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..
Drug Discovery with Dynamic Goal-aware Fragments
Authors: Seul Lee, Seanie Lee, Kenji Kawaguchi, Sung Ju Hwang
ICML 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally demonstrate that GEAM effectively discovers drug candidates through the generative cycle of the three modules in various drug discovery tasks. Our code is available at https://github.com/ Seul Lee05/GEAM. The experimental results show that GEAM significantly outperforms existing state-of-the-art methods, demonstrating its effectiveness in addressing real-world drug discovery problems. |
| Researcher Affiliation | Collaboration | 1KAIST 2National University of Singapore 3Deep Auto.ai. |
| Pseudocode | Yes | The single generation cycle of GEAM is described in Algorithm 1 in Sec. A. |
| Open Source Code | Yes | Our code is available at https://github.com/ Seul Lee05/GEAM. |
| Open Datasets | Yes | We used ZINC250k (Irwin et al., 2012) to train FGIB to predict Y and extract initial fragments. |
| Dataset Splits | Yes | Following Yang et al. (2021), Lee et al. (2023b) and Gao et al. (2022), we used the ZINC250k (Irwin et al., 2012) dataset with the same train/test split used by Kusner et al. (2017) in all the experiments. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory amounts, or specific cloud instance types used for running the experiments. |
| Software Dependencies | No | The paper mentions software like RDKit (Landrum et al., 2016) and Quick Vina 2 (Alhossary et al., 2015) but does not provide specific version numbers for these or any other software components, which is required for reproducibility. |
| Experiment Setup | Yes | Regarding the architecture of FGIB, we set the number of message passing in the MPNN to 3 and the number of layers of the MLP to 2. FGIB was trained to 10 epochs in each of the task with a learning rate of 1e 3 and β of 1e 5. The initial vocabulary size was set to K = 300. Regarding the dynamic vocabulary update, the maximum vocabulary update in a single cycle was set to 50 and the maximum vocabulary size was set to L = 1,000. We set the termination number of atoms in the SAC to n SAC = 40, so that an episode ends when the size of the current molecule exceeds 40. The population size of the GA was set to P = 100 and the mutation rate was set to 0.1. The minimum number of atoms of generated molecules was set to 15. |