Comparison of Protein Design Tools For Engineering Non-heme Iron Dependent Dioxygenases

NOV1 (from Novosphingobium aromaticivorans DSM124444) and CsO2 (from Caulobacter segnis dioxygenase) are non-heme iron-dependent dioxygenases, belonging to the carotenoid cleavage dioxygenases enzyme family. Both NOV1 and CsO2 have a symmetric seven-bladded β-propeller structure and contain a mononuclear iron cofactor at their active sites. These enzymes possess the remarkable ability to catalyze the cleavage of interphenyl Cα-Cβ double bonds in a wide range of stilbene-like compounds with a 4’-OH group [1-2]. Vanillin, an important aroma chemical, used extensively in food and cosmetics, commands an enormous global demand. Vanillin production can be achieved through a one-step reaction, converting isoeugenol to vanillin, catalyzed by NOV1 [3]. Alternatively, vanillin can be obtained through a two-step biotransformation process: decarboxylation of ferulic acid into 4-vinylguaiacol, and subsequent oxidative alkene cleavage to vanillin catalyzed by CsO2 [4]. No dioxygenase was found to be directly active on ferulic acid, which is a more economically viable substrate for its direct transformation into vanillin. Our goal is to design NOV1 and CsO2 enzymes capable of catalyzing the reaction from ferulic acid to vanillin directly. We propose three distinct strategies to achieve this purpose. Firstly, employing the Rosetta protein design tool [5], a physics-based method, we redesigned the enzyme active site to stabilize the reaction transition states. This involves the identification of the reaction transition states through quantum mechanics calculations; using RosettaMatch algorithm to build a theozyme based on the transition states and key residues, and subsequently refining the enzyme active site to optimize this theozyme. Secondly, utilizing the graph neural network-based deep learning protein design tool, proteinMPNN [6-7]. We designed the sequences of NOV1 and CsO2, while fixing the active site residues and highly conserved residues within this enzyme family, to enhance the enzymes catalytic properties. Lastly, using RFdiffusion [8], a diffusion model-based deep learning protein design tool. We redesigned the binding pocket for the NOV1 and CsO2 enzymes to improve their catalytic activity. The results of these enzyme design strategies will be presented to compare the characterizations of these powerful protein design tools. To date, no publications using these protein design tools on non-heme iron-dependent enzymes are available. This advancement on the design of non-heme iron-dependent enzymes represents a practical application of deep learning protein design tools in biotechnology and industrial processes, providing valuable insights for future developments.