Explainable AI for metabolic engineering

Abstract

Metabolic engineering is an important field in biotechnology, aimed at optimizing cellular processes to produce desired compounds. In this thesis, we focus on predicting the metabolome from the proteome, as understanding this relationship is crucial for understanding cellular metabolism. We investigate the usage of additional biological information like protein-protein interactions and cellular stoichiometry to improve the predictive performance of metabolome prediction models. We also employ explanation algorithms to gain key insights into the regulatory processes of a yeast cell.

We demonstrate the effectiveness of our approach by predicting the metabolic fold-change of multiple yeast kinase knockouts. Our results show that incorporating additional biological information does not significantly improve the accuracy of the metabolome prediction models. Furthermore, we identified enzymes that are relevant for all metabolites used in this study, which indicates the existence of a global set of regulatory enzymes.

Overall, our study shows that through careful manipulation of the limit amount of data decent performance can be expected when predicting the metabolome. We apply a broad spectrum of machine learning algorithms to identify optimal model architecture. The methods and insights presented in this thesis could be used for creating a general pipeline for predicting a broad spectrum of metabolites from the proteome.