Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
REBIND: Enhancing Ground-state Molecular Conformation Prediction via Force-Based Graph Rewiring
Authors: Taewon Kim, Hyunjin Seo, Sungsoo Ahn, Eunho Yang
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that REBIND significantly outperforms state-of-the-art methods across various molecular sizes, achieving up to a 20% reduction in prediction error. The versatility of our proposed framework is demonstrated through benchmarks on both small-scale datasets, i.e., QM9 (Ramakrishnan et al., 2014) and Molecule3D (Xu et al., 2021c), and a large-scale GEOM-DRUGS (Axelrod & Gomez-Bombarelli, 2022) dataset. |
| Researcher Affiliation | Collaboration | Taewon Kim1,2 , Hyunjin Seo1,2 , Sung Soo Ahn1, Eunho Yang1,3 Korea Advanced Institute of Science and Technology (KAIST)1, Polymerize2, AITRICS3 EMAIL |
| Pseudocode | No | The paper describes the architecture and methodology (Section 4) but does not present any explicitly labeled 'Pseudocode' or 'Algorithm' block, nor does it format any part of its method description as a code-like algorithm. |
| Open Source Code | Yes | The code is available in: https://github.com/holymollyhao/ReBIND |
| Open Datasets | Yes | We evaluated REBIND on well-established benchmark datasets, including QM9 (Ramakrishnan et al., 2014), Molecule3D (Xu et al., 2021c), and GEOM-DRUGS (Axelrod & Gomez-Bombarelli, 2022). |
| Dataset Splits | Yes | Molecule3D is a large-scale dataset of molecular structures, for which we employed two distinct splitting strategies: random split and scaffold split. The scaffold split groups molecules based on their core substructures, enabling a more realistic evaluation. Additionally, GEOM-DRUGS comprises large-size molecules relevant to drug discovery, providing a challenging benchmark for assessing the scalability of our framework on complex molecular structures. Since the original dataset includes multiple stable conformations, we choose the most stable conformation with respect to the Boltzmann energy for each molecule. Detailed descriptions of each dataset are provided in Appendix C. ... We adopted the evaluation protocols and train/validation/test splits from Xu et al. (2024) for the QM9 and Molecule3D datasets and from Jing et al. (2022) for the GEOM-DRUGS dataset. |
| Hardware Specification | Yes | The experiments were conducted on RTX Titan and RTX 3090 (24GB) GPU machines. |
| Software Dependencies | No | All GNN architectures were implemented using PyTorch Paszke et al. (2019) and PyTorch Geometric Fey & Lenssen (2019). The paper mentions software like PyTorch and PyTorch Geometric, along with RDKit, but does not specify their version numbers. |
| Experiment Setup | Yes | We set the hidden dimension to 512 and the number of layers to 8; for these references, we adopted their optimal configurations. We employed the AdamW optimizer with a batch size of 100 and no weight decay. Learning rates were initially warmed up from 0 and then fixed based on the best validation performance on the QM9 dataset, within the range of [3e-5, 5e-5, 7e-5, 9e-5]. The number of attention heads was set to 8. We used a seed of 42 for all experiments and trained all models for 20 epochs, following the configuration in Xu et al. (2024). |