Efficient Evolutionary Search Over Chemical Space with Large Language Models
Authors: Haorui Wang, Marta Skreta, Cher-Tian Ser, Wenhao Gao, Lingkai Kong, Felix Strieth-Kalthoff, Chenru Duan, Yuchen Zhuang, Yue Yu, Yanqiao Zhu, Yuanqi Du, Alan Aspuru-Guzik, Kirill Neklyudov, Chao Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive empirical studies on both commercial and open-source models on multiple tasks involving property optimization, molecular rediscovery, and structure-based drug design, demonstrating that the joint usage of LLMs with EAs yields superior performance over all baseline models across single- and multi-objective settings. |
| Researcher Affiliation | Collaboration | 1Georgia Institute of Technology, 2University of Toronto, 3Vector Institute, 4Massachusetts Institute of Technology, 5University of Wuppertal, 6Deep Principle Inc., 7University of California, Los Angeles 8Cornell University, 9Université de Montréal, 10Mila Quebec AI Institute |
| Pseudocode | Yes | This process is outlined in Algorithm 1. |
| Open Source Code | Yes | Our code is available at https://github.com/zoom-wang112358/MOLLEO. |
| Open Datasets | Yes | We evaluate MOLLEO on 26 tasks from two molecular generation benchmarks, Practical Molecular Optimization (PMO) (Gao et al., 2022) and Therapeutics Data Commons (TDC) (Huang et al., 2021). [...] For the initial population of molecules, we randomly sample 120 molecules from ZINC 250K (Sterling & Irwin, 2015). [...] We took ZINC20 (Irwin et al., 2020), a database of 1.4 billion compounds that were used to generate the training set for Bio T5, and Pub Chem (Kim et al., 2023)( 250K molecules), which was used to generate the training set for Molecule STM, and checked if the final molecules for the JNK3 task from each model appeared in the respective datasets. |
| Dataset Splits | No | The paper describes an optimization algorithm that uses datasets like ZINC 250K for initial molecule pools and PMO/TDC tasks for evaluation. However, it does not specify conventional training/test/validation splits (e.g., percentages or exact counts) for reproducing dataset partitioning in a machine learning model training context. The focus is on the optimization process and its performance over oracle calls, rather than on fixed data splits for model generalization. |
| Hardware Specification | Yes | Our experiments were computed on NVIDIA A100-SXM4-80GB and T4V2 GPUs. |
| Software Dependencies | Yes | All GPT-4 checkpoints were hosted on Microsoft Azure3. 2.https://platform.openai.com/docs/models 3 *.openai.azure.com |
| Experiment Setup | Yes | For the choice of hyperparameters, we use the best practices from Graph-GA (Jensen, 2019), the baseline genetic algorithm that we build our method upon. We kept the best hyperparameters that were determined in (Gao et al., 2022). In each iteration, Graph-GA samples two molecules with a probability proportional to their fitnesses for crossover and mutation and then randomly mutates the offspring with probability pm = 0.067. This process is repeated to generate 70 offspring. The fitnesses of the offspring are measured, and the top 120 most fit molecules in the entire pool are kept for the next generation. For docking experiments, we reduce the number of generated offspring to 7 and the population size to 12 due to long experiment runtimes. We set the maximum number of oracle calls to 10,000 for all experiments except docking, where we set it to 1,000. We kept the default early-stopping criterion the same as in PMO (Gao et al., 2022), which is that we terminate the algorithm if the mean score of the top 100 molecules does not increase by at least 1e-3 within five epochs. |