Retrosynthesis Prediction via Search in (Hyper) Graph
Zixun Lan, Binjie Hong, Jiajun Zhu, Zuo Zeng, Zhenfu Liu, Limin Yu,, Fei Ma
TL;DR
This paper introduces RetroSiG, a semi-template-based search method in hypergraphs for retrosynthesis prediction, capable of handling complex reactions with improved interpretability and accuracy.
Contribution
RetroSiG is a novel framework that addresses limitations of previous methods by enabling prediction of complex reactions through graph search mechanisms.
Findings
RetroSiG achieves competitive results on retrosynthesis benchmarks.
It effectively predicts complex reactions with multiple reaction centers.
Ablation studies confirm the importance of the one-hop constraint and hypergraph modeling.
Abstract
Predicting reactants from a specified core product stands as a fundamental challenge within organic synthesis, termed retrosynthesis prediction. Recently, semi-template-based methods and graph-edits-based methods have achieved good performance in terms of both interpretability and accuracy. However, due to their mechanisms these methods cannot predict complex reactions, e.g., reactions with multiple reaction center or attaching the same leaving group to more than one atom. In this study we propose a semi-template-based method, the \textbf{Retro}synthesis via \textbf{S}earch \textbf{i}n (Hyper) \textbf{G}raph (RetroSiG) framework to alleviate these limitations. In the proposed method, we turn the reaction center identification and the leaving group completion tasks as tasks of searching in the product molecular graph and leaving group hypergraph respectively. As a semi-template-based…
Peer Reviews
Decision·ICLR 2024 Conference Withdrawn Submission
(S1): The retrosynthesis prediction problem is generally important, and the authors correctly note how some of the existing models can be too constrained with respect to either the form of the reaction centre or multiset of leaving groups. (S2): The high-level approach is reasonable (although I have many comments about specific details – see (W1)).
(W1): Several parts of the method are either not clear to me, or they are clear but I am not sure if they are necessary. - (a) How does RetroSiG choose where the leaving groups are attached in the graph, and also how many times each group is used? The action space used by the hypergraph agent suggests each node (leaving group) is only chosen once, so the count of how many it is applied has to be selected separately. - (b) The authors mention that the final prediction is produced by assemblin
S1. The paper demonstrates a level of originality in addressing the challenge of retrosynthesis prediction in organic synthesis. While semi template-based and graph-edits-based methods have been previously explored, the paper introduces a novel approach called RetroSiG, which combines a semi-template-based method with a search mechanism in the product molecular graph and leaving group hypergraph. This integration aims to overcome the limitations of existing methods in predicting complex reaction
W1. Incorporate more diverse and complex reaction datasets: You could improve the generalizability of your method by incorporating more diverse and complex reaction datasets. This would help to ensure that the method can handle a wider range of reactions and produce more accurate predictions. W2. Provide a more detailed analysis of the experimental results: While you provide some experimental results, you could provide a more detailed analysis of the results to help readers better understand th
1. The writing is clear. 2. The method is novel to me. The method transforms the template prediction into a search problem. 3. The method can handle the reaction with multiple reaction centers, which is ignored by the previous semi-templated-based methods.
1. First of all, I do think the illustration figures are too small for me to understand their meaning. 2. In essence, this paper introduces a method for template prediction. However, it still faces challenges with generalization. If a template from the test set isn't present in the training set, its associated leaving groups won't be linked by hyperedges within the hypergraph. Consequently, when employing a search limited by a one-hop constraint, searching the relevant leaving groups becomes ch
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Bioinformatics and Genomic Networks · Data Mining Algorithms and Applications