AI-driven and rule-based synthesis planning approaches are frequently portrayed as dichotomous paradigms. At a high level both try to tackle the vastness of the organic synthetic space - reaction types, structural and functional diversity and potential reaction conditions. Given the high dimensionality of the synthetic challenge, they each also try to address the sparsity of available data in many regions of this space. While machine learning approaches leverage big-data techniques and compensate for missing data by identifying patterns across different reaction types, rule-based systems resort to chemical knowledge to make generalizations and to define the scope and limitations of the rules. The two, however, are not incompatible, and their combination provides opportunities to increase the accuracy, comprehensiveness and diversity of predicted synthetic plans. Retrosynthesis in SciFinder-n is based on reaction rules that are generated automatically based on the full corpus of single-step reactions available at CAS. Chemical knowledge regarding reactivity, structural integrity, functional groups, stereochemistry and other areas is incorporated into the rule-generation algorithms. The reaction rules capture the essence of the transformations and the spectrum of substrates to which the rule may be applied. But there are facets of organic chemistry that are hard to capture and do not always emerge from the reactions that are associated with each rule. Most notably, regio-, chemo- and stereo-selectivity aspects are inconsistently manifested in the derived rules. While some rules may be based on thousands of examples, and have enough information to derive selectivity properties, the majority of rules have hundreds of examples or less, which may not be sifficient to learn such properties. In this talk we will describe our work on using machine learning techniques to compensate for limitations related to selectivity, as well as on other machine learning models that may be used to guide the retrosynthetic search in a way that balances diversity, comprehensiveness and efficiency.

Enhancing rule-based retrosynthesis with machine learning techniques

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.