Engineering faster transglycosidases and their acceptor specificity

Article abstract of DOI:10.1039/c9gc00621d

Transglycosidases are enzymes that have the potential to catalyze the synthesis of a wide range of high-value compounds starting from biomass-derived feedstocks. Improving their activity and broadening the substrate range are important goals to enable the widespread application of this family of biocatalysts. In this work, we engineered 20 mutants of the rice transglycosidase Os9BGlu31 and evaluated their catalysis in 462 reactions over 18 different substrates. This allowed us to identify mutants that expanded their substrate range and showed high activity, including W243L and W243N. We also developed double mutants that show very high activity on certain substrates and exceptional specificity towards hydrolysis, such as L241D/W243N. In order to guide a more general use of Os9BGlu31 variants as transglycosylation catalysts, we built cheminformatics models based on topological descriptors of the substrates. These models showed useful predictive potential on the external validation set and are allowing the identification of efficient catalytic routes to novel phytohormone and antibiotic glucoconjugates of interest.

Full text of DOI:10.1039/c9gc00621d

View Article Online
View Journal
Green
Chemistry
Accepted Manuscript
This article can be cited before page numbers have been issued, to do this please use: V. Blay, L. T. Tran,
S. Luang, C. Eurtivong, S. Choknud, H. Gonzalez-Diaz and J. R. Ketudat Cairns, Green Chem., 2019, DOI:
10.1039/C9GC00621D.
This is an Accepted Manuscript, which has been through the
Volume 18 Number
7 7 April 2016 Pages 1821–2242
Royal Society of Chemistry peer review process and has been
accepted for publication.
Green
Chemistry
Accepted Manuscripts are published online shortly after
Cutting-edge research for a greener sustainable future
www.rsc.org/greenchem
acceptance, before technical editing, formatting and proof reading.
Using this free service, authors can make their results available
to the community, in citable form, before we publish the edited
article. We will replace this Accepted Manuscript with the edited
and formatted Advance Article as soon as it is available.
You can find more information about Accepted Manuscripts in the
author guidelines.
Please note that technical editing may introduce minor changes
to the text and/or graphics, which may alter content. The journal’s
standard Terms & Conditions and the ethical guidelines, outlined
in our author and reviewer resource centre, still apply. In no
event shall the Royal Society of Chemistry be held responsible
for any errors or omissions in this Accepted Manuscript or any
consequences arising from the use of any information it contains.
ISSN 1463-9262
CRITICAL REVIEW
G. Chatel et al.
Heterogeneous catalytic oxidation for lignin valorization into valuable
chemicals: what results? What limitations? What trends?
rsc.li/green-chem

Page 1 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Engineering Faster Transglycosidases and their Acceptor Specificity
Linh T. Trana,b,†,§, Vincent Blayc,d,§,*, Sukanya Luange, Chatchakorn Eurtivongf, Sunaree Choknuda,b,
Humbert González-Díazg,h, James R. Ketudat Cairnsa,b,g,*
aSchool of Chemistry, Institute of Science, Suranaree University of Technology, Nakhon Ratchasima 30000,
Thailand.
bCenter for Biomolecular Structure, Function and Application, Suranaree University of Technology, Nakhon
Ratchasima 30000, Thailand.
cFisher College of Business, The Ohio State University, Gerlach Hall, 2108 Neil Ave. Columbus, OH 43210,
United States.
dMolecular Topology & Drug Design Research Unit, Departamento de Química Física, Universitat de
València, Av. V. A. Estellés, s/n, 46100 Burjassot, Spain.
eDepartment of Biochemistry, Faculty of Medicine, Khon Kaen University, Khon Kaen 40002, Thailand.
fProgram of Chemical Biology, Chulabhorn Graduate Institute, Chulabhorn Royal Academy of Science,
Bangkok 10210, Thailand.
gDepartment of Organic Chemistry II, University of the Basque Country UPV/EHU, 48940, Leioa, Spain.
hIKERBASQUE, Basque Foundation for Science, 48011, Bilbao, Spain.
gLaboratory of Biochemistry, Chulabhorn Research Institute, Bangkok 10210, Thailand.
ABSTRACT. Transglycosidases are enzymes that have the potential to catalyze the synthesis of a wide
range of high-value compounds starting from biomass-derived feedstocks. Improving their activity and
broadening the substrate range are important goals to enable the widespread application of this family of
biocatalysts. In this work, we engineered 20 mutants of the rice transglycosidase Os9BGlu31 and evaluated
their catalysis in 462 reactions over 18 different substrates. This allowed us to identify mutants that expanded
their substrate range and showed high activity, including W243L and W243N. We also developed double
mutants that show very high activity on certain substrates and exceptional specificity towards hydrolysis,
like L241D/W243N. In order to guide a more general use of Os9BGlu31 variants as transglycosylation
catalysts, we built cheminformatics models based on topological descriptors of the substrates. These models
showed useful predictive potential on the external validation set and are allowing the identification of
efficient catalytic routes to novel phytohormone and antibiotic glucoconjugates of interest.
Keywords: Catalysis; Transglycosylation; Protein Engineering; Biocatalysis; Regioselectivity; Machine
Learning; Neural Networks; Docking.
1

Green Chemistry
Page 2 of 29
View Article Online
DOI: 10.1039/C9GC00621D
1. INTRODUCTION
Glycosylation is an essential mechanism for building structural and bioactive components of cells and their
matrix and for providing blocking groups that restrain reactive and bioactive molecules from disrupting the
metabolism of the cells in which they are synthesized.1 Understanding the roles of the various glycosides in
nature requires their synthesis or purification from natural sources, where they are often found in minute
quantities. Moreover, the glycosylation of some bioactive compounds may modulate their uptake and
bioactivity, which holds interest for the fine chemical and pharmaceutical industries. The large-scale
synthesis of these glucoconjugates may require complicated strategies or expensive starting materials, while
their extraction and purification may require large amounts of solvents. Therefore a simple and efficient
system to produce such compounds in aqueous solvents is highly desirable.
Generally, glycosylation begins with the activation of a sugar by its attachment to a nucleotide, yielding
nucleotide sugars, such as UDP-glucose (uridine diphosphate-α-D-glucopyranoside). The sugar can then be
transferred by a Leloir-type glycosyltransferase (GT) to a carbohydrate or noncarbohydrate molecule to make
a glycoside.2 Glycosides and carbohydrates are broken down by cleavage of the glycosidic bond created by
the GTs by glycoside hydrolases (GHs), usually by hydrolysis. However, many GHs are able to catalyze
transglycosylation, that is, the transfer of the glycosyl moiety to another carbohydrate or aglycon moiety
(Figure 1). Some enzymes related to retaining GH act as transglycosidases (TGs, also referred to as non-
Leloir-type GTs), which preferentially catalyze transglycosylation rather than hydrolysis.
The transglycosidases that have been studied to date belong to families of retaining GHs and are believed to
catalyze transglycosylation by the same mechanism used by the related GHs, shown in Figure 1. The
retaining mechanism of GH (and GT) starts with a nucleophilic residue in the active site attacking the
anomeric carbon. At the same time, the departure of the aglycon moiety is facilitated by a general acid in the
active site. Since the glucosyl moiety becomes covalently bound to the nucleophilic residue, this first step is
called glycosylation.3-4 In the second step (deglycosylation), a nucleophilic acceptor molecule can attack
from the same side as the aglycon departed. This is a ping-pong bi-bi mechanism, in which two SN2 reactions,
with inversion of stereochemistry at each step, return the anomeric carbon to its original stereochemical
configuration. In the case of hydrolysis, water acts as the nucleophile in the deglycosylation step, whereas in
transglycosylation another molecule acts as the nucleophile.
2

Page 3 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Figure 1. Mechanism of transglycosylation and hydrolysis reactions in retaining GH and GT. The enzyme
complexes are labeled with the symbols used in the general nomenclature for a ping-pong bi-bi mechanism.
The glycosyl intermediate is labeled E*. The second substrate is water in the case of hydrolysis or a hydroxyl
group (R2OH) in the case of transglycosylation. The glycon moiety shown is glucose, in line with the
transglucosidase activity described in the current paper.
As noted above, TGs are related to GHs and have been included in the GH families grouped by amino acid
sequence similarity at the carbohydrate active enzyme database (www.cazy.org). Most of the
transglycosidases that have been described catalyze synthesis and remodeling of oligo- and polysaccharides.
For instance, xyloglucan endotransferases (XET) belong to the family GH16, which also contains xyloglucan
endohydrolases. Differences in a surface loop in the XET and the xyloglucan endohydrolases have been
shown to contribute to the transferase versus hydrolase specificity5, but this principle cannot be readily
transferred to other families. Several attempts have been made to engineer GH to make TG, including the
production of a β-transglucosidase from a bacterial GH1 β-glycosidase6, a trans-α-L-arabinofuranosidase
from GH517, as well as α-glucosyl transferases for the production of oligosaccharide antigens8,9. In each of
these cases, the target products were oligosaccharides, whereas relatively little work has been done on the
use of TG to synthesize glycosides.
The use of directed evolution approaches allowed the identification of changes that led to higher
transglycosylation to hydrolysis ratios in the GH that were converted to TG.6,7 While it was anticipated that
decreasing the binding of water and increasing sugar binding could increase transglycosylation, it was
3

Green Chemistry
Page 4 of 29
View Article Online
surmised that several of the mutations actually worked by decreasing the function of the catalytic nucleophile
DOI: 10.1039/C9GC00621D
or acid/base. This was hypothesized to prolong the lifetime of the intermediate by destabilizing the transition
states of the two reaction steps (Figure 1) and by decreasing the ability of water, a weaker nucleophile that
requires deprotonation, to compete with sugars as the acceptor in the deglycosylation step. This strategy
resembles the action of glycosynthase and thioglycoligase mutations of retaining GH, in which the
nucleophile and acid/base, respectively, are mutated to non-nucleophilic, non-ionizable residues, thereby
allowing the transfer of the glycon from activated donor glycosides to suitable sugar acceptors without
hydrolysis10,11
.
In the last several years, some GH1 enzymes that catalyze the transglycosylation of lipids, anthocyanins, and
other noncarbohydrate small molecules have been described12-16. The sensitive to freezing 2 (SFR2) gene in
Arabidopsis thaliana was found to encode a galactolipid/galactolipid galactosyltransferase. Several
anthocyanin glucosyltransferases were found to actually be GH1 TG13,15,16, whereas rice (Oryza sativa)
Os9BGlu31 was found to be a general TG that can transfer glucose between phenolic acids, phytohormones,
and flavonoids and it also appears to deglycosylate fatty acids and other substrates in the plant14,17. Little
engineering of these types of enzymes has been reported, although we recently demonstrated that mutation
of the active site cleft residue Trp243 (W243) to Asn (W243N) increased the activity and broadened the
specificity of Os9BGlu3118.
Promiscuous GT have been used to glycosylate several medicinal compounds to modulate their solubility
and bioactivity19. For instance, several microbial antibiotics are glycosides and these promiscuous GTs have
allowed the transfer of unusual sugars in glycodiversification to develop new or more robust activities20,21
.
From an industrial point of view, the development of transglycosidase (bio)catalysts would be desirable
because they do not require a nucleotide sugar substrate or intermediate in the transfer of sugars to
compounds of interest. Ideally, the sugars could be obtained from biomass22,23. Recently, cyclodextrin
glucosyltransferases from Thermoanaerobacterium sp. were reported to transfer glucose onto aryl
glucopyranosides and furanosides, achieving an unusual substrate specificity toward alkyl furanosides24.
Although most of the transglycosidases studied to date show a high specificity for the transfer of the glucosyl
moiety to certain sugars or aglycon groups, the promiscuity of Os9BGlu31 makes it especially promising to
aid in the synthesis of a wide range of compounds14,18. Nonetheless, wild-type Os9BGlu31 cannot efficiently
transfer glucosyl groups to all acceptors, so new variants with broader or different specificity are of interest
to increase its potential as a biosynthetic tool. The development of catalysts that work under mild conditions
4

Page 5 of 29
Green Chemistry
View Article Online
with environmentally friendly reagents would greatly facilitate the production of compDoOuI:n1d0.s10f3o9/rC9sGtuC0d0i6e2s1Dof
phytohormone metabolism1 and antibiotic glycosides with modified pharmacokinetic properties19.
In this work, we have explored the potential of rice Os9BGlu31 active site cleft mutations to broaden its
acceptor potential by engineering and evaluating 14 different amino acids at residue position 243. Moreover,
we combined the mutation W243N with hydrophilic mutations of residues in positions that could interact
with water during hydrolysis and evaluated their effect on the hydrolysis and transglycosylation specificity.
Since the structure for Os9BGlu31 has not been solved yet, homology models of the covalent glycosyl-
enzyme intermediate were made to evaluate the interaction of acceptor substrates with the active site.
Although docking into such models provides insight into possible modes of binding, it could not
quantitatively predict activity differences, due to the low resolution of the models. Therefore, we have
applied cheminformatics tools to obtain further insights and anticipate the behavior of the different
transglycosidases when presented new substrates.
There are two main possible approaches to build a cheminformatics model on experimental results. In the
absence of a mechanistic hypothesis, it is possible to compute many molecular descriptors and then select
statistically a phenomenological route that may be possible to interpret a posteriori.25,26 Alternatively, one
can hypothesize a mechanistic route that selects a reduced number of physicochemical descriptors
beforehand to build a model with them.27,28 In this work, we explored both approaches and built artificial
neural network (ANN) models on a subset of physicochemical descriptors that can predict the activity of
multiple enzymes against new substrates of interest.
2. MATERIALS AND METHODS
2.1 Construction of pET32a/DEST/TEV/Os9BGlu31
The wild-type pET32a/DEST/TEV/Os9BGlu31 expression vector was constructed as follows. The cDNA
fragment encoding the mature Os9BGlu31 gene was amplified with the AK121679F and AK121679R
primers indicated in Table S1 by Pfu DNA polymerase with the AK121679 clone plasmid as the template.
The PCR product (~1.5 kb) was purified from the agarose gel and cloned into pENTR/TEV/D-TOPO
Gateway entry vector (Invitrogen) and incubated at 22 °C for 18 h. The entry clone size was checked by SacI
restriction endonuclease digestion. Then, the pENTR/TEV/D-TOPO/Os9BGlu31 was recombined into the
pET32a/DEST destination vector29 by Gateway LR Clonase (Invitrogen) reaction and cloned in Escherichia
5

Green Chemistry
Page 6 of 29
View Article Online
coli strain DH5α, selected on 50 μg ml-1 ampicillin LB agar. The recombinant expressiDoOnI:v10e.1c0t3o9r/Cc9GloCn00e62w1Das
verified to contain the inserted gene by digestion with EcoRI restriction endonuclease and DNA sequencing
(Macrogen Corp.).
2.2 Site-directed mutagenesis of Os9BGlu31
The tryptophan residue W243 in wild-type Os9BGlu31 was changed to 14 different residues by QuikChange
site-directed mutagenesis (Agilent) with the pET32DEST/TEV/Os9BGlu31 vector as the template and the
primers shown in Table S1. These mutants comprise cysteine (C), glutamic acid (E), glycine (G), histidine
(H), isoleucine (I), lysine (K), leucine (L), asparagine (N), proline (P), glutamine (Q), arginine (R), serine
(S), threonine (T) and valine (V). Furthermore, double mutants I172T/W243N, L183Q/W243N and
L241D/W243N were prepared by introducing one of the mutations I172T, L183Q or L241D in the vector,
in addition to W243N. The sequences of all mutant clones were confirmed by DNA sequencing (Macrogen
Corp.).
2.3 Expression and purification of Os9BGlu31 and its mutants
The pET32a/DEST/TEV/Os9BGlu31 wild-type plasmid and all the mutants were transformed into Origami
B(DE3) and the proteins expressed, extracted, and purified by an initial immobilized metal ion affinity
chromatography (IMAC) step as previously described for the pET32a/DEST/Os9BGlu31 expression
vector14,18. All of the purified fractions were assayed in 150 μl total volume with 2 mM 4-nitrophenyl β-D-
glucopyranoside (4NPGlc) in 50 mM acetate buffer, pH 4.5, at 30 °C for 30 min. The reactions were stopped
by adding 75 μl of 2 M Na2CO3 and the absorbance at 405 nm was then measured. The fractions showing
activity with 4NPGlc were pooled and the imidazole was removed by buffer exchange with equilibration
buffer (150 mM NaCl in 20 mM Tris HCl, pH 8.0) in 30 kDa molecular weight cutoff (MWCO) centrifugal
filters. The N-terminal fusion tag was removed by cleavage with 1 mg TEV protease per 50 mg of the fusion
protein at 4 °C for 16 h. Then, the digested proteins were loaded onto a second IMAC column in equilibration
buffer, and the flow-through fractions showing activity to cleave 4NPGlc were pooled and concentrated with
30 kDa MWCO centrifugal filters.
2.4 Relative activities of Os9BGlu31 wild type and mutants
The activities of Os9BGlu31 and its mutants were compared upon a range of acceptor substrates for
transglycosylation and water as the acceptor for hydrolysis in 50 mM citrate buffer (pH 4.5). Unless stated
otherwise, the enzymatic assays were set up with 0.25 mM acceptor, 2.5 vol.% dimethyl sulfoxide
6

Page 7 of 29
Green Chemistry
View Article Online
(DMSO), 5 mM 4NPGlc as the donor, and varying amounts of wild-type Os9BGlu31 and its mutants so
DOI: 10.1039/C9GC00621D
that conversion of the limiting substrate was below 10% (Supporting Information files SI01.xlsx and
SI02.xlsx). The reactions were conducted at 30 °C for 15 min, during which the concentration of products
evolved linearly with time, and they were stopped by the addition of 1% formic acid. The reaction mixtures
were centrifuged at 10,000 g for 10 min to remove the enzymes, and the supernatants were evaluated by
reverse-phase UPLC, as previously described18. Briefly, 2 μL of the reaction mixtures were injected into a
ZORBAX SB-C18 (1.8 µm, 2.1 x 150 mm) column equilibrated in 95% solvent A (0.2% formic acid in
water) and 5% solvent B (0.2% formic acid in acetonitrile) in an Agilent 1290 UPLC with a diode-array
detector (DAD). The compounds were eluted by a linear gradient from 5% to 50% B (v/v) for 13 min,
50% to 70% B (v/v) for 1 min, and 70% to 5% B (v/v) for 2 min, at a flow rate of 0.3 ml min-1. Relative
activities were evaluated from the absorbance at 360 nm of the 4-nitrophenol (4NP) released, which eluted
at 10.5 min.
2.5 Transglycosylation of substrates to multiple glucoconjugates by Os9BGlu31 variants
To identify differences in glycosylation on one acceptor that has multiple nucleophilic groups, the products
of transglycosylation by Os9BGlu31 wild type and W243 mutants (C, I, L, N, T, V) were determined in the
reaction with 5 mM 4NPGlc as the donor, 0.5 mM acceptor, and 5 µg of proteins in 50 mM citrate buffer
(pH 4.5). The reactions were conducted at 30 °C for 1 hour and then stopped with 1% formic acid. The
mixtures were injected into the UPLC-DAD as described in section 2.4. Absorbance was measured at
wavelengths between 190 and 500 nm to detect the glucoconjugate compounds.
2.6 Linear cheminformatics model. The glucosyl acceptors were characterized by 1D, 2D, and 3D
molecular descriptors of different classes (constitutional, molecular format, autocorrelation, Basak, burden,
connectivity, topological, charge descriptors, etc.). 1132 descriptors were computed using the software
ChemDes30, which integrates multiple state-of-the-art packages (Pybel, CDK, RDKit, Chemopy, etc.). 1824
additional molecular descriptors were computed with the newly released software Mordred31, which
computes 3D conformers by MM optimization and implements fully revised algorithms for popular indices
found in Dragon and PaDEL.
To select the descriptors that will participate in the model, we first removed any index with zero variance
across the range of substrates studied. We then reduced the set of descriptors under consideration by looking
at the pairwise correlation between all of them. The correlation coefficient matrix was processed by the
findCorrelation function in the caret R package32. We established the cutoff that no pair of descriptors had
7

Green Chemistry
Page 8 of 29
View Article Online
a correlation coefficient > 0.8. When a pair of descriptors has a correlation above the cutoff, findCorrelation
DOI: 10.1039/C9GC00621D
computes the mean absolute correlation of each descriptor in the pair to all others and removes the one with
the largest mean absolute correlation. This step reduces the redundancy and the number of descriptors under
consideration. In a second step, we ranked the descriptors using as the criterion the information gain as
defined in the FSelector package33 and the Pearson correlation of each individual descriptor to the average
enzymatic activity measured in triplicate experiments. The top descriptors were then considered by a forward
stepwise selection algorithm (StepAIC, MASS package) to build the multiple linear regression model34. A
scheme of the pipeline devised is shown in Figure S1.
2.7 Nonlinear ANN models. The molecular descriptors used in the nonlinear models are D1 = ALOGP (an
estimate of logP), D2 = MR (molar refractivity), and D3 = TPSA (topological polar surface area), which can
be found in file SI03.xslx. STATISTICA 10 was used to implement Artificial Neural Networks (ANN) as
nonlinear ML models. The ANNs tested are Multi-Layer Perceptrons (MLP) with one dense hidden layer.
The MLPs models have up to 7 hidden neurons in the hidden layer. Different activation functions were
evaluated for the hidden and output layers. The models were trained within 200 cycles of the Broyden-
Fletcher-Goldfarb-Shanno (BFGS) backpropagation algorithm.
2.8 Docking study. The Os9BGlu31 model was calculated in MODELLER 9.1935 based on the structure of
Os3BGlu6 in covalent complex with 2-fluoroglucoside (PDB ID: 3GNR)36. To understand the structure of
Os9BGlu31 in complex with glucose, the structure of Os3BGlu6 in complex with 2-fluoroglucoside structure
was modified by replacing the fluorine at the C2 with oxygen. Five models each of Os9BGlu31 and
Os9BGlu31/glucose complexes were generated from the Os3BGlu6 covalent complex with 2-
fluoroglucoside. The models with the lowest scores of the MODELLER Objective Function (MOF)37 were
selected. Os9BGlu31 mutants with and without glucose were generated from the Os9BGlu31 wild type
models using FoldX38 after insertion of the covalently linked α-D-glucoside coordinates into the models with
glucose. The final models were evaluated for quality using ProSa200339 and PROCHECK software40.
The acceptor substrate 3D structures were created with the drawing and manipulation tools in the Scigress
3.3.2 software suite. The homology models were protonated. The acceptor substrate structures and the newly
added protons, glucoside, and mutated residues were then subjected to energy minimization using the MM2
force-field. Genetic Optimization for Ligand Docking (GOLD) software version 5.6.2 was used for the
docking calculations. The center of binding was defined at coordinates (x = -0.581, y = -6.381, z = -21.102)
with a 10 Å radius. Thirty docking runs were allowed for each ligand with default search efficiency (100%).
8

Page 9 of 29
Green Chemistry
Lysine and arginine were defined as protonated; aspartic and glutamic acid were consiDdOeIr: e10d.10d3Ve9i/epCwr9AoGrttCico0len0O6an2tl1ineDed.
The GoldScore scoring function was used to evaluate the predicted binding modes41.
3. RESULTS AND DISCUSSION
3.1 Activity of wild-type Os9BGlu31 and mutants
Previously, we showed that wild-type Os9BGlu31 and a several mutants are able to catalyze
transglycosylation, transferring a glucosyl moiety from a donor substrate such as 4NPGlc to an acceptor14,18
.
In this study, we studied the effect of amino acid substitutions in the position W243 that complete the set of
20 natural amino acids at this position, when combined with our previous work18. These included the W243N
mutant that had highest activity in our previous study and 13 new mutations (W243C, E, G, H, I, K, L, P, Q,
R, S, T, and V). We also introduced the second point mutations I172T, L183Q, and L241D in the W243N
mutant with the goal to test whether multiple hydrophilic substitutions increase the hydrolysis to
transglycosylation ratio and broaden the substrate/product specificity to offer novel biosynthetic possibilities.
The Os9BGlu31 wild type and its mutants were evaluated in the glycosylation of 22 acceptors, including
water (for hydrolysis). The relative activities of Os9BGlu31 wild type and its mutants are measured by
quantifying the release of 4NP from the reaction of 4NPGlc with the different acceptors. The activity results
are represented in Figure 2. Details on the activities are provided in the Supporting Information file SI01.xlsx.
Among the compounds tested, the preferred acceptor for the wild-type Os9BGlu31 is ferulic acid, while
caffeic acid and 1-naphthaleneacetic acid are the substrates preferred by other Os9BGlu31 variants (Figure
2). Most of the Os9BGlu31 variants generated yield higher rates of 4NP release with phenolic acceptors than
with buffer alone, which results in hydrolysis by water. Among the single mutants, W243L is arguably the
most active overall. W243N also displays an increased activity with most of the substrates, although none is
particularly favored, which agrees with previous results18. These mutations increase the activity of the
enzyme and its potential use for the glycosylation of substrates of interest.
The mutants I172T, L183Q, and L241D show lower activity and stability than the wild type or W243N. The
less soluble L241D mutant, in particular, shows the lowest performance over multiple acceptors (Figure 2).
Remarkably, the introduction of L241D in the Os9BGlu31 W243N variant caused a significant activity
increase for most substrates but the flavonoids (Table 1). Epistasis is the non-linear interaction between
9

Green Chemistry
Page 10 of 29
View Article Online
mutations such that their effects over function, in this case transglycosidation activity, are significantly higher
DOI: 10.1039/C9GC00621D
or lower than expected by simple addition of their inidividual effects. In Table 1, we observe several
examples of positive epistasis42-44 (higher improvements in function than expected) against several
substrates. Notice that the double mutant L241D/W243N also exhibits the highest transglycosylation rate
observed in this work using syringic acid as the acceptor. This is an example of sign epistasis42-44 (the point
mutation L241D has an adverse effect on the activity of the single mutant, but it has a positive effect on the
double mutant L241D/W243N).
Another second mutation was introduced to generate the mutant L183Q/W243N. Although the second
mutation barely affected the transglycosylation rate of phenolic compounds over W342N, it increased the
activity with other acceptors and positive epistasis was observed in several cases (Table 1). Very prominent
sign epistasis was also observed in this double mutant against the substrate 1-naphthol. Lastly, the
introduction of the second mutation W243N in the mutant I172T could barely restore any of the activity lost
upon the first mutation, suggesting that conservation of the amino acid I172 is important to catalyze the
transglycosylation reaction. The entries in Table 1 with no annotation on epistasis indicate that those
mutations may be purely additive or that the epistasis was not strong given the expected uncertainty in the
data. Details on how the thresholds for the epistatic interactions were defined are provided in the Supporting
Information file SI01.xlsx.
10

Page 11 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Figure 2. Heatmap of the 4NP release (in nmol min-1 mg-1) catalyzed by 21 Os9BGlu31 variants when
glycosylating 22 different acceptors (462 reactions). The dendrograms were constructed using the average
Euclidian distance between clusters as the linkage method. For reaction conditions see SI01.xlsx.
Table 1. Reaction rates (in nmol min-1 mg-1) and types of epistatic interactions observed for the double
mutants in this work. For reaction conditions see SI01.xlsx.
Wild
type
L183Q/W243N L241D/W243N I172T/W243N
/ Epistasis / Epistasis / Epistasis
Acceptor
L183Q L241D I172T W243N
1-Naphthaleneacetic
acid
185
40
16
68
3300
1344 / -
2750
108
1-Naphthol
84
12
26
29
14
4
33
82
82
43
1186
3108
2592
1703
2104 / + (sign) 1316
88
4-Coumaric acid
4-Hydroxybenzoic acid
6-Hydroxyflavone
275
324
286
15
15
14
1482 / -
1066 / -
1996
3515
2869
1877
85
85
134
11

Green Chemistry
Page 12 of 29
View Article Online
Abscisic acid
Apigenin
60
165
252
55
17
12
29
10
18
37
8
15
0
45
33
114
61
43
99
34
34
89
25
53
35
82
81
72
92
41
861
603
911
1208 / + (sign) 78
DOI: 10.1039/C9GC00621D
1004 / + (sign) 407
83
Caffeic acid
Chloramphenicol
Chrysin
15
0
2544
1005
1655
3759
479
1081 / -
3957 / + (sign) 94
1401 / + (sign) 993
156
134
115
67
14
24
15
0
1996
1877
4674 / +
Ferulic acid
Gibberellin A4
Hesperetin
550
77
1088 / -
591 / +
1167
1029 / + (sign) 110
957 / - 84
2977 / + (sign) 135
64
1913 / + (sign) 3348 / + (sign) 225
92
24
27
12
19
17
32
28
20
36
11
1521
2152
321
Indole-1-acetic acid
Kaempferol
Luteolin
295
210
141
138
251
312
275
318
103
0
2092
0
653 / + (sign) 105
0
708
Salicylic acid
Sinapic acid
Syringic acid
trans-Cinnamic acid
Vanillic acid
Water
15
0
1674
2276
3051
1617
2812
727
1050 / -
1198 / -
897 / -
1131
1638
78
4220 / + (sign) 82
3933 / + (sign) 62
2188 / + (sign) 92
6
6
15
6
1482 / -
3515
85
968 / + (sign) 1316 / + (sign) 75
In addition to the mutants I172T, L183Q, and L241D, five W243 mutants (H, K, P, Q, and R) also showed
lower transglycosylation and hydrolysis activities than the wild type. The rest of the mutants showed
intermediate activities on the different substrates (Figure 2). Notably, the function of Os9BGlu31 could be
affected by both single and double mutants, but none of the mutants studied showed a higher activity on all
the substrates. Based on the results obtained, we selected the following 8 Os9BGlu31 variants with high
activity for further study, along with the wild type as the reference: W243C, W243I, W243L, W243N,
W243T, W243V, L183Q/W243N, and L241D/W243N.
The preceding results were obtained with different concentrations of DMSO, which were adjusted based on
the substrate solubility (0.25 vol.% for phenolic compounds, 1.25 vol.% for phytohormones and flavonoids,
0 vol.% for water). We suspected that the DMSO used as co-solvent might have a slight effect on the catalytic
activity of the enzymes. Therefore, we performed the activity study of the 9 high activity enzymes keeping
the concentration of DMSO fixed at 2.5% by volume. These reactions were repeated in triplicate (file
SI02.xlsx). The results are presented in Figure 3 and the activity trends are in agreement with the previous
discussion.
12

Page 13 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Figure 3. Relative rates of 4NP release by Os9BGlu31wild-type enzyme, W243 mutants, and
L183Q/W243N and L241D/W243N double mutants against 22 different acceptors, including water. For
reaction conditions, see SI02.xlsx.
It was found that the conversion of several hydrophilic residues did not improve substantially the hydrolysis
activity of Os9BGlu3118. Remarkably, in the present study we have identified several high activity mutants
that can hydrolyze 4NPGlc faster than the wild-type (Figure 4). In particular, Os9BGlu31 W243N shows the
highest hydrolysis activity overall among the single mutant enzymes. This suggested that the mutation could
be used in combination with additional hydrophilic mutations to further increase the hydrolysis rate of the
13

Green Chemistry
Page 14 of 29
View Article Online
enzyme. In fact, the double mutants engineered catalyze both reactions, hydrolysis and transglycosylation,
DOI: 10.1039/C9GC00621D
with the L183Q/W243N variant showing the highest overall ratio of hydrolysis to transglycosylation and
L241D/W243N showing the second highest. These findings may validate the idea that hydrophobic residues
near the putative water binding site contribute to transglycosylation activity, which was not apparent in
previous studies6,7,18. The selectivity to each of the two reactions can be determined by monitoring the
products being synthesized (section 3.2).
Figure 4. Rate of release of 4NP with the best substrate and with water for Os9BGlu31 wild type and high-
activity mutants. The line inside the bar indicates the median activity over 18 different substrates. For reaction
conditions see SI02.xlsx.
3.2 Effect of substrate on product distribution
To further understand the function of Os9BGlu31, we analyzed the different transglycosylation products
during longer reaction times with a diverse group of substrates and enzyme mutants. Interestingly, several
acceptors containing more than one hydroxyl group can accept glucose from the W243 mutants in different
positions to produce multiple glucoconjugate compounds. These acceptors are the phenolic acids 4-
hydroxybenzoic acid, ferulic acid, and vanillic acid and the flavonoids apigenin, luteolin, and kaempferol
(Figure 5).
14

Page 15 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Figure 5. UPLC chromatograms of products of reactions catalyzed by Os9BGlu31 W243L variant using
acceptors with multiple -OH groups and 4NPGlc as glucose donor (absorbance at λ=360 nm). The structures
of the acceptors are shown at the top of the corresponding chromatogram. The reaction mixtures contained 5
mM 4NPGlc as the donor, 0.5 mM acceptor, and 5 µg of Os9BGlu31 W243L in 50 mM citrate buffer (pH
4.5), and were incubated at 30 °C for 1 hour and then stopped with 1% formic acid. The components were
separated over a C18 UPLC column with the solvent and gradient conditions described in the Methods. The
large peaks of 4NPGlc at 6.7 min and 4NP at 10.6 min were used to assess the total enzymatic activity. Other
15

Green Chemistry
Page 16 of 29
View Article Online
peaks are marked according to their previous identification18. FA is ferulic acid, FAG is ferulic acid 1-O-
DOI: 10.1039/C9GC00621D
glucose ester, 4NP is 4-nitrophenol, 4NPG is short for 4NPGlc (4NP-beta-D-glucoside), 4HBA is 4-
hydroxybenzoic acid, 4HBA-GE is 4-HBA 1-O-glucose ester, VA is vanillic acid, VA-GE indicates VA
glucose ester.
Notably, the wild-type enzyme transferred the sugar exclusively to one hydroxyl group position of the
phenolic acids (Figures S2 to S4) or the flavonoids, which is a good example of regioselectivity. Mutants
W243 C, I, L, and N produced more than two products when provided with ferulic acid, including the bis-
glucoside at 5.3 min, the glucoside at 6.4 min, and the glucose ester at 7.1 min (based on our previous
characterization18), while only one more product (ferulic acid glucose ester, FAG) was formed by W243T
and W243V (Figure S2). The W243 mutants produced different amounts of glucoconjugates, with the
glucosyl moiety being transferred to different positions to a variable extent. For instance, while the wild type
and mutants yielded significant amounts of the 4HB-glucose ester, these yields were even higher for the
mutants W243N and W243L. Since both glucosides and glucose esters of these phenolic acids have been
reported in nature45, the ability to synthesize them will be useful to study their biological roles.
We measured the product range formed from flavonoid substrates, some of which contain several hydroxyl
groups (apigenin, luteolin, and kaempferol). Kaempferol and luteolin have similar structures with four
hydroxyls (positions 3, 5, 7, and 4’ in kaempferol and 5, 7, 3’, and 4’ in luteolin). W243N was shown to
produce multiple kaempferol glucosides and bis-glucosides by transglycosylation18. In this study, we also
observed various glycosides with the newly generated mutants (Figure S5) and were able to extend this to
other flavonoids with different hydroxyl group positions.
The chromatograms in Figures S6 and S7 show the presence of multiple products in reactions catalyzed by
Os9BGlu31 mutants. Four additional products were eluted in most of the reactions with apigenin catalyzed
by Os9BGlu31 mutants. The product eluting at 9.45 min (apigenin 7-O-glucoside)18 was synthesized by both
wild-type and mutant enzymes but no other products were produced by the wild type, while additional peaks
at 7.5, 9.55 and 11.3 min were seen in the mutants. The peak at 7.5 min that was particularly prominent in
the reactions with the W243L and W243N variants is likely to be a bis-glucoside, based on the more
hydrophilic elution position and previous observations with kaempferol glucosides18. This suggests that
Os9BGlu31 W243L and W254N are efficient at transglycosylating two positions on apigenin, resulting in
high bis-glucoside production. Besides, the presence of three products eluting in the range of monoglucosides
(9-12 min) suggests that the selected mutants can glycosylate all three hydroxyl groups on apigenin.
16

Page 17 of 29
Green Chemistry
Luteolin could also accept the transfer of glucose by Os9BGlu31 transglucosidase variaDnOtsI:.10W.1023V94i/e3Cw9LAGrtCiac0lpe0Op6n2el1inaDers
more active than the other mutants, based on the abundance of new compounds, especially the one eluting at
7.4 min. As noted for the corresponding apigenin glucoconjugate, this compound is likely to be a bis-
glucoside, based on its relatively early elution, which might result from the high activity of W243L towards
two hydroxyl positions. W243L also showed a high yield to the product eluting at 9.4 min, in addition to the
peak at 8.7 min, which was high in most variants. W243T showed a lower yield of the peak at 8.7 min, as it
is more selective to the compound eluting at 9.4 min. The relatively high peak at 10.6 min in the W243T
reaction and relatively low hydrolysis rate in Figures 3 and 4 suggest it may be producing another glucoside
hidden in the p-nitrophenol peak, but we were unable to resolve this putative glucoside. This data suggest
that most variants could be used to produce the glycoside at 8.7 min, while W243L and W243T could be
used to produce the bis-glucoside and other luteolin glucosides.
3.3 Structural model of the ternary complex including the acceptor substrate
Previously, we produced a homology model of the Os9BGlu31 wild type structure18, but the apo structure is
not appropriate to evaluate the reactivity toward acceptor substrates, since they react with the glucosyl-
enzyme intermediate (Figure 1). To obtain a better understanding of the glycosyl transfer mechanism and
visualize the binding orientations of the acceptor substrates within the Os9BGlu31 active site, we modeled
the wild type and mutant proteins as their covalent glucosyl intermediates and docked the potential acceptors
substrates on them. In these models, the glucoside is deeply buried in the active site and the access path is
narrow, but small molecules can access it. This could be related to the lower activities of larger flavonoid
substrates which were evident in Figure 3, as the narrow binding site would difficult the access of very large
substrates. Molecular docking revealed the substrates are oriented on the opposite face of the glucose from
the catalytic nucleophile, E387. The binding mode of the substrates is exemplified by ferulic acid and 1-
naphthaleneacetic acid in Figure 6. Although ferulic acid was positioned somewhat distant from the anomeric
carbons in the wild type active site, the models shown are in line with the predicted SN2 mechanism proposed,
as this requires the nucleophilic substrates to be positioned on the opposite face of the leaving group, E387,
which ultimately results in the re-inversion of stereochemistry to form the β-anomeric products. Moreover,
as shown in Table S2, the binding scores for the acceptor substrates on the glycosylated enzyme intermediate
are generally higher for the W243L mutant compared to the wild type, suggesting that a higher free energy
of binding of the acceptor may be contributing to lower the activation energy of the deglycosylation step and
increase the kinetic rate.
17

Green Chemistry
Page 18 of 29
View Article Online
DOI: 10.1039/C9GC00621D
Figure 6. Graphical representations of the binding modes of (A) ferulic acid to wild-type Os9BGlu31 and
(B) 1-naphthaleneacetic acid to W243L mutant covalent glucosyl intermediates. The blue ribbons depict the
secondary structure of the proteins. The acceptor substrates are shown in a ball-and-stick representation.
E387-glucoside, W243, and L243 are shown as sticks only. The solid green lines indicate the distances
measured between the nucleophilic carboxylate and the electrophilic C1 on the glucoside.
18

Page 19 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
3.4 Cheminformatics
3.4.1 Linear, phenomenological model
In order to generalize the findings of this work to other possible substrates, we constructed different
cheminformatics models. To build a proper model we require a set of descriptors that is relevant (i.e., is
highly related to the enzymatic activity that we want to predict) and at the same time minimizes the
redundancy of information captured by the different indices under consideration (i.e., the descriptors are not
totally correlated to each other).
We considered the results with all 9 high-activity enzymes altogether on all substrates except water. Catalysis
on water is different from the rest of substrates and its small size prevents the calculation of several molecular
descriptors. For instance, RotRatio is undefined for molecules that do not have a rotatable bond; HybRatio
cannot be computed for molecules with no carbon atoms; AATSp, AATSCp, MATSp, and GATSp are
undefined when p > number of atoms, etc.
Firstly, we computed a large number of molecular features describing each of the substrates (2956
descriptors). We then applied a redundancy-reduction filter so that no pair of molecular descriptors had a
correlation above 0.8. This step reduced the number of descriptors under consideration to 100, which suggests
that some molecular descriptors that have been proposed in the literature along the years may be more
correlated to others than one might desire46. In a second step, we ranked the descriptors using as the criterion
the Pearson correlation of each individual descriptor to the average enzymatic activity. The 12 most relevant
descriptors according to this criterion are presented in Figure 7a. Their names are indicated in Table S3. We
also considered another criterion for the ranking: in Figure 7b, the descriptor importance is measured in terms
of information gain, an entropy-based metric.
In Figure 7, we can observe some interesting results. The descriptor most correlated to the enzymatic activity
on its own is largestChain. This is a 1D constitutional descriptor from the RDK package and, as its names
suggests, it quantifies the number of atoms in the largest chain on the molecular graph. MATS3dv is the
Moran coefficient of lag 3 weighted by valence electrons. This is a 2D centered autocorrelation descriptor
computed by the Mordred software. This descriptor being selected as important suggests that relatively
symmetrical acceptors, such as aromatic-containing substrates47, may favor faster transglucosylation rates.
19

Green Chemistry
Page 20 of 29
View Article Online
ATSC5dv and ATS5c are the centered Moreau-Broto autocorrelation descriptors ofDlOaI:g10.5103w9/Ce9igGCh0t0e6d21Dby
valence electrons and by charges, respectively.
If we look at the maximum information gain as the criterion for selecting descriptors, we also find multiple
autocorrelation descriptors: MATSv5 and MATSp6 are the Moran autocorrelation coefficients of lag 5
weighted by van der Waals atomic volumes, and of lag 6 weighted by atomic polarizability, respectively.
GATSp6 and GATS1Z are the Geary autocorrelation coefficients of lag 6 weighted by atomic polarizability
and of lag 1 weighted by atomic number, respectively. These two descriptors are considered relevant by both
filtering algorithms. The fact that autocorrelation descriptors with different lags and different weightings are
considered relevant by the algorithm suggests that the symmetry of the substrate, both in terms of
connectivity and atomic makeup, would favor higher transglucosylation rates.
It is also interesting to see that, thanks to the pairwise-correlation reduction step, descriptors of diverse nature
are considered in the ranking. For instance, Mor29 and Mor13 are the unweighted 3D-MoRSE descriptors
of distance d = 29 and 13, as computed by Mordred. The large value of the scattering parameter d should
make these descriptors insensitive to atom pairs situated at large distance but highly sensitive to short
distances (<3 Å)48. This index may thus be capturing the effect of heteroatoms and multiple bonds (and thus
bond distances) on the ability of the substrates to accept the glucosyl moiety. Mor06m is another 3D-MoRSE
descriptor that is considered relevant. Notably, the topological charge indices JGI7 and JGI10 are also
considered important by the information-based criterion. These indices have been shown to capture the
charge distribution (e.g., the dipolar moment) within a molecule based solely on its topology, thus bypassing
the complexity of 3D-optimizing the molecular structure49.
A major limitation of most filter methods is the fact that they barely consider the correlation between
features50. Thanks to our filter strategy, however, this limitation could be greatly reduced. In Table S4, we
can see that in a few cases the correlation between some of the indices shortlisted using the correlation to
the enzymatic rate is close to the cutoff of 0.8. On average, however, the correlation between features is
much lower, around 0.5, thanks to the design of findCorrelation algorithm. Interestingly, the correlation
among features when using the information gain as the criterion is even lower, but the correlation to the
enzymatic rate is more diverse, in the range R = 0.05-0.39 vs. R = 0.49-0.61 when using the correlation
criterion (Table S4).
20

Page 21 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Figure 7. Top feature selection results using as criteria the linear correlation and the information gain
between the descriptors and the average enzymatic activity measured across all enzymes.
With the subset of 21 different molecular descriptors shortlisted in Figure 7, it becomes possible to train a
variety of models to seek a relationship between the molecular descriptors, the type of enzyme, and the
enzymatic activity over a given acceptor substrate. One possibility is to build a linear model. The filtering
algorithm has reduced the dimensionality of the problem by rejecting variables that are not very informative.
Next, we conduct a variable selection to build a model (notice that we still have 22 variables in our
consideration set: 21 continuous molecular descriptors and 1 categorical variable or factor representing the
enzyme type). One way to select variables is to use a stepwise selection method. To this end, we used the
function stepAIC in the MASS R package. This algorithm adds or removes one variable at a time so that the
new model leads to the highest reduction in the Akaike Information Criterion (AIC) relative to the previous
21

Green Chemistry
Page 22 of 29
View Article Online
model considered. AIC is a statistic that allows evaluating the quality of a model: it is cDloOsI:e10ly.10r3e9l/Ca9teGdC0t0o62t1hDe
correlation coefficient R, but it introduces a penalty for every degree of freedom consumed upon introduction
of a new variable34. This way, we selected the following model:
Rate (nmolmin ―1 mg ―1) = Intercept(Enzyme) + 89.44·largestChain + 2.31·ATSC5dv
n = 153 R2 = 0.68 R2 adj = 0.65 F(10, 142) = 29.5 AIC = 1665.6 p < 0.01
in which the independent term takes the values indicated in Table 2 depending on the query enzyme.
Not surprisingly, the algorithm has identified the type of enzyme and the largestChain descriptor as the most
relevant variables, and it has also selected ATSC5dv. One advantage of linear models is their ease of
interpretability. On the one hand, the molecular descriptors largestChain and ATSC5dv have a positive effect
on the turnover of a substrate, with largestChain having predominant effect. On the other hand, differences
between values in Table 2 correspond to the expected differences in activity between enzymes when
presented a given substrate. It thus becomes clear that mutants W243L and W234N are much more active on
average than the wild-type enzyme.
Table 2. Coefficients for the independent term as a function of the enzyme in the linear cheminformatics model.
Level
WT
Intercept (nmol min-1 mg-1)
8.69
L183Q/W243N
W243C
79.74
121.46
216.75
248.40
330.22
462.87
519.52
562.16
W243V
W243I
W243T
L241D/W243N
W243N
W243L
The linear model does not overfit the data, as indicated by the close values of R2 and adjusted R2, the
relatively low value of the AIC, and the model and every variable are considered highly significant by the
corresponding F-tests. Remarkably, this simple linear model yields a high determination coefficient, R2 =
0.68. One factor that may contribute to this is that, in our study, most of the enzyme variants considered are
single mutants at the same position 243. It seems thus conceivable that these mutations have a localized
22

Page 23 of 29
Green Chemistry
steric effect within the active site cleft of the enzyme. This effect can thus be represenDteOdI: 1r0e.1a0s3V9oi/enCw9aAGbrtCicl0lye0O6wn2l1ineDell
by a constant factor dependent on the specific residue introduced and a variable related to the substrate chain
size, which is in agreement with the docking study in the previous section. On the other hand, this model has
been built and trained considering all the data (except for the hydrolysis reactions). To increase the predictive
power of the model, in the next section we explore more flexible models and assess their performance against
validation data.
3.4.2. Nonlinear, mechanistic models
In this section, we report a model based only on the molecular descriptors classically used in Hansch’s
method. The method has the advantage of using a few pre-selected descriptors to seek the models, making
unnecessary the exploration of large spaces of chemical descriptors51. Hansch’s model is based on an extra-
thermodynamic linear free energy relationship (LFER). The LFER model decomposes a biological
mechanism into a series of several steps. Then, it seeks a linear combination or weighted summation
(additive model) of the free energies of the different steps. It considers as free energy contributions
parameters like the negative logarithm of the ionization constant (pKi = -logKi) and the logarithm of the
partition coefficient (logP). These parameters are related to the free energy of the ionization process and to
the posterior membrane transport (partition) steps of the biological compound, respectively. Specifically, P
is the partition coefficient of the biological compound in the system n-octanol/water and it measures the
compound’s lipophilicity. Hansch’s model also considers other steric and electronic properties of the
compounds, like their polar surface area or molar refractivity, which increases the flexibility of the method
to describe enzymatic processes like the one investigated. The molecular descriptors used in this work are
D1 = ALOGP (an estimate of logP), D2 = MR (molar refractivity), and D3 = TPSA (topological polar surface
area). Moreover, as shown in Table S4, these different parameters show a limited correlation between them.
We evaluated multiple non-linear models using Artificial Neural Network (ANN) algorithms. The models
use as input variables the type of enzyme and the 3 molecular descriptors calculated for the query compound.
The networks built are Multi-Layer Perceptrons (MLP) with one hidden layer comprising 3, 6 or 9 neurons.
Different combinations of identity, logistic, and exponential activation functions were considered. Table 3
summarizes the results obtained. On the one hand, we can see that the performance of the ANN is satisfactory
(the coefficient of variation of experimental repeats is around 5-15%), with values of the correlation
coefficient around R = 0.75-0.86 for the training and validation sets. In this case, 4/17 of the compounds
were assigned randomly to the validation set across all the enzymes (see SI03.xlsx). Thus, the activity results
for 1-naphtol, 4-coumaric acid, kaempferol, and trans-cinnamic acid were not used in training the ANNs.
23

Green Chemistry
Page 24 of 29
View Article Online
Notably, the predictions for these compounds achieved a similar accuracy to those used to train the networks
DOI: 10.1039/C9GC00621D
(Table 3, validation).
On the other hand, we observe that there is not a great advantage in using a very flexible model over using a
linear model with appropriate descriptors in our problem. In fact, the ANN that only uses identity functions
shows a limited performance, which stems from the descriptors specified. Abruzzo et al. also reported that
variable selection had a greater impact than the selection of the model in a comparative study of
transcriptomics data classifiers52. Furthermore, we observed that the training of larger dense hidden layers is
more prone to encounter local optima and it may also increase the chances of undesired overfitting. Taken
together, these results also suggest that the true relationship between inputs and output is not highly nonlinear.
Table 3. Performance of 10 different artificial neural networks predicting transglycosylation activity with multiple
substrates and enzymes (for details, see SI03.xlsx).
Network
Training
RSE
Validation
RSE
Hidden
activation
Output
activation
Topology
R
R
MLP 12-3-1
MLP 12-5-1
MLP 12-7-1
Logistic
Logistic
Logistic
Identity
Identity
Identity
Identity
Identity
Identity
Identity
Identity
Identity
Identity
33438
24856
30873
36041
20831
18715
24504
17788
18145
46663
0.7264
0.8063
0.7554
0.7016
0.8403
0.8574
0.8170
0.8665
0.8639
0.5822
33494
23629
27619
31832
27899
30130
23676
23239
25942
32346
0.7397
0.8241
0.7786
0.7381
0.7996
0.7978
0.8017
0.8120
0.7997
0.7488
MLP 12-3-1 Exponential
MLP 12-5-1 Exponential
MLP 12-7-1 Exponential
MLP 12-3-1 Exponential
MLP 12-5-1 Exponential
MLP 12-7-1 Exponential
MLP 12-7-1
Identity
Importantly, the models built can be used in practice to estimate the enzymatic activity of new acceptors over
the different enzymes and thus used to screen libraries of potential substrates computationally. Given the
reasonable performance of the different networks in Table 3, we averaged the individual predictions by all
of them to reach an ensemble or consensus prediction. In Table S5, we report the results of exploring in silico
the transglucosylation rates of a new set of putative acceptors by the wild-type Os9BGlu31 and different
mutants. Although several of these compounds are expected to exhibit low activitiy, the results suggest that
24

Page 25 of 29
Green Chemistry
3,4-dihydroxybenzoic acid, 6-benzylaminopurine, indole-3-acetic acid, and serotonin DmOaI:y10.1b0e3V9i/peCwa9ArGrttCiicc0le0uO6ln2al1inrDely
reactive transglucosylation substrates over the enzyme variants W243L, W243N, and L241D/W243N. Aided
by these predictions, we are already advancing the production of novel antibiotic and phytohormone
glucoconjugates of interest, such as abscisic acid glucoside (preliminary results shown in Figure S8), on
which we recently submitted a patent application.
In these results, glucose was transferred from 4NPGlc to a variety of compounds and acceptor positions,
allowing us to prepare a broad array of valuable glucoconjugates. Other glucose donors have also been
proposed53. For example, β-glucosyl fluoride was recently shown to be an effective donor for certain Leloir-
type glycosyltransferases in the presence of catalytic amounts of UDP54. With regards to the large-scale
production of glucoconjugates, more atom-efficient donors could be pursued, which opens up another area
of potential research. The cheminformatics pipeline proposed in this work could also be extended to new
combinations of substrates and catalysts, thus enabling increasing possibilities for transglycosidases and
other enzymes.
4. CONCLUSION
Glucoconjugate molecules have great interest to the fine chemistry and phytochemistry industries, and many
of these compounds could be synthesized from feedstocks derived from biorefineries. However, the
production of complex glucoconjugates using conventional synthetic approaches may not be sustainable or
economical, as multiple synthetic steps of limited yield would be involved. In this work, we studied 16 new
Os9BGlu31 variants and compared them to the wild-type Os9BGlu31 to identify new high-activity catalysts,
like W243L, and to correlate their activities with properties of the substrates being converted. Furthermore,
the combination of the mutations designed to increase the hydrophilicity of the active site, L183Q and
L241D, with the high-activity W243N mutation led to catalysts with high hydrolytic activity compared to
their transglycosylation activity. These mutations exhibited positive epistasis on other substrates as well,
which may be related to their proximity in the protein fold. We went a step further and used computational
methods to facilitate a greener catalyst selection depending on the target substrate with a pipeline that can
be used or adapted by other researchers. Cheminformatics models can help minimize the resources consumed
to investigate new substrates and maximize the throughput of new catalytic processes. Our models suggest
that steric constraints in the active site have a crucial effect on the activity of a substrate, which agrees with
the docking simulations. Careful variable selection can afford linear models with reasonable performance,
25

Green Chemistry
Page 26 of 29
although higher accuracy could be attained with nonlinear artificial neural networksD. OTI:h10e.1s0e3V9mi/eCw9oAGrdtCice0lel0sO6n2la1inDree
proving useful to identify substrates and efficiently produce new glucoconjugates of interest, such as
phytohormone glucosides for enzymatic and plant physiology studies or antibiotic glycosides with improved
solubility and pharmacological properties.
Supporting Information
Figure S1 illustrates the data analysis pipeline used. Chromatograms in Figures S2 to S7 show the transfer
of glucose to different acceptors by Os9BGlu31 wild type and its mutants. Figures S8 demonstrates the
synthesis of the phytohormone glucoconjugate abscisic acid glucose ester enabled by this study. Table S1
presents the sequences of the oligonucleotides used as primers in the site-directed mutagenesis of
Os9BGlu31. Table S2 indicates the docking scores on the wild type and W243L mutant. Table S3 indicates
the names of the descriptors shortlisted. Table S4 is the correlation matrix between selected descriptors.
Table S5 presents ensemble activity predictions for potential new transglucosylation acceptors. File
SI01.xlsx contains the activity results in the screening of all mutants and the epistasis analysis. File SI02.xlsx
presents the data from the high activity-mutants used to train the ML models. File SI03.xlsx contains
cheminformatics data (descriptors and results of the ANNs).
Author Information:
*E-mail: blayroger.1@osu.edu (V.B.); cairns@sut.ac.th (J.R.K.C).
§These authors contributed equally.
†Current Address: Research Institute for Interdisciplinary Science, Okayama University, Okayama, Japan.
Notes
The authors declare no competing financial interests.
26

Page 27 of 29
Green Chemistry
View Article Online
DOI: 10.1039/C9GC00621D
Acknowledgements
This work was supported by grants from Suranaree University of Technology (SUT) and the National
Research University Project from the Commission on Higher Education (Thailand) to SUT. Ministerio de
Economía y Competitividad (FEDER CTQ2016-74881-P and CTQ2013-41229-P) and Basque Government
(IT1045-16) are gratefully acknowledged for their financial support. V. B. thanks the support from The Ohio
State University and the Center for Biomolecular Structure, Function and Application (SUT).
REFERENCES
1. J. R. Ketudat Cairns, B. Mahong, S. Baiya and J. S. Jeon, Plant. Sci., 2015, 241, 246–259.
2. L. L. Lairson and S. J. Withers, Chem. Comm., 2004, 20, 2243–2248.
3. D. E. Koshland, Biol. Rev., 1953, 28, 416–436.
4. S. G. Withers, R. A. J. Warren, I. P. Street, K. Rupitz, J. B. Kempton and R. Aebersold, J. Am.
Chem. Soc., 1990, 112, 5887-5889.
5. M. J. Baumann, J. M. Eklöf, G. Michel, A. M. Kallas, T. T. Teeri, M. Czjzek and H. Brumer, Plant
Cell, 2007, 19, 1947-1963.
6. D. Teze, J. Hendrickx, M. Czjzek, D. Ropartz, Y. H. Sanejouand, V. Tran, C. Tellier and M. Dion,
PEDS, 2014, 27, 13–19.
7. B. Bissaro, J. Durand, X. Biarnés, A. Planas, P. Monsan, M. J. O’Donohue and R. Faure, ACS
Catal. 2015, 5, 4598–4611.
8. E. Champion, I. André, C. Moulis, J. Boutet, K. Descroix, S. Morel, P. Monsan, L. A. Mulard and
M. Remaud-Siméon, J. Am. Chem. Soc., 2009, 131, 7379−7389.
9. E. Champion, F. Guérin, C. Moulis, S. Barbe, T. H. Tran, S. Morel, K. Descroix, P. Monsan, L.
Mourey, L. A. Mulard, S. Tranier, M. Remaud-Siméon and I. André, J. Am. Chem. Soc., 2012, 134,
18677–18688.
10. L. F. Mackenzie, Q. P. Wang, R. A. J. Warren and S. G. Withers, J. Am. Chem. Soc., 1998, 120,
5583–5584.
11. M. Jahn, J. Marles, R. A. J. Warren and S. G. Withers, Angew. Chem. Int. Ed., 2003, 42, 532–534.
12. E. R. Moellering, B. Muthan, C. Benning, Science, 2010, 330, 226-228.
13. Y. Matsuba, N. Sasaki, M. Tera, M. Okamura, Y. Abe, E. Okamoto, H. Funabashi, M. Takatsu, M.
Saito, H. Matsuoka, K. Nagasawa and Y. Ozeki, Plant Cell, 2010, 22, 3374-3389.
14. S. Luang, J. L. Cho, B. Mahong, R. Opassiri, T. Akiyama, K. Phasai, J. Komvongsa, N. Sasaki, Y.
Hua, Y. Matsuba, Y. Ozeki, J. S. Jeon and J. R. Ketudat Cairns, J. Biol. Chem., 2013, 288, 10111–
10123.
15. T. Miyahara, R. Sakiyama, Y. Ozeki and N. Sasaki, J. Plant Physiol. 2013, 170, 619–624.
27

Green Chemistry
Page 28 of 29
View Article Online
16. Y. Nishizaki, M. Yasunaga, E. Okamoto, M. Okamoto, Y. Hirose, M. Yamaguchi, Y. Ozeki and N.
DOI: 10.1039/C9GC00621D
Sasaki, Plant Cell, 2013, 25, 4150–4165.
17. J. Komvongsa, B. Mahong, K. Phasai, Y. Hua, J. S. Jeon and J. R. Ketudat Cairns, J. Agric. Food
Chem., 2015, 63, 9764–9769.
18. J. Komvongsa, S. Luang, J. V. Marques, K. Phasai, L. B. Davin, N. G. Lewis and J. R. Ketudat
Cairns, Biochim. Biophys. Acta, 2015, 1850, 1405–1414.
19. R. W. Gantt, P. Peltier-Pain and J. S. Thorson, Nat. Prod. Rep., 2011, 28, 1811–1853.
20. C. Zhang, B. R. Griffith, Q. Fu, C. Albermann, X. Fu, I.-K. Lee and L. Li, Science, 2006, 313,
1291–1294.
21. V. Křen and T. Řezanka, FEMS Microbiol. Rev., 2008, 32, 858–889.
22. I. Finore, A. Poli, P. Di Donato, L. Lama, A. Trincone, M. Fagnano, M. Mori, B. Nicolaus and A.
Tramice, Green Chem., 2016, 18, 2460–2472.
23. R. A. Sheldon, ACS Sustainable Chem. Eng., 2018, 6, 4464–4480.
24. A. Pennec, L. Legentil, L. Herrera-Estrella, V. Ferrières, A.-L. Chauvin and C. Nugier-Chauvin,
Green Chem., 2014, 16, 3803–3809.
25. V. Blay, R. Garcia-Domenech and J. Galvez, ChemTexts, 2017, 3, 2.
26. V. Blay, J. Gullon-Soleto, M. Galvez-Llompart, J. Galvez and R. Garcia-Domenech, ACS Sustain.
Chem. Eng., 2016, 4, 4224–4231.
27. B. Bhhatarai, R. Garg and P. Gramatica, Mol. Inform., 2010, 29, 511–522.
28. C. B. Santiago, J.-Y. Guo and M. S. Sigman, Chem. Sci., 2018, 9, 2398–2412.
29. R. Opassiri, B. Pomthong, T. Onkoksoong, T. Akiyama, A. Esen and J. R. Ketudat Cairns, BMC
Plant Biol., 2006, 6, 33.
30. J. Dong, D. S. Cao, H. Y. Miao, S. Liu, B. C. Deng, Y. H. Yun, N. N. Wang, A. P. Lu, W. B. Zeng
and A. F. Chen, J. Cheminform., 2015, 7, 60.
31. H. Moriwaki, Y. S. Tian, N. Kawashita and T. Takagi, J Cheminform., 2018, 10, 4.
32. M. Kuhn, J. Stat. Softw., 2008, 28, 1–26.
33. T. Cheng, Y. Wang and S. H. Bryant, Bioinformatics, 2012, 28, 2851–2852.
34. T. Hastie, R. Tibshirani and J. Friedman. The Elements of Statistical Learning: Data Mining,
Inference, and Prediction, Second ed., Springer, 2009.
35. A. Sali and T. L. Blundell, J. Mol. Biol., 1993, 234, 779–815.
36. S. Seshadri, T. Akiyama, R. Opassiri, B. Kuaprasert and J. K. Cairns, Plant Physiol., 2009, 151,
47–58.
37. M. Y. Shen and A. Sali, Protein Sci., 2006, 15, 2507–2524.
38. J. W. Schymkowitz, F. Rousseau, I. C. Martins, J. Ferkinghoff-Borg, F. Stricher and L. Serrano,
Proc. Natl. Acad. Sci. USA, 2005, 102, 10147–10152.
39. M. J. Sippl, Proteins, 1993, 17, 355–362.
40. R. A. Laskowski, M. W. Macarthur, D. S. Moss and J. M. Thornton, J. Appl. Cryst., 1993, 26, 283–
291.
28

Page 29 of 29
Green Chemistry
41. G. Jones, P. Willett, R. C. Glen, A. R. Leach and R. Taylor, J. Mol. Biol., 1997,D2O6I:71,0.71023V97i/e–Cw97AG4rtCic80le.0O6n2l1inDe
42. J. F. Storz. Curr. Opin. Struc. Biol., 2018, 50, 18-25.
43. T. N. Starr and J. W. Thornton. Protein Sci., 2016, 25, 1204-1218.
44. M. Camps, A. Herman, E. Loh and L. A. Loeb, Crit. Rev. Biochem. Mol. Biol., 2007, 42, 313–326.
45. S. Klick and K. Herrmann, Phytochemistry, 1988, 27, 2177-2180.
46. R. Todeschini and V. Consonni. Handbook of Molecular Descriptors. Wiley 2008. p. 692. ISBN
9783527613106.
47. B. Hollas, J. Math. Chem., 2003, 33, 91–101.
48. O. Devinyak, D. Havrylyuk and R. Lesyk, J. Mol. Graph. Model., 2014, 54, 194–203.
49. J. Galvez, R. Garcia, M. T. Salabert and R. Soler, J. Chem. Inf. Comput. Sci., 1994, 34, 520–525.
50. G. Chandrashekar and F. A. Sahin, Comput. Electr. Eng., 2014, 40, 16–28.
51. P. Abeijon, X. Garcia-Mera, O. Caamano, M. Yanez, E. Lopez-Castro, F. J. Romero-Duran, H.
Gonzalez-Diaz, Curr. Drug. Targets, 2017, 18, 511–521.
52. L. V. Abruzzo, L. L. Barron, K. Anderson, R. J. Newman, W. G. Wierda, S. O’Brien, A. Ferrajoli,
M. Luthra, S. Talwalkar, R. Luthra, D. Jones, M. J. Keating and K. R. Coombes, J. Mol. Diagn.,
2007, 9, 546–555.
53. S. Shoda, Proc. Jpn. Acad., Ser. B, 2017, 93, 125-145.
54. A. Lepak, A. Gutmann and B. Nidetzky, ACS Catal., 2018, 8, 9148-9153.
29