Graph-Based Forward Synthesis Prediction of Biocatalyzed Reactions

Peter Mikhael, Itamar Chinn, Regina Barzilay

Track: Machine learning: computational method and/or computational results

Keywords: biosynthesis, enzyme, metabolism, chemical reactions, graph neural networks, forward synthesis

TL;DR: We present a novel approach that harnesses graph-based deep learning to predict the primary products of enzyme-catalyzed reactions, considering both the protein sequence and substrates involved.

Abstract: The identification of biocatalyzed reaction products plays a critical role in enzyme function prediction, drug discovery, and metabolic engineering. Uncovering the products of biocatalyzed reactions experimentally is both time-consuming and costly, which underscores the urgent need for computational methods. Previous machine learning methods have largely focused on spontaneous, non-biocatalyzed reactions but do not perform well when applied to biocatalyzed reactions specifically. We present a novel approach that harnesses graph-based deep learning to predict the primary products of enzyme-catalyzed reactions, considering both the protein sequence and substrates involved. On the recently published dataset EnzymeMap, we find that our method based on graph-editing outperforms existing transformer-based approaches.

Submission Number: 51