A Theoretically-Principled Sparse, Connected, and Rigid Graph Representation of Molecules
Authors: Shih-Hsin Wang, Yuhao Huang, Justin Baker, Yuan-En Sun, Qi Tang, Bao Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We substantiate the effectiveness and efficiency of our proposed graphs in various molecular modeling benchmarks. Code is available at https://github.com/shihhsinwang0214/SCHull. |
| Researcher Affiliation | Academia | Shih-Hsin Wang1 , Yuhao Huang1 , Justin Baker2, Yuan-En Sun3, Qi Tang4 & Bao Wang1 1Department of Mathematics and Scientific Computing and Imaging (SCI) Institute University of Utah, Salt Lake City, UT 84102, USA 2Department of Mathematics, UCLA, Los Angeles, CA 90095, USA 3Department of Biochemistry, University of Utah, Salt Lake City, UT 84102, USA 4School of Computational Science and Engineering, Georgia Tech, Atlanta, GA 30332, USA |
| Pseudocode | No | The paper does not contain explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured, numbered steps. While it describes processes in sections like 'SCHULL A NEW GRAPH CONSTRUCTION' and presents equations for message passing, these are not formatted as pseudocode. |
| Open Source Code | Yes | Code is available at https://github.com/shihhsinwang0214/SCHull. |
| Open Datasets | Yes | We substantiate the effectiveness and efficiency of the proposed SCHull graph using a few molecular modeling benchmarks, including predicting atomic forces for small chemicals, protein fold classification, enzyme reaction classification, and protein-ligand binding affinity prediction; see details in Section 4. Examples include a single instance from the Fold biomolecule dataset (Murzin et al., 1995; Hou et al., 2018). We evaluate how effectively the SCHull framework enhances different models in learning equivariant features by predicting atomic forces on the MD17 dataset (Chmiela et al., 2017). For this task, we use the SCHull-integrated models to predict LBA. The dataset is sourced from PDBbind (Wang et al., 2004; Liu et al., 2015) along with the experimental protocols established by (Jing et al., 2020). |
| Dataset Splits | Yes | Both the training and validation sets consist of 1,000 samples each, with the remaining data reserved for testing. For this task, 12,312 proteins are used for training, 736 for validation, 718 for Fold, 1,254 for Superfamily, and 1,272 for Family. The dataset is divided into 29,215 proteins for training, 2,562 for validation, and 5,651 for testing. The curated dataset of 3,507 complexes is split into train/val/test splits based on a 30% sequence identity threshold to verify the model generalization ability for unseen proteins. |
| Hardware Specification | Yes | Experiments are conducted on a single NVIDIA RTX 3090 GPU, and T4 and A100 GPUs provided by Google Colab (Google Colaboratory, 2023). |
| Software Dependencies | No | The paper mentions 'Our implementation relies on the Py Torch Geometric (Fey & Lenssen, 2019) and Sci Py (Virtanen et al., 2020) frameworks.' but does not specify version numbers for these software components or the programming language used. |
| Experiment Setup | Yes | Training setup: The training setups for different tasks are available in Appendices I.2 and I.3. Detailed experimental and hyperparameter settings are provided in Appendices H.2 and I.2, respectively. Model and training hyperparameters for the MD17 task are listed in Table 11. Model and training hyperparameters search space for the protein tasks are listed in Table 12. The training procedure uses the Adam optimizer to minimize the mean squared error (MSE) loss between the model predictions and ground truth graph labels. The training uses the Adam optimizer with a learning rate of 1e-4 and a learning rate scheduler, which reduces by a factor of 0.9, has a patience of 25, and a minimum lr of 1e-5. The model is trained for 100 epochs, with the test loss reported for the epoch with the best validation loss. |