PMechRP: Interpretable Deep Learning for Polar Reaction Prediction

Ryan J Miller, Brayden Rudisill, Pierre Baldi, David Van Vranken

Keywords: chemistry, deep learning, interpretable, transformers, reaction prediction, mechanisms

Abstract: In recent years, machine learning based methods for chemical reaction prediction have garnered significant interest due to the time consuming and resource intensive nature of designing synthetic pathways. However, with the majority of models being trained on the US Patent Office dataset, many proposed architectures lack interpretability by modeling chemical reactions as overall transformations. These models map directly from reactants to products, and provide minimal insight into the underlying driving forces of a reaction. In order to improve interpretrability and provide insight into the causality of a chemical reaction, we train various machine learning frameworks on the PMechDB dataset. This dataset contains polar elementary steps, which model chemical reactions as a sequence of steps associated with movements of electrons. Through training on PMechDB, we have created a new system for polar mechanistic reaction prediction: PMechRP. Our findings indicate that PMechRP is able to provide both accurate and interpretrable predictions, with a novel two-step transformer based method achieving the highest top-5 accuracy at 89.9%.

Primary Area: Machine learning for physical sciences (for example: climate, physics)

Submission Number: 16792