Computer-aided design of multi-enzyme networks

The use of enzymes to manufacture chemicals has enjoyed increasing popularity during the last years, providing an eco-friendly alternative to conventional synthesis. The project targeted the increased use of enzymatic reactions through the computer-aided prediction of the reactivity and substrate range of various enzymes, as well as enzymatic synthesis pathways to a given chemical. An extensive dataset of enzymatic reactions was curated and validated from literature, which is now building the foundation of many heuristic and machine learning models of enzymatic reactivity. This work allowed, for the first time, an accurate prediction of enzymatic reactions across many different enzyme and reaction types, demonstrating that enzymatic reactions, and thus enzymatic synthesis pathways, can be predicted on the computer. The models underlying these predictions were mainly based on machine learning, where new reaction representations and model architectures were developed to facilitate the prediction of enzymatic reactivity. To this aim, the existing concept of graph-based machine learning for molecules, where each molecule is represented as a graph (a collection of nodes representing atoms, and edges connecting the nodes representing bonds between atoms) was extended to reactions. In this way, the concept of graph-convolutional neural networks, a type of machine learning model where the property of a molecule is learned from the local information in the molecular graph, became applicable to enzymatic reactions for the prediction of the reactivity and selectivity of enzymes. Also, existing frameworks for computer-aided retrosynthesis, i.e. the task of teaching a computer how to identify possibly synthesis routes to a specified chemical, were adapted to enzymatic reactions. Together with the new database described above, this yielded accurate models for enzymatic synthesis planning for the first time. Further work incorporated the quantification of errors and uncertainties in such machine learning models, so that a user can be informed about uncertain, and thus possibly wrong predictions, by the model. The project thus enabled the data-driven, accurate prediction of enzymatic reactions on the computer, and advanced this important field of research, namely to harness machine learning and deep learning to enable a more eco-friendly, sustainable chemical synthesis.