Session: COMP Poster Session:

Rapid forward and retrosynthetic organic reaction prediction remains an active area of artificial intelligence research in chemistry. IBM’s freely available predictive AI, RXN for Chemistry (https://rxn.res.ibm.com/), treats reaction prediction as a machine translation problem, converting a sequence of reactant molecules, in the form of SMILES strings, to products. In this way, prediction is done by finding a conversion from the “language” of reactants to the “language” of products, without considering the minutia of chemical transformations or reaction mechanisms. Though this method circumvents the arduous task of encoding all the fundamental “rules” of organic chemistry, it does not necessarily obey conservation laws or other physical principles, and thus is potentially susceptible to “alchemy” and other pathologies. This work aims to illuminate areas where transformer models fall short and to consider how to best remedy these cases where the fundamentals of organic chemistry are violated. By using a data set of 100 organic chemistry reactions sourced from an undergraduate level textbook, we evaluated the performance of the RXN for Chemistry model on classic reaction classes in chemistry and identify areas where this model is generally successful, as in substitution reactions, and unsuccessful, as in Diels-Alder or Robinson annulation reactions, at predicting outcomes. We also compared the change in molecular complexity between predicted and human retrosynthesis strategies using an additive measure of molecular complexity. We found that specific reaction classes, such as elimination, Diels-Alder, or Robinson annulations, have characteristically low predictive accuracies and tend to instead favor other reaction mechanisms, such as substitution, where the model’s accuracy is superior. Moreover, forward predictions replicate a similar distribution of changes in molecular complexity as the literature, whereas retrosynthetic predictions do not.

Joshua Schrier

Fordham University

Appearances