariant).
   • A DDPM-style latent diffusion process is trained on these latents. The denoising network is conditioned on the receptor pocket context and the online topology signals generated by the cyclization heads.
   • The training objective combines the standard DDPM noise-prediction loss with a supervised loss for the predicted cyclization topology (linkage matrix and type).

3. Property-Informed Bayesian Posterior Guidance:

   • To optimize drug-relevant properties, APCyc employs a gradient-based guidance strategy in the latent space.
   • Differentiable energy-based surrogates are trained to predict properties (solubility, permeability, protease resistance, immunogenicity, binding affinity) from the noisy latent states.
   • During the reverse diffusion process, the sampling is steered using Bayesian posterior guidance. The score function is augmented with gradients from the property surrogates, allowing for the joint optimization of multiple objectives. An adaptive weighting scheme balances the gradients of competing objectives to prevent one property from dominating the others.

Key Contributions

• Automated Pocket-Conditioned Cyclization: Unlike prior methods requiring predefined cyclization strategies, APCyc learns to select both cyclization types and linkage sites directly from the target binding-pocket context, treating topology as a design variable.
• Property-Informed Controllable Generation: The framework utilizes Bayesian posterior guidance to steer generation across multiple therapeutic properties (affinity, permeability, protease resistance, solubility, immunogenicity), achieving state-of-the-art scores in permeability proxies and protease resistance.
• End-to-End Design: APCyc unifies cyclization selection, all-atom generation, and property-guided optimization. It achieves leading structural quality (e.g., Rosetta total scores and consistency) while maintaining strong binding affinity.

Experimental Results
The model was evaluated on the CPCore dataset (a refined subset of CPSea containing ~72k cyclic peptide-protein complexes) and tested on a non-redundant set of 56 targets.

• Property Optimization: APCyc variants demonstrated explicit control over property targets. The "Membrane-permeable" setting achieved the highest permeability proxy (0.107) among all methods, while the "Multi-property" setting achieved the best protease resistance score (-1.474).
• Structural Quality: The generated peptides exhibited high stability and consistency. The multi-property variant achieved the best Rosetta total score (-758.545) and consistency (0.971).
• Target-Level Performance: Case studies on specific targets (3rc4 and 4xal) showed that APCyc consistently outperformed baseline methods (PepGLAD, PepFlow, CPComposer) in binding scores and multi-property profiles, significantly improving membrane permeability on individual targets.
• Ablation Studies: Removing the guidance mechanism resulted in lower performance on protease resistance and immunogenicity, confirming the necessity of the property-guided component.

Significance and Claims
The paper claims that APCyc addresses the critical limitation of existing generative models by integrating cyclization topology selection directly into the generative process rather than treating it as a post-processing step. By explicitly modeling the coupling between target geometry, cyclization strategy, and drug-relevant properties, APCyc enables the controllable design of cyclic peptides that satisfy complex therapeutic constraints. The authors position this work as a robust foundation for next-generation therapeutic peptide design, offering a framework that balances structural feasibility with multi-objective drug-likeness. The authors note that the method is intended to support therapeutic discovery by proposing candidates for expert review and experimental validation, rather than serving as a direct source of clinical recommendations.