Retrosynthetic Planning with Dual Value Networks

Authors: Guoqing Liu, Di Xue, Shufang Xie, Yingce Xia, Austin Tripp, Krzysztof Maziarz, Marwin Segler, Tao Qin, Zongzhang Zhang, Tie-Yan Liu

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the widely-used USPTO dataset, our PDVN algorithm improves the search success rate of existing multi-step planners (e.g., increasing the success rate from 85.79% to 98.95% for Retro*, and reducing the number of model calls by half while solving 99.47% molecules for Retro Graph).To demonstrate the effectiveness of our PDVN algorithm, we conduct extensive experiments on the widely used USPTO dataset (Lowe, 2012; Chen et al., 2020).
Researcher Affiliation Collaboration 1Microsoft Research AI4Science 2National Key Laboratory for Novel Software Technology, Nanjing University 3Renmin University of China 4University of Cambridge.
Pseudocode No The paper describes the algorithm steps in detailed text and provides a flowchart in Figure 2, but it does not include a formal pseudocode block or algorithm listing.
Open Source Code No The paper does not provide an explicit statement or a link for the availability of its source code.
Open Datasets Yes On the widely-used USPTO dataset (Lowe, 2012; Chen et al., 2020). For the training target molecule dataset Dtrain, we follow the procedure from (Chen et al., 2020; Kim et al., 2021) and construct synthesis routes based on the publicly available reaction dataset extracted from the United States Patent Office (USPTO) (Lowe, 2012).
Dataset Splits No The paper mentions using "Dtrain" for generating simulated experiences and "Dtest" for evaluation, but it does not specify explicit train/validation/test dataset splits with percentages, counts, or predefined citations for model development or hyperparameter tuning.
Hardware Specification Yes The whole training process takes about 18 hours on a server with four NVIDIA TITAN Xp and 48 CPU cores (using 15 parallel workers).
Software Dependencies No The paper mentions "RDKit" and "Adam optimizer" but does not provide specific version numbers for these or any other software dependencies needed to replicate the experiments.
Experiment Setup Yes During the Planning phase, the batch size of sampled target molecules is 1024. We set crxn(s, a) = 0.1 and cdead = 5. During the Updating phase, the Adam optimizer (Kingma & Ba, 2014) with a mini-batch of size 128 and a learning rate of 0.001 is used for all models. Table 6. Hyper-parameters for PDVN planning. C (PUCT) 1.0, alpha (Synthesizability penalty) 0.8, MCTS depth 15, Number of simulations 100, cdead 5.0, crxn(s, a) 0.1 Table 7. Hyper-parameters for PDVN training. Training dataset size 299202, Batch size 1024, Optimizer Adam, Learning rate 1e-3, Dropout rate 0.1, Mini-batch size 128, SL epochs 8.