GRASP: Navigating Retrosynthetic Planning with Goal-driven Policy
Authors: Yemin Yu, Ying Wei, Kun Kuang, Zhengxing Huang, Huaxiu Yao, Fei Wu
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on the benchmark Pistachio dataset and a chemists-designed dataset demonstrate that the framework outperforms existing state-of-the-art approaches by up to 32.2% on search efficiency and 5.6% on quality. Remarkably, our user studies show that GRASP successfully plans pathways that accomplish the goal prescribed with a goal (building block materials). |
| Researcher Affiliation | Academia | Yemin Yu1,2, Ying Wei1 , Kun Kuang2, Zhengxing Huang2, Huaxiu Yao3, Fei Wu2,4 1City University of Hong Kong, 2Zhejiang University 3Stanford University, 4Shanghai Artificial Intelligence Laboratory 1{yeminyu2-c,yingwei}@cityu.edu.hk 2{kunkuang,zhengxinghuang,wufei}@zju.edu.cn 3{huaxiu}@cs.stanford.edu |
| Pseudocode | Yes | Algorithm 1 GRASP |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, providing a repository link, or including code in supplementary materials for the described methodology. |
| Open Datasets | Yes | We use the Pistachio reaction dataset (Ver. 18.11.19) [16] as our benchmark dataset for training our single-step models, and the implementation details are listed in Appendix A.1. ... We use the complete 231M commercially available molecules presented in e Molecules 1 for the building block molecule set. 1http://downloads.emolecules.com/free/2019-11-01/ |
| Dataset Splits | Yes | The dataset is further split randomly into train/val/test sets following 90%/5%/5% proportions. ... Specifically, we obtained 61554 pathways with an average length of 3.66. We split the dataset into 40000 training pathways, 21354 validation pathways, and 200 test pathways for Retro* value network training. |
| Hardware Specification | No | The paper mentions computational time (e.g., 'single-step inference ( 2s per iter)'), but it does not provide specific details about the hardware (e.g., GPU/CPU models, memory, or cloud instances) used for running the experiments. |
| Software Dependencies | No | The paper refers to using "molecular transformer (MT)" and the TD3 algorithm, but it does not specify any software names with their version numbers (e.g., Python 3.x, PyTorch 1.x) that are needed to replicate the experiments. |
| Experiment Setup | Yes | Since the relabeling probability pr is an important hyperparameter to balance between general and goal-driven planning, we will further examine the effect of different pr on the planning performance in the experiment. ... We choose top-k=100 reactions ranked according to the confidence score predicted by the single-step model as the available single-step candidate set for a given molecule since top-k=100 is sufficient to represent feasible single-step reaction space for a molecule. ... Finally, γ is the discount factor, and H is the maximum horizon (length) for the pathway. |