UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules
Authors: Ziyang Yu, Wenbing Huang, Yang Liu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The paper includes a dedicated "4. Experiments" section which details an "Experimental Setup" (4.1) covering "Datasets", "Baselines", and "Metrics". It further presents evaluations on different molecular domains (4.2-4.5) and conducts an "Ablation Study" (E.1), providing results in tables (e.g., Table 1, 2, 3) and figures (e.g., Figure 3, 4, 5). |
| Researcher Affiliation | Academia | The affiliations listed are: "1Department of Computer Science and Technology, Tsinghua University, Beijing, China", "2Gaoling School of Artificial intelligence, Renmin University of China, Beijing, China", "3Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE", and "4Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China". All these institutions are academic. |
| Pseudocode | No | The paper describes its methodology through narrative text and mathematical equations in sections like "3. Overall Workflow of Uni Sim", "3.2. Unified Pretraining", and "3.3. Vector Field Model for Dynamics", but it does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Section A. Reproducibility states: "Our code is available at https://github.com/yaledeus/UniSim." |
| Open Datasets | Yes | The paper lists and cites several datasets in Section 4.1 'Datasets', including: "PCQM4Mv2 (Hu et al., 2021)", "ANI1x (Smith et al., 2020)", "Pep MD (Yu et al., 2024)", "Protein monomers processed by Jiao et al. (2024), a subset of PDB (Berman et al., 2000)", "ATLAS (Vander Meersche et al., 2024)", and "Solvated Protein Fragments (SPF) (Unke & Meuwly, 2019)". |
| Dataset Splits | Yes | In Section 4.1 'Datasets', the paper states: "We randomly split the above pretraining datasets for training and validation by 4:1." and for ATLAS: "obtaining 790/14 as the train/test splits.". Section 4.5 'Case Study: Alanine-Dipeptide' mentions: "assigned as the training/validation/test trajectories." and "200,000 data pairs randomly sampled for training and validation from corresponding trajectories, respectively.". Table 7 'Details of the trajectory datasets' provides specific counts for training, validation, and test trajectories and pairs. |
| Hardware Specification | Yes | Section F. Computing Infrastructure states: "Uni Sim was trained on 8 NVIDIA Ge Force RTX 3090 GPUs within a week. The inference procedures were performed on one NVIDIA Ge Force RTX 3090 GPU." |
| Software Dependencies | No | The paper mentions software tools like "Torch MD-NET" (Section 3.2) and "Open MM" (Section C.2) but does not specify their version numbers or any other software dependencies with version information. |
| Experiment Setup | Yes | Section C.3 'Hyperparameters' includes Table 4 which provides detailed information on various hyperparameters for model construction, training, and inference, such as 'Hidden dimension 256', 'Learning rate 5e-4', 'Optimizer Adam', 'SDE steps T [15,25,50]', and 'Guidance strength α [0.05,0.06,0.07]'. |