Graph neural networks can predict ketosynthase substrate specificity

Abstract

Abstract Modular polyketide synthases are a class of biosynthetic machinery responsible for the production of a chemically diverse range of natural products, including essential medicines. The modular architecture of this system presents as inherently rearrangeable; however, manipulation of these pathways to produce novel polyketides often results in a reduced output or the pathway stalling entirely. One cause of such breakdown is the ability of ketosynthases (KS), an essential component of the biosynthetic modules, to proofread the substrate produced by preceding steps and reject unexpected chemistries. Here we explore the KS proofreading phenomenon using Graph Neural Networks (GNN) trained on KS AlphaFold structures, as a means to predict when proofreading occurs and how to mitigate it. Our GNN models can distinguish six out of ten β-carbon reduction state pairwise combinations (81–92% AUC). The α-methylation state on the polyketide substrate can also be partially predicted from the KS structure (79% AUC). To explore the utility of the GNNs as a co-pilot in mitigating proofreading, we ran a proof-of-concept experiment, where we show that the GNN models can be used to predict beneficial mutations to a KS receiving a non-native substrate. We also present a pilot study on acyltransferase domains, which identifies novel regions potentially associated with substrate specificity.