By applying the principles of Darwinian natural selection in the laboratory, directed evolution has become a powerful practical approach to study enzymes and optimize them to catalyze industrially relevant transformations. In this thesis, I applied this strategy to the engineering of amino acid dehydrogenases for biocatalysis of chiral amines, focusing on two crucial features for successful directed evolution experiments. A first key aspect is the development of technologies allowing the screening of large libraries of enzyme variants to explore sequence space efficiently. Massive scale-down of assay volumes by compartmentalization of library members in water-in-oil emulsions has recently led to the development of ultrahigh-throughput screening platforms that allow sorting of more than 106 variants per hour. So far, these microfluidic droplet sorters have relied exclusively on fluorescent readouts. To further extend the range of applications toward enzymes for which no fluorescent assays are available, I successfully developed a sorting module based on absorbance detection. Using this new module, microdroplets could be sorted based on an absorbance readout at rates of up to 1 million droplets per hour. To demonstrate the utility of this module for protein engineering, three rounds of directed evolution were performed to improve a poorly stable NAD+ dependent phenylalanine dehydrogenase (PheDH) toward its native substrate. Five hits showed increased activity (improved up to 10-fold in lysate; kcat increased >3.5-fold), soluble protein expression levels (>2.5-fold) and thermostability (Tm, 8 °C higher). To increase the sensitivity of the device (3–4 orders of magnitude lower than fluorescence assays) for detection of enzymes with limited stability and low turnovers, an extra step of growth in droplets from single cell encapsulation, followed by piconinection of substrates and lysis agents was implemented. As a result, a fivefold signal enhancement over background was achieved, for an amine dehydrogenase (AmDH) reaction shown to be undetectable in a droplet single cell assay. Second, I investigated how mutational robustness may correlate with protein stability and lead to successful hits after mutagenesis and screening. To examine this issue, I initially investigated various approaches (including ancestral resurrection and computational design) to identify stabilized PheDH variants. One such variant (dubbed Pross 4) showed increased expression levels (>3.3-fold) and thermostability (Tm, 13 °C higher) compared to the wild-type PheDH. I further compared the mutational tolerance and the hit rate between PheDH and Pross 4 by generating variant libraries focused on key active site residues and screening them for improved AmDH activity. The Pross 4 background generated 6.4 times more active variants than the PheDH background, the best hits displaying increased activity (up to 2.5-fold in lysate; kcat/KM increased up to 8-fold) compared to previously engineered AmDHs with the PheDH scaffold. In conclusion, this work highlights how directed evolution experiments could be designed for increased success rates, by combining reliable high-throughput screens with careful choice of evolutionary robust starting points.