A critical challenge for computer-assisted synthesis design is that proposed reaction steps may fail when attempted in the laboratory, despite initially seemingly viable. The true measure of success for any synthesis program is whether the predicted outcome matches what is observed experimentally. To address this, a neural network model is trained on experimental data from the USPTO and Reaxys to provide qualitative predictions of organic reaction outcomes in silico. Our approach involves a data augmentation strategy whereby reaction databases are supplemented with chemically-plausible negative reaction examples to overcome the literature bias towards successful reactions. Traditional reaction templates are used to generate a list of candidate outcomes for the machine learning model to score. In this manner, reactivity rules describing functional group compatibility and selectivity are implicitly learned rather than encoded. A new reaction representation is developed to focus on the fundamental transformation at the reaction site rather than constituent reactant and product fingerprints. The overall workflow can be performed without any manual curation using actual reactions reported in the patent and journal literature. We will describe the reaction prediction model and its role in a broader effort to automate synthesis design and, ultimately, automate the synthesis itself. Specific examples of automatically-generated synthetic routes will be shown with in silico and experimental validation. Opportunities for improvement will also be discussed.

Applying machine learning to synthesis design: Prediction of organic reaction outcomes

254th National Meeting (2017)

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