Recent advances in artificial intelligence for retrosynthesis

Abstract

Retrosynthesis is the cornerstone of organic chemistry, providing chemists in material and drug manufacturing access to poorly available and brand-new molecules. Conventional rule-based or expert-based computer-aided synthesis has obvious limitations, such as high labor costs and limited search space. In recent years, dramatic breakthroughs driven by artificial intelligence have revolutionized retrosynthesis. Here we aim to present a comprehensive review of recent advances in AI-based retrosynthesis. For single-step and multi-step retrosynthesis both, we first list their goal and provide a thorough taxonomy of existing methods. Afterwards, we analyze these methods in terms of their mechanism and performance, and introduce popular evaluation metrics for them, in which we also provide a detailed comparison among representative methods on several public datasets. In the next part we introduce popular databases and established platforms for retrosynthesis. Finally, this review concludes with a discussion about promising research directions in this field.

Models Matter: the impact of single-step retrosynthesis on synthesis planning

“…Compared to the alternative public dataset USPTO-Full, 52 the performance of all single-step models is much higher on USPTO-PaRoutes-1M, where LocalRetro has a more than +25% top-50 accuracy improvement. 8 The difference in single-step performance between USPTO-PaRoutes-1M and USPTO-Full and the equal performance on USPTO-PaRoutes-1M might be explainable by the underlying data sources and their respective preprocessing. USPTO-PaRoutes-1M is a superset of USPTO-Full, where the rst contains USPTO grants and applications (3 748 191 total reactions) and the latter only USPTO grants (1 808 938 total reactions).…”

“…7 Template-based methods use reaction templates, an abstraction of the reactions in the data, which summarize the underlying pattern of these reactions. There are different approaches to extracting templates, though in all cases these processes aim to represent the atom and bond structures required to perform a reaction, 8 where a single template will represent multiple reactions. Templatebased methods consider single-step prediction as a classication problem where the task is to predict the appropriate template for the target molecule/product.…”

“…On the other hand, template-free approaches commonly treat retrosynthesis prediction as a sequence-to-sequence prediction problem, 8 employing methods seen in natural language processing such as language translation tasks. Instead of extracting and predicting the corresponding templates, the approach aims to learn the underlying reactions to directly predict reactants.…”

The Strategy of Strengthening the Country Through Higher Education in China: a Historical Policy Perspective

“…2 Recently, deep learning-based approaches have substantially expedited the process of retrosynthesis planning, known as Computer-Assisted Synthesis Planning (CASP). 3,4 For example, ASKCOS, 5 an open-source platform, can easily generate hundreds of retrosynthesis routes for a given target molecule. However, it is important to note that not every generated route is guaranteed to be feasible, as the predicted reactants may not yield the expected product in actual lab scenarios.…”

When Yield Prediction Does Not Yield Prediction: An Overview of the Current Challenges

“…Such tools typically fall under the umbrella of computer-assisted synthesis planning and include many different tools and models that can help chemists with several tasks. Retrosynthesis models suggest how to break a compound, either as a single-step prediction or multistep prediction, which provides a sequence of steps for how to synthesize a compound from simpler starting material. − Furthermore, there are a range of product prediction models, or forward models that predict what the product of two or more reactants will be, , or can provide guidance on regioselectivity issues. , There are also condition or reagent models suggesting suitable catalysts, solvents, temperatures, etc. , Finally, there are yield or reactivity models estimating the success of a reaction, which is the topic of this perspective and will be reviewed below. Although many encouraging studies have been reported, ML models for chemistry are not without critique. , Furthermore, while many studies emphasize general reaction properties, such as yield prediction in regression and classification tasks, properties tied to physical chemistry, such as reaction rates and activation energies, have received less attention.…”