Designing Cyclic Peptides via Harmonic SDE with Atom-Bond Modeling

Authors: Xiangxin Zhou, Mingyu Li, Yi Xiao, Jiahan Li, Dongyu Xue, Zaixiang Zheng, Jianzhu Ma, Quanquan Gu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that the cyclic peptides designed by our method exhibit reliable stability and affinity. ... We use Rosetta (Chaudhury et al., 2010) to compute the total energy of reference ligands, linear peptides designed by baseline methods, and cyclic peptides engineered by our approaches. ... The results are shown in Table 1.
Researcher Affiliation Collaboration 1Byte Dance Seed (Work was done during Xiangxin s internship at Byte Dance Seed.) 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3New Laboratory of Pattern Recognition (NLPR), State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA) 4Institute for AI Industry Research, Tsinghua University 5School of Medicine, Shanghai Jiao Tong University 6Department of Electronic Engineering, Tsinghua University. Correspondence to: Quanquan Gu <EMAIL>.
Pseudocode Yes Algorithm 1 Routed Sampling Input: number of residues within the ligand peptide N, cyclization Type O, 3D receptor structures T , SDE solver time interval dt, infinitesimal constant ̈ Output: cyclic peptide with its all-atom coordinates x0, chemical graph G0, amino acid sequence A0
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of the described methodology.
Open Datasets Yes The small molecule dataset is sourced from PDBBind (Wang et al., 2005) and has 14,348 protein-ligand complexes. The peptide dataset is derived from RCSB PDB (Burley et al., 2023), Propedia (Martins et al., 2023) and Pep BDB (Wen et al., 2019).
Dataset Splits Yes Samples are clustered by receptor sequence identity of 0.3 to split the dataset into training and validation sets. ... Furthermore, clustering at 40% sequence identity is applied to separate training and test data, removing any training complexes that share clusters with the test set. The final dataset comprises 4,157 training complexes, 114 for validation, and 93 for testing.
Hardware Specification Yes ATOMSDE converges within 48 hours and RESROUTER converges within 18 hours on 8 NVIDIA H100 GPUs. ... We benchmark the average time of generating one peptide for all co-design baselines and our methods on a single NVIDIA A100-SXM4-80GB GPU. ... All simulations were run on RTX 4090 GPUs using the CUDA implementation of particle-mesh Ewald (PME) molecular dynamics in Amber22 (Salomon-Ferrer et al., 2013).
Software Dependencies Yes All simulations were run on RTX 4090 GPUs using the CUDA implementation of particle-mesh Ewald (PME) molecular dynamics in Amber22 (Salomon-Ferrer et al., 2013). ... RDKit 2 does not always accurately produce chemical bonds during conversion ... the Fast Relax protocol in Py Rosetta (Chaudhury et al., 2010) is employed to relax each complex...
Experiment Setup Yes We use the same optimizer setting for both ATOMSDE and RESROUTER: Adam W (Loshchilov, 2017) optimizer with constant learning rate 0.0001, beta1 0.9, beta2 0.999, and weight decay 0.01. For beta schedule, we use ̈(t) = (̈max ̈min)t+̈min, where ̈min = 0.01 and ̈max = 3.0. ... In the first stage, 2,500 steepest descent and 2,500 conjugate gradient cycles were applied to all atoms, with constraints on water molecules and counterions.