Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows
Authors: Xiangxin Zhou, Yi Xiao, Haowei Lin, Xinheng He, Jiaqi Guan, Yang Wang, Qiang Liu, Feng Zhou, Liang Wang, Jianzhu Ma
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare our models with three representative baselines for SBDD: Pocket2Mol (Peng et al., 2022) generate 3D molecules by sequentially placing atoms around a given protein pocket; Target Diff (Guan et al., 2023) generates atom coordinates and atom types based on a diffusion model and bonds are determined with a post-processing algorithm; IPDiff (Huang et al., 2023) incorporated protein-ligand interaction priors into both the forward and reverse processes to enhance the diffusion models. ... We report the properties of designed ligand molecules in Table 1. ... Fig. 5 shows the Cover Ratio and minimum RMSD against holo states along the number of samples. ... Table 2 shows that our refined pocket conformation can effectively improve the performance of SBDD methods with rigid pocket inputs. Fig. 4 shows the distribution in number differences of NCIs of apo/our pockets and molecules designed by Targetdiff... |
| Researcher Affiliation | Collaboration | 1School of Artificial Intelligence, University of Chinese Academy of Sciences 2New Laboratory of Pattern Recognition (NLPR), State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA) 3School of Information and Communication Technology, Griffith University 4Beijing Stone Wise Technology Co Ltd 5Institute for Artificial Intelligence, Peking University 6Institute for AI Industry Research, Tsinghua University 7Department of Computer Science, University of Illinois Urbana-Champaign 8Department of Electronic Engineering, Tsinghua University |
| Pseudocode | No | The paper describes the methodology in text and illustrates the model architecture in figures (e.g., Figure 2 and Figure 3). However, it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the source code for the described methodology, nor does it provide any links to a code repository. |
| Open Datasets | Yes | We curate our dataset based on MISATO dataset (Siebenmorgen et al., 2024). The original dataset contains approximately 20,000 protein-ligand complexes with associated 8ns molecular dynamics (MD) simulation trajectories. ... The development of MISATO dataset is based on the PDBbind database (Wang et al., 2005) which contains around 20,000 complexes structures and experimental binding affinities measurement. |
| Dataset Splits | Yes | We select 50 complexes that have no overlap with the training set as the test set. |
| Hardware Specification | Yes | We benchmark the inference time of baselines and our methods for generating 10 ligand molecules given the same pocket on 1 Tesla V100-SXM2-32GB. |
| Software Dependencies | No | The paper mentions tools like GROMACS (Van Der Spoel et al., 2005) and RDKit (Landrum et al., 2020) for data processing, and AdamW (Loshchilov, 2017) as an optimizer. However, it does not provide specific version numbers for core deep learning libraries (e.g., Python, PyTorch, CUDA) or other crucial software dependencies required to reproduce the experimental setup. |
| Experiment Setup | Yes | There are 7 individual losses: 4 continuous flow matching losses for residue frames translation (Eq. (5)), rotation (Eq. (7)), torsion angles (Eq. (8)) and ligand atom position (same as Eq. (5)), 2 discrete flow matching losses for ligand atom and bond types (Eq. (14)), and interaction loss (Eq. (18)). There are first averaged across all residues or atoms in a training sample and then simply weighted summed with weights: 2.0, 1.0, 1.0, 4.0, 1.0, 1.0, 0.5. We use Adam W (Loshchilov, 2017) as the optimizer with learning rate 0.0002, beta1 0.95, and beta2 0.999. γ controls the stochasticity of the stochastic flow (see Eqs. (19) to (21)). We use 2.0, 0.005, 1.0, 2.0 as the values of γ for residue frames translation, rotation, torsion angles, and ligand atom positions. |