Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints

Authors: Dapeng Jiang, Xiangzhe Kong, Jiaqi Han, Mingyu Li, Rui Jiao, Wenbing Huang, Stefano Ermon, Jianzhu Ma, Yang Liu

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our model, despite trained with linear peptides, is capable of generating diverse target-binding cyclic peptides, reaching success rates from 38% to 84% on different cyclization strategies. 4. Experiments
Researcher Affiliation Academia 1 Institute for AI Industry Research (AIR), Tsinghua 2 Xingjian College, Tsinghua University 3 Department of Computer Science and Technology, Tsinghua University 4 Stanford University 5 Gaoling School of Artificial Intelligence, Renmin University of China 6 Beijing Key Laboratory of Research on Large Models and Intelligent Governance 7 Department of Electronic Engineering, Tsinghua University.
Pseudocode Yes Algorithm 1 Training Procedure of CP-Composer Algorithm 2 Inference Procedure of CP-Composer
Open Source Code Yes F. Code Availability The codes for our CP-Composer is provided at the link https://github.com/jdp22/CP-Composer_final.
Open Datasets Yes Dataset. We utilize Pep Bench and Prot Frag datasets (Kong et al., 2024) for training and validation, with the LNR dataset (Kong et al., 2024; Tsaban et al., 2022) for testing.
Dataset Splits Yes We utilize Pep Bench and Prot Frag datasets (Kong et al., 2024) for training and validation, with the LNR dataset (Kong et al., 2024; Tsaban et al., 2022) for testing. Pep Bench contains 4,157 protein-peptide complexes for training and 114 complexes for validation, with a target protein longer than 30 residues and a peptide binder between 4 to 25 residues.
Hardware Specification Yes We train CP-Composer on a 24G memory RTX 3090 GPU with Adam W optimizer. ...Simulations are conducted using the Amber22 package with the CUDA implementation of particle-mesh Ewald (PME) MD and executed on Ge Force RTX 4090 GPUs (Salomon-Ferrer et al., 2013).
Software Dependencies Yes Simulations are conducted using the Amber22 package with the CUDA implementation of particle-mesh Ewald (PME) MD and executed on Ge Force RTX 4090 GPUs (Salomon-Ferrer et al., 2013). ...For system preparation, the ff14SB force field is applied to proteins and peptides (Maier et al., 2015).
Experiment Setup Yes The initial learning rate is set to 10 4 and is reduced by a factor of 0.8 if the validation loss does not improve for 5 consecutive epochs. Regarding the diffusion model, we train for no more than 1000 epochs. The learning rate is 10 4 and decay by 0.6 and early stop the training process if the validation loss does not decrease for 10 epochs. During the training process, we set the guidance strength as 1 for sampling at the validation stage. ...For the RBF kernel, we use 32 feature channels.