Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search

Authors: Binghong Chen, Chengtao Li, Hanjun Dai, Le Song

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on benchmark USPTO datasets show that, our proposed method outperforms existing state-of-the-art with respect to both the success rate and solution quality, while being more efficient at the same time.
Researcher Affiliation Collaboration 1College of Computing, Georgia Institute of Technology 2Galixir 3Google Research, Brain Team 4Ant Financial.
Pseudocode Yes Algorithm 1: Retro (t)
Open Source Code Yes Retro*1, for chemical retrosynthesis planning. In our method, ... 1Available at https://github.com/binghong-ml/ retro_star
Open Datasets Yes We use the publicly available reaction dataset extracted from United States Patent Office (USPTO) to train one-step model and extract synthesis routes.
Dataset Splits Yes The remaining 1.3M reactions are further split randomly into train/val/test sets following 80%/10%/10% proportions.
Hardware Specification No The paper mentions 'Time is measured by the number of one-step model calls, with a hard limit of 500' and '0.3s per call', indicating computational cost, but does not provide specific hardware details such as GPU/CPU models or memory.
Software Dependencies No The paper mentions using 'RDChiral' for extracting reaction templates and training a 'template-based MLP model (Segler & Waller, 2017)' for one-step retrosynthesis, but it does not specify versions for these or other general software dependencies (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes Throughout all experiments, we take the top-50 templates predicted by MLP model and apply them on each product to get corresponding reactant lists. The cost function is defined as the negative log-likelihood of the reaction. ... We set it to be 1 by default. ... We compare Retro* against DFPN-E (Kishimoto et al., 2019), MCTS (Segler et al., 2018) and greedy Depth First Search (DFS) on product molecules in test route dataset ... We use the number of calls to the one-step model ( 0.3s per call) as a surrogate of time ... under the same time limit.