UC Irvine

Achieving human-level performance at predicting chemical reactions remains an open prob- lem with broad potential applications. Here we describe a deep learning-based tool for chemical reaction prediction and product identification. Significant efforts were made to curate and refine a new, high-quality data set of hand-selected chemical reactions written at the level of elementary electron movements. Using deep artificial neural networks trained on this data, we demonstrate a high degree of accuracy at predicting real-world reactions. Because predictions are made at the elementary step level, they can be chained together to form multi-step reaction pathway searches, to help identify unknown side products.

We also present a computational brewing application, COBRA, capable of simulating com- plex chemical mixtures. We demonstrate its efficacy at modeling both the photooxidation of isoprene, and the oxidation of squalene in the presence of ozone, by comparing predicted results with results obtained from high-resolution mass spectrometry.

In addition, we address the problem of atom-mapping for chemical reactions, by designing a new atom-mapping algorithm that can be used to annotate unmapped reactions.