Cyclic peptides have become a new focus in drug discovery due to their ability to bind challenging targets, including “undruggable” protein-protein interactions, with low toxicity. Despite their potential, general methods for de novo design of cyclic peptide ligands based on target protein structures remain limited. Here, we developed CYC_BUILDER, a reinforcement learning based fragment growing method for efficient assembly of peptide fragments and cyclization to generate diverse cyclic peptide binders for target proteins. CYC_BUILDER employs a Monte Carlo Tree Search (MCTS) framework to integrate seed fragment exploration, fragment fusion based peptide growth, structure optimization, evaluation and peptide cyclization. It supports peptide cyclization through both head-to-tail amide bond and disulfide bond formation. We first validated CYC_BUILDER on known protein-cyclic peptide complexes, demonstrating its ability to accurately re-generate cyclic peptide binders in terms of both sequences and binding poses. We then applied it to design cyclic peptide inhibitors for TNF α, a key mediator in inflammation-related diseases. Among the nine experimentally tested designed peptides, four showed potent binding to TNF α and inhibited its cellular activity. CYC_BUILDER provides an efficient tool for cyclic peptide drug design, offering significant potential for addressing challenging therapeutical targets.
Many diseases have been found to involve multifactorial aspects and complex networks of protein-protein interactions (PPIs). These interfaces are usually large with buried solvent accessible surface areas ranging from 1500 to 3000 Å 2, which are difficult to be covered by small molecules (usually with solvent accessible surface areas between 150-500 Å2). In contrast, peptides are well-suited to block PPIs. Peptides offer several advantages over traditional small-molecule drugs. Their biosynthesis and solid-phase synthesis methods are well-established, and their degradation products, amino acids,are minimally toxic. Despite these advantages, linear peptides face significant challenges, including poor membrane permeability, rapid degradation and reduced binding strength due to entropic losses upon binding. Cyclization of linear peptides has emerged as a highly effective solution to overcome these limitations. A number of cyclic peptide drugs have been already used to treat diseases, such as T-cell lymphoma, lupus pneumonitis, and excessive blood loss.
Machine learning has significantly advanced small molecule drug and protein design. Early efforts have extended to the design of cyclic peptides. The availability of cyclic peptide conformation databases, such as CREMP, has enabled efficient cyclic peptide conformation sampling using deep learning algorithms, such as StrEAM, RINGER, have achieved efficient sampling of cyclic peptide conformations. Several groups have modified Alphafold to optimize the backbone structure of free cyclic peptides or cyclic peptide-protein complex and design sequences for cyclic peptide binders. However, due to the limited number of crystal structures of cyclic peptide-protein complexes, de novo design of cyclic peptide ligands solely based on the target protein structures remains largely unexplored.
This study introduces a target protein structure based cyclic peptide binders design method, CYC_BUILDER, using Monte Carlo Tree Search (MCTS) framework and reinforcement learning algorithms. Cyclic peptides are grown in the protein binding site using fragment-based building blocks to give stabe and natrual-like conformations. By incorporating reward functions that account for both binding strength and structure rationality, CYC_BUILDER efficiently generates cyclic peptide binders for diverse protein targets and exhibits good generalization capability. Compared to the anchor extension strategy proposed by Hosseinzadeh et al using Rosetta and other deep learning methods such as AfCycDesign and RFpeptide, our algorithm explores high-quality cyclic peptide scaffolds rapidly and guides the intermediate structures towards multiple cyclization patterns (both head-to-tail and disulfide bonds) using potential score functions. Sequence and backbone diversity of the generated cyclic peptide ligands are also promising. We have applied this method to design and experimentally validated cyclic peptide binders for tumor necrosis factor α (TNF α), a key player in inflammation related diseases, and obtained potent peptide inhibitors.
CYC_BUILDER offers a general fragment growing strategy for structure based de novo design of cyclic peptide binders. Based on the fragment database extracted from the proteinprotein interface, our program efficiently constructs head-to-tail or disulfide cyclized peptides targeting protein-protein interfaces, including both surface and pocket regions. Monte Carlo Tree Search drives peptide growth and cyclization, progressively refining fragment selection for optimal conformations without reliance on complex structural training data. With advanced scoring methods, the tool assesses conformation stability, binding affinity, and cyclization propensity, utilizing grid-based interaction scoring and enhanced intramolecular hydrogen bond analysis. CYC_BUILDER excels in creating peptide binders with strong affinity and diversity, making it well-suited for complex protein-protein interactions.
The main framework consists of 4 modules, including fragment sampling, fragment growth, structure optimization and backbone closure. The growing process starts from a seed fragment, which serves as the root node of the Monte Carlo tree. The seed peptide fragment can be extracted according to the hotspots of an existing protein or peptide binder, or it can be generated by the Seed seeking algorithm. Originating from the root node, each growth action in our model corresponds to the sampling of a new set of fragments from the tripeptide structure library and the fragment fusion to the current peptide. To enhance the sampling efficiency, we categorized the tripeptide fragments in the library based on bockbone conformation, the orientation of middle residue, and the polarity and size of residues (20 natural amino acid residues were classified into 7 categories). The probability matrix for sampling fragments is initialized according to the type of the fragments. The sampled fragments are then assembled onto the already constructed peptide using a flexible assembly algorithm.
Each growth action (conducted by fragment splice algorithm) that creates a new leaf is evaluated by the scoring function, which contains four terms: the binding energy, the properties of the interaction interface, the rationality and stability of the cyclic peptide structure, and the propensity for cyclization. Subsequently, sample probability matrix and state values are updated through backpropagation. In the state of the current leaf node, the model assesses whether the termination conditions can be meet. If the peptide meets the termination criteria, the model executes a cyclization operation. Otherwise, the peptide continues to grow.
In the traditional MCTS algorithm, the rollout/playout operation in the simulation process is carried out by the Monte Carlo method. That is, it extends the leaf node from the current state until the search termination condition is met, referred to as an episode. Then, the score of the simulation result is backpropagated to the current node, stored as a part of the reward of this simulation in the current State. The sampling scores of the current state in combination with information such as the number of visits were used to evaluate the value of this action (denoted by Q). To ensure that the simulation component of the model converges rapidly to reasonable values, CYC_BUILDER growth module adopts a time-dependent (TD) algorithm rollout method. The simulation explores the next state for the benifit of avoiding the issue of episodes failing to terminate (peptide cyclized). Moreover, the sampling matrix and scoring weights are optimized by policy gradients along the growing process. Max sampling depth and a maximum child nodes number can be specified by users.
In the structure optimization module, during the growing process, the peptide backbone was perturbed first and then the side chains were repacked. Efficient scoring functions were developed for the evaluation of peptide-protein target affinity, backbone conformation stablity and cyclization tendency. The output complex structure of generated cyclic peptides are optimized and filtered using Rosetta.
We trimmed the ADCP cyclic peptide redock test-set which contain the X-ray crystal structures of proteins complexed with cyclic peptides for testing whether our program CYC_BUILDER can generate cyclic peptide binders with native interactions and similar binding affinity. Cyclic peptides containing 6-20 natural amino acid residues in the redock test-set were collected and redundant structures with peptide sequence similarity ≤ 40% (calculated by Biopython) were removed. For comp
