Virtual Bioprospecting of Interfacial Enzymes: Relating Sequence and Kinetics

Abstract

Deposition of enzyme sequences greatly outruns any possibility of thorough experimental characterization. There seems to be a particular shortage of quantitative kinetic data, and this limits both structure–function analyses and the selection of biocatalysts for technical use. In this study, we present a virtual screening approach, which takes advantage of empirical scaling relations for interfacial enzymes in order to predict kinetic parameters from sequences. As an example, we analyzed an industrially important group of enzymes, namely, fungal cellulases from glycoside hydrolase family 7 (GH7). We screened this family and selected three previously uncharacterized enzymes, which were predicted to have high substrate-binding strength (a property that is desirable for biomass deconstruction). Generally, we found good agreement between the predicted and experimental kinetic parameters. In addition, one of the enzymes, Cel7C from Acremonium thermophilum, showed an unprecedented substrate-binding strength and outperformed the model enzyme, Cel7A from Trichoderma reesei by 50%, when tested on real biomass. We conclude that the method provides a means of computing kinetic parameters for hundreds of GH7 cellulases based only on the enzyme sequence, and surmise that similar approaches could be useful for other groups of enzymes within both engineering and discovery.

Binding prediction of multi-domain cellulases with a dual-CNN

“…We used simulated binding data from two previous studies of family GH6 and GH7 to pre-train a regression model. 5,6 Subsequently, we used a smaller set of experimental values 4 to fine-tune the regression model, resulting in a model with an estimated RMSE of 1.53 kJ/mol, outperforming previous computational screening methods for cellulases. Once trained, this regression model is also orders of magnitude faster than the previous MDbased method.…”

“…In two recent studies, we performed virtual screenings with MD-based binding estimates for two cellulase families, GH6 and GH7. 5,6 Family GH7 consisted of approx. 400 representative enzymes, family GH6 of about 1'200.…”

“…2 Previous work showed that the binding of the whole complex could be approximated by the binding of the CD towards a threaded chain and binding of the CBM on the crystalline surface (right). [3][4][5] Cellulases are typically modular, consisting of a catalytic domain (CD) and a carbohydratebinding module (CBM) connected by a flexible linker (see Fig. 1).…”

“…1). [3][4][5] This simplifies the setup drastically, as crystal structures of the CD in complex with threaded substrate exists and the interactions of the CBM towards the crystal surface are well studied. 2 Building up on this approximation, we recently proposed a cheap computational method based on molecular dynamics (MD) simulation.…”

“…4 While this method already has proven useful to screen through smaller enzyme families to discover new promising enzymes, it is still not feasible to compute properties of larger families or explore larger sequence space. 5,6 Here, machine learning (ML) techniques promise to capture the complexity of the underlying interactions, providing reasonable binding estimates from either sequence or structure in an even shorter time. [7][8][9] Additionally, when at least partially based on real experimental results, machine learning could not only lead to faster predictions but also to more accurate ones than compared to the semi-empirical MD-based 2D-LIE screening method.…”