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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
3D Interaction Geometric Pre-training for Molecular Relational Learning
Authors: Namkyeong Lee, Yunhak Oh, Heewoong Noh, Gyoung S. Na, Minkai Xu, Hanchen Wang, Tianfan Fu, Chanyoung Park
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on various tasks using real-world datasets, including out-of-distribution and extrapolation scenarios, demonstrate the effectiveness of 3DMRL, showing up to a 24.93% improvement in performance across 40 tasks. Our code is publicly available at https://github.com/Namkyeong/3DMRL. In Section 5, the paper details "Experiments", "Experimental Setup", "Experimental Results", "Ablation Studies", and "Molecule Collision Analysis" with performance tables including RMSE and AUROC values, along with standard deviations. |
| Researcher Affiliation | Collaboration | Namkyeong Lee1, Yunhak Oh1, Heewoong Noh1, Gyoung S. Na1,2, Minkai Xu3, Hanchen Wang3,4 , Tianfan Fu5, Chanyoung Park1 1 KAIST 2 KRICT 3 Stanford University 4 Genentech 5 Nanjing University. The affiliations include academic institutions (KAIST, KRICT, Stanford University, Nanjing University) and an industry entity (Genentech), indicating a collaboration. |
| Pseudocode | Yes | The overall framework is depicted in Figure 2, and the pseudocode for the entire framework is provided in Appendix F. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/Namkyeong/3DMRL. |
| Open Datasets | Yes | We utilize three distinct datasets, i.e., Chromophore, Combi Solv, and DDI, to pre-train 3DMRL for each downstream task as described in Section 5. Specifically, we use the Chromophore dataset for downstream tasks involving the optical properties of chromophores, the Combi Solv dataset for tasks related to the solvation free energy of solutes, and the DDI dataset, which we created for the drug-drug interaction task. The Chromophore dataset [16] consists of 20,236 combinations... The Combi Solv dataset [39] contains 10,145 combinations... For the DDI dataset, we compile drug-drug pairs from the Zhang DDI [48], Ch Ch Miner [49], and Deep DDI [33] datasets. |
| Dataset Splits | Yes | Following Pathak et al. [32], for the molecular interaction prediction task, we evaluate the models under a 5-fold cross-validation scheme. The dataset is randomly split into 5 subsets and one of the subsets is used as the test set, while the remaining subsets are used to train the model. ... For the DDI prediction task [21], we conduct experiments on two different out-of-distribution scenarios, namely molecule split and scaffold split. ... In both scenarios, we split the data into training, validation, and test sets with a ratio of 60/20/20%. |
| Hardware Specification | Yes | Computational resources. We perform all pre-training on a 40GB NVIDIA A6000 GPU, whereas all downstream tasks are executed on a 24GB NVIDIA GeForce RTX 3090 GPU. |
| Software Dependencies | Yes | Software configuration. Our model is implemented using Python 3.7, PyTorch 1.9.1, RDKit 2020.09.1, and Pytorch-geometric 2.0.3. |
| Experiment Setup | Yes | For model optimization during Pre-training stage, we employ the Adam optimizer with an initial learning rate of 0.0005 for the chromophore task, 0.0001 for the solvation free energy task, and 0.0005 for the DDI tasks. The model is optimized over 100 epochs during pre-training. In the downstream tasks, the learning rate was reduced by a factor of 10-1 after 20 epochs of no improvement in model performance in validation set, following the approach in a previous work [32], with the initial learning rate of 0.005 for the chromophore task, 0.001 for the solvation free energy task, and 0.0005 for the DDI tasks. ... We utilize a hidden dimension of 56 for molecular interaction tasks and 300 for drug-drug interaction tasks, employing the ReLU activation function for both. For the 3D virtual environment encoder f3D, we utilize SchNet [34], ... For both molecular interaction and drug-drug interaction tasks, we configure SchNet with 128 hidden channels, 128 filters, 6 interaction layers, and a cutoff distance of 5.0. |