Learning Topology-Specific Experts for Molecular Property Prediction
Authors: Suyeon Kim, Dongha Lee, SeongKu Kang, Seonghyeon Lee, Hwanjo Yu
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Top Expert has boosted the performance for molecular property prediction and also achieved better generalization for new molecules with unseen scaffolds than baselines. |
| Researcher Affiliation | Academia | 1Pohang University of Science and Technology (POSTECH), South Korea 2University of Illinois at Urbana-Champaign (UIUC), United States |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/kimsu55/ToxExpert. |
| Open Datasets | Yes | We use 8 benchmark datasets (Wu et al. 2018) for molecular property prediction, and the statistics of the datasets are summarized in Table 1. |
| Dataset Splits | Yes | For each dataset, we follow the scaffold splitting protocol (Hu et al. 2020), which sorts all the molecules by their scaffold indices and then splits them into training/validation/testing sets with a ratio of 80:10:10, respectively. |
| Hardware Specification | No | The paper mentions that "All computations required for multiple experts are executed in parallel by GPU processors." but does not specify any particular GPU model or other hardware details. |
| Software Dependencies | No | The paper mentions using "RDKit" to extract molecular features, but it does not provide a specific version number for this software or any other software dependencies. |
| Experiment Setup | Yes | Each model is trained at most 200 epochs, and the training process is terminated when the validation ROC-AUC does not increase for 50 successive epochs. We train all models ten times with different random seeds and report the average score with its standard deviation. We search for the best hyperparameter configuration through a grid search based on the validation ROC-AUC. The number of experts is chosen from K {3, 5, 7, 10}, and the loss balancing parameters are selected from α, β {5, 1, 0.1, 0.01}. For the temperature annealing of Gumbel-Softmax, the initial temperature T0 is set to 10, and the final temperature TE is chosen from {0.01, 0.1, 1}. |