Marian Menke

• Enzymes harbor immense potential for application in industrial synthesis processes, however the need for improvement in activity, selectivity, a broadened substrate scope or an increased thermal and organic solvent tolerance to meet the required non-natural industrial conditions requires optimization. Different protein engineering approaches are known and undergone to optimize specific properties of these biocatalysts. Machine learning emerged as a novel approach to alter the properties through algorithmic pattern recognition in data and has the potential for a paradigm shift in protein engineering. The complementary role of machine learning for protein engineering is exemplified for amine transaminases in Article I and Article II. Based on the key observations, a machine learning model was built that incorporated information regarding steric and electronic properties of amine substrates and active site residues. The predictor was improved in iterative computational prediction and experimental validation cycles and was used for the correct prediction of variants with improved activity and accurate activity estimation for a novel substrate. Article II envisioned further exploration of the potential of the machine learning predictor design with a focus on more challenging bulky amine substrates, prevalent in active pharmaceutical ingredient (API) precursors. Simultaneously, the key steps involved in the design of machine learning-guided protein engineering were formulated in a protocol. The predictor was successfully used for the prediction of higher active variants. In concept Article III the results of Article I are placed into context of the lack of machine learning descriptors for substrate scope enhancements. In addition, current limitations and potential improvements for data-driven protein engineering in the future are further elaborated on. Lastly, Article IV describes a developed high-throughput assay for S-adenosyl-L-methionine-dependent methyltransferases to address one of the restrictions mentioned in Article III. This facilitated data generation should accelerate methyltransferase engineering campaigns and could enable machine learning-based investigations in the future.