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

# APCyc: Property-Informed Design of Cyclic Peptides via Automated Cyclization
Source: [https://arxiv.org/html/2606.12991](https://arxiv.org/html/2606.12991)
Yifan Zhao[yzhao642@connect\.hkust\-gz\.edu\.cn](https://arxiv.org/html/2606.12991v1/mailto:[email protected])AI ThrustThe Hong Kong University of Science and Technology \(Guangzhou\)GuangzhouChinaLang Qin[lqin969@connect\.hkust\-gz\.edu\.cn](https://arxiv.org/html/2606.12991v1/mailto:[email protected])AI ThrustThe Hong Kong University of Science and Technology \(Guangzhou\)GuangzhouChinaandJintai Chen[jintaichen@hkust\-gz\.edu\.cn](https://arxiv.org/html/2606.12991v1/mailto:[email protected])AI\-Peptide Drug Design Joint LaboratoryThe Hong Kong University of Science and Technology \(Guangzhou\)GuangzhouChina

\(2026\)

###### Abstract\.

Cyclic peptides represent a promising class of therapeutic compounds in modern drug discovery, often offering improved stability and binding affinity\. However, thede novodesign of cyclic peptides remains challenging because methods must identify pocket\-adaptive cyclization patterns and linkage sites while simultaneously controlling drug\-relevant properties\. This challenge is particularly pronounced for recent generative models trained predominantly on linear peptide data, which may fail to capture cyclization\-specific constraints\. To address the limitation, we introduceAPCyc, a target\-awarede novocyclic peptide generation framework that explicitly models cyclization and jointly optimizes multiple essential physicochemical properties\. By using an expandedresidue vocabularyand explicitly encodingcyclization\-site and linkage\-type information, APCyc learns cyclization\-aware representations and leveragesBayesian posterior guidanceto steer sampling toward cyclic peptides satisfying multiple property objectives\. Experimental results demonstrate that our model learns target\-dependent cyclization preferences, and enables effective and controllable multi\-property optimization for cyclic peptide design\. The source code of this paper is available at https://github\.com/HKUSTGZ\-ML4Health\-Lab/APCyc\.

Cyclic Peptide; Generative Model; Diffusion Model

††journalyear:2026††copyright:cc††conference:Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2; August 9–13, 2026; Jeju Island, Republic of Korea\.††booktitle:Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.2 \(KDD 2026\), August 9–13, 2026, Jeju Island, Republic of Korea††isbn:979\-8\-4007\-2259\-2/2026/08††doi:10\.1145/3770855\.3818908††ccs:Applied computing Bioinformatics††ccs:Applied computing Computational biology††ccs:Computing methodologies Artificial intelligence††ccs:Computing methodologies Machine learningAccepted at the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining \(KDD 2026\)\. DOI:[10\.1145/3770855\.3818908](https://doi.org/10.1145/3770855.3818908)\.

## 1\.Introduction

Cyclic peptides are short and synthetically accessible amino acid sequences characterized by macrocyclic structural constraints that pre\-organize the peptide backbone into bioactive conformations\(Fosgerau and Hoffmann,[2015](https://arxiv.org/html/2606.12991#bib.bib17); Ji et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib22)\), thereby minimizing the entropic cost of binding to target proteins and making them attractive for therapeutic applications\(Vlieghe et al\.,[2010](https://arxiv.org/html/2606.12991#bib.bib54); Lau and Dunn,[2018](https://arxiv.org/html/2606.12991#bib.bib30); Driggers et al\.,[2008](https://arxiv.org/html/2606.12991#bib.bib15); Zorzi et al\.,[2017](https://arxiv.org/html/2606.12991#bib.bib59); Liu et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib37)\)\. The advantages have motivated the development of computational methods for cyclic peptide discovery, with recent AI\-based approaches increasingly used to model sequence–structure–property relationships and prioritize candidates from large combinatorial sequence spaces\. However, designing therapeutically useful cyclic peptides remains challenging because candidate molecules must satisfy multiple coupled drug\-relevant constraints beyond target binding, including solubility, permeability, protease resistance, metabolic stability, and safety\-related properties\(Otvos Jr and Wade,[2014](https://arxiv.org/html/2606.12991#bib.bib40)\)\. Thus, cyclic peptide design is inherently a multi\-objective optimization problem that requires methods capable of balancing therapeutic constraints rather than optimizing activity alone\.

Recent target\-aware peptide design methods have primarily focused on generating linear peptides conditioned on target \(receptor\) sequence and pocket geometry\. Within this paradigm, geometry\-aware generative models, including latent diffusion approach\(Kong et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib28)\)or multimodal torsional flow matching\(Lin et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib36); Li et al\.,[2024b](https://arxiv.org/html/2606.12991#bib.bib33)\), have demonstrated the capability to perform full\-atomic linear peptide design at target interfaces\. Concurrently, approaches utilizing autoregressive generation\(Li et al\.,[2024a](https://arxiv.org/html/2606.12991#bib.bib32)\), reinforcement learning paradigms\(Oh et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib39)\), or pocket\-conditioned diffusion models\(Sayuti et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib47)\), have furthered the effective design of linear peptide binders across diverse targets\(Bhat et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib8)\)\. However, these frameworks generally treat cyclization as a post\-hoc structural modification rather than an integrated design variable\(Li et al\.,[2024c](https://arxiv.org/html/2606.12991#bib.bib34); Rettie et al\.,[2025a](https://arxiv.org/html/2606.12991#bib.bib44)\), failing to capture the complex, target\-specific topological requirements inherent in high\-affinity binding\(Rettie et al\.,[2025a](https://arxiv.org/html/2606.12991#bib.bib44),[b](https://arxiv.org/html/2606.12991#bib.bib45); Wang et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib55)\)\.

Beyond this lack of structural integration\(Jiang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib23)\), existing frameworks rely on rigid, a priori cyclization heuristics\(Zhou et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib58)\)\. In these methods, the cyclization type and linkage site positions are typically predefined based on heuristics, prior assumptions, or trial\-and\-error, rather than inferred directly from the structural requirements of the target binding pocket\. This decoupling is fundamentally flawed, as binding sites exhibit substantial geometric diversity, ranging from deep and groove\-like pockets to flat or shallow surfaces\(Guo et al\.,[2015](https://arxiv.org/html/2606.12991#bib.bib19); Garcia Jimenez et al\.,[2023](https://arxiv.org/html/2606.12991#bib.bib18)\)that inherently dictates the optimal peptide topology\(Keeling et al\.,[2016](https://arxiv.org/html/2606.12991#bib.bib25)\)\. As a result, identifying the pocket\-adaptive cyclization strategy in practical cyclic peptide discovery often requires costly enumeration of candidate cyclization strategies and iterative screening\(Li et al\.,[2022](https://arxiv.org/html/2606.12991#bib.bib35); Kim et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib26)\), motivating the development of automated models for selecting cyclization strategies conditioned on the target binding site\. Meanwhile, key drug\-relevant properties such as permeability are tightly coupled to cyclization strategy, with the optimal choice varying across targets\(Chandramohan et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib9); Ji et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib22)\)\. In present methods, essential drug\-like properties of cyclic peptides, including membrane permeability, solubility, protease resistance, and immunogenicity, are largely overlooked, posing additional challenges for therapeutic peptide design\.

To tackle the above challenges jointly, we proposeAPCyc, a generative framework that enablesAutomated generation of cyclization linkage types and sites, as well asProperty\-informed optimization forCyclic peptide design\. Unlike existing approaches that rely on fixed cyclization strategies\(Zhou et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib58); Jiang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib23)\), APCyc explicitly formulates cyclization topology selection as a discrete, target\-conditioned decision\-making problem\. This design enables the automated generation of cyclization types and residue\-level linkage sites conditioned on the receptor pocket context\. Specifically, APCyc decomposes cyclization topology into residue\-level participation signals and pairwise linkage signals, allowing the model to explicitly learn cyclized residues, linkage sites, and linkage types during training\. We first extend the peptide residue vocabulary by explicitly distinguishing residues involved in cyclization from non\-cyclized residues, even when they share the same amino\-acid identity\. These cyclization\-aware embeddings are then incorporated into the denoising network of a latent diffusion model\(Kong et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib28)\), enabling supervised cyclization signals to condition and guide the denoising trajectory\. For property\-informed optimization, we train energy\-based surrogates in latent space, which guide the diffusion process via gradients toward joint multi\-property objectives\. Figure[1](https://arxiv.org/html/2606.12991#S1.F1)summarizes the high\-level design concept of APCyc\. We summarize our contributions as follows:

- •Automated pocket\-conditioned cyclizationUnlike prior methods that require predefined cyclization types or linkage sites, APCyc learns to select both directly from the target binding\-pocket context\.
- •Property\-informed controllable generationAPCyc uses Bayesian posterior guidance to steer cyclic peptide generation across affinity, permeability, protease resistance, solubility, and immunogenicity, achieving the best permeability proxy \(0\.1070\.107\) and protease\-resistance score \(−1\.474\-1\.474\)\.
- •End\-to\-end cyclic peptide designAPCyc unifies cyclization selection, all\-atom generation, and property\-informed guidance\. Across guidance settings, APCyc achieves leading structural quality, including the best Rosetta total score \(−758\.545\-758\.545\) and consistency \(0\.9710\.971\), while maintaining strong binding affinity\.

![Refer to caption](https://arxiv.org/html/2606.12991v1/x1.png)Figure 1\.APCyc Concept\.\(A\)The receptor pocket provides the structural context for target\-conditioned cyclic peptide generation\.\(B\)APCyc explicitly models cyclization as a discrete decision process by predicting both cyclization sites and linkage types\. A cyclization modeling matrix encodes pairwise site probabilities, while chemically distinct linkage modes, including amide, disulfide, and isopeptide bonds, are incorporated into the generation process\. Multi\-objective property guidance further steers sampling toward desirable therapeutic profiles, such as membrane permeability, protease resistance, and low immunogenicity\.\(C\)APCyc generates diverse candidate cyclic peptides under different design property objectives\.Teaser illustration showing receptor pocket context, APCyc cyclization and property guidance mechanisms, and candidate cyclic peptides optimized for affinity, membrane permeability, and multi\-property balance\.
## 2\.Related Work

### 2\.1\.Target\-aware cyclic peptide design

Traditional cyclic peptide discovery has relied largely on experimental screening methods, such as phage display technologies\(Heinis et al\.,[2009](https://arxiv.org/html/2606.12991#bib.bib20)\), and rational macrocyclization guided by medicinal\-chemistry heuristics\(Driggers et al\.,[2008](https://arxiv.org/html/2606.12991#bib.bib15); Zorzi et al\.,[2017](https://arxiv.org/html/2606.12991#bib.bib59); Deyle et al\.,[2017](https://arxiv.org/html/2606.12991#bib.bib13)\)\. However, these approaches are labor\-intensive and offer limited ability to jointly explore sequence, structure, and cyclization\-topology spaces\(Newman and Cragg,[2016](https://arxiv.org/html/2606.12991#bib.bib38)\)\. With the rise of deep learning and advances in protein structure prediction models such as AlphaFold2\(Jumper et al\.,[2021](https://arxiv.org/html/2606.12991#bib.bib24)\), computational methods for cyclic peptide design have begun to emerge\. AfCycDesign\(Rettie et al\.,[2025a](https://arxiv.org/html/2606.12991#bib.bib44)\)adapts AlphaFold2 to cyclic peptides by encoding N\-to\-C terminal cyclization through modified positional encodings, enabling structure prediction, sequence redesign, andde novohallucination of cyclic peptides\. In parallel, diffusion\-based generative models built on equivariant architectures\(Satorras et al\.,[2021](https://arxiv.org/html/2606.12991#bib.bib46)\)have improved geometric modeling for cyclic peptide generation\. RFpeptides\(Rettie et al\.,[2025b](https://arxiv.org/html/2606.12991#bib.bib45)\)further extends deep generative design to protein\-binding macrocycles through a denoising diffusion\-based pipeline\. DiffPepBuilder\(Wang et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib55)\)applies SE\(3\)\-equivariant diffusion, with cyclization handled by post hoc disulfide bonding\. Furthermore, CPSDE\(Zhou et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib58)\)utilizes chemical graph\-based harmonic SDEs with atom–bond modeling, while CP\-Composer\(Jiang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib23)\)performs cyclic peptide design through geometric constraint composition\. While effective, these methods are mainly trained on linear peptides and usually require cyclization constraints to be fixed a priori\. In contrast, APCyc treats cyclization topology as a target\-conditioned design variable, allowing it to infer cyclization types and linkage sites from the receptor pocket context\.

### 2\.2\.Gradient\-guided Bayesian diffusion

Diffusion models support controllable generation by modifying the reverse denoising process through guidance signals\. Early work introduced classifier guidance\(Dhariwal and Nichol,[2021](https://arxiv.org/html/2606.12991#bib.bib14)\), which steers sampling using gradients from a separately trained classifier that predicts the desired condition or attribute\.Ho and Salimans \([2022](https://arxiv.org/html/2606.12991#bib.bib21)\)proposed classifier\-free guidance, which enables conditional generation by combining conditional and unconditional score estimates without requiring an explicit classifier\. Beyond categorical conditioning, guidance has been extended to energy\-based formulations, where surrogate model gradients guide sampling toward desired objectives under Bayesian posterior interpretation\(Bao et al\.,[2022](https://arxiv.org/html/2606.12991#bib.bib6); Chung et al\.,[2022](https://arxiv.org/html/2606.12991#bib.bib11)\)\. Further extensions have explored multi\-constraint guidance, where multiple objectives or constraints are combined through score composition\. For example,Bansal et al\.\([2023](https://arxiv.org/html/2606.12991#bib.bib5)\)proposed universal guidance, which controls diffusion sampling using arbitrary guidance functions or off\-the\-shelf auxiliary networks, while MolJO\(Qiu et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib42)\)applies gradient\-based guidance to joint structure\-based molecular optimization\. Our work utilizes regressor\-based gradient guidance, combining gradients from multiple physicochemical property surrogates as posterior guidance terms for joint optimization of drug\-relevant properties\.

![Refer to caption](https://arxiv.org/html/2606.12991v1/x2.png)Figure 2\.Sampling workflow of APCyc for pocket\-conditioned cyclic peptide design\.APCyc extracts a local receptor pocket from the input receptor structure and uses it as a fixed condition\. Cyclic peptide variables, including residue identities, 3D coordinates, and cyclization topology, are encoded into residue and coordinate latents\. A latent diffusion process with an AM\-EGNN denoiser performs reverse sampling under the fixed pocket condition, with topology injection and energy guidance jointly steering cyclic peptide generation\. The decoder reconstructs the final cyclic peptide ligand in the fixed receptor pocket\.Overview diagram of APCyc encoding receptor\-peptide complexes, denoising latent variables with topology injection and posterior guidance, and decoding full\-atom cyclic peptides\.

## 3\.Method

In this section, we presentAPCyc, a joint sequence–structure latent diffusion framework that enables automated cyclization\-topology modeling and property\-informed generation for cyclic peptides\. Section[3\.1](https://arxiv.org/html/2606.12991#S3.SS1)introduces the notation and formalizes target\-conditioned cyclic peptide generation, then presents the joint sequence–structure latent diffusion formulation\. Building on this formulation, Section[3\.2](https://arxiv.org/html/2606.12991#S3.SS2)introducesAutomated Cyclization Pair Injection, which injects cyclization\-aware pair representations into the denoising network, enabling cyclization types and linkage sites to be inferred from the target context during generation\. We then give the diffusion process in detail in Section[3\.3](https://arxiv.org/html/2606.12991#S3.SS3)\. To promote desired drug\-relevant property profiles, Section[3\.4](https://arxiv.org/html/2606.12991#S3.SS4)presentsBayesian Posterior Guidance, which combines gradients from property surrogate models to steer the diffusion process toward joint multi\-property objectives\. Finally, Section[3\.5](https://arxiv.org/html/2606.12991#S3.SS5)describes the training objectives and the complete inference algorithm\. The overview of our method is shown in Figure[2](https://arxiv.org/html/2606.12991#S2.F2)\.

### 3\.1\.Preliminaries

Geometric Graph RepresentationWe represent the protein–peptide complex as a residue\-level geometric graph𝒢=\(𝒩,ℰ\)\\mathcal\{G\}=\(\\mathcal\{N\},\\mathcal\{E\}\), where𝒩\\mathcal\{N\}denotes the set of residue nodes from both the receptor and the peptide, andℰ\\mathcal\{E\}denotes edges encoding sequential, spatial, and covalent relationships between residues\. Each nodei∈𝒩i\\in\\mathcal\{N\}is associated with a tuple𝐧i=\(si,𝐗i\)\\mathbf\{n\}\_\{i\}=\(s\_\{i\},\\mathbf\{X\}\_\{i\}\), wheresis\_\{i\}denotes the residue token,𝐗i∈ℝK×3\\mathbf\{X\}\_\{i\}\\in\\mathbb\{R\}^\{K\\times 3\}stores padded full\-atom coordinates\. Here,KKis the maximum number of atoms per residue\.

Following previous work\(Kong et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib28); Lin et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib36)\), we decompose the complex into a receptor graphℛ\\mathcal\{R\}and a peptide graph𝒫\\mathcal\{P\}\. The receptor context𝒞⊆ℛ\\mathcal\{C\}\\subseteq\\mathcal\{R\}is defined as the binding\-pocket residues within 10 Å of the reference peptide during training, and is treated as a fixed conditioning input during generation\. We represent the cyclic peptide as𝒫=\(𝐬,𝐗\)\\mathcal\{P\}=\(\\mathbf\{s\},\\mathbf\{X\}\), where𝐬\\mathbf\{s\}denotes the peptide sequence,𝐗\\mathbf\{X\}denotes the full\-atom peptide structure\.

In addition to sequence and structure, we denote the desired drug\-relevant property targets as𝐘=\(y1,…,yM\)∈ℝM\\mathbf\{Y\}=\(y\_\{1\},\\ldots,y\_\{M\}\)\\in\\mathbb\{R\}^\{M\}, where eachymy\_\{m\}is a normalized scalar score for a physicochemical or developability\-related property, such as membrane permeability, solubility, or protease resistance\. Given a fixed receptor context𝒞\\mathcal\{C\}and desired property targets𝐘\\mathbf\{Y\}, conditional cyclic peptide generation aims to sample peptide graphs fromp​\(𝒫∣𝒞,𝐘\)p\(\\mathcal\{P\}\\mid\\mathcal\{C\},\\mathbf\{Y\}\), where sequence, structure, and cyclization linkage are generated jointly\.

Latent Diffusion ModelingFollowing prior work on antibody and linear peptide design\(Kong et al\.,[2023](https://arxiv.org/html/2606.12991#bib.bib27),[2024](https://arxiv.org/html/2606.12991#bib.bib28)\), we adopt a dyMEAN\-style autoencoder to map peptide graphs into residue\-level latent representations\. Specifically, an encoderℰϕ\\mathcal\{E\}\_\{\\phi\}embeds the peptide graph into latent variables𝒵=\{\(zi,z~i\)\}i=1L\\mathcal\{Z\}=\\\{\(z\_\{i\},\\tilde\{z\}\_\{i\}\)\\\}\_\{i=1\}^\{L\}, whereziz\_\{i\}is E\(3\)\-invariant andz~i\\tilde\{z\}\_\{i\}is E\(3\)\-equivariant\. We then learn a diffusion model over𝒵\\mathcal\{Z\}using an equivariant GNN denoiser, and decode the final latents back to peptide sequence and structure via a decoder𝒟ξ\\mathcal\{D\}\_\{\\xi\}\.

Cyclization\-Enhanced Vocabulary ExpansionMotivated by explicit cyclization annotations in cyclic peptide datasets such as CPSea\(Yang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib57)\), we extend the standard residue alphabet𝒜std\\mathcal\{A\}\_\{\\mathrm\{std\}\}to𝒜ext=𝒜std∪𝒜cyc\\mathcal\{A\}\_\{\\mathrm\{ext\}\}=\\mathcal\{A\}\_\{\\mathrm\{std\}\}\\cup\\mathcal\{A\}\_\{\\mathrm\{cyc\}\}, where𝒜cyc\\mathcal\{A\}\_\{\\mathrm\{cyc\}\}contains specialized tokens for residues participating in specific cyclization linkages, such as CYS\_SS for disulfide\-linked cysteine and LYS\_ISO/ASP\_ISO for residues involved in isopeptide bonds\. This extension allows the model to distinguish residues involved in cyclization from their non\-cyclized counterparts at the token level, even when they share the same standard amino\-acid identity\.

### 3\.2\.Automated Cyclization Pair Injection

At a discrete diffusion timestept∈\{0,…,T\}t\\in\\\{0,\\ldots,T\\\}, the noisy state𝒵t=\{\(zi,t,z~i,t\)\}i=1L\\mathcal\{Z\}\_\{t\}=\\\{\(z\_\{i,t\},\\tilde\{z\}\_\{i,t\}\)\\\}\_\{i=1\}^\{L\}is sampled from the closed\-form marginal of the DDPM\-style latent diffusion process defined in Sec\.[3\.3](https://arxiv.org/html/2606.12991#S3.SS3):

\(1\)zi,t=αt​zi,0\+σt​ϵi,z~i,t=αt​z~i,0\+σt​ϵ~i,z\_\{i,t\}=\\alpha\_\{t\}z\_\{i,0\}\+\\sigma\_\{t\}\\epsilon\_\{i\},\\quad\\tilde\{z\}\_\{i,t\}=\\alpha\_\{t\}\\tilde\{z\}\_\{i,0\}\+\\sigma\_\{t\}\\tilde\{\\epsilon\}\_\{i\},whereϵi∼𝒩​\(𝟎,𝐈\)\\epsilon\_\{i\}\\sim\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\)is Gaussian noise for the invariant latent,ϵ~i∼𝒩​\(𝟎,𝐈\)\\tilde\{\\epsilon\}\_\{i\}\\sim\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\)is isotropic Gaussian noise with the same shape asz~i,0\\tilde\{z\}\_\{i,0\}, andαt\\alpha\_\{t\}andσt\\sigma\_\{t\}are the cumulative signal and noise coefficients satisfyingσt2=1−αt2\\sigma\_\{t\}^\{2\}=1\-\\alpha\_\{t\}^\{2\}\.

Pair Representation ConstructionTo facilitate cyclization inference, we construct a high\-dimensional pairwise tensor𝐇tpair∈ℝL×L×dp\\mathbf\{H\}^\{\\text\{pair\}\}\_\{t\}\\in\\mathbb\{R\}^\{L\\times L\\times d\_\{p\}\}\. We leverage both the invariant node latentszi,tz\_\{i,t\}and the equivariant coordinatesz~i,t\\tilde\{z\}\_\{i,t\}\. The pairwise feature vector for residues\(i,j\)\(i,j\)is computed as:

\(2\)𝐡i​j,tpair=ϕenc​\(Concat​\(zi,t,zj,t,zi,t⊙zj,t,ψ​\(‖z~i,t−z~j,t‖\)\)\)\\mathbf\{h\}\_\{ij,t\}^\{\\text\{pair\}\}=\\phi\_\{\\text\{enc\}\}\\left\(\\text\{Concat\}\\left\(z\_\{i,t\},\\,z\_\{j,t\},\\,z\_\{i,t\}\\odot z\_\{j,t\},\\,\\psi\(\\\|\\tilde\{z\}\_\{i,t\}\-\\tilde\{z\}\_\{j,t\}\\\|\)\\right\)\\right\)where⊙\\odotdenotes the Hadamard product andψ​\(⋅\)\\psi\(\\cdot\)represents a sinusoidal RBF embedding\(Schütt et al\.,[2018](https://arxiv.org/html/2606.12991#bib.bib49)\)\. This representation captures both the semantic compatibility \(viazz\) and geometric proximity \(viaz~\\tilde\{z\}\) required for bond formation\. To disentangle geometric modeling from cyclization reasoning, we project𝐡i​j,tpair\\mathbf\{h\}\_\{ij,t\}^\{\\text\{pair\}\}into two task\-specific subspaces:

\(3\)𝐡i​j,tstruct=ϕstruct​\(𝐡i​j,tpair\),𝐡i​j,tcyc=ϕcyc​\(𝐡i​j,tpair\)\.\\mathbf\{h\}\_\{ij,t\}^\{\\text\{struct\}\}=\\phi\_\{\\text\{struct\}\}\(\\mathbf\{h\}\_\{ij,t\}^\{\\text\{pair\}\}\),\\quad\\mathbf\{h\}\_\{ij,t\}^\{\\text\{cyc\}\}=\\phi\_\{\\text\{cyc\}\}\(\\mathbf\{h\}\_\{ij,t\}^\{\\text\{pair\}\}\)\.whereϕstruct\\phi\_\{\\text\{struct\}\}andϕcyc\\phi\_\{\\text\{cyc\}\}are separate projection heads \(MLPs\)\. The resulting𝐡i​jstruct\\mathbf\{h\}\_\{ij\}^\{\\text\{struct\}\}is utilized for geometric updates, while𝐡i​jcyc\\mathbf\{h\}\_\{ij\}^\{\\text\{cyc\}\}serves as the basis for predicting cyclization probability\.

We then predict the cyclization type distribution𝐩τ,t\\mathbf\{p\}\_\{\\tau,t\}and the linkage matrix𝐒t\\mathbf\{S\}\_\{t\}at timesteptt\. Chemical constraints, such as Cys–Cys compatibility, are enforced through a validity mask𝐌valid∈\{0,1\}L×L\\mathbf\{M\}\_\{\\mathrm\{valid\}\}\\in\\\{0,1\\\}^\{L\\times L\}\. Distinct from row\-wise attention, the linkage probabilities are computed by a global normalization over all valid residue\-pair entries\. Specifically, the cyclization head predicts pairwise site logits𝐋site,t∈ℝL×L\\mathbf\{L\}\_\{\\mathrm\{site\},t\}\\in\\mathbb\{R\}^\{L\\times L\}from𝐡i​j,tcyc\\mathbf\{h\}\_\{ij,t\}^\{\\mathrm\{cyc\}\}\. We first compute a masked, pre\-symmetrized probability matrix𝐏t\\mathbf\{P\}\_\{t\}as

\(4\)vec​\(𝐏t\)=Softmax​\(vec​\(𝐋site,t\+ℳ​\(𝐌valid\)\)\),\\mathrm\{vec\}\(\\mathbf\{P\}\_\{t\}\)=\\mathrm\{Softmax\}\\left\(\\mathrm\{vec\}\\left\(\\mathbf\{L\}\_\{\\mathrm\{site\},t\}\+\\mathcal\{M\}\(\\mathbf\{M\}\_\{\\mathrm\{valid\}\}\)\\right\)\\right\),wherevec​\(⋅\)\\mathrm\{vec\}\(\\cdot\)flattens anL×LL\\times Lmatrix into a vector, and the mask operatorℳ\\mathcal\{M\}is defined element\-wise as

\(5\)ℳ​\(𝐌valid\)i​j=\{0,Mvalid,i​j=1,−∞,Mvalid,i​j=0\.\\mathcal\{M\}\(\\mathbf\{M\}\_\{\\mathrm\{valid\}\}\)\_\{ij\}=\\begin\{cases\}0,&M\_\{\\mathrm\{valid\},ij\}=1,\\\\ \-\\infty,&M\_\{\\mathrm\{valid\},ij\}=0\.\\end\{cases\}Thus invalid entries receive zero probability after the softmax\. The final undirected linkage matrix is obtained by symmetrizing𝐏t\\mathbf\{P\}\_\{t\}:

\(6\)Si​j,t=\{Pi​j,t\+Pj​i,t,i≠j,0,i=j\.S\_\{ij,t\}=\\begin\{cases\}P\_\{ij,t\}\+P\_\{ji,t\},&i\\neq j,\\\\ 0,&i=j\.\\end\{cases\}
The cyclization type distribution is predicted as

\(7\)𝐩τ,t=Softmax​\(ϕtype​\(Pool​\(\{𝐡i​j,tcyc\}i<j\)\)\)\.\\mathbf\{p\}\_\{\\tau,t\}=\\mathrm\{Softmax\}\\left\(\\phi\_\{\\mathrm\{type\}\}\\left\(\\mathrm\{Pool\}\\left\(\\\{\\mathbf\{h\}\_\{ij,t\}^\{\\mathrm\{cyc\}\}\\\}\_\{i<j\}\\right\)\\right\)\\right\)\.
Adaptive Topology Injection via Edge Message ModulationGiven the predicted cyclization type distribution𝐩τ,t\\mathbf\{p\}\_\{\\tau,t\}and the symmetric linkage matrix𝐒t\\mathbf\{S\}\_\{t\}, we inject the online topology signal into the AM\-EGNN denoising network through edge\-feature augmentation and message modulation\. The injection strength is controlled by timestep\-dependent schedulesλfeat​\(t\)\\lambda\_\{\\mathrm\{feat\}\}\(t\),λbias​\(t\)\\lambda\_\{\\mathrm\{bias\}\}\(t\), andλgate​\(t\)\\lambda\_\{\\mathrm\{gate\}\}\(t\)\. In our implementation, these schedules are linear functions of the diffusion timestep and are used to gradually adjust the topology\-injection strength during denoising\.

Since𝐒t\\mathbf\{S\}\_\{t\}has already been symmetrized in Eq\. \([6](https://arxiv.org/html/2606.12991#S3.E6)\), we directly useSi​j,tS\_\{ij,t\}as the undirected linkage probability between residuesiiandjj\. We first fuse the linkage probability and the cyclization type prediction into the edge features:

\(8\)𝐞i​j,tinj=ϕedge​\(Concat​\(𝐞i​j,tbase,𝐩τ,t,λfeat​\(t\)​Si​j,t\)\),\\mathbf\{e\}\_\{ij,t\}^\{\\mathrm\{inj\}\}=\\phi\_\{\\mathrm\{edge\}\}\\left\(\\mathrm\{Concat\}\\left\(\\mathbf\{e\}\_\{ij,t\}^\{\\mathrm\{base\}\},\\mathbf\{p\}\_\{\\tau,t\},\\lambda\_\{\\mathrm\{feat\}\}\(t\)S\_\{ij,t\}\\right\)\\right\),where𝐞i​j,tbase\\mathbf\{e\}\_\{ij,t\}^\{\\mathrm\{base\}\}denotes the base geometric RBF edge feature at timesteptt, andλfeat​\(t\)​Si​j,t\\lambda\_\{\\mathrm\{feat\}\}\(t\)S\_\{ij,t\}is treated as a scalar topology feature\.

To further modulate message passing, we compute an additive bias and a multiplicative gate from the structural pair representation and the injected topology signal:

\(9\)𝐛i​j,t\\displaystyle\\mathbf\{b\}\_\{ij,t\}=ϕbias​\(𝐡i​j,tstruct\)\+λbias​\(t\)​Si​j,t​𝟏dm,\\displaystyle=\\phi\_\{\\mathrm\{bias\}\}\\left\(\\mathbf\{h\}\_\{ij,t\}^\{\\mathrm\{struct\}\}\\right\)\+\\lambda\_\{\\mathrm\{bias\}\}\(t\)S\_\{ij,t\}\\mathbf\{1\}\_\{d\_\{m\}\},\(10\)𝐠i​j,t\\displaystyle\\mathbf\{g\}\_\{ij,t\}=σ​\(ϕgate​\(𝐡i​j,tstruct\)\)⊙\[σ​\(λgate​\(t\)​Si​j,t\)​𝟏dm\],\\displaystyle=\\sigma\\left\(\\phi\_\{\\mathrm\{gate\}\}\\left\(\\mathbf\{h\}\_\{ij,t\}^\{\\mathrm\{struct\}\}\\right\)\\right\)\\odot\\left\[\\sigma\\left\(\\lambda\_\{\\mathrm\{gate\}\}\(t\)S\_\{ij,t\}\\right\)\\mathbf\{1\}\_\{d\_\{m\}\}\\right\],where𝟏dm\\mathbf\{1\}\_\{d\_\{m\}\}is an all\-one vector with the same dimension as the edge message, so that the scalar linkage probability is broadcast to the message dimension\. Hereσ​\(⋅\)\\sigma\(\\cdot\)denotes the sigmoid function and⊙\\odotdenotes element\-wise multiplication\.

Within each AM\-EGNN layer, let𝐦i​j,tbase\\mathbf\{m\}\_\{ij,t\}^\{\\mathrm\{base\}\}denote the base message from nodejjto nodeiicomputed using the injected edge feature𝐞i​j,tinj\\mathbf\{e\}\_\{ij,t\}^\{\\mathrm\{inj\}\}\. The topology\-aware message is then defined as

\(11\)𝐦i​j,tmod=\(𝐦i​j,tbase\+𝐛i​j,t\)⊙𝐠i​j,t\.\\mathbf\{m\}\_\{ij,t\}^\{\\mathrm\{mod\}\}=\\left\(\\mathbf\{m\}\_\{ij,t\}^\{\\mathrm\{base\}\}\+\\mathbf\{b\}\_\{ij,t\}\\right\)\\odot\\mathbf\{g\}\_\{ij,t\}\.This modulation allows the predicted cyclization topology to guide denoising while preserving the flexibility of the AM\-EGNN to refine local geometry through the original geometric edge features\.

###### Proposition 3\.1 \(SE\(3\)\-Equivariance\)\.

LetFθF\_\{\\theta\}denote the denoising update with topology injection\. ThenFθF\_\{\\theta\}is SE\(3\)\-equivariant, i\.e\.,

Fθ​\(g⋅𝒵t\)=g⋅Fθ​\(𝒵t\),∀g∈SE​\(3\)\.F\_\{\\theta\}\(g\\cdot\\mathcal\{Z\}\_\{t\}\)=g\\cdot F\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\}\),\\quad\\forall g\\in\\mathrm\{SE\}\(3\)\.

During the late denoising phase, we switch to a deterministic strategy by forming𝐒tinj\\mathbf\{S\}^\{\\mathrm\{inj\}\}\_\{t\}from the discretized linkage matrix𝐒t\\mathbf\{S\}\_\{t\}\. This ”hard” linkage is treated as an explicit covalent edge, enhancing precise geometric closure in the final structure\. Figure[3](https://arxiv.org/html/2606.12991#S3.F3)illustrates the online topology injection module\.

![Refer to caption](https://arxiv.org/html/2606.12991v1/x3.png)Figure 3\.Cyclization topology information injection\.APCyc predicts cyclization linkages and uses the topology signal to modulate AM\-EGNN message passing\.Diagram of online topology injection where pairwise representations produce a linkage matrix and cyclization features that modulate message passing before the AM\-EGNN update\.
### 3\.3\.Geometry Latent Diffusion Process

FollowingKong et al\.\([2024](https://arxiv.org/html/2606.12991#bib.bib28)\), we model cyclic peptide generation with a DDPM\-style latent diffusion process over residue\-level latent variables\. The encoder outputs𝒵0=\{\(zi,0,z~i,0\)\}i=1L\\mathcal\{Z\}\_\{0\}=\\\{\(z\_\{i,0\},\\tilde\{z\}\_\{i,0\}\)\\\}\_\{i=1\}^\{L\}are treated as clean latents, wherezi,0z\_\{i,0\}is the invariant residue latent andz~i,0\\tilde\{z\}\_\{i,0\}is the equivariant geometric latent\.

Lett∈\{0,…,T\}t\\in\\\{0,\\ldots,T\\\}be a discrete diffusion timestep\. Given a variance\-preserving noise schedule\{βt\}t=1T\\\{\\beta\_\{t\}\\\}\_\{t=1\}^\{T\}, we define

ηt=1−βt,η¯t=∏s=1tηs,αt=η¯t,σt=1−η¯t\.\\eta\_\{t\}=1\-\\beta\_\{t\},\\quad\\bar\{\\eta\}\_\{t\}=\\prod\_\{s=1\}^\{t\}\\eta\_\{s\},\\quad\\alpha\_\{t\}=\\sqrt\{\\bar\{\\eta\}\_\{t\}\},\\quad\\sigma\_\{t\}=\\sqrt\{1\-\\bar\{\\eta\}\_\{t\}\}\.The forward perturbation admits the closed\-form marginal

\(12\)zi,t=αt​zi,0\+σt​ϵi,z~i,t=αt​z~i,0\+σt​ϵ~i,z\_\{i,t\}=\\alpha\_\{t\}z\_\{i,0\}\+\\sigma\_\{t\}\\epsilon\_\{i\},\\quad\\tilde\{z\}\_\{i,t\}=\\alpha\_\{t\}\\tilde\{z\}\_\{i,0\}\+\\sigma\_\{t\}\\tilde\{\\epsilon\}\_\{i\},whereϵi\\epsilon\_\{i\}andϵ~i\\tilde\{\\epsilon\}\_\{i\}are independent standard Gaussian noise terms with shapes matchingzi,0z\_\{i,0\}andz~i,0\\tilde\{z\}\_\{i,0\}, respectively\.

The denoising networkϵθ​\(𝒵t,𝒞,t\)\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)predicts the noise added in Eq\. \([12](https://arxiv.org/html/2606.12991#S3.E12)\)\. It is implemented as the cyclization\-aware AM\-EGNN described in Sec\.[3\.2](https://arxiv.org/html/2606.12991#S3.SS2)\. During generation, we start from𝒵T∼𝒩​\(𝟎,𝐈\)\\mathcal\{Z\}\_\{T\}\\sim\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\)and apply the standard DDPM reverse update

\(13\)𝒵t−1=1ηt​\(𝒵t−βtσt​ϵθ​\(𝒵t,𝒞,t\)\)\+β~t​𝝃,𝝃∼𝒩​\(𝟎,𝐈\),\\mathcal\{Z\}\_\{t\-1\}=\\frac\{1\}\{\\sqrt\{\\eta\_\{t\}\}\}\\left\(\\mathcal\{Z\}\_\{t\}\-\\frac\{\\beta\_\{t\}\}\{\\sigma\_\{t\}\}\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)\\right\)\+\\sqrt\{\\tilde\{\\beta\}\_\{t\}\}\\boldsymbol\{\\xi\},\\quad\\boldsymbol\{\\xi\}\\sim\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\),whereβ~t=1−η¯t−11−η¯t​βt\\tilde\{\\beta\}\_\{t\}=\\frac\{1\-\\bar\{\\eta\}\_\{t\-1\}\}\{1\-\\bar\{\\eta\}\_\{t\}\}\\beta\_\{t\}, and𝝃\\boldsymbol\{\\xi\}is set to zero whent=1t=1\.

Training ObjectiveAt training time, we uniformly sample a timestept∈\{1,…,T\}t\\in\\\{1,\\ldots,T\\\}, construct𝒵t\\mathcal\{Z\}\_\{t\}using Eq\. \([12](https://arxiv.org/html/2606.12991#S3.E12)\), and optimize the DDPM noise\-prediction loss

\(14\)ℒdiff=𝔼t,𝒵0,ϵ​\[‖ϵ−ϵθ​\(𝒵t,𝒞,t\)‖22\],\\mathcal\{L\}\_\{\\mathrm\{diff\}\}=\\mathbb\{E\}\_\{t,\\mathcal\{Z\}\_\{0\},\\boldsymbol\{\\epsilon\}\}\\left\[\\left\\\|\\boldsymbol\{\\epsilon\}\-\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)\\right\\\|\_\{2\}^\{2\}\\right\],whereϵ=\{\(ϵi,ϵ~i\)\}i=1L\\boldsymbol\{\\epsilon\}=\\\{\(\\epsilon\_\{i\},\\tilde\{\\epsilon\}\_\{i\}\)\\\}\_\{i=1\}^\{L\}\.

We further supervise the online topology prediction using the pre\-symmetrized linkage probability matrix𝐏t\\mathbf\{P\}\_\{t\}from Eq\. \([4](https://arxiv.org/html/2606.12991#S3.E4)\) and the cyclization type distribution𝐩τ,t\\mathbf\{p\}\_\{\\tau,t\}\. Let𝐒∗\\mathbf\{S\}^\{\*\}denote the ground\-truth linkage matrix andτ∗\\tau^\{\*\}denote the ground\-truth cyclization type\. We define

Yi​j∗=Si​j∗​Mvalid,i​j∑u,vSu​v∗​Mvalid,u​v,Y\_\{ij\}^\{\*\}=\\frac\{S\_\{ij\}^\{\*\}M\_\{\\mathrm\{valid\},ij\}\}\{\\sum\_\{u,v\}S\_\{uv\}^\{\*\}M\_\{\\mathrm\{valid\},uv\}\},and use

\(15\)ℒtopo=𝔼t,𝒵0,ϵ​\[−∑i,jYi​j∗​log⁡\(Pi​j,t\)\+λtype​CE​\(𝐩τ,t,τ∗\)\]\.\\mathcal\{L\}\_\{\\mathrm\{topo\}\}=\\mathbb\{E\}\_\{t,\\mathcal\{Z\}\_\{0\},\\boldsymbol\{\\epsilon\}\}\\left\[\-\\sum\_\{i,j\}Y\_\{ij\}^\{\*\}\\log\(P\_\{ij,t\}\)\+\\lambda\_\{\\mathrm\{type\}\}\\mathrm\{CE\}\(\\mathbf\{p\}\_\{\\tau,t\},\\tau^\{\*\}\)\\right\]\.The final objective is

\(16\)ℒtotal=ℒdiff\+λtopo​ℒtopo\.\\mathcal\{L\}\_\{\\mathrm\{total\}\}=\\mathcal\{L\}\_\{\\mathrm\{diff\}\}\+\\lambda\_\{\\mathrm\{topo\}\}\\mathcal\{L\}\_\{\\mathrm\{topo\}\}\.

### 3\.4\.Property Optimization via Bayesian Posterior Guidance

To steer generation toward peptide candidates satisfying desired therapeutic criteria, such as binding affinity, permeability, solubility, or protease resistance, we perform property\-guided sampling in the latent space\. Given the target property vector𝐘=\(y1,…,yM\)∈ℝM\\mathbf\{Y\}=\(y\_\{1\},\\ldots,y\_\{M\}\)\\in\\mathbb\{R\}^\{M\}and receptor context𝒞\\mathcal\{C\}, the goal is to sample from the property\-conditioned posteriorpt​\(𝒵t∣𝒞,𝐘\)p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\},\\mathbf\{Y\}\)rather than only from the receptor\-conditioned priorpt​\(𝒵t∣𝒞\)p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\}\)\.

By Bayes’ rule, the posterior score can be decomposed as

\(17\)∇𝒵tlog⁡pt​\(𝒵t∣𝒞,𝐘\)=∇𝒵tlog⁡pt​\(𝒵t∣𝒞\)\+∇𝒵tlog⁡pt​\(𝐘∣𝒵t,𝒞\)\.\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\log p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\},\\mathbf\{Y\}\)=\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\log p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\}\)\+\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\log p\_\{t\}\(\\mathbf\{Y\}\\mid\\mathcal\{Z\}\_\{t\},\\mathcal\{C\}\)\.The first term is represented by the DDPM denoising network throughsθ​\(𝒵t,𝒞,t\)≈−ϵθ​\(𝒵t,𝒞,t\)/σts\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)\\approx\-\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)/\\sigma\_\{t\}\. The second term is intractable because drug\-relevant properties are not directly defined on noisy latent states\. We therefore introduce a time\-conditioned differentiable surrogateΨψ\\Psi\_\{\\psi\}and define the property energy

\(18\)ℰψ​\(𝒵t,t;𝐘,𝒞\)=∑m=1Mλm​ℓm​\(y^m,t,ym\),𝐘^t=Ψψ​\(𝒵t,𝒞,t\),\\mathcal\{E\}\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},t;\\mathbf\{Y\},\\mathcal\{C\}\)=\\sum\_\{m=1\}^\{M\}\\lambda\_\{m\}\\ell\_\{m\}\\left\(\\hat\{y\}\_\{m,t\},y\_\{m\}\\right\),\\quad\\hat\{\\mathbf\{Y\}\}\_\{t\}=\\Psi\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\),where𝐘^t=\(y^1,t,…,y^M,t\)\\hat\{\\mathbf\{Y\}\}\_\{t\}=\(\\hat\{y\}\_\{1,t\},\\ldots,\\hat\{y\}\_\{M,t\}\)is the predicted property vector at timesteptt,ℓm\\ell\_\{m\}measures the deviation from themm\-th desired property target, andλm\\lambda\_\{m\}controls its relative importance\. This energy serves as a differentiable proxy for the negative log\-likelihood−log⁡pt​\(𝐘∣𝒵t,𝒞\)\-\\log p\_\{t\}\(\\mathbf\{Y\}\\mid\\mathcal\{Z\}\_\{t\},\\mathcal\{C\}\)\. The resulting guided posterior is written as

\(19\)pt​\(𝒵t∣𝒞,𝐘\)∝pt​\(𝒵t∣𝒞\)​exp⁡\(−γt​ℰψ​\(𝒵t,t;𝐘,𝒞\)\),p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\},\\mathbf\{Y\}\)\\propto p\_\{t\}\(\\mathcal\{Z\}\_\{t\}\\mid\\mathcal\{C\}\)\\exp\\left\(\-\\gamma\_\{t\}\\mathcal\{E\}\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},t;\\mathbf\{Y\},\\mathcal\{C\}\)\\right\),whereγt\\gamma\_\{t\}is a timestep\-dependent guidance scale\. Therefore, the guided score is approximated by

\(20\)sθ,ψguide​\(𝒵t,𝒞,𝐘,t\)=−ϵθ​\(𝒵t,𝒞,t\)σt−γt​∇𝒵tℰψ​\(𝒵t,t;𝐘,𝒞\)\.s\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},\\mathbf\{Y\},t\)=\-\\frac\{\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)\}\{\\sigma\_\{t\}\}\-\\gamma\_\{t\}\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\mathcal\{E\}\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},t;\\mathbf\{Y\},\\mathcal\{C\}\)\.Equivalently, under the DDPM noise\-prediction parameterization, we use the guided noise estimate

\(21\)ϵθ,ψguide=ϵθ\+σt​γt​∇𝒵tℰψ​\(𝒵t,t;𝐘,𝒞\)\.\\boldsymbol\{\\epsilon\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}=\\boldsymbol\{\\epsilon\}\_\{\\theta\}\+\\sigma\_\{t\}\\gamma\_\{t\}\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\mathcal\{E\}\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},t;\\mathbf\{Y\},\\mathcal\{C\}\)\.During sampling, we replaceϵθ\\boldsymbol\{\\epsilon\}\_\{\\theta\}withϵθ,ψguide\\boldsymbol\{\\epsilon\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}in the DDPM reverse update:

\(22\)𝒵t−1=1ηt​\(𝒵t−βtσt​ϵθ,ψguide\)\+β~t​𝝃\.\\mathcal\{Z\}\_\{t\-1\}=\\frac\{1\}\{\\sqrt\{\\eta\_\{t\}\}\}\\left\(\\mathcal\{Z\}\_\{t\}\-\\frac\{\\beta\_\{t\}\}\{\\sigma\_\{t\}\}\\boldsymbol\{\\epsilon\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}\\right\)\+\\sqrt\{\\tilde\{\\beta\}\_\{t\}\}\\boldsymbol\{\\xi\}\.Thus, the energy gradient jointly guides the invariant residue latents\{zi,t\}i=1L\\\{z\_\{i,t\}\\\}\_\{i=1\}^\{L\}and the equivariant geometric latents\{z~i,t\}i=1L\\\{\\tilde\{z\}\_\{i,t\}\\\}\_\{i=1\}^\{L\}, optimizing property with structural refinement during generation\.

![Refer to caption](https://arxiv.org/html/2606.12991v1/x4.png)Figure 4\.Overall comparison across different methods on therapeutic properties and generative metrics\.Radar chart comparing PepGLAD, CPComposer, PepFlow, and APCyc guidance variants across solubility, protease resistance, membrane permeability, immunogenicity, binding affinity, cyclization success, diversity, and consistency metrics\.Algorithm 1APCyc Inference with Topology and Property Guidance1:Input:Denoiser

ϵθ\\boldsymbol\{\\epsilon\}\_\{\\theta\}, decoder

𝒟ξ\\mathcal\{D\}\_\{\\xi\}, receptor context

𝒞\\mathcal\{C\}, property targets

𝐘=\(y1,…,yM\)\\mathbf\{Y\}=\(y\_\{1\},\\ldots,y\_\{M\}\), property surrogate

Ψψ\\Psi\_\{\\psi\}, base weights

𝐰=\{wm\}m=1M\\mathbf\{w\}=\\\{w\_\{m\}\\\}\_\{m=1\}^\{M\}, validity mask

𝐌valid\\mathbf\{M\}\_\{\\mathrm\{valid\}\}, aggregation operator

Agg​\(⋅\)\\mathrm\{Agg\}\(\\cdot\), norm operator

Norm​\(⋅\)\\mathrm\{Norm\}\(\\cdot\), guidance scale

γt\\gamma\_\{t\}, hard\-linkage threshold

thardt\_\{\\mathrm\{hard\}\}
2:Initialize:Sample

𝒵T=\{\(zi,T,z~i,T\)\}i=1L∼𝒩​\(𝟎,𝐈\)\\mathcal\{Z\}\_\{T\}=\\\{\(z\_\{i,T\},\\tilde\{z\}\_\{i,T\}\)\\\}\_\{i=1\}^\{L\}\\sim\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\)
3:for

t=T,…,1t=T,\\ldots,1do

4:

𝐩τ,t,𝐏t,𝐒t←CyclizationHeads​\(𝒵t,𝒞,t;𝐌valid\)\\mathbf\{p\}\_\{\\tau,t\},\\mathbf\{P\}\_\{t\},\\mathbf\{S\}\_\{t\}\\leftarrow\\mathrm\{CyclizationHeads\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t;\\mathbf\{M\}\_\{\\mathrm\{valid\}\}\)
5:

𝐒tinj←\{SymOneHot​\(arg⁡maxi<j:Mvalid,i​j=1⁡Si​j,t\),t≤thard,𝐒t,t\>thard\.\\mathbf\{S\}^\{\\mathrm\{inj\}\}\_\{t\}\\leftarrow\\begin\{cases\}\\mathrm\{SymOneHot\}\\left\(\\arg\\max\_\{i<j:\\,M\_\{\\mathrm\{valid\},ij\}=1\}S\_\{ij,t\}\\right\),&t\\leq t\_\{\\mathrm\{hard\}\},\\\\ \\mathbf\{S\}\_\{t\},&t\>t\_\{\\mathrm\{hard\}\}\.\\end\{cases\}
6:

𝐘^t←Ψψ​\(𝒵t,𝒞,t\)\\hat\{\\mathbf\{Y\}\}\_\{t\}\\leftarrow\\Psi\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\)
7:for

m=1,…,Mm=1,\\ldots,Mdo

8:

E^m,t←ℓm​\(y^m,t,ym\)\\hat\{E\}\_\{m,t\}\\leftarrow\\ell\_\{m\}\(\\hat\{y\}\_\{m,t\},y\_\{m\}\)
9:

𝐠m,t←∇𝒵tE^m,t\\mathbf\{g\}\_\{m,t\}\\leftarrow\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\hat\{E\}\_\{m,t\}
10:

Nm,t←Norm​\(𝐠m,t\)N\_\{m,t\}\\leftarrow\\mathrm\{Norm\}\(\\mathbf\{g\}\_\{m,t\}\)
11:endfor

12:

λm​\(t\)←stopgrad​\(wm⋅Agg​\(\{Nj,t\}j=1M\)Nm,t\+εbal\),m=1,…,M\\lambda\_\{m\}\(t\)\\leftarrow\\mathrm\{stopgrad\}\\left\(w\_\{m\}\\cdot\\frac\{\\mathrm\{Agg\}\(\\\{N\_\{j,t\}\\\}\_\{j=1\}^\{M\}\)\}\{N\_\{m,t\}\+\\varepsilon\_\{\\mathrm\{bal\}\}\}\\right\),\\quad m=1,\\ldots,M
13:

E^bal​\(𝒵t,t\)←∑m=1Mλm​\(t\)​E^m,t\\hat\{E\}\_\{\\mathrm\{bal\}\}\(\\mathcal\{Z\}\_\{t\},t\)\\leftarrow\\sum\_\{m=1\}^\{M\}\\lambda\_\{m\}\(t\)\\hat\{E\}\_\{m,t\}
14:

𝐠prop,t←∇𝒵tE^bal​\(𝒵t,t\)\\mathbf\{g\}\_\{\\mathrm\{prop\},t\}\\leftarrow\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\hat\{E\}\_\{\\mathrm\{bal\}\}\(\\mathcal\{Z\}\_\{t\},t\)
15:

ϵ^θ←ϵθ​\(𝒵t,𝒞,t;𝐒tinj,𝐩τ,t\)\\hat\{\\boldsymbol\{\\epsilon\}\}\_\{\\theta\}\\leftarrow\\boldsymbol\{\\epsilon\}\_\{\\theta\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t;\\mathbf\{S\}^\{\\mathrm\{inj\}\}\_\{t\},\\mathbf\{p\}\_\{\\tau,t\}\)
16:

ϵ^θ,ψguide←ϵ^θ\+γt​σt​𝐠prop,t\\hat\{\\boldsymbol\{\\epsilon\}\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}\\leftarrow\\hat\{\\boldsymbol\{\\epsilon\}\}\_\{\\theta\}\+\\gamma\_\{t\}\\sigma\_\{t\}\\mathbf\{g\}\_\{\\mathrm\{prop\},t\}
17:

𝒵t−1←DDPMReverseStep​\(𝒵t,ϵ^θ,ψguide,t\)\\mathcal\{Z\}\_\{t\-1\}\\leftarrow\\mathrm\{DDPMReverseStep\}\\left\(\\mathcal\{Z\}\_\{t\},\\hat\{\\boldsymbol\{\\epsilon\}\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\},t\\right\)
18:endfor

19:

𝐩τ,0,𝐏0,𝐒0←CyclizationHeads​\(𝒵0,𝒞,0;𝐌valid\)\\mathbf\{p\}\_\{\\tau,0\},\\mathbf\{P\}\_\{0\},\\mathbf\{S\}\_\{0\}\\leftarrow\\mathrm\{CyclizationHeads\}\(\\mathcal\{Z\}\_\{0\},\\mathcal\{C\},0;\\mathbf\{M\}\_\{\\mathrm\{valid\}\}\)
20:

𝒫^←𝒟ξ​\(𝒵0\)\\hat\{\\mathcal\{P\}\}\\leftarrow\\mathcal\{D\}\_\{\\xi\}\(\\mathcal\{Z\}\_\{0\}\)
21:return

𝒫^,arg⁡max⁡𝐩τ,0,𝐒0\\hat\{\\mathcal\{P\}\},\\arg\\max\\mathbf\{p\}\_\{\\tau,0\},\\mathbf\{S\}\_\{0\}

### 3\.5\.Training and Inference

APCyc is trained in a decoupled two\-stage paradigm\. We first train a joint latent autoencoder to learn residue\-level invariant and equivariant latents, and then train a DDPM\-style latent diffusion model over the frozen latent space\.

Stage 1: Joint Latent AutoencodingWe train the encoderℰϕ\\mathcal\{E\}\_\{\\phi\}and decoder𝒟ξ\\mathcal\{D\}\_\{\\xi\}to map a peptide graph𝒫\\mathcal\{P\}into clean residue\-level latents𝒵0=\{\(zi,0,z~i,0\)\}i=1L\\mathcal\{Z\}\_\{0\}=\\\{\(z\_\{i,0\},\\tilde\{z\}\_\{i,0\}\)\\\}\_\{i=1\}^\{L\}and reconstruct the full\-atom peptide\. The autoencoding objective is

\(23\)ℒAE=1L​∑i=1L\(ℒrec​\(i\)\+ℒlatent​\(i\)\)\.\\mathcal\{L\}\_\{\\mathrm\{AE\}\}=\\frac\{1\}\{L\}\\sum\_\{i=1\}^\{L\}\\left\(\\mathcal\{L\}\_\{\\mathrm\{rec\}\}\(i\)\+\\mathcal\{L\}\_\{\\mathrm\{latent\}\}\(i\)\\right\)\.The reconstruction term is

\(24\)ℒrec​\(i\)=CE​\(si,s^i\)\+‖𝐗i−𝐗^i‖F2\+γphys​ℒphys​\(i\),\\mathcal\{L\}\_\{\\mathrm\{rec\}\}\(i\)=\\mathrm\{CE\}\(s\_\{i\},\\hat\{s\}\_\{i\}\)\+\\\|\\mathbf\{X\}\_\{i\}\-\\hat\{\\mathbf\{X\}\}\_\{i\}\\\|\_\{F\}^\{2\}\+\\gamma\_\{\\mathrm\{phys\}\}\\mathcal\{L\}\_\{\\mathrm\{phys\}\}\(i\),wheresis\_\{i\}ands^i\\hat\{s\}\_\{i\}denote the ground\-truth and reconstructed residue types,𝐗i\\mathbf\{X\}\_\{i\}and𝐗^i\\hat\{\\mathbf\{X\}\}\_\{i\}denote the ground\-truth and reconstructed full\-atom coordinates, andℒphys\\mathcal\{L\}\_\{\\mathrm\{phys\}\}enforces valid local peptide geometry, such as bond lengths and bond angles\. The latent regularization termℒlatent\\mathcal\{L\}\_\{\\mathrm\{latent\}\}regularizes the invariant latentzi,0z\_\{i,0\}and the equivariant geometric latentz~i,0\\tilde\{z\}\_\{i,0\}; the detailed autoencoder objective is provided in Appendix[B](https://arxiv.org/html/2606.12991#A2)\.

Stage 2: Latent Diffusion TrainingIn the second stage, the autoencoder is frozen and the denoising networkϵθ\\boldsymbol\{\\epsilon\}\_\{\\theta\}is optimized using the joint objectiveℒtotal\\mathcal\{L\}\_\{\\mathrm\{total\}\}in Sec\.[3\.3](https://arxiv.org/html/2606.12991#S3.SS3)\. This stage uses the DDPM noise\-prediction objective over noisy latents𝒵t\\mathcal\{Z\}\_\{t\}, together with auxiliary topology supervision for the Online Topology Injection module in Sec\.[3\.2](https://arxiv.org/html/2606.12991#S3.SS2)\. Thus, the denoiser learns both latent geometry denoising and timestep\-dependent cyclization topology prediction\.

Inference with Adaptive Balanced GuidanceFor property\-guided generation, we use the target property vector𝐘=\(y1,…,yM\)\\mathbf\{Y\}=\(y\_\{1\},\\ldots,y\_\{M\}\)defined in Sec\.[3\.1](https://arxiv.org/html/2606.12991#S3.SS1)\. At each reverse diffusion steptt, the property surrogateΨψ\\Psi\_\{\\psi\}predicts

\(25\)𝐘^t=Ψψ​\(𝒵t,𝒞,t\),𝐘^t=\(y^1,t,…,y^M,t\)\.\\hat\{\\mathbf\{Y\}\}\_\{t\}=\\Psi\_\{\\psi\}\(\\mathcal\{Z\}\_\{t\},\\mathcal\{C\},t\),\\quad\\hat\{\\mathbf\{Y\}\}\_\{t\}=\(\\hat\{y\}\_\{1,t\},\\ldots,\\hat\{y\}\_\{M,t\}\)\.For each property target, we define a property\-wise energy

\(26\)E^m,t=ℓm​\(y^m,t,ym\),m=1,…,M,\\hat\{E\}\_\{m,t\}=\\ell\_\{m\}\(\\hat\{y\}\_\{m,t\},y\_\{m\}\),\\quad m=1,\\ldots,M,whereℓm\\ell\_\{m\}measures the deviation from the desiredmm\-th property target\.

To balance multiple objectives, we compute the gradient norm of each property\-specific energy:

\(27\)𝐠m,t=∇𝒵tE^m,t,Nm,t=Norm​\(𝐠m,t\)\.\\mathbf\{g\}\_\{m,t\}=\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\hat\{E\}\_\{m,t\},\\quad N\_\{m,t\}=\\mathrm\{Norm\}\(\\mathbf\{g\}\_\{m,t\}\)\.The adaptive weight for themm\-th property is

\(28\)λm​\(t\)=stopgrad​\(wm⋅Agg​\(\{Nj,t\}j=1M\)Nm,t\+εbal\),\\lambda\_\{m\}\(t\)=\\mathrm\{stopgrad\}\\left\(w\_\{m\}\\cdot\\frac\{\\mathrm\{Agg\}\(\\\{N\_\{j,t\}\\\}\_\{j=1\}^\{M\}\)\}\{N\_\{m,t\}\+\\varepsilon\_\{\\mathrm\{bal\}\}\}\\right\),wherewmw\_\{m\}is the base importance weight,Agg​\(⋅\)\\mathrm\{Agg\}\(\\cdot\)is an aggregation operator over gradient norms, andεbal\\varepsilon\_\{\\mathrm\{bal\}\}is a small constant for numerical stability\. The balanced property energy and its guidance gradient are then

\(29\)E^bal​\(𝒵t,t\)=∑m=1Mλm​\(t\)​E^m,t,𝐠prop,t=∇𝒵tE^bal​\(𝒵t,t\)\.\\hat\{E\}\_\{\\mathrm\{bal\}\}\(\\mathcal\{Z\}\_\{t\},t\)=\\sum\_\{m=1\}^\{M\}\\lambda\_\{m\}\(t\)\\hat\{E\}\_\{m,t\},\\quad\\mathbf\{g\}\_\{\\mathrm\{prop\},t\}=\\nabla\_\{\\mathcal\{Z\}\_\{t\}\}\\hat\{E\}\_\{\\mathrm\{bal\}\}\(\\mathcal\{Z\}\_\{t\},t\)\.Following Sec\.[3\.4](https://arxiv.org/html/2606.12991#S3.SS4), we form the guided noise estimate as

\(30\)ϵθ,ψguide=ϵθ\+γt​σt​𝐠prop,t,\\boldsymbol\{\\epsilon\}\_\{\\theta,\\psi\}^\{\\mathrm\{guide\}\}=\\boldsymbol\{\\epsilon\}\_\{\\theta\}\+\\gamma\_\{t\}\\sigma\_\{t\}\\mathbf\{g\}\_\{\\mathrm\{prop\},t\},whereγt\\gamma\_\{t\}is the timestep\-dependent guidance scale\. The guided noise is then used in the DDPM reverse update in Eq\. \([22](https://arxiv.org/html/2606.12991#S3.E22)\)\. The complete inference procedure is summarized in Algorithm[1](https://arxiv.org/html/2606.12991#alg1)\.

Table 1\.Property metrics \(mean over valid samples\)\. Higher is better for Solubility, Protease resistance, and membrane permeability \(Pproxy\); lower is better for Immunogenicity\. Top\-1 and Top\-2 are highlighted byboldandunderline, respectively\.

## 4\.Experiment

### 4\.1\.Experimental Setup

DatasetIn this work, we utilizeCPCore, a high\-quality refined subset of theCPSeadataset\(Yang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib57)\)\. CPSea is a large\-scale collection of 2\.71 million cyclic peptide–protein complexes curated from the AlphaFold Database \(AFDB\)\.CPCoreis constructed through intersectional filtering to ensure biophysical fidelity and essential properties\. To be specific, the dataset consists of71,867unique complexes, a scale suitable for end\-to\-end training of generative models from scratch\. It covers diverse and classical cyclization topologies, includingmainchainamide bonds,disulfidebonds, and side\-chainisopeptide\(specifically Lys\-Asp/Asn\) linkages\. FollowingYang et al\.\([2025](https://arxiv.org/html/2606.12991#bib.bib57)\), we partition the dataset based on FoldSeek structural clusters to prevent data leakage\. Of the 30,819 clusters in CPCore, we randomly reserve 616 clusters \(∼\\sim2%\) for validation and use the remainder for training\. For evaluation, we employ the Large Non\-Redundant \(LNR\) dataset\. After filtering for sequence identity \(<<40% via MMseqs2\) and structural suitability, the final test set comprises 56 unique targets\. Concretely, after the cluster\-based split, the CPCore training set contains 70,284 complexes and the validation set contains 1,583 complexes\.

Property CollectionTo facilitate multi\-objective optimization and ensure therapeutic viability, we developed a comprehensive property evaluation pipeline\. This pipeline integrates a diverse suite of computational tools to characterize the physicochemical, pharmacological, and binding profiles of the cyclic peptides\. The following provides a detailed overview of the specialized property prediction tools, and we provide implementation details for property collection in Appendix[C](https://arxiv.org/html/2606.12991#A3)\.

- •Physicochemical Descriptors: Hydrophobicity is assessed via the Grand Average of Hydropathy \(GRAVY\) using the Kyte\-Doolittle scale\(Kyte and Doolittle,[1982](https://arxiv.org/html/2606.12991#bib.bib29)\)\. Additional rational descriptors, including partition coefficient \(logP\)\(Wildman and Crippen,[1999](https://arxiv.org/html/2606.12991#bib.bib56)\), topological polar surface area \(TPSA\)\(Ertl et al\.,[2000](https://arxiv.org/html/2606.12991#bib.bib16)\), and atom\-normalized TPSA \(rTPSA\)\(Ertl et al\.,[2000](https://arxiv.org/html/2606.12991#bib.bib16)\)are calculated via RDKit\.
- •Solubility: Intrinsic solubility is predicted via the CamSol method\(Sormanni et al\.,[2015](https://arxiv.org/html/2606.12991#bib.bib50)\)\.
- •Protease resistance: Proteolytic stability is assessed via ProsperousPlus\(Li et al\.,[2023](https://arxiv.org/html/2606.12991#bib.bib31)\)to estimate the susceptibility of the cyclic backbone to enzymatic cleavage\.
- •Immunogenicity: Potential binding and presentation to MHC molecules are predicted using BigMHC\(Albert et al\.,[2023](https://arxiv.org/html/2606.12991#bib.bib2)\)\.
- •Binding Affinity: Interface binding energetics are characterized by AutoDock Vina score\(Trott and Olson,[2010](https://arxiv.org/html/2606.12991#bib.bib52)\)and Rosetta interface energy \(rosetta\_dG\)\(Alford et al\.,[2017](https://arxiv.org/html/2606.12991#bib.bib3)\)using the interface\_analyze protocol\.

Baselines\.We compare our method with several representative state\-of\-the\-art peptide generative models\. PepGLAD\(Kong et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib28)\)performs target\-aware full\-atom peptide design using geometric latent diffusion with a variational autoencoder for joint sequence–structure generation\. Building upon the PepGLAD architecture, CP\-Composer\(Jiang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib23)\)enables zero\-shot cyclic peptide generation by decomposing cyclization into composable geometric constraints\. Since CP\-Composer focuses on the transferability from linear to cyclic peptides, we directly employ its original linear peptide generation weights and apply head\-to\-tail cyclization constraints for our comparisons\. Additionally, we evaluate PepFlow\(Li et al\.,[2024b](https://arxiv.org/html/2606.12991#bib.bib33)\), a multi\-modal conditional flow\-matching model that jointly models residue orientation, torsion angles, and sequence for full\-atom co\-design\. To ensure a fair and rigorous comparison, PepGLAD and PepFlow were retrained from scratch on the CPCore dataset using identical data splits and training protocols, while CP\-Composer was implemented in its intended zero\-shot capacity\.

Table 2\.Main results \(mean over valid samples\)\. Top\-1 and Top\-2 are highlighted byboldandunderline, respectively\.MetricsTo evaluate the efficacy of our method and the therapeutic potential of the designed cyclic peptides, we first assess their pharmacological viability and structural rationality\. Following the property evaluation pipeline described previously, we calculate a suite ofdrug\-like properties, including physicochemical descriptors \(LogP, TPSA, GRAVY\), intrinsic solubility, protease resistance and immunogenicity\. We defineAffinityusing the AutoDock Vina score\(Trott and Olson,[2010](https://arxiv.org/html/2606.12991#bib.bib52)\), which measures empirical docking fitness, and the Rosetta interface energy \(d​GdG\)\(Alford et al\.,[2017](https://arxiv.org/html/2606.12991#bib.bib3)\), which quantifies the binding free energy between the peptide and its target\.

Furthermore, we evaluate the generative performance using success rates and ensemble\-level distributional metrics\. A generated sample is consideredCyclization Successif the distance between the cyclization residues falls within the range of\[3\.0,8\.0\]\[3\.0,8\.0\]Å, whileEnergy Successis defined by a negative interface binding energy \(d​G<0dG<0\)\. In addition to these criteria, we adopt the Rosetta total score\(Chaudhury et al\.,[2010](https://arxiv.org/html/2606.12991#bib.bib10)\)as a measure ofStability, reflecting the physical plausibility and energetic favorability of the peptide’s three\-dimensional binding conformation\. To quantifyDiversity, we utilize the co\-diversity metricDc​o=Ds​e​q×Ds​t​r​u​c​tD\_\{co\}=\\sqrt\{D\_\{seq\}\\times D\_\{struct\}\}\(Kong et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib28)\), whereDs​e​qD\_\{seq\}andDs​t​r​u​c​tD\_\{struct\}denote the fractions of clusters over valid samples obtained via single\-linkage hierarchical clustering based on sequence distance \(1−1\-similarity, threshold0\.40\.4\) andCαC\_\{\\alpha\}RMSD \(threshold4\.04\.0Å\), respectively\. The correspondence between sequence and structure spaces is further evaluated byConsistency, measured via Cramér’s V\(Cramér,[1999](https://arxiv.org/html/2606.12991#bib.bib12)\), which captures the correlation between sequence and structural cluster assignments\. The detailed metrics are in Appendix[C](https://arxiv.org/html/2606.12991#A3)\.

### 4\.2\.Main Results

For each receptor in test set, we generate1010candidate peptides\. Samples that pass both the cyclization and energy criteria are regarded asvalid\. For each metric, we first compute the mean over valid samples within the same receptor, and then average across receptors to obtain the final reported performance\.

Since our method enables demand\-driven cyclic peptide design, we report several representative therapeutic design objectives, including affinity\-optimized peptides, protease\-resistant peptides, highly membrane\-permeable peptides, and multi\-property jointly optimized peptides\. We provide the detailed guidance proportion in Appendix[D](https://arxiv.org/html/2606.12991#A4)\. As shown in Table[1](https://arxiv.org/html/2606.12991#S3.T1), our method allows explicit control over multiple key physicochemical objectives, including solubility, protease resistance, membrane permeability, and immunogenicity\. Property\-specific optimization successfully steers generated peptides toward the intended design targets\. Figure[4](https://arxiv.org/html/2606.12991#S3.F4)visualizes these normalized cross\-metric trends, showing that APCyc variants occupy complementary regions of the therapeutic–generative trade\-off space while improving their intended guidance targets\. For instance, the membrane\-permeable setting achieves the highest permeability proxy among all methods while maintaining comparable solubility and immunogenicity, indicating that permeability enhancement does not substantially compromise other safety\-related attributes\. Meanwhile, the multi\-property joint optimization yields the best protease resistance and competitive immunogenicity, demonstrating the capability of our framework to balance trade\-offs across multiple therapeutic constraints rather than optimizing a single metric in isolation\. Protease resistance\- and affinity\-oriented variants further preserve favorable physicochemical profiles, suggesting that targeted optimization remains compatible with overall molecular feasibility\.

Notably, we also observe a mild decrease in solubility under guided optimization\. This trend likely arises from the intrinsic physicochemical coupling between membrane permeability, hydrophobicity, and structural compactness, where promoting permeability oriented features can partially counteract solubility\. Such behavior is consistent with known trade\-offs in peptide drug design and further highlights the necessity of controllable multi\-objective optimization\(Ramelot et al\.,[2023](https://arxiv.org/html/2606.12991#bib.bib43)\)\. Together, these results verify that our framework enables controllable and therapeutically meaningful modulation of peptide properties beyond baseline generative behavior\.

Table 3\.Ablation on protease resistance and immunogenicity\.Beyond property\-level improvements, Table[2](https://arxiv.org/html/2606.12991#S4.T2)shows that the generated peptides also retain strong performance under standard generative evaluation criteria\. Specifically, our guided variants achieve consistently strongStability, outperforming the baselines in most cases, withAffinity\-optimizedbeing the only exception that is slightly weaker than CPComposer\. In addition, theMembrane\-permeablevariant attains the highestConsistencyscore, indicating improved sequence–structure agreement under permeability\-oriented guidance\. Meanwhile, theProtease res\.\-optimizedvariant achieves the best binding affinity among our models, yielding the most favorableVinascore and interface free energy \(d​GdG\)\. Finally, our method maintains competitiveDiversityacross guidance settings, demonstrating that improved stability and affinity do not come at the expense of ensemble\-level variation\. This suggests that enhanced therapeutic property control remains compatible with generative quality and binding effectiveness\. Overall, the proposed framework achieves balanced improvements across both functional therapeutic objectives and conventional generative metrics, supporting its suitability for practical cyclic peptide design\.

### 4\.3\.Target\-level Case Study

Beyond the aggregate benchmark results, we further conduct target\-level case studies to examine whether APCyc can generate cyclic peptides with favorable binding and multi\-property profiles on individual targets\. Specifically, we select two representative targets, 3rc4 and 4xal, and compare APCyc\-generated cyclic peptides with baseline references on the same targets\.

As shown in Table[4](https://arxiv.org/html/2606.12991#S4.T4), APCyc consistently achieves stronger binding scores and more favorable multi\-property profiles than the PepFlow/PepGLAD mean on both targets\. For Vina docking score and immunogenicity, where lower values are preferred, APCyc obtains substantially lower scores\. For membrane permeability and protease resistance, where higher values are preferred, APCyc also achieves better or comparable results\. In particular, APCyc improves membrane permeability from0\.8060\.806to2\.4082\.408on 3rc4, and from0\.9170\.917to3\.0153\.015on 4xal\. These target\-level results suggest that the overall benchmark improvements of APCyc are reflected not only in averaged metrics, but also in representative individual cases\.

Table 4\.Target\-level case study on two representative targets\. Baseline mean denotes the PepFlow/PepGLAD mean\.
### 4\.4\.Ablation

To evaluate the contribution of the proposed guidance mechanism, we conduct an ablation study by comparing the full APCyc model with a variant where the guidance component is removed\. We further analyze the impact of guidance on key therapeutic properties through targeted ablations\. As shown in Table[3](https://arxiv.org/html/2606.12991#S4.T3), guided optimization achieves competitive or improved protease resistance and immunogenicity compared with the unguided baseline, demonstrating controllable regulation of stability\- and safety\-related properties\. Meanwhile, Table[5](https://arxiv.org/html/2606.12991#S4.T5)shows that the membrane\-permeable variant substantially increases the permeability proxy relative to the baseline, confirming the effectiveness of guidance in steering permeability\-related behavior\. Overall, these results highlight that explicit guidance is essential for directing cyclic peptide generation toward desired physicochemical objectives\.

Table 5\.Ablation on membrane permeability proxy\. Higher values indicate better permeability\.

## 5\.Limitations and Ethical Considerations

Data and supervision constraintsAPCyc is trained on CPCore\(Yang et al\.,[2025](https://arxiv.org/html/2606.12991#bib.bib57)\), a curated CPSea subset derived from large\-scale structure prediction resources\. Although filtering and clustering improve data quality, the complexes may still inherit prediction noise, imperfect pocket geometries, and target\-dependent biases\. In addition, CPCore mainly covers classical cyclization chemistries, including backbone amide, disulfide, and isopeptide linkages\. Generalization to alternative crosslinkers, bicyclic topologies, or non\-canonical stapling chemistries may therefore require additional structural data and broader chemical annotations\.

Intended use, dual\-use, and uncertaintyAPCyc is intended to support benign therapeutic discovery by proposing cyclic peptide candidates for expert review and experimental validation, rather than serving as a source of clinical recommendations\. Because generative models for bioactive molecule design may be misused to search for harmful activity\(Urbina et al\.,[2022](https://arxiv.org/html/2606.12991#bib.bib53)\), practical deployment should remain within controlled research settings, follow biomedical governance and biosafety procedures, and include risk assessment, domain\-expert oversight, and safety screening before synthesis\(Topol,[2019](https://arxiv.org/html/2606.12991#bib.bib51); Schneider et al\.,[2020](https://arxiv.org/html/2606.12991#bib.bib48); Pannu et al\.,[2024](https://arxiv.org/html/2606.12991#bib.bib41)\)\. In addition, safety\-related objectives such as immunogenicity rely on computational predictors and may be uncertain or miscalibrated\(Begoli et al\.,[2019](https://arxiv.org/html/2606.12991#bib.bib7)\)\. These predictions should therefore be treated as preliminary signals and complemented with additional models, expert assessment, experimental validation, and uncertainty reporting where possible\(Amrhein et al\.,[2019](https://arxiv.org/html/2606.12991#bib.bib4)\)\.

## 6\.Conclusions

We present APCyc, a target\-aware latent diffusion framework that enables automated cyclization and property\-informed design for cyclic peptides\. APCyc dynamically infers pocket\-adaptive linkage types and sites conditioned on the receptor context, and simultaneously optimizes multiple drug\-like properties using the proposed Bayesian Posterior Guidance\. As a result, APCyc addresses the critical challenge of controllable cyclic peptide generation under complex property constraints, providing a robust foundation and new insights for the design of next\-generation therapeutic peptides\.

###### Acknowledgements\.

This work was supported by the Guangdong Basic and Applied Basic Research Foundation \(2026A1515011793\), and the Youth S&T Talent Support Programme of Guangdong Provincial Association for Science and Technology \(SKXRC2025467\)\.

## GenAI Disclosure

Parts of the manuscript preparation, including language refinement and editing, were assisted by generative AI tools\. All scientific content, experimental design, implementation, and conclusions were developed and verified solely by the authors\.

## References

- \(1\)
- Albert et al\.\(2023\)Benjamin Alexander Albert, Yunxiao Yang, Xiaoshan M Shao, Dipika Singh, Kellie N Smith, Valsamo Anagnostou, and Rachel Karchin\. 2023\.Deep neural networks predict class I major histocompatibility complex epitope presentation and transfer learn neoepitope immunogenicity\.*Nature machine intelligence*5, 8 \(2023\), 861–872\.
- Alford et al\.\(2017\)Rebecca F Alford, Andrew Leaver\-Fay, Jeliazko R Jeliazkov, Matthew J O’Meara, Frank P DiMaio, Hahnbeom Park, Maxim V Shapovalov, P Douglas Renfrew, Vikram K Mulligan, Kalli Kappel, et al\.2017\.The Rosetta all\-atom energy function for macromolecular modeling and design\.*Journal of chemical theory and computation*13, 6 \(2017\), 3031–3048\.
- Amrhein et al\.\(2019\)Valentin Amrhein, Sander Greenland, and Blake McShane\. 2019\.Scientists rise up against statistical significance\.*Nature*567, 7748 \(2019\), 305–307\.
- Bansal et al\.\(2023\)Arpit Bansal, Hong\-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein\. 2023\.Universal guidance for diffusion models\. In*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*\. 843–852\.
- Bao et al\.\(2022\)Fan Bao, Min Zhao, Zhongkai Hao, Peiyao Li, Chongxuan Li, and Jun Zhu\. 2022\.Equivariant energy\-guided sde for inverse molecular design\.*arXiv preprint arXiv:2209\.15408*\(2022\)\.
- Begoli et al\.\(2019\)Edmon Begoli, Tanmoy Bhattacharya, and Dimitri Kusnezov\. 2019\.The need for uncertainty quantification in machine\-assisted medical decision making\.*Nature Machine Intelligence*1, 1 \(2019\), 20–23\.
- Bhat et al\.\(2025\)Suhaas Bhat, Kalyan Palepu, Lauren Hong, Joey Mao, Tianzheng Ye, Rema Iyer, Lin Zhao, Tianlai Chen, Sophia Vincoff, Rio Watson, et al\.2025\.De novo design of peptide binders to conformationally diverse targets with contrastive language modeling\.*Science Advances*11, 4 \(2025\), eadr8638\.
- Chandramohan et al\.\(2024\)Arun Chandramohan, Hubert Josien, Tsz Ying Yuen, Ruchia Duggal, Diana Spiegelberg, Lin Yan, Yu\-Chi Angela Juang, Lan Ge, Pietro G Aronica, Hung Yi Kristal Kaan, et al\.2024\.Design\-rules for stapled peptides with in vivo activity and their application to Mdm2/X antagonists\.*Nature communications*15, 1 \(2024\), 489\.
- Chaudhury et al\.\(2010\)Sidhartha Chaudhury, Sergey Lyskov, and Jeffrey J Gray\. 2010\.PyRosetta: a script\-based interface for implementing molecular modeling algorithms using Rosetta\.*Bioinformatics*26, 5 \(2010\), 689–691\.
- Chung et al\.\(2022\)Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye\. 2022\.Diffusion posterior sampling for general noisy inverse problems\.*arXiv preprint arXiv:2209\.14687*\(2022\)\.
- Cramér \(1999\)Harald Cramér\. 1999\.*Mathematical methods of statistics*\. Vol\. 9\.Princeton university press\.
- Deyle et al\.\(2017\)Kaycie Deyle, Xu\-Dong Kong, and Christian Heinis\. 2017\.Phage selection of cyclic peptides for application in research and drug development\.*Accounts of chemical research*50, 8 \(2017\), 1866–1874\.
- Dhariwal and Nichol \(2021\)Prafulla Dhariwal and Alexander Nichol\. 2021\.Diffusion models beat gans on image synthesis\.*Advances in neural information processing systems*34 \(2021\), 8780–8794\.
- Driggers et al\.\(2008\)Edward M Driggers, Stephen P Hale, Jinbo Lee, and Nicholas K Terrett\. 2008\.The exploration of macrocycles for drug discovery—an underexploited structural class\.*Nature Reviews Drug Discovery*7, 7 \(2008\), 608–624\.
- Ertl et al\.\(2000\)Peter Ertl, Bernhard Rohde, and Paul Selzer\. 2000\.Fast calculation of molecular polar surface area as a sum of fragment\-based contributions and its application to the prediction of drug transport properties\.*Journal of medicinal chemistry*43, 20 \(2000\), 3714–3717\.
- Fosgerau and Hoffmann \(2015\)Keld Fosgerau and Torsten Hoffmann\. 2015\.Peptide therapeutics: current status and future directions\.*Drug discovery today*20, 1 \(2015\), 122–128\.
- Garcia Jimenez et al\.\(2023\)Diego Garcia Jimenez, Vasanthanathan Poongavanam, and Jan Kihlberg\. 2023\.Macrocycles in drug discovery—learning from the past for the future\.*Journal of medicinal chemistry*66, 8 \(2023\), 5377–5396\.
- Guo et al\.\(2015\)Zuojun Guo, Bo Li, Li\-Tien Cheng, Shenggao Zhou, J Andrew McCammon, and Jianwei Che\. 2015\.Identification of protein–ligand binding sites by the level\-set variational implicit\-solvent approach\.*Journal of Chemical Theory and Computation*11, 2 \(2015\), 753–765\.
- Heinis et al\.\(2009\)Christian Heinis, Trevor Rutherford, Stephan Freund, and Greg Winter\. 2009\.Phage\-encoded combinatorial chemical libraries based on bicyclic peptides\.*Nature chemical biology*5, 7 \(2009\), 502–507\.
- Ho and Salimans \(2022\)Jonathan Ho and Tim Salimans\. 2022\.Classifier\-free diffusion guidance\.*arXiv preprint arXiv:2207\.12598*\(2022\)\.
- Ji et al\.\(2024\)Xinjian Ji, Alexander L Nielsen, and Christian Heinis\. 2024\.Cyclic peptides for drug development\.*Angewandte Chemie*136, 3 \(2024\), e202308251\.
- Jiang et al\.\(2025\)Dapeng Jiang, Xiangzhe Kong, Jiaqi Han, Rui Jiao, Wenbing Huang, Stefano Ermon, Jianzhu Ma, and Yang Liu\. 2025\.Zero\-Shot Cyclic Peptide Design via Composable Geometric Constraints\. In*Proceedings of the 42nd International Conference on Machine Learning*\.
- Jumper et al\.\(2021\)John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al\.2021\.Highly accurate protein structure prediction with AlphaFold\.*nature*596, 7873 \(2021\), 583–589\.
- Keeling et al\.\(2016\)Kelly L Keeling, Okki Cho, Denis B Scanlon, Grant W Booker, Andrew D Abell, and Kate L Wegener\. 2016\.The key position: influence of staple location on constrained peptide conformation and binding\.*Organic & biomolecular chemistry*14, 41 \(2016\), 9731–9735\.
- Kim et al\.\(2025\)Taegwan Kim, Eunbee Baek, and Jonghoon Kim\. 2025\.Exploring Macrocyclic Chemical Space: Strategies and Technologies for Drug Discovery\.*Pharmaceuticals*18, 5 \(2025\), 617\.
- Kong et al\.\(2023\)Xiangzhe Kong, Wenbing Huang, and Yang Liu\. 2023\.End\-to\-end full\-atom antibody design\.*arXiv preprint arXiv:2302\.00203*\(2023\)\.
- Kong et al\.\(2024\)Xiangzhe Kong, Yinjun Jia, Wenbing Huang, and Yang Liu\. 2024\.Full\-atom peptide design with geometric latent diffusion\.*Advances in Neural Information Processing Systems*37 \(2024\), 74808–74839\.
- Kyte and Doolittle \(1982\)Jack Kyte and Russell F Doolittle\. 1982\.A simple method for displaying the hydropathic character of a protein\.*Journal of molecular biology*157, 1 \(1982\), 105–132\.
- Lau and Dunn \(2018\)Jolene L Lau and Michael K Dunn\. 2018\.Therapeutic peptides: Historical perspectives, current development trends, and future directions\.*Bioorganic & medicinal chemistry*26, 10 \(2018\), 2700–2707\.
- Li et al\.\(2023\)Fuyi Li, Cong Wang, Xudong Guo, Tatsuya Akutsu, Geoffrey I Webb, Lachlan JM Coin, Lukasz Kurgan, and Jiangning Song\. 2023\.ProsperousPlus: a one\-stop and comprehensive platform for accurate protease\-specific substrate cleavage prediction and machine\-learning model construction\.*Briefings in Bioinformatics*24, 6 \(2023\), bbad372\.
- Li et al\.\(2024a\)Jiahan Li, Tong Chen, Shitong Luo, Chaoran Cheng, Jiaqi Guan, Ruihan Guo, Sheng Wang, Ge Liu, Jian Peng, and Jianzhu Ma\. 2024a\.Hotspot\-driven peptide design via multi\-fragment autoregressive extension\.*arXiv preprint arXiv:2411\.18463*\(2024\)\.
- Li et al\.\(2024b\)Jiahan Li, Chaoran Cheng, Zuofan Wu, Ruihan Guo, Shitong Luo, Zhizhou Ren, Jian Peng, and Jianzhu Ma\. 2024b\.Full\-atom peptide design based on multi\-modal flow matching\.*arXiv preprint arXiv:2406\.00735*\(2024\)\.
- Li et al\.\(2024c\)Jianan Li, Keisuke Yanagisawa, and Yutaka Akiyama\. 2024c\.CycPeptMP: enhancing membrane permeability prediction of cyclic peptides with multi\-level molecular features and data augmentation\.*Briefings in Bioinformatics*25, 5 \(2024\)\.
- Li et al\.\(2022\)Xinting Li, Timothy W Craven, and Paul M Levine\. 2022\.Cyclic peptide screening methods for preclinical drug discovery: miniperspective\.*Journal of Medicinal Chemistry*65, 18 \(2022\), 11913–11926\.
- Lin et al\.\(2024\)Haitao Lin, Odin Zhang, Huifeng Zhao, Dejun Jiang, Lirong Wu, Zicheng Liu, Yufei Huang, and Stan Z Li\. 2024\.Ppflow: Target\-aware peptide design with torsional flow matching\.*arXiv preprint arXiv:2405\.06642*\(2024\)\.
- Liu et al\.\(2024\)Lei Liu, Liu Yang, Suqi Cao, Zhigang Gao, Bin Yang, Guoqing Zhang, Ruixin Zhu, and Dingfeng Wu\. 2024\.CyclicPepedia: a knowledge base of natural and synthetic cyclic peptides\.*Briefings in Bioinformatics*25, 3 \(2024\), bbae190\.
- Newman and Cragg \(2016\)David J Newman and Gordon M Cragg\. 2016\.Natural products as sources of new drugs from 1981 to 2014\.*Journal of natural products*79, 3 \(2016\), 629–661\.
- Oh et al\.\(2025\)Doogie Oh, Yongdoo Park, and Jongseong Kim\. 2025\.De novo generation of peptide binders with desired properties by deep generative models reinforced through enrichment of focused sets for iterative fine\-tuning\.*Chemical Communications*\(2025\)\.
- Otvos Jr and Wade \(2014\)Laszlo Otvos Jr and John D Wade\. 2014\.Current challenges in peptide\-based drug discovery\.62 pages\.
- Pannu et al\.\(2024\)Jaspreet Pannu, Doni Bloomfield, Alex Zhu, Robert MacKnight, Gabe Gomes, Anita Cicero, and Thomas V Inglesby\. 2024\.Prioritizing high\-consequence biological capabilities in evaluations of artificial intelligence models\.*arXiv preprint arXiv:2407\.13059*\(2024\)\.
- Qiu et al\.\(2024\)Keyue Qiu, Yuxuan Song, Jie Yu, Hongbo Ma, Ziyao Cao, Zhilong Zhang, Yushuai Wu, Mingyue Zheng, Hao Zhou, and Wei\-Ying Ma\. 2024\.Empower Structure\-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks\.*arXiv preprint arXiv:2411\.13280*\(2024\)\.
- Ramelot et al\.\(2023\)Theresa A Ramelot, Jonathan Palmer, Gaetano T Montelione, and Gaurav Bhardwaj\. 2023\.Cell\-permeable chameleonic peptides: Exploiting conformational dynamics in de novo cyclic peptide design\.*Current opinion in structural biology*80 \(2023\), 102603\.
- Rettie et al\.\(2025a\)Stephen A Rettie, Katelyn V Campbell, Asim K Bera, Alex Kang, Simon Kozlov, Yensi Flores Bueso, Joshmyn De La Cruz, Maggie Ahlrichs, Suna Cheng, Stacey R Gerben, et al\.2025a\.Cyclic peptide structure prediction and design using AlphaFold2\.*Nature Communications*16, 1 \(2025\), 4730\.
- Rettie et al\.\(2025b\)Stephen A Rettie, David Juergens, Victor Adebomi, Yensi Flores Bueso, Qinqin Zhao, Alexandria N Leveille, Andi Liu, Asim K Bera, Joana A Wilms, Alina Üffing, et al\.2025b\.Accurate de novo design of high\-affinity protein\-binding macrocycles using deep learning\.*Nature Chemical Biology*\(2025\), 1–9\.
- Satorras et al\.\(2021\)Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling\. 2021\.E \(n\) equivariant graph neural networks\. In*International conference on machine learning*\. PMLR, 9323–9332\.
- Sayuti et al\.\(2025\)Tiara Natasha Binte Sayuti, Kakuly Mittal, Tan Lai Heng, and Jagath C Rajapakse\. 2025\.Target\-aware latent diffusion model for design of apoptosis\-inducing anticancer peptides\.*Computers in Biology and Medicine*199 \(2025\), 111286\.
- Schneider et al\.\(2020\)Petra Schneider, W Patrick Walters, Alleyn T Plowright, Norman Sieroka, Jennifer Listgarten, Robert A Goodnow Jr, Jasmin Fisher, Johanna M Jansen, José S Duca, Thomas S Rush, et al\.2020\.Rethinking drug design in the artificial intelligence era\.*Nature reviews drug discovery*19, 5 \(2020\), 353–364\.
- Schütt et al\.\(2018\)Kristof T Schütt, Huziel E Sauceda, P\-J Kindermans, Alexandre Tkatchenko, and K\-R Müller\. 2018\.Schnet–a deep learning architecture for molecules and materials\.*The Journal of chemical physics*148, 24 \(2018\)\.
- Sormanni et al\.\(2015\)Pietro Sormanni, Francesco A Aprile, and Michele Vendruscolo\. 2015\.The CamSol method of rational design of protein mutants with enhanced solubility\.*Journal of molecular biology*427, 2 \(2015\), 478–490\.
- Topol \(2019\)Eric J Topol\. 2019\.High\-performance medicine: the convergence of human and artificial intelligence\.*Nature medicine*25, 1 \(2019\), 44–56\.
- Trott and Olson \(2010\)Oleg Trott and Arthur J Olson\. 2010\.AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading\.*Journal of computational chemistry*31, 2 \(2010\), 455–461\.
- Urbina et al\.\(2022\)Fabio Urbina, Filippa Lentzos, Cédric Invernizzi, and Sean Ekins\. 2022\.Dual use of artificial\-intelligence\-powered drug discovery\.*Nature machine intelligence*4, 3 \(2022\), 189–191\.
- Vlieghe et al\.\(2010\)Patrick Vlieghe, Vincent Lisowski, Jean Martinez, and Michel Khrestchatisky\. 2010\.Synthetic therapeutic peptides: science and market\.*Drug discovery today*15, 1\-2 \(2010\), 40–56\.
- Wang et al\.\(2024\)Fanhao Wang, Yuzhe Wang, Laiyi Feng, Changsheng Zhang, and Luhua Lai\. 2024\.Target\-specific de novo peptide binder design with DiffPepBuilder\.*Journal of Chemical Information and Modeling*64, 24 \(2024\), 9135–9149\.
- Wildman and Crippen \(1999\)Scott A Wildman and Gordon M Crippen\. 1999\.Prediction of physicochemical parameters by atomic contributions\.*Journal of chemical information and computer sciences*39, 5 \(1999\), 868–873\.
- Yang et al\.\(2025\)Ziyi Yang, Hanyuan Xie, Yinjun Jia, Xiangzhe Kong, Jiqing Zheng, Ziting Zhang, Yang Liu, Lei Liu, and Yanyan Lan\. 2025\.CPSea: Large\-scale cyclic peptide–protein complex dataset for machine learning in cyclic peptide design\. In*Advances in Neural Information Processing Systems*\.Datasets and Benchmarks Track\.
- Zhou et al\.\(2025\)Xiaoyu Zhou, Meng Li, Yifan Xiao, Jing Li, Dong Xue, Zhen Zheng, Jian Ma, and Quanquan Gu\. 2025\.Designing Cyclic Peptides via Harmonic Stochastic Differential Equations with Atom–Bond Modeling\. In*Proceedings of the 42nd International Conference on Machine Learning*\.
- Zorzi et al\.\(2017\)Alessandro Zorzi, Kaycie Deyle, and Christian Heinis\. 2017\.Cyclic peptide therapeutics: past, present and future\.*Current opinion in chemical biology*38 \(2017\), 24–29\.

## Appendix AProof of[Proposition3\.1](https://arxiv.org/html/2606.12991#S3.Thmtheorem1)

###### Proof\.

Letg∈SE​\(3\)g\\in\\mathrm\{SE\}\(3\)denote a rigid transformation acting on equivariant coordinates asg​\(𝐱\)=𝐑𝐱\+𝐭g\(\\mathbf\{x\}\)=\\mathbf\{R\}\\mathbf\{x\}\+\\mathbf\{t\}with𝐑∈SO​\(3\)\\mathbf\{R\}\\in\\mathrm\{SO\}\(3\)\. Invariant latent features remain unchanged\.

#### Invariant topology prediction

The pair representation in \([2](https://arxiv.org/html/2606.12991#S3.E2)\) depends only on invariant node latents and Euclidean distances, which are preserved under rigid motions\. Therefore the predicted linkage matrix𝐒^t\\hat\{\\mathbf\{S\}\}\_\{t\}obtained via the masked Softmax in \([4](https://arxiv.org/html/2606.12991#S3.E4)\) is invariant:

𝐒^t​\(g⋅𝒵t\)=𝐒^t​\(𝒵t\)\.\\hat\{\\mathbf\{S\}\}\_\{t\}\(g\\cdot\\mathcal\{Z\}\_\{t\}\)=\\hat\{\\mathbf\{S\}\}\_\{t\}\(\\mathcal\{Z\}\_\{t\}\)\.

#### Invariant Topology Injection

The injected edge features, bias, and gating terms defined in Eqs\. \([8](https://arxiv.org/html/2606.12991#S3.E8)\)–\([10](https://arxiv.org/html/2606.12991#S3.E10)\) depend only on invariant distances,𝐒^t\\hat\{\\mathbf\{S\}\}\_\{t\}, and scalar schedules\. Hence all modulation coefficients remain invariant under SE\(3\) transformations\.

#### Equivariance of coordinate updates

The AM\-EGNN coordinate update takes the standard form

z~i,t′=z~i,t\+∑jαi​j,t​\(z~i,t−z~j,t\),\\tilde\{z\}^\{\\prime\}\_\{i,t\}=\\tilde\{z\}\_\{i,t\}\+\\sum\_\{j\}\\alpha\_\{ij,t\}\\,\(\\tilde\{z\}\_\{i,t\}\-\\tilde\{z\}\_\{j,t\}\),where the scalar weightsαi​j,t\\alpha\_\{ij,t\}depend only on invariant quantities\. Applyingggto both sides yields

z~i,t′​\(g⋅𝒵t\)=g⋅z~i,t′​\(𝒵t\),\\tilde\{z\}^\{\\prime\}\_\{i,t\}\(g\\cdot\\mathcal\{Z\}\_\{t\}\)=g\\cdot\\tilde\{z\}^\{\\prime\}\_\{i,t\}\(\\mathcal\{Z\}\_\{t\}\),which proves SE\(3\)\-equivariance\.

#### Conclusion

Since topology prediction is invariant and coordinate updates are equivariant, the overall topology injection mechanism preserves SE\(3\)\-equivariance\. ∎

## Appendix BVariational Autoencoder Training Loss

FollowingKong et al\.\([2024](https://arxiv.org/html/2606.12991#bib.bib28)\), we train a joint sequence–structure variational autoencoder that maps a peptide𝒫=\{\(si,𝐗i\)\}i=1L\\mathcal\{P\}=\\\{\(s\_\{i\},\\mathbf\{X\}\_\{i\}\)\\\}\_\{i=1\}^\{L\}to residue\-level latents𝒵=\{\(zi,z~i\)\}i=1L\\mathcal\{Z\}=\\\{\(z\_\{i\},\\tilde\{z\}\_\{i\}\)\\\}\_\{i=1\}^\{L\}, whereziz\_\{i\}isE​\(3\)\\mathrm\{E\}\(3\)\-invariant andz~i\\tilde\{z\}\_\{i\}isE​\(3\)\\mathrm\{E\}\(3\)\-equivariant\. The decoder reconstructss^i\\hat\{s\}\_\{i\}and𝐗^i\\hat\{\\mathbf\{X\}\}\_\{i\}\.

#### Overall objective\.

Following Sec\.[3\.5](https://arxiv.org/html/2606.12991#S3.SS5), the autoencoder is optimized by

\(31\)ℒAE=1L​∑i=1L\(ℒrec​\(i\)\+ℒKL​\(i\)\)\.\\mathcal\{L\}\_\{\\text\{AE\}\}=\\frac\{1\}\{L\}\\sum\_\{i=1\}^\{L\}\\big\(\\mathcal\{L\}\_\{\\text\{rec\}\}\(i\)\+\\mathcal\{L\}\_\{\\text\{KL\}\}\(i\)\\big\)\.

#### Reconstruction loss\.

We use cross\-entropy for residue tokens and MSE for full\-atom coordinates, together with an auxiliary physical regularization:

\(32\)ℒrec​\(i\)=H​\(si,s^i\)\+‖𝐗i−𝐗^i‖22\+γ​ℒphys​\(i\)\.\\mathcal\{L\}\_\{\\text\{rec\}\}\(i\)=H\(s\_\{i\},\\hat\{s\}\_\{i\}\)\+\\\|\\mathbf\{X\}\_\{i\}\-\\hat\{\\mathbf\{X\}\}\_\{i\}\\\|\_\{2\}^\{2\}\+\\gamma\\,\\mathcal\{L\}\_\{\\text\{phys\}\}\(i\)\.

#### Auxiliary physical loss\.

To better preserve realistic geometry,ℒphys\\mathcal\{L\}\_\{\\text\{phys\}\}supervisesCα\\mathrm\{C\}\_\{\\alpha\}coordinates, bond lengths, and side\-chain dihedral angles:

\(33\)ℒphys​\(i\)=λCA​ℒCA​\(i\)\+λbond​ℒbond​\(i\)\+λangle​ℒangle​\(i\)\.\\mathcal\{L\}\_\{\\text\{phys\}\}\(i\)=\\lambda\_\{\\mathrm\{CA\}\}\\mathcal\{L\}\_\{\\mathrm\{CA\}\}\(i\)\+\\lambda\_\{\\mathrm\{bond\}\}\\mathcal\{L\}\_\{\\mathrm\{bond\}\}\(i\)\+\\lambda\_\{\\mathrm\{angle\}\}\\mathcal\{L\}\_\{\\mathrm\{angle\}\}\(i\)\.

#### Latent regularization\.

The invariant latentziz\_\{i\}is regularized toward a standard Gaussian, while the equivariant latentz~i\\tilde\{z\}\_\{i\}is anchored to the referenceCα\\mathrm\{C\}\_\{\\alpha\}coordinate𝐫i\\mathbf\{r\}\_\{i\}:

\(34\)ℒKL​\(i\)=DKL​\(qϕ​\(zi\)∥𝒩​\(𝟎,𝐈\)\)\+η​‖z~i−𝐫i‖22\.\\mathcal\{L\}\_\{\\text\{KL\}\}\(i\)=D\_\{\\mathrm\{KL\}\}\\\!\\left\(q\_\{\\phi\}\(z\_\{i\}\)\\,\\\|\\,\\mathcal\{N\}\(\\mathbf\{0\},\\mathbf\{I\}\)\\right\)\+\\eta\\,\\\|\\tilde\{z\}\_\{i\}\-\\mathbf\{r\}\_\{i\}\\\|\_\{2\}^\{2\}\.

#### Default weights\.

We setλCA=1\.0\\lambda\_\{\\mathrm\{CA\}\}=1\.0,λbond=1\.0\\lambda\_\{\\mathrm\{bond\}\}=1\.0, andλangle=0\.5\\lambda\_\{\\mathrm\{angle\}\}=0\.5in experiments, and use a smallη\\etato stabilize equivariant latent learning\.

## Appendix CMetrics Implementation

### C\.1\.Inputs and identifiers

Each candidate is provided as a complex PDB \(optionally gzipped\)\. We assume a two\-chain convention in the complex: the peptide is chainLand the receptor/target protein is chainR\. We use the PDB filename stem as the peptide identifierid\. All sequence\-based predictors operate on the peptide\-chain sequence extracted from chainL; structure\-based descriptors operate on the chain\-Lstructure; affinity evaluates theL–Rinterface\.

### C\.2\.Solubility \(CamSol intrinsic\)

To assess solubility, we employ the CamSol\-intrinsic predictor on the peptide sequence and record the returned scalarcamsol\_score\(higher indicates better solubility\):

\(35\)Ssol:=camsol\_score\.S\_\{\\mathrm\{sol\}\}:=\\texttt\{camsol\\\_score\}\.

### C\.3\.Enzymatic stability / proteolysis susceptibility \(ProsperousPlus\)

To quantify proteolysis susceptibility, we employ ProsperousPlus in prediction mode to estimate cleavage propensity under a fixed panel of proteases

\(36\)𝒫=\{S01\.001,A01\.001,C01\.060,M10\.003,M10\.008\}\.\\mathcal\{P\}=\\\{\\texttt\{S01\.001\},\\texttt\{A01\.001\},\\texttt\{C01\.060\},\\texttt\{M10\.003\},\\texttt\{M10\.008\}\\\}\.Because ProsperousPlus expects a fixed\-length context, we convert each peptide sequence into a set of 8\-mer windows\. Specifically, after uppercasing and replacing non\-canonical characters with\-, we generate length\-8 windows with stride 2, ensuring the C\-terminal window is included; if the number of windows exceeds3030, we subsample evenly to at most3030windows\. For each window set, ProsperousPlus outputs per\-position cleavage probabilities \(columnpro\) for each protease\. We aggregate these raw probabilities by averaging within each protease and then summing across proteases:

\(37\)proteolysis\_sumscore:=∑p∈𝒫\(1Np​∑k=1Npprop,k\),\\texttt\{proteolysis\\\_sumscore\}~:=~\\sum\_\{p\\in\\mathcal\{P\}\}\\left\(\\frac\{1\}\{N\_\{p\}\}\\sum\_\{k=1\}^\{N\_\{p\}\}\\texttt\{pro\}\_\{p,k\}\\right\),whereNpN\_\{p\}is the number of probability entries produced for proteaseppacross all windows/positions\. Smaller values imply lower cleavage propensity \(i\.e\., higher stability\), and we therefore define a higher\-is\-better stability score as

\(38\)Sprot:=−proteolysis\_sumscore\.S\_\{\\mathrm\{prot\}\}:=\-\\,\\texttt\{proteolysis\\\_sumscore\}\.

### C\.4\.Immunogenicity proxy \(BigMHC EL/IM\)

To assess immunogenicity, we run BigMHC on a specified MHC\-I allele \(default:HLA\-A\*02:01\) and obtain two outputs per peptide:el\(presentation likelihood\) andim\(immunogenicity likelihood\)\. Peptides of length 8–11 are scored directly\. For peptides longer than 11 residues, we enumerate sliding windows of lengths8,9,10,118,9,10,11, score each window, and average the top\-KKwindow scores \(we useK=3K\{=\}3\) for EL and IM separately\. We report the raw immunogenicity log\-score, where lower values indicate lower predicted immunogenicity, and use its sign\-flipped form for internal surrogate training and guidance:

\(39\)Iimm\\displaystyle I\_\{\\mathrm\{imm\}\}:=log⁡\(max⁡\(el,ϵ\)\)\+log⁡\(max⁡\(im,ϵ\)\),\\displaystyle:=\\log\(\\max\(\\texttt\{el\},\\epsilon\)\)\+\\log\(\\max\(\\texttt\{im\},\\epsilon\)\),\(40\)Simm\\displaystyle S\_\{\\mathrm\{imm\}\}:=−Iimm\.\\displaystyle:=\-I\_\{\\mathrm\{imm\}\}\.Thus all guidance objectives follow a higher\-is\-better convention\.

### C\.5\.Physicochemical descriptors from peptide\-chain structure

To obtain structure\-derived physicochemical descriptors, we analyze the peptide chainLextracted from the complex PDB\. We compute: \(i\)GRAVYas the average Kyte–Doolittle hydropathy over the extracted peptide sequence; \(ii\)TPSAandlogPusing RDKit descriptors computed from the peptide\-chain PDB block; \(iii\) the heavy\-atom\-normalized polar surface area

\(41\)rTPSA:=TPSA\#​heavy atoms\.\\texttt\{rTPSA\}:=\\frac\{\\texttt\{TPSA\}\}\{\\\#\\text\{heavy atoms\}\}\.

### C\.6\.Permeability proxy score

We consolidate the above descriptors into a single permeability proxy\. Since the raw descriptors have different scales, we first normalize each descriptor by dataset\-levelzz\-scoring:

\(42\)z​\(x\)=x−μ​\(x\)σ​\(x\),z\(x\)=\\frac\{x\-\\mu\(x\)\}\{\\sigma\(x\)\},whereμ​\(⋅\)\\mu\(\\cdot\)andσ​\(⋅\)\\sigma\(\\cdot\)are computed over the dataset \(ignoring missing values; using population standard deviation\)\. We then define

\(43\)Pproxy:=z​\(logP\)\+z​\(GRAVY\)−z​\(TPSA\)−z​\(rTPSA\),P\_\{\\mathrm\{proxy\}\}:=z\(\\texttt\{logP\}\)\+z\(\\texttt\{GRAVY\}\)\-z\(\\texttt\{TPSA\}\)\-z\(\\texttt\{rTPSA\}\),which corresponds to higher hydrophobicity and lower polar surface area being more favorable\.

### C\.7\.Binding affinity: Vina score\-only evaluation

Binding affinity is estimated by evaluating the*given*complex pose without redocking\. ChainsLandRare extracted, converted to PDBQT via MGLTools\. AutoDock Vina is executed in\-\-score\_onlymode:

vina \-\-receptor R\.pdbqt \-\-ligand L\.pdbqt \-\-center\_\* c\_\* \-\-size\_\* s\_\* \-\-score\_only

The scoring box is defined from ligand bounds with paddingpp\(default1010Å\):

\(44\)cx=xmin\+xmax2,sx=max⁡\(\(xmax−xmin\)\+p,1\.0\),c\_\{x\}=\\tfrac\{x\_\{\\min\}\+x\_\{\\max\}\}\{2\},\\quad s\_\{x\}=\\max\\\!\\big\(\(x\_\{\\max\}\-x\_\{\\min\}\)\+p,\\,1\.0\\big\),and analogously fory,zy,z\. The reported affinity \(kcal/mol\) is recorded asvina\_score\.

### C\.8\.Binding affinity: RosettaScripts interface

In addition to Vina, we compute a Rosetta interface energy proxy with RosettaScripts\. Unlike the example FastRelax workflow, we do not perform additional relaxation in this pipeline; instead, we directly evaluate the input complex using theref2015scorefunction and a minimal protocol \(interface\_analyze\.xml\) that applies a singleDdgfilter withjump=1,repeats=1,repack=0\(no repacking/minimization\), and scorefunctionref2015\. We run rosetta\_scripts\.\* in score\-only mode and extract the scalardgreported by the filter asrosetta\_dG\.

### C\.9\.Affinity normalization and combined score

Because Vina and Rosetta energies have different scales, we convert each to a higher\-is\-better normalized term viazz\-scoring and sign flip\. We then define

\(45\)Aros\\displaystyle A\_\{\\mathrm\{ros\}\}:=−z​\(rosetta\_dG\),Avina:=−z​\(vina\_score\),\\displaystyle:=\-z\(\\texttt\{rosetta\\\_dG\}\),\\qquad A\_\{\\mathrm\{vina\}\}:=\-z\(\\texttt\{vina\\\_score\}\),\(46\)Aaff\\displaystyle A\_\{\\mathrm\{aff\}\}:=12​Aros\+12​Avina,\\displaystyle:=\\tfrac\{1\}\{2\}A\_\{\\mathrm\{ros\}\}\+\\tfrac\{1\}\{2\}A\_\{\\mathrm\{vina\}\},and useAaffA\_\{\\mathrm\{aff\}\}asaffinityproperty to train the surrogate\.

### C\.10\.Unified optimization direction

For surrogate training and gradient\-based guidance, all property scores are converted to a*higher\-is\-better*convention to ensure a consistent optimization direction across objectives\. This transformation is used only for internal optimization\. Reported results follow the standard biochemical interpretation of each metric\.

## Appendix DGuidance Weights for Therapeutic Objectives

APCyc uses five Bayesian posterior\-guided sampling configurations,Base,Aff\.,Prot\.,Perm\., andMulti, by assigning property\-specific guidance weights\. Table[6](https://arxiv.org/html/2606.12991#A4.T6)lists the coefficients applied to the joint energy function in Sec\.[3\.4](https://arxiv.org/html/2606.12991#S3.SS4)\.

Table 6\.Guidance weights for different therapeutic design objectives\.