Enzymes, which act as catalysts for specific bioreactions under unique conditions, are valuable biological macromolecules that were originally found in living organisms. With the changing environment, some organisms have evolved differently to adapt to new environment conditions. However, not all organisms can overcome environmental stress, which results in the loss of enzyme activity. Thus, in this thesis, I first study soluble starch synthase I in wheat, which is largely affected by global climate change and causes a worldwide reduction in wheat yield. The first chapter characterizes the thermostability of 18 soluble starch synthase I enzymes from different crops, and through a thermal inactivation assay, identifies a more thermostable soluble starch synthase I compared to wheat. The results of this chapter can be applied for future genome editing to improve starch yield in wheat under high-temperature conditions. In addition, exploring the application of enzymes and applying it to our daily lives could largely be beneficial to humans’ daily applications. Many enzymes have been used in manufacturing for production or biodegradation. Therefore, the second chapter explores the application of carboxylic acid reductase and aldehyde-deformylating oxygenase in the biodegradation of perfluorooctanoic acid. The substrate promiscuity of carboxylic acid reductase was studied by measuring its kinetic parameters against different substrates.

Peptides are the important molecules in food protein that play a significant role in enhancing its flavor and taste. By controlling the generation of peptides, the flavor of food can be rebalanced to fit specific applications. Enzymatic hydrolysis of food proteins is a way to achieve the desired control on peptides, but it requires a careful selection of proteases and a controlled hydrolysis to maximize its potential.The main aim of this master thesis is to study S53 protease family for its applications in food sector. A previous study done in our lab suggested that this family consists of proteases with broad specificity (studied with synthetic substrates). Therefore, we hypothesized that they may not just degrade plant proteins but could also provide us with a diverse range of peptides (with varied sequences), where each peptide will have its own unique application in the food industry. To work on this hypothesis, firstly, we conducted gel electrophoresis experiments using the ten proteases (which could be expressed efficiently) of this family to investigate their proteolytic action towards four different proteins sources (chickpea, pea, sunflower, pumpkin). This initial study included various stages, starting from the identification of the dominant storage protein (in all the substrates) using SDS-PAGE followed by their digestion to produce hydrolysates. These hydrolysates were further studied by gel electrophoresis. Once, the proteases were tested for their activity, a bi-protease treatment was given to the protein sources with an aim to enhance the peptide generation. A total of 45 combinations were considered and an increase in peptide production in the resulting hydrolysates was studied using a florescence assay. A difference in relative fluorescent unit (RFU) values (obtained from the assay) was linearly related to the increase in peptide production from multi-protease treatment. SDS-PAGE results of protein sources showed various storage proteins in the range 15 to 75 kDa. All the S53 proteases were able to digest plant proteins, however with different proteolysis degree, signaling that different amount and composition of peptides were produced. Results from the florescence assay showed some protease combinations with RFU values greater than that of single-protease treatment. These findings helped us in screening some of the best protease blends for producing peptides which could be further studied to support new strategies for controlling the flavor of food proteins.

1,4 dicarbonyls are very important building blocks for many biologically and pharmaceutically active compounds including pharmaceuticals, pigments and pesticides. However, the synthesis of 1,4 dicarbonyls is a major bottleneck in industry. The conventional method for 1,4 dicarbonyl synthesis is the Stetter reaction, which features the addition of an aldehyde donor into an enone acceptor. One of the various drawbacks of this method is non-specific addition and creation side products such as 1,2 dicarbonyls. Enzymes are known to be highly specific and so far, two types of thiamine pyrophosphate (TPP) dependent enzymes have shown the ability to synthesize 1,4 dicarbonyl compounds and are termed stetterases: MenD and PigD (along with close homologs HapD and SeAAS). In this work, a major step is taken towards discovery of novel stetterases. A library of 77 thiamine pyrophosphate dependent enzymes is screened and assayed for stetterase activity with various donor and acceptor substrates. The substrate range of MenD is expanded and a novel stetterase is discovered. Homology modelling and catalytic site analysis is performed on known stetterases as a step towards stetterase design for specific substrates.

Chapter 1. Evaluating Protein Engineering Thermostability Prediction Tools Using an Independently Generated Dataset Engineering proteins to enhance thermal stability is a widely utilized approach for creating industrially relevant biocatalysts. The development of new experimental datasets and computational tools to guide these engineering efforts remains an active area of research. Thus, to complement the previously reported measures of T50 and kinetic constants, we are reporting an expansion of our previously published dataset of mutants forβ-glucosidase to include both measures of TM and ΔΔG. For a set of 51mutants, we found thatT50andTMare moderately correlated, with a Pearson correlation coefficient and Spearman’s rank coefficient of 0.58 and 0.47, respectively, indicating that the two methods capture different physical features. The performance of predicted stability using nine computational tools was also evaluated on the dataset of 51 mutants, none of which are found to be strong predictors of the observed changes inT50, TM, or ΔΔG. Furthermore, the ability of the nine algorithms to predict the production of isolatable soluble protein was examined, which revealed that RosettaΔΔG, FoldX, DeepDDG, PoPMuSiC, and SDM were capable of predicting if a mutant could be produced and isolated as a soluble protein. These results further highlight the need for new algorithms for predicting modest, yet important, changes in thermal stability as well as a new utility for current algorithms for prescreening designs for the production of mutants that maintain fold and soluble production properties.

Chapter 2. Construction and biophysical evaluation of 277 β-glucosidase mutants to elucidate structure, and function relationships Enzyme engineering revolutionizes biotech, pharmaceuticals, and therapeutics by improving stability and catalytic efficiency. Computational tools accelerate this, but quantitative predictions remain challenging. We expanded a Carlin mutant dataset to 277 diverse β-glucosidase B samples, uncovering independent engineering of thermostability and kinetics. Benchmarking revealed Rosetta and amino acid features predict expression, and Evolutionary Scale Modeling (ESM) predicts enzyme turnover rate. However, for more robust machine learning predictions, it's essential to emphasize the need for ongoing dataset expansion to enhance quantitative predictions and advance protein engineering tools.

Chapter 3. Computational design of human N-acetylglucosaminyltransferases hGnT-I mutants with improved solubility and expression in E. coliThis study aimed to enhance the soluble expression and stability of critical enzyme, human N-acetylglucosaminyltransferases GnT-1 (hGnT-I), which are involved in processing glycoprotein N-glycans. N-glycosylation is a vital quality attribute of glycoprotein and glycopeptide therapeutics, affecting their solubility, stability, safety, efficacy, and immunogenicity. However, producing these enzymes from mammalian sources in E. coli expression systems is challenging due to their complex post-translational modifications and folding. To overcome this hurdle, we explored the use of truncation and PROSS (Protein Repair One-Stop Shop), a computational approach that uses evolutionary information to suggest mutations that improve the soluble expression and stability of hGnT-1. Our designs successfully improved stability by enhancing molecular interactions and packing, leading to increased expression and activity. This study offers a novel approach to enhance GTs' expression and stability, providing a valuable resource for glycoprotein engineering and modification.

Biotechnology has the potential to deliver solutions to many global problems in medicine, materials science, nutrition, agriculture, natural resource preservation, and energy. Engineered cells and enzymes can perform chemical transformations that are rare or unknown in nature, and even catalyze reactions not accessible to traditional synthetic chemistry while also operating at gentler, more environmentally friendly conditions. Decreases in the price of DNA sequencing and synthesis has led to generation of vast databases that can be screened for any imaginable function. These sequence databases are even more powerful now due to the development of software to enable rapid generation of 3D protein structures, like AlphaFold2. However, tools to predict the function of these proteins or their performance in engineered cells are not yet robust, leading to long development times and limited successful applications to date. New tools and methods must be developed for the true potential of biotechnology to be unlocked.For my thesis, I explore metabolic pathway construction and screening, protein design, and protein sequence-structure-function relationships across broad contexts with the objective of tool and knowledge development for future efforts in biotechnology. My first chapter discusses the introduction of the triazine degradation pathway, of interest for remediation of contaminated sites, into E. coli and methods for characterizing pathway flux and identifying bottlenecks to guide engineering efforts and limit accumulation of metabolic intermediates. My second chapter focuses on methods for the design of supercharged proteins, which have many interesting potential applications, and parameters that increase the likelihood of successful design of these proteins. My third chapter regards the generation and characterization of a library of single point mutations in the enzyme B-glucosidase B and use of kinetic data to predict the effects of changes in sequence on enzyme function. These seemingly disparate topics all serve to improve tools for protein screening, production, functional prediction, and application, addressing several gaps toward improved development timelines and success rates for biocatalysts.

Humans have long possessed knowledge about the functions of proteins, even before the discovery of their sequences and structures. This ancient understanding is evident in various applications such as food, brewing, drugs, and more. However, it was not until the 20th century that biochemistry advanced significantly due to a deeper understanding of protein sequences and structures.

A widely accepted notion in the field is that a protein’s structure dictates its function. The precise orientation of amino acids, predetermined by the protein sequence, controls the biochemical properties of the protein. The study of protein function through sequence, structure, and biomolecular chemistry is known as protein function prediction, and the inverse problem in which the sequence is optimized for a target function and/or structure is named protein design.

The history of protein evolution is encoded in sequences, and this information can be extracted through conservation analysis. Position-specific scoring matrices (PSSMs) can be used as a proxy for evolutionarily critical functions, while sequence logos are commonly used to identify key functional residues. By overlaying consensus profiles onto structures, it is possible to gain a more comprehensive understanding of the relationship between sequence and structure, whereas, at the atomic level, domain experts identify functional residues, propose mechanisms for function, and verify hypothesesexperimentally based on their biochemical knowledge. All of these efforts are guided, at least in part, by human experts.

To automatically sample sequence and structure space in molecular modeling (MM), David Baker and his team introduced Rosetta. Rosetta was first developed for de novo structure prediction but has since expanded into homology modeling, protein design and docking. Like other molecular modeling techniques, Rosetta describes biomolecular interactions through score function. It captures the physical interactions between atoms and their statistics, and can be used to design for biological functions such as stability, enzymatic activity, protein-ligand affinity and protein-protein interaction. To refine the resolution further, machine learning (ML) models can be trained on energetics and other arbitrary features and deployed to model the functional landscape.

In contrast, without relying on external features, end-to-end models learn from sequence and/or structure directly. Protein language models (pLMs) identify patterns in protein sequences as universal sequence-to-sequence approximators, and are shown to be higher-order generalization of site-specific and pairwise conservation. The likelihood of recovering a masked amino acid in pLMs, similar to PSSMs, serves as an indicator of evolutionary preference and can link to protein functions. Language models trained on text completion can generate articles by predicting the next word iteratively, and likewise for protein sequences, pLMs are also shown to generate novel designs, sometimes conditioned on structures, species, family, E.C. numbers, and more.

This thesis touches on protein design problems in both the perspectives of molecular modeling and machine learning. The first chapter focuses on evaluating existing physics-based and machine-learning-based methods on predicting thermal stability, using beta-glucosidase as an independent case study. The second chapter dives into engineering RuBisCO oligomeric state through evolutionary analysis and molecular modeling in the context of protein-protein interface. The third chapter extends the investigation of protein-protein interactions, and presents a machine translation approach in pLM for designing the pairing partner when given either the heavy or light chain of an antibody. These findings demonstrate the possibility of protein design for thermal stability, oligomeric state and antibody engineering, and take a small step forward in potential applications in enzymatic manufacturing, carbon capture, and antibody therapeutics.

This work explores the vast field of computational enzyme design. Chapter 1 emphasizes the significance of characterizing proteins, with a focus on soluble starch synthases, an essential enzyme in wheat starch deposition. Chapter 2 introduces enzymatic cascades for discovering a novel approach to producing rare sugars. The utilization of computational modeling helps unveil the key enzymes in the cascade and elucidates a mechanism to identify phosphatases, driving the energetically favorable step for releasing these sugars. In the concluding Chapter 3, the combination of previous characterization data and computational calculations leads to the development of a novel oxidase with potential applications in biosensors. By integrating characterization information and computational workflows, the goal is to contribute to the understanding of how computational enzyme engineering shapes the future of biotechnology and advances various industries.

Enzyme discovery through mining nature’s biocatalysis dataset, also known as enzyme genome mining, has gained a lot of popularity in recent years, mainly because of the fast-growing number of sequenced genes available resulted from the development of DNA sequencing technologies. The newly found enzymes with novel activity or disparate specificity can drastically expand enzyme repertoire. However, when a target reaction is identified, how to recognize the candidates from a large pool of enzymes remain a problem. For my thesis, I started on a real-world problem, citrus bitterness. I explored the possible glycosyltransferases to solve the problem. Secondly, I proposed an integrative pipeline of functional profiling and machine learning models to predict and decipher the complicated glycosyltransferase’s acceptor specificity puzzle. In my third project, I genome mined ketoacid decarboxylases for oxaloacetate decarboxylating reaction, which is the key reaction to build a useful biosynthetic pathway of 3-HP. For the fourth and final project I have focused on, I applied computational enzyme design techniques to tailor the enzyme active site to utilize an unnatural redox cofactor. Overall, I will demonstrate the effectiveness of the combination of enzyme genome mining and enzyme engineering in food processing, natural product formation, and synthetic biology. Chapter 1. A major problem in the orange industry is “delayed” bitterness, which is caused by limonin, a bitter compound developing from its non-bitter precursor limonoate A-ring lactone (LARL) during and after extraction of orange juice. The glucosidation of LARL by limonoid UDP-glucosyltransferase (LGT) to form non-bitter glycosyl-limonin during orange maturation has been demonstrated to be the natural way to debitter by preventing the formation of limonin. Here the debittering potential of heterogeneously expressed maltose-binding protein (MBP) fused-cuLGT from Citrus unishiu Marc in Escherichia coli is evaluated. An LC-MS method was established to determine the concentration of limonin and its derivatives. The protocols to obtain its potential substrates, LARL and limonoate (limonin with both A and D ring open), were also developed. Surprisingly, MBP-cuLGT didn’t exibit any detectable effect on limonin degradation when Navel orange juice was used as the substrate and was unable to biotransform either LARL or limonoate as purified substrates. However, it was found that MBP-cuLGT displayed a broad activity spectrum towards flavonoids, confirming the enzyme produced was active under the conditions evaluated in vitro. Our results demonstrate that cuLGT functionality was incorrectly identified, and highlight the need for further efforts to identify the enzyme responsible for LGT activity in order to develop biotechnology based approaches for producing orange juice from varietals that traditionally have a delayed bitterness. Chapter 2. Accurate prediction of enzyme function and substrate specificity remains a long-lasting problem. Sequence similarity- and structure- based function annotation methods provide a coarse prediction ability, but it is not sufficient to distinguish the nuanced substrate specificity and scope among the closely related enzymes. Previous efforts on glycosyltransferase family 1 (GT1) function prediction such as glycosyl acceptor selectivity perform poorly in that GT1 orthologues display diverse function profiles utilizing a broad spectrum of substrates and available computational algorithm failed to recapitulate the complex sequence-function relationship. Here, throughout characterization of 22 selected evolutionary distant GT1s representing the entire GT1 protein space are performed against 47 structurally and functionally diverse acceptors. The acquired activity dataset enables us to train and evaluate different machine learning models with features encoding multiple aspects of the reactions thought bioinformatics and cheminformatic analysis. The random forest-based model, named with GtRF, has the best performance, as gauged by various statistic indicators, with advantage of interpretability and generality. GtRF also shows the comparable accuracy and recall when it is adopted for forward prediction on novel enzyme-substrate pairs. Finally, we proposed an extensive usage of the predication models to annotate the whole GT1 family. Chapter 3. 3-Hydroxypropionate (3-HPA) is a 3-carbon precursor to many commodity and specialty chemicals. The construction of workhorse microorganisms with economically viable biosynthetic pathways is essential for the production of 3-HPA. Despite different pathways reported, all suffer from a variety of drawbacks, including low efficiency and yield, using expensive substrates and generating multiple side products. Here, a novel routine to produce 3-HPA using glucose as a feedstock is proposed. This pathway converts phosphoenolpyruvate, one of the metabolites in glycolysis, into oxaloacetate. Decarboxylation reaction of oxaloacetate is then involved to obtain 3-oxopropanoate, followed by a reduction reaction to produce 3-HPA. However, no enzymes have been reported to catalyze the decarboxylation reaction of oxaloacetate to 3-oxopropanoate. To find an active enzyme towards oxaloacetate, we applied 2 rounds of genome enzyme mining in thiamine diphosphate (TPP) dependent ketoacid decarboxylases and discovered a panel of enzymes with promiscuity towards oxaloacetate. The most active candidate D1Y3P7 from Pyramidobacter piscolens was then rationally engineered; However, we did not achieve increased catalytic efficiency. We also screened and characterized the dehydrogenases and pyruvate carboxykinases that are favorable for the pathway construction. All of these enzymes were tested in in-vitro system and current work served as the basis for next step fermentation experiment for the production of 3-HPA. Chapter 4. Unnatural redox cofactors which operate in an orthogonal manner to NAD(P) + in vivo may mediate specific reducing power delivery in whole-cell biomanufacturing. Furthermore, using less expensive analogs of NAD(P) + in vitro may greatly reduce the cost of cell-free biotransformation. Here, we have developed an unnatural redox cofactor system based on NMN+ . We explored its applications in lower-cost redox cycling in vitro by accomplishing two proof-of-concept example enzymes, Bs GDH and Ec GapA. This was enabled by a computationally designed methods and achieved resulting specificity switch towards an unnatural nicotinamide cofactor: GDH design with a 107 -fold cofactor specificity switch towards NMN + over NAD(P) while GapA double mutant A180S-G10R with ~200-fold cofactor specificity switch from NAD + to NMN+ . Overall, this work opened opportunities to design biochemistry using unnatural redox cofactors.

Enzymes play a monumental role in medicine, materials science, food science, agriculturaltechnology, and energy. These natural catalysts allow us to explore chemical reactions that often exceed the capabilities of traditional synthetic chemistry, with the advantage of usually operating under mild conditions and temperatures. Even so, there valuable reactions for which either there exists no natural enzyme or we haven’t discovered it yet. To address this problem, we have taken notes from nature and used its powerful tool of evolution to alter the functionality of enzymes to impose the activity we want. Now, with genome sequencing being practically free, and the sequences of millions of proteins available at our fingertips, the possibilities for engineering any reaction we want in an enzyme are limited only by our imaginations. These tools combined with the unprecedented power of computers to explore sequence space, and to detect patterns humans can’t even see present the opportunity to design novel enzymes and rapidly re-engineer enzymes to service nearly any industry. For my thesis, I used the power of enzyme engineering to explore the production of flavonoids, a secondary metabolite found in plants, and specifically anthocyanins, a colorful class of compounds sought after as a natural source of food coloring by food manufacturers aiming for cleaner labels. Flavonoids are a massive class of compounds that are highly similar, often differing in a single sugar or glucose group. These similarities make them an ideal target to exploit and engineer to regio- and stereo- specificity of enzymes to biosynthesize these products for use in medicine and material sciences. I explore using enzymes to make them in several ways. My first chapter looks at building them from the ground up through the investigation of glycosyltransferases, a massive and extremely valuable class of enzymes. I then look at isolating single anthocyanins from complex mixtures via selective degradation with engineered enzymes, a process that could be easy to implement and valuable for manufacturing these compounds at an industrial scale.