Gradient GA: Gradient Genetic Algorithm For Drug Molecular Design
Authors: Debadyuti Mukherjee, Chris Zhuang, Yingzhou Lu, Tianfan Fu, Ruqi Zhang
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method significantly improves both convergence speed and solution quality, outperforming cutting-edge techniques. The proposed method has shown up to a 25% improvement in the Top 10 score over the vanilla genetic algorithm. |
| Researcher Affiliation | Academia | Debadyuti Mukherjee EMAIL Department of Computer Science Purdue University Chris Zhuang EMAIL Department of Computer Science Purdue University Yingzhou Lu EMAIL Stanford Medicine School Standford University Tianfan Fu EMAIL Department of Computer Science Rensselaer Polytechnic Institute Ruqi Zhang EMAIL Department of Computer Science Purdue University |
| Pseudocode | Yes | The overall pipeline is illustrated in Figure 1, and the corresponding pseudocode is provided in Algorithm 1. |
| Open Source Code | Yes | The code is available at https://github.com/debadyuti23/Gradient GA. |
| Open Datasets | Yes | For all methods, we use the ZINC 250K database (Irwin et al., 2012) to select the initial molecule population, extract chemical fragments, and perform pretraining. ZINC is a free database of commercially available compounds for virtual screening. |
| Dataset Splits | No | The paper states: "The initial molecule population is randomly sampled from drug-like molecule databases, such as ZINC (Sterling & Irwin, 2015)." and "The size of the validation set is 10% of the respective sets." in Appendix D for validating the predictive model M. However, it does not explicitly provide specific train/test/validation splits for the main molecular design task or the ZINC 250K database as a whole. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory amounts) used for running the experiments. It only mentions computational cost comparisons in terms of time. |
| Software Dependencies | No | The paper mentions "Mean Squared Error (MSE) and Adam (Kingma & Ba, 2014) with a learning rate of 0.001 as the loss function and optimizer, respectively, when training M." and describes the MPNN architecture, but does not specify version numbers for any software, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | We select Mean Squared Error (MSE) and Adam (Kingma & Ba, 2014) with a learning rate of 0.001 as the loss function and optimizer, respectively, when training M. The model is trained for 200 epochs for each round of training. The MPNN has 2 layers of convolution and GRU with 1 layer of Set2Set. There are n = 16 hidden features. The upper bound of the generated molecule k is set to 70. The threshold criterion τ for a good score of a molecule is set as the maximum metric score 0.001. |