Abstract:In recent years, with the rapid advance of computer science, various modern intelligent algorithms have successively emerged. Transformer, based on multi-head attention mechanism, is one of the most favored AI models among in this century. The introduction of these algorithms leads to dramatic progress in retrosynthesis prediction. Unlike conventional retrosynthesis prediction models, retrosynthesis prediction based on intelligent algorithms can automatically extract chemistry knowledge from chemical reaction datasets to predict retrosynthesis routes. In this review, we provide a comprehensive overview of retrosynthesis prediction based on modern intelligent algorithms, particularly artificial intelligence algorithm. After introducing the related deep learning model, the existing chemical reaction datasets and molecular representations are presented. Subsequently, the current state-of-the art of AI-assisted retrosynthesis prediction models in recent years is discussed, including template-based models, template-free models, and semi-template-based models. Additionally, we conclude by comparing retrosynthesis prediction models across different categorizations. Finally, several challenges and limitations of these current methods are summarized, with a view to promising directions for future research.

What problem does this paper attempt to address?

This paper mainly discusses the application of modern intelligent algorithms in retrosynthetic prediction. Retrosynthetic analysis is a commonly used method in organic synthesis design, which deduces the synthetic route from the target compound. With the rapid development of computer science, especially the emergence of artificial intelligence (AI) models such as Transformer based on multi-head attention mechanism, significant progress has been made in retrosynthetic prediction. Unlike traditional retrosynthetic models, models based on intelligent algorithms can automatically extract knowledge from chemical reaction datasets to predict retrosynthetic pathways. The paper first introduces deep learning models, existing chemical reaction datasets, and molecular representations. Then, the latest progress of AI-assisted retrosynthetic prediction models in recent years is discussed, including template-based models, template-free models, and semi-template-based models. In addition, different categories of retrosynthetic prediction models are compared, and the challenges and limitations of current methods are summarized, providing directions for future research. The research points out that early rule-based models had unsatisfactory performance due to limitations in computing power and data. However, modern AI models, especially deep learning models, can discover new reaction pathways that are not constrained by existing knowledge libraries, although they may have low interpretability and high computational complexity issues. The paper also covers the application of deep learning algorithms such as sequence generation models, graph neural networks, reinforcement learning, and search algorithms in retrosynthetic prediction. In summary, this paper attempts to address how to use modern intelligent algorithms to automate and optimize retrosynthetic analysis in organic synthesis, improve efficiency and accuracy, and overcome the limitations of existing methods.

Application of Modern Intelligent Algorithms in Retrosynthesis Prediction