Enzymes promote challenging chemical transformations in a highly specific manner and under mild reaction conditions. For this reason, they have found applications in various industries and will play a central role in the transition to a sustainable bioeconomy. However, despite the immense diversity of natural enzymes, there are many industrially relevant reactions for which no enzyme has been identified so far. This limitation has spurred research into the design of artificial enzymes that employ catalytic mechanisms not found in nature. A powerful design strategy to this end is to incorporate new metal ions or organometallic cofactors into proteins, as many (transition) metals are important catalysts in organic chemistry but absent from the repertoire of natural enzymes. Such artificial metalloenzymes (ArMs) could prove to be a crucial component of biocatalytic schemes for producing products ranging from fuels to pharmaceuticals in a sustainable fashion. Moreover, they would enable the design of entirely new metabolic pathways and thus allow for the creation of organisms with novel capabilities. However, as of today, such applications are still largely out of reach. This can be attributed to three major challenges that the field is facing: First, the reaction scope of ArMs still falls short of the diverse toolbox of synthetic chemistry. In order for ArMs to have a tangible impact, their reaction scope needs to be extended towards new reaction mechanisms, with a focus on those that are of particular relevance in organic chemistry. Second, most ArMs reported to date have a relatively low activity compared to natural enzymes. Therefore, they will need to be subjected to extensive protein engineering. As this is typically a cumbersome process that is largely governed by chance, there is a need for efficient and broadly applicable engineering strategies. Third, research in the field has focused on in vitro reactions, and only few ArMs have been shown to function in a cellular environment. Genuine in vivo application, for example for the design of novel pathways, can even be considered an outstanding challenge. This thesis addresses these challenges in the context of the biotin-streptavidin technology, which represents a popular and versatile strategy for creating ArMs. It utilizes the exceptionally strong interaction of the vitamin biotin and the homotetrameric protein streptavidin (Sav) to embed biotinylated organometallic cofactors in the Sav protein scaffold. This strategy has been used to create ArMs for several new-to-nature reactions. To bring additional reactions within the reach of biocatalysis, we developed the first ArMs for goldcatalyzed hydroamination and hydroarylation (Chapters 2&3). Neither the reaction mechanisms nor gold catalysis in general have been observed among natural enzymes. They are, however, of great interest in synthetic chemistry. Moreover, we created the first enzyme for atroposelective metathesis (Chapter 4). Atropisomerism is a type of axial chirality that is frequently encountered in medicinal chemistry. Therefore, this type of reaction could provide an attractive route for the synthesis of various pharmaceuticals. Relying on a biotinylated Hoveyda-Grubbs catalyst as a cofactor, we identified ArM variants that promote the formation of an atropisomeric binaphthalene compound by ring-closing metathesis with an enantiomeric ratio of up to 81:19, while the free cofactor delivers a racemic mixture. Although the yield of the reaction remained low, this represents the first example of atroposelective metathesis in an aqueous environment. Abstract II To address the low activity and limited biocompatibility of ArMs, we engineered ArMs towards higher activity in whole-cell assays. At the same time, we developed broadly applicable methods for systematic and data-driven enzyme engineering. Our first contribution in this respect is a screening platform that enables the rapid discovery of active ArMs for a wide range of reactions (Chapter 2). To this end, we created a sequence-defined library of 400 Sav double mutants by targeting two crucial amino acid residues close to the cofactor. In combination with a simple and adaptable screening assay based on whole-cell catalysis in 96-well plates, this enabled the discovery of active ArMs for five newto-nature reactions. Depending on the reaction, the best variants reached a five- to fifteen-fold higher cell-specific activity than the wild type. Given the high success rate and moderate screening effort, this screening platform is an attractive starting point for ArM engineering. Based on insights we gathered in the aforementioned project, we developed strategiesthat are meant to make more extensive engineering campaigns more efficient and accessible (Chapter 3). Such campaigns that target many amino acid positions at the same time are inevitable to obtain industrially competitive biocatalysts. However, as the search space grows exponentially with every additional position, finding optimized variants while keeping the experimental effort manageable becomes challenging. In recent years, machine learning has received increasing attention as a potential solution to this problem, but acquiring large and informative sequence-function data sets to train machinelearning models is often a serious obstacle. We demonstrate how targeted library design, lab automation and next-generation sequencing can be harnessed to gather such data sets in an efficient and inexpensive way. Relying on these technologies, we used active learning to simultaneously engineer five positions in an ArM for gold-catalyzed hydroamination. The machine-learning model increased the hit rate of our screening 14-fold compared to random mutagenesis, which underscores the potential of data-driven approaches to ArM engineering. The best ArM variant displayed an 18- fold higher cell-specific activity than the wild type, and 3-fold higher than the best variant identified in our previous screening. Lastly, we address the challenge of using ArMs for metabolic engineering (Chapter 5). Whole-cell biotransformation, as used for the aforementioned screenings, represents only a first step on the way to fully biocompatible ArMs. The latter calls for mutual compatibility of ArM catalysis and cell growth, as well as the ability to incorporate ArMs into metabolic pathways. We aimed to take a step in this direction by constructing an artificial tryptophan synthesis pathway in E. coli. Upon deactivation of the native biosynthetic pathway, this would result in a strain that relies on a new-to-nature, ArM-catalyzed reaction for growth. Besides the conceptual advance, this would also be highly interesting for protein and strain engineering, as these could benefit from the extremely high throughput of growth-based selection schemes. Gratifyingly, we observed that an artificial pathway incorporating an ArM-catalyzed deallylation indeed provided a growth advantage to an otherwise tryptophan-auxotrophic strain. Despite this encouraging result, selective enrichment of bacteria with active ArMs did not occur in mixed cultures. We were able to attribute this observation with a high likelihood to the release of Sav from the cell, resulting in ArM catalysis in the medium and the loss of the genotype-phenotype linkage. While this means that growth-based selection is not possible using the current system, it points to potential solutions for the future. All in all, this project can be seen as the first step towards metabolic engineering with Sav-based ArMs.