Computational Design of Enantiocomplementary Epoxide Hydrolases for Asymmetric Synthesis of Aliphatic and Aromatic Diols

Abstract The use of enzymes in preparative biocatalysis often requires tailoring enzyme selectivity by protein engineering. Herein we explore the use of computational library design and molecular dynamics simulations to create variants of limonene epoxide hydrolase that produce enantiomeric diols from meso‐epoxides. Three substrates of different sizes were targeted: cis‐2,3‐butene oxide, cyclopentene oxide, and cis‐stilbene oxide. Most of the 28 designs tested were active and showed the predicted enantioselectivity. Excellent enantioselectivities were obtained for the bulky substrate cis‐stilbene oxide, and enantiocomplementary mutants produced (S,S)‐ and (R,R)‐stilbene diol with >97 % enantiomeric excess. An (R,R)‐selective mutant was used to prepare (R,R)‐stilbene diol with high enantiopurity (98 % conversion into diol, >99 % ee). Some variants displayed higher catalytic rates (k cat) than the original enzyme, but in most cases K M values increased as well. The results demonstrate the feasibility of computational design and screening to engineer enantioselective epoxide hydrolase variants with very limited laboratory screening.


Introduction
The use of biocatalysis in chemistry is an attractive option for many synthetic processes, especially for preparing fine chemicals and bioactive compounds. [1][2][3][4][5][6][7] Although nature provides an enormous diversity of industrially useful enzymes, they often must be engineered to meet industrial process requirements. [8,9] In case of pharmaceuticalsynthesis, of special importance are chemoselectivity,c ompatibilityw ith harsh reaction conditions and product enantiopurity. [1,10] Consequently,e xtensive studies have been carried out on controlling and improving enzyme selectivity by protein engineering, often through directede volution, [11] which led to enzymesw ith improved selectivity in kinetic resolution of enantiomers andb etter performance in asymmetric transformation of prochiral compounds. [8,10,12] Whereas directed evolution is very successful, it requires high-throughput screening methods. In case of enzyme enantioselectivity,s creening is possible by chiral chromatography or by the use of quasi enantiomers in NMR or MS [13] but this may be time-consuming and expensive. Directed evolutionb ecomes complicated when no high-throughput expression is available, such as in case of enzymes that mustb ep roduced in fungi.S everalm ethods have been proposed to overcome these bottlenecks, such as optimizing strategies for library construction, for example, by focusing mutations in different combinations around the active site, [14][15][16] andb yi ncorporating structural [9] or phylogenetic information. [17,18] Another option is the use of computational tools to design improved enzymes. [19][20][21][22][23] Statistical methods have also been used. [24,25] Biophysics-based computational protocols have emerged as powerful platforms for the engineering of thermostable and organic-solvent compatible enzymev ariants. [26][27][28][29][30][31] Multiple mutations can be explored simultaneously,a llowing for larger jumps in sequence space than directed evolution and structure-based rational mutagenesis. Furthermore, in silico screening of enzyme variants by docking and high-throughput molecular dynamics simulations makes it possible to predict enzyme properties and decrease librarys ize for experimental evaluation from thousands to dozens. [32,33] In this study,w ee xplore ac omputational framework (catalytic selectivity by computational design, CASCO) [32] for obtaining enantiocomplementary epoxide hydrolases. This framework uses the Rosetta scoring function and search algorithm [19,34] to generate libraries of primary designs. Next, high-throughput molecular dynamics (MD) simulations with scoring the frequency of occurrenceo fr eactive (or near-attack)c onformations The use of enzymes in preparative biocatalysis often requires tailoringe nzymes electivity by protein engineering. Herein we explore the use of computational library design and molecular dynamics simulationst oc reate variants of limonene epoxide hydrolaset hat produce enantiomeric diols from meso-epoxides. Three substrates of different sizes were targeted: cis-2,3butene oxide,c yclopentene oxide, and cis-stilbene oxide.M ost of the 28 designst ested were activeand showedthe predicted enantioselectivity.E xcellent enantioselectivities wereo btained for the bulky substrate cis-stilbeneo xide,a nd enantiocomple-mentarym utants produced (S,S)-and (R,R)-stilbene diolw ith > 97 %e nantiomeric excess. An (R,R)-selective mutant was used to prepare (R,R)-stilbened iol with high enantiopurity (98 %c onversioni nto diol, > 99 % ee). Some variants displayed higher catalytic rates (k cat )t han the original enzyme, but in most cases K M values increased as well. The results demonstrate the feasibility of computational design and screening to engineere nantioselective epoxideh ydrolase variantsw ith very limitedl aboratory screening.
(NACs) [35][36][37] are used for ranking and to select as mall set of variants that qualify for laboratory testing. The frequency of reactive poses during MD simulations can explain the selectivity of computationally designed enzymev ariants. [38][39][40] For that purpose, generating multiple MD trajectories with independent assignment of initial atom velocities gives much better sampling of accessible conformational space than the use of a single trajectory runningo ver al ong simulation time. [41][42][43][44][45] Accordingly,t oe nable screening of thousands of Rosetta designs by MD, CASCO uses 20-80 of such short MD simulations for scoring conformational stability of designedr eactive enzymesubstrate complexes. The MD step thus examines if the conformation of the enzymes ubstrate complex, which is partially constrained during the Rosetta design step (NAC),w ill be maintained in short MD runs, or whether that reactive conformation is immediately lost, for example by movement of the substrate to an on-reactive pose. We used this approachearlier to predict enantioselectivity in kinetic resolutions catalyzed by haloalkane dehalogenases. [46] The potentialo ft his computational approach was illustrated for limonenee poxideh ydrolase (LEH) redesign in ap revious study [32] where we observedt hat the performance of the best enzymesi na37-variant libraryo btained by the CASCO framework was similar to that of the best variants obtained by screening approximately 4700 variants generated by the CASTing strategy for directed evolution. [47] Nevertheless,c omputational library design for enzyme engineering has serious challenges,o fw hich reliability of predictions is an important example. To further explore the possibilities and limitations of computational redesign,w ed esigned and examined an ovel set of enantioselective limonene epoxide hydrolase (LEH) variants.
LEH catalyzes the hydrolysis of epoxides by activating a bound water molecule for nucleophilic attack directly on one of the substrate's oxirane carbons. [48] The water is positioned by H-bonds to Asn55 and Tyr53 while Asp132 abstractsa proton from the water,w hich enables nucleophilic attack as a hydroxy ion (Scheme 1). At the same time, Asp101 protonates the epoxide oxygen, whichm akes it ab etter leavingg roup. The reaction is concerted. [49] In case of meso-epoxides, the ste-reochemical outcome is determined by regioselectivity of the attack and the products are enantiomers. LEH has been extensively used as am odel system for exploring the use of directed evolution strategies to engineer enantioselectivity,a nd many variants have been described. [47,[50][51][52][53][54][55] We previously examined the use of computationaldesign and screening to improve stability and control enantioselectivity. [26,29,32] In this work, we investigated three small sets of enantiocomplementary epoxide hydrolase variants forc onverting meso-epoxides( cis-2,3-butene oxide, cyclopentene oxide, and cis-stilbene oxide)t ot heir corresponding (R,R)-or (S,S)-diols.F or each target enantiomer,f ive top-ranked new variants were experimentally characterized to explore how challenging it is to position each of the substrates uniquelyi nt he active site. We also measured k cat and K M values for selected variants.T his showedt hat the k cat of somev ariants was highert han of the thermostable LEH-P variant, from which all mutants were derived. However,t he Michaelis constants (K M )w ere almost always highera sw ell, which decreases catalytic efficiency.T he results confirmed the workingh ypothesis that obtaining unique binding orientations,r eflected in high enantioselectivity,w as easier with the bulkier substrates. Mutants converted cis-stilbeneo xide to diols with an enantiomeric excess (ee)o f > 99 %. Small-scale preparative reactions were carried out.

Computational design
The enantioselectivity of limonenee poxide hydrolase is dependent on the regioselectivity of water attack within the active site. To design epoxideh ydrolase variants for enantioselective conversion of the three substrates (1a,2a, 3a,S cheme 1), we performed as eries of design calculations for these substrates. Substrates were docked in the active site and placed in ar eactive configuration using restraints. This included ah ydrogen bond between the leaving oxygen and D101, ac lose distance between then ucleophilic water and the attacked oxirane carbon,a nd close to linear orientationo ft he nucleophilic water,t he oxirane carbon, and the epoxide oxygen. Next, the Monte Carlo searcha lgorithm of Rosetta was used to optimize the identity and side chain geometries of amino acids surrounding the active site for either proRR or proSS attack of the nucleophilic water on the epoxide carbon (Scheme 1). The reactive geometries were essentiallyd efined as an eara ttack conformation ( Figure 1A). The explored sequence space was createdb yt argeting 11 selected positions around the active site ( Figure 1B)w ith randomization to any of the nine hydrophobic residues (AFGILMPVW). Thisw ay,R osetta generated variants forming al ow energy complex with the target substrate either in the proRR or proSS orientation,r esulting in thousands of possible proRR and proSS designs per substrate (Table 1). These sets of primary Rosettad esignsr epresent mutant libraries enriched in the desired phenotype.
To computationally screen these libraries in an orthogonal manner by the likelihood of showingt he correctr egioselectivity of water attack, we performed molecular dynamics simula-tions. From MD trajectories, enantioselectivities were predicted by scoring the fraction of time that the enzyme-substrate complex is in the same reactive proRR or proSS conformation used during the Rosetta design step (near-attack conformations, NACs, Figure 1A). Instead of as ingle or af ew long MD simulations, we performed al arge number of parallel MD simulations with independently assigned initial atom velocities (HTMI-MD), because this gives more extensive conformational sampling and better agreement with experimental results than single long simulations. [41][42][43][44][45][46] For each of the 15 000 Rosetta designs, at least five independently initialized MD runs of 10 ps were performed ( Table 1). Designs that passed initial selection rounds weres ubjected to more MD simulations, up to 80 10 ps and 5 100 ps. NACs were counted on the fly and their frequency was averaged for each design.T he ratio between averaged NAC frequencies for proRR and proSS attack conformations was used as ap redictor for enantioselectivityu sing Equation (1). This MD screening of the primary libraries de-   (Table 1). It was noticedd uring in silico screening by MD that al arger fraction of the stilbene oxide designs displayed high NAC frequencies and high predicted enantioselectivities than what was found with cyclopentene and butene oxide designs. As a result, more stilbene oxide designs survived the final selection (7.1 %) than designs for cyclopentene oxide 1a (1.5 %) and butene oxide 2a (3.8 %), even thought he selection criteria were set significantly more strict for stilbene oxide than for the other substrates (Table 1). This suggests ab etter occupancy of reactive orientationsf or stilbene oxide, and more restricted conformations of the stilbene oxide designs than of the designs with the two smaller substrates, at least during MD simulations.
Visual inspection of the top-ranked designs was carried out to verify that there weren on oticeable structuralp roblems that would decrease catalytic activity or enantioselectivity.F or every target enantiomer,t he variants predicted to have ah igh ee pred for that product were ranked, those with the highest [NAC] pref first, and variants were inspected until five variants were identified for each target enantiomer.I tw as noticedt hat for cyclopentene oxide 1a and butene oxide 2a there were few designsw ith obvious errors. In 25 inspected designsf or those two substrates there were only five with recognizable structurale rrors (one was unusually flexible, three had such a spacious active site that reorientation of the substrate seemed likely,a nd in one mutant there was anew H-bond to the epoxide oxygen, Ta ble S1).I nc ontrast, 10 of the 20 inspected variants for stilbene oxide displayed structural problems. Fourd esigns appeared too spacious (Supporting Information) and six designsh ad aw ater positioned such that attack on the unintended carbon atom of the epoxide seemed likely,e ven thoughthe NAC analysisdid not suggest this.

Experimental characterization
The top five designs for each product enantiomer of the three substrates werei nvestigated experimentally.T wo designs (26A, 45A) were selected twice, both for proRR hydrolysis of cyclopentene oxide 1a and butene oxide 2a (Table 2). We also included three designsp redicted to have proRR selectivity by molecular dynamics simulation while they were originally designed using Rosetta to have proSS selectivity.T he 28 new LEH variants werec onstructed in the thermostable LEH-P template describede arlier [56] by sequential rounds of QuikChange mutagenesis. The use of at hermostable parent enzyme increases the chance that protein functioni sm aintained upon introduction of mutations that may be too destabilizing in am esostable template and is also reported for directed evolution protocols. [57][58][59] The LEH template used here (LEH-P, T m,app = 70 8C, PDB 4R9K) is not the most stable enzyme from our previous work, as it lacks the disulfide bonds of the mosts table variant (T m,app = 85 8C, PDB 4R9L). [56] After sequence verification genes were expressed in E. coli To p10 or 10b for enzymeproduction. All mutantswere well expressed and could be isolated by His-tag metal affinity chro-matography.T ypical yields were 50-150 mg per liter of culture. The redesigned LEHs were very stable and could be stored for over 2years at À80 8Cw ithoutl oss of activity.M ostv ariants showedasomewhat lower apparent melting temperatureb ut more stable variants were also found ( Table 2). Overall thermostabilityw as well maintained with an average T m,app of the redesigned enzymes of 62.7 8Ca sc ompared with 70 8Cf or the template LEH-P and 50 8Cf or the wild-type LEH. Activity assays were done by mixing purified enzyme with epoxide and measuring diol formation.O ft he obtained 28 designs, 26 showed catalytic activity on the substrate they wered esigned for.O f these active variants,6 6% have an activity that is between 10-164 %o ft he wild-type activity for their respective substrate.
Chiral analysiso ft he formed diols revealed that 21 (77 %) of the designed activev ariants showedt he enantioselectivity predicted by MD simulations (10 % ee cutoff, Ta ble 2). The cyclopentene oxide (1a)d esigns were generated to examine if the design and selectionp rotocolg ave similarr esultst oa ne arlier study in which 34 LEH variants were tested for the same substrate. [32] Of the 10 new cyclopentene oxide designs, nine were active and had the predicted enantioselectivity.T he inactive designw as 45A, which was also selected as ad esign for butene oxide 2a,f or which it did have al ow catalytic activity. The five proSS designs produced (S,S)-diol 1b with ee values of 56-85 %, while lower ee values were obtained with proRR designs for (R,R)-1b.T he highere nantioselectivity of proSS designs is in agreement with previous observations. [32] The butene oxide designs were generated to test the possibility to control regioselectivity of water attack with av ery small prochiral epoxide substrate. All 10 variants were active and eight of them had the predicted enantioselectivity.A s with cyclopentene oxide, the proSS designsp erformed better than the proRR designs, with the best mutant (32A) providing (S,S)-diol 2b with 77 % ee The wrong predicted designsh ad very low enantioselectivity (ee of 2-6 %). Thus,f or both of the small substrates the MD simulations were an effective tool for predicting LEH enantioselectivity.
For the largest substrate, stilbene oxide (3a), six of the eight active variants showed the predicted enantioselectivity.This included two variants (51A, 60A) that were originally designed using Rosetta to have (S,S)-3b selectivity but for which the HTMI-MDs creening predictedp referential formation of (R,R)diol, which was in agreement with what waso bserved experimentally.I nt hese cases,M Dc orrected the Rosetta design prediction. On the other hand, the combination of Rosetta and MD for design and prediction of enantioselectivity still gave two mismatches between prediction and experiment in case of 3a designs. Of these, variant 52A was exceptional because it was predicted to give (R,R)-diol 3b whereas experimentally it produced (S,S)-diol with high ee (97 %).
Despite the two prediction errors, the results show that highly enantioselective variants could also be designed computationally for stilbene oxide.T he thermostablet emplate enzyme had 92 %( R,R)-diol selectivity,a nd four of the eight active new variants also displayed very high (R,R)-preference (> 91 % ee)w hereas two other designed variants displayed high (S,S)-diol preference (> 91 % ee). The more extensive pro-tein-substrate interactions with ab ulky substrate likely result in more restricted reactivec onformations of enzyme-substrate complexes,a ccompanied by high product enantioselectivity. Of the experimentally characterized variants for all three substrates, three have ar elatively large (> 100 3 )i ncrease in volumeo fa ctive site, as calculated from the decreased side chain volume of introduced amino acids, andi ndeed these three variants have low or no catalytic activity (see Supporting Information). When variants with ap redicted increase of the active site volume of > 100 3 would have been removed from the libraries, only the primary Rosetta libraries for cis-stilbene oxide would have shrunken (elimination of 178 of the 442 variants listedi nT able 1).
Catalytic propertieso ft he best variants. To examineh ow the use of Rosetta for redesign of LEH toward production of a specific diol enantiomer influences catalytic activity,w ee xamined the kineticp roperties (k cat and K M )o ft he best mutants ( Table 3). The LEH variants were produced using 1L cultures, giving again 50-150 mg purified protein per liter of broth, which was similart ot he yield of the parentt hermostable enzyme. Variants RR8 and SS16 were selected earlier as the best variants from as et of 37 designs for cyclopentene oxide 1a [32] and were included for comparison ( Table 3). The results showed that 46C and RR8 had catalytic rate constants( k cat )f or cyclopentene oxide that were similar to that of the thermostable template enzyme LEH-P.V ariants 43A and 24A had lower catalytic constants (2.5-and 7-fold, respectively). Variant SS16 variant displayed an almost 2-fold higher k cat than the template LEH-P.T hus, catalytic rates were quite well maintained. However,i nm ost cases the K M values were higher( 8-to 76-fold) for  Table 3). The kinetic measurements showt hat the main cause of the lower catalytic activities of mostd esigns found during initial tests (Table 2) is due to ahigher K M ,not al ower k cat .
The results for the variants designed to convert the other small substrate, cis-butene oxide 2a, are similar. For this substrate, we examined 32A, which has the highest enantioselectivity for producing (S,S)-diol and invertede nantioselectivity relative to the template. It showeda ni nsignificant increasei n k cat ,b ut the K M was much higher, resulting in ad rop in k cat /K M in comparison with the template LEH-P.
The resultsw ere different for the variantsd esigned to convert the bulky epoxide stilbene oxide 3a to (R,R)-or (S,S)-diol. Here, k cat values were lower than with wild-type, whereas K M values were better.The observation that k cat values are lowered in the designsf or 3a whereas NAC percentages duringM D simulations ( Table 2) werefine indicates that prediction of reactivities across different substrates using such short MD simulations is troublesome. This is not unexpected, as MD does not account for energy barriers along reactioncoordinates.
The high enantioselectivities obtained for stilbene oxide variants, relative to the designs for the two other substrates, are likely due am ore restricted conformational freedom in case of the bulky stilbene oxide.T he better K M values relative to those with small substrates might be due to the designp rocedure making the active site too spacious for small substrates, preventingasnug fit with good hydrophobic bindingi nteractions. Stilbene oxide is also more bulky than the natural substrate limoneneepoxide and the mutationsc reated enough additional space (ca. 67 3 for mutant 60A, calculated from decreased side chain volumes) for tighter binding and al ow K M .C onsequently,t he 3-fold lower k cat of mutant 60A with 3a wasa ccompanied by a6 -fold better K M ,l eadingt oa ni mproved catalytic efficiency in 60A. Furthermore, the tight substrate binding caused the specificity constants (k cat /K M )t ob ehigher for stil-bene oxide 3a than for cyclopentene oxide 1a and cis-2,3butene oxide 2a ( Table 3). The improved K M valuesf or 3a relative to wild-type were observed with all three tested stilbene oxide designs.

Preparative scale conversions
The stilbeneo xide enantioselectivities of designs4 1B and 52A are higher than reported for other LEH variants tested on this substrate. [50][51][52] To examine if these redesigned LEHs could be used in preparative scale conversions,the conversion of cis-stilbene oxide 3a to the (R,R)-and (S,S)-diols by variants 60A and 41B was examined under different reactionc onditions, including varying temperatures and cosolvents (Table 4). Cosolvents were tested because the solubility of the substrates and products in water is low.E ven in the presence of 10 %d ioxanei n 50 mm HEPES, pH 8.0, both cis-stilbene oxide and the diols were only partially soluble when added at 50 mm.U nder these conditions, the cis-stilbeneo xide remained visible as globular crystalsw hile the (R,R)-diol and the (S,S)-diol formed needles. The LEH variants 41B and 60A were active in this suspension. Additiono fc osolvents 1,4-dioxane and THF and the presence of ab iphasic system with al ayer of ethyl acetate were studied. No conversion was observed with variant 60Ai nt he biphasic system with ethyl acetate. Addition of 10 %d ioxaneg ave the best conversion, yielding up to 78 %d iol in 44 ha t3 08C. Remarkably,a dding THF as co-solvent (similar properties as 1,4dioxane) gave lower conversion than reaction conditions without cosolvento rw ith 10 %d ioxane. The conversion of cis-stilbene oxide by variant 60A was furtheri mproved (63 to 80 %) by increasing the reaction temperature from 30 to 40 8C. The best conditions (10 %d ioxane and 40 8C) were combined and gave ac onversion of 86 and 63 %f or variants 60A and 41B,r espectively.I ncreasing the dioxane concentration to 15 %w as beneficial for the proRR variant 60A, yielding ac onversion of 98 %, but drastically decreasedc onversion of 3a by the proSS variant 41B. The enantiomeric excess of the stilbene diol products wasa nalyzed by chiral HPLC (Figure S2-S4). For the best conversionst he following resultsw ere obtained: > 99 % ee for

Structural origin of enantioselectivity
To provide as tructurale xplanation for the observed mutant enantioselectivities we inspected the Rosetta designed structures andthe average HTMI-MDstructures of selected mutants. According to these models, the nucleophilic water molecule stays in virtually the same position ( Figure 1, Figure 2). This agrees with the X-ray structure, in which the catalytic water has an unusually low B-factor [48] indicating aprecise orientation due to H-bonds from Tyr53, Asn55,and Asp132.The enantioselectivity of the enzyme is therefore determined by the positioning of substrate relative to this water.I na ll of the modeled structures, the positional differences that influence enantioselectivity can globally be described as as liding motion of the epoxide carbon atoms in front of the nucleophilic water molecule ( Figure 2). Mutationsc ausing the substrate to reside more toward the center of the dimeric enzyme (i.e.,n ear b strands b4, b5, b6, and helix H4, see legend of Figure 1f or residue numbers) will lead to attack on the (R)-configured carbon of the epoxide ring, resulting in an (S,S)-diol. Vice versa, proRR attack will dominate if the substrate is positioned more toward the peripheral side (i.e.,n ear b strand b3a nd helicesH 1a nd H3).
In the modelso ft he proSS-selective variants that convert substrates 1a and 2a with high enantioselectivity (ee > 75 %), the dominants ubstrate orientationsa re achieved by steric hindrance introduced by mutations at the proRR side (e.g.,L 35W/ F, L74W/F,M 78F,a nd I80W/F). These mutationsw ill promote positioning of the substrate more toward the central (proSS) side. At the same time, space-creating mutations on the proSS side (I116V and F139L) will furtheri ncrease the preference for (S,S)-diol formation. Indeed, designsf or substrate 1a and 2a carrying both L35W and I116V gave an (S,S)-diol preference of > 73 % ee (Table 2). Furthermore, also in mutant 32A the steric hindrance mutationsL 35F and L74F on the peripheral( proRR)   side are accompanied by I116V,l eading to (S,S)-butanediol formation with the best ee of 77 %. The origin of the opposite (R,R)-selectivity of variants with substrates 1a and 2a can be explained in as imilarw ay.I ti s likely that increased sterich indrance due to mutations on the central side of the substrate binding cavity (e.g.,m utations L114W,I 116F/M) encourage positioning of the substrate closer to the peripheral( proRR)s ide of the binding pocket. This is visible in mutant4 6C for substrate 1a (e.g.,m utation I116F) and possibly in 47B for substrate 2a (e.g.,m utation I116M). However,s uch as teric effect is not clear in all predicted proRR mutant structures, and some designs indeed show low enantioselectivity (e.g.,4 5A, 50A). The weakly (R,R)-diol selective variant 63B contains mutation I116V,c reating space on the central (proSS)s ide, but it is accompanied by L114W,r educing that space. The combinedeffect of such mutations appears difficult to rationalize in view of effectso fs ide chain interactions and dynamics.
The mutants designed for stilbene oxide 3a have high product enantioselectivities but the mutations that cause them appear to not only involves terice ffects. The majority of the 3a designs showed (R,R)-preference,i ncludingv ariant 63B designed for (S,S)-enantioselectivity.S urprisingly,t he two most (S,S)-diol selective mutantsf or 3a (52A, 41B) carried mutation I116F,amutation that introduces steric hindrance at the central side and in case of substrates 1a and 2a favors (R,R)-selectivity (see above). The unexpected (S,S)-diol preference might be due to p-p interactions between the substrate and the newly introduced aromatic ring of Phe116. The same selectivity by attraction might hold for one of the best (R,R)-selectivem utants:v ariant 60A (ee > 99 %) carries the I80W mutation at the peripheral( proRR)s ide and no steric hindrance introducing mutationo nt he proSS site. For the other strongly (R,R)-selective mutant, 38A with ee > 99 %, the introduction of sterich indrance at the proSS side (F134W) can explain the improvement in enantioselectivity along the same lines as for substrates 1a and 2a.T he wild-type (ee > 90 %( R,R)-diol) also has aromatic functionality with Phe134 at the peripheral (proRR)r egion. Thus, it appears possible that enantiomeric preference with stilbene oxide is partially determined by an influence of attractive p-p interactions on binding orientationso ft he substrate.
Instead of at ranslational shift in the position of substrate, the stereoselectivity of LEH variants could also be influenced by rotating the substrate in the binding pocket by 1808 along an axis formed by the epoxide oxygen and the spot in between the two epoxidec arbon atoms.T his would switch the orientation of the epoxide carbon atoms relative to the nucleophilic water.H owever,s uch substrate rotations were not observed in any of the modelso rd uring MD simulations. Also attempts by us to dock substrates into productiveo rientations featuring such arotationwere unsuccessful.

Discussion
Changing enantioselectivity of limonene epoxide hydrolase by directede volution has been extensively investigated by Reetz, Sun and co-workers, focusing on improving directed evolution strategies including optimizing target positions and positional diversity in libraries. [47,50,52] Such well-optimized directed evolution protocols still rely on substantial experimental screening by chiral chromatography,w hich triggered us to examine if computational methods could be developed to enrich libraries and replace mosto ft he laboratory screening by computational screening of protein libraries. [60] In silico screening methods have been used earlier to design cocaine hydrolyzing enzymes with improved catalytic efficiency, [61,62] but also to design libraries of ac ytochromeP 450 harboring variants with controlled selectivity, [63] and to increasea midase activity in an esterase. [64] The results show that design of small sets of mutants with Rosetta and screening by MD simulations could indeed generate LEH variantsw ith desired enantioselectivity.R osettad esign targeted 11 positions at the same time, with an ine-residued iversityp er position. MD screeningw as done by multiple ultrashort simulations with on-the-fly scoring of reactive conformations and allowed to screen thousands of Rosetta designs.T his CASCOp rotocol [32] decreased the size of the library required, and for each substrate only five variants per desired enantiomer were used to find mutants with enhancede nantioselectivity.I nc ase of cyclopentene oxide 1a,d irected evolution [47,[50][51][52] and previous computational design [32] gave (S,S)-selective LEH variants producing the diol with similar ee as found here (60-95 % ee,T able 2).
Ac omparison of predicted enantioselectivities as calculated from NAC percentages ( Table 2) shows that there is good overall agreement only in qualitative terms, that is, (R,R)-or (S,S)diol stereopreference wasc orrectly predicted for 77 %o ft he variants using multiple short MD simulations with independent initialization.O nthe other hand, within as et of five designs, experimental activities for individual variants did not correlate with their computed NAC%, showing that the MD simulations as performed here do not provideq uantitative information on catalytic rates. Note, however,t hat rates shown in Ta ble 2a re strongly influencedb yb oth K M values and do not reflect k cat . Furthermore, in ab roaders ense, for each of the three substrates examined, the set of designs that gave the highest NAC% in MD also showedt he highest average activity.T hus, on average the proRR designs for 3a were on average more active than its proSS designs and also gave the highest NAC percentages for 3a.For the other two substrates proSS designs were more active and gave higherNAC%.
In this study,t he best enantioselectivities were clearly obtained with stilbene oxide 3a.Whereasm ost designs produced (R,R)-diol 3b (with ee > 99 %f or two variants), enantiocomplementarym utantsy ielding (S,S)-3b were also found, with a highest ee of 97 %. With butene oxide 1a; [51,52] the obtained variants showed high andmodest enantioselectivity in the productiono f( R,R)-and (S,S)-diols, respectively.T he LEH mutants found for stilbene oxide could be used to produce enantiopure (R,R)-diol and (S,S)-diol at preparative scale, indicating the potentialo ft his approach to generateapractically useful biocatalyst. Ring openingo fs tilbene oxide was tested earlier using mutantso ptimized on cyclopentene oxide and cyclohexene oxide, which resulted in highly (R,R)-selective variants ChemBioChem 2020ChemBioChem , 21,1893ChemBioChem -1904 www.chembiochem.org 2020 The Authors. Publishedb yWiley-VCH Verlag GmbH &Co. KGaA, Weinheim (with ee 99 %) but only modest (S,S)-selective variants (ee 44 %). [65] The resultss uggest that the likelihood of obtaining high enantioselectivity is muchb etter with the bulkier stilbene oxide than with the smaller substrates, with the two phenyl rings of stilbene oxide offeringm ore opportunities for steric hindrance as well as van der Waals and p-p interactions.
Similar to what we found with redesigned aspartases catalyzing asymmetric hydroamination of acrylates, [33] we observed that the lower activity found with most designsf or 1a and 2a as comparedw ith the template ( Table 2) was not due to decreased k cat ,but due to an increase in K M (Table 3). From apractical point of view,t his is less disturbing than the opposite, because for reasonso fp rocess economy preparative-scale applications must be carriedo ut at high substrate concentrations anyway.R ecently,S un et al. [65] attributed low activityo fL EH variants obtained by directed evolution to misalignmento ft he nucleophilic water,e poxide carbon and epoxide oxygen for an S N 2r eaction, causedb yi ncreased flexibility in the active site, but withoutr eportingk inetic data. This explanation likely does not hold for the computationally redesigned enzymes tudied by us because such am isalignment would decrease k cat and probably also K M ,w hich is not what we observe, except for some stilbeneo xide designs (Table 3). Proper alignment of the nucleophilic water and reacting substrate atoms forS N 2i so ne of the constraints in the Rosetta designp rocess and aN AC criterion during MD screening.
Surprisingly,f our of the observed enantioselectivitiesw ith stilbeneo xide 3a were opposite to the Rosetta design target ( Table 2). Twoo ft hese werec orrected in MD simulations. To understand how thesei ncorrect designsc ould emerge, we examined sequences and structures of Rosetta-optimized enzyme-substrate complexes of all substrates (see above). For substrates 1a and 2a,t he observed enantioselectivitiesc ould be explained by the combination of steric hindrance and space-creating mutations from the peripherala nd central side of the substrate bindingp ocket, acting together to steer the substrate in a proRR or proSS binding mode. This did not hold for several of the mutants designed for stilbene oxide 3a.F or both variants that are strongly proSS selective for 3a the only bulk introducing mutation is I116F,w hich would be expected to introduce steric hindrance at the centrals ide of the cavity and thus stimulate proRR selectivity. Also, for one of the two most proRR selective variants (60A, ee > 99 %) the I80W mutation would be expected to decreases pace at the peripheral side and thereby stimulate proSS selectivity.T hus, it appears that effects of mutationso nb ulkiness are poorly related to stereopreference in case of aromatic substrate 3a,s uggesting that electronic effectst hat are not well modeled, such as p-p interactions, dominate over stericf actors in determining substrate positioning or reactivity.
Standard computational design and MD simulations do not explicitly account for p-p interactions to save computation time. MD simulations can give more realisticr esults in cases where aromatic interactions play ar ole when the force field is adaptedw ith an additional noncovalent interaction term. [66] Whether the main effect of p-p interactions is on binding, conformational dynamics, or reactivityo fb ound substrates is uncleara tp resent. Recently,Z augg et al. [67] investigated the origin of enantioselectivity of Aspergillus niger epoxide hydrolase in the conversiono ft he chiral substrate phenyl glycidyl ether (PGE). Molecular dynamics simulations suggested that the protein does not differentiate enantiomersb ased on binding mode, and free energy calculations did not show significant differences between (R)-and (S)-PGE bindinge ither.T he authors suggested that the enantioselectivity is due to kinetic differences. For such an a/b-hydrolase fold epoxideh ydrolase acomputational analysis is more complicated due to the multiplicity of reaction pathways and chemical steps. Earlier,L au et al. [68] studied murine epoxide hydrolase with (1S,2S)-trans-2methylstyrene oxide using ab initio and density functional calculations,a nd suggested the importance of interactions between the substrate's phenylg roup and aromatic residues in the binding pocket.M oreover,L ind and Himo [69] published the reactionm echanism of as oluble epoxide hydrolase (StEH1) converting styrene oxide. They proposed coplanarity of the oxirane C1-C2 carbonsw ith the substrate's phenyl substituent, and p-p interactions between this phenylg roup and ah istidine and phenylalanine to be important for the stabilization of the transition state and for the selectivity of the enzyme. Rinaldi et al. [70] proposed that substrate-dependent LEH regioselectivity is relatedt or eorganization of the actives ite toward each ligand.B ased on QM/MM calculations, they confirmedt hat substrate-specific LEH regioselectivity is due to both conformational and electronic parameters.

Conclusions
We conclude that computational design and MD simulations are well able to predict and screen enantioselectivity of LEH variants in case of smalla liphatic substrates. Whereas highly selectivev ariants forp roduction of aromatic diols can be obtained, prediction accuracyi sl ower.I nv iew of the effect of interactions involving aromatic groups on epoxide hydrolase enantioselectivity, rapid scoring methods that more accurately include effectso fp-p interactions appear necessary to further improvec omputational screeningo fL EH variants acting on aromatic substrates.

Experimental Section
Materials. The meso-epoxides and their corresponding diols, oligonucleotides for mutagenesis, organic solvents and glycerol were purchased from Sigma-Aldrich. Restriction enzymes and PfuUltra Hotstart PCR Master Mix were obtained from New England Biolabs and Agilent, respectively.N i-NTAr esin was purchased from GE Healthcare Life Sciences. SYPRO orange was obtained from Ther-moFisher Scientific. Complete protease inhibitor cocktail tablets were bought from Roche. Media components were obtained from Difco (BD Biosciences).
Computational design. To design LEH variants for production of highly enantioenriched diols from meso-epoxides the previously developed CASCO strategy was used with only minor modifications. [32] The X-ray structure of the wild-type LEH (Protein Databank 1NWW) was used for computational design. Eleven positions around the active site (M32, L35, L74, M78, I80, V83, L103, L114, ChemBioChem 2020ChemBioChem , 21,1893ChemBioChem -1904 www.chembiochem.org 2020 The Authors. Publishedb yWiley-VCH Verlag GmbH &Co. KGaA, Weinheim I116, F134 and F139) were selected to mutate simultaneously to any of the nine hydrophobic residues (AFGILMPVW). Each of the three substrates was docked in the enzyme active site, either in a proRR or proSS conformation using Rosetta enzyme design [34,71] Catalytically productive binding modes were defined using ac onstraint file as previously. [32] This geometric description of how the substrate should be bound included the obligation to form Hbonds between the epoxide oxygen and D101 and between the nucleophilic water and D132, Y53, and N55. Furthermore, the water oxygen had to be close (1.8 )t ot he attacked carbon atom while the angle of nucleophilic attack (i.e.,f rom water oxygen, attacked carbon, and epoxide oxygen) should be close to 1808.A nother constraint was that the distance between the nucleophilic water and the non-attacked epoxide carbon atom should be > 3.8 .T oh inder undesired substrate-binding orientations, a bulky residue (W,For Y) was introduced at one of the eleven target positions, as this may reduce binding poses not contributing to the desired selectivity. [32] Rosetta enzyme design was used to simultaneously mutate the remaining ten residues to any of the nine hydrophobic residues and sequence-conformational space was searched for substrate-bound structures with low energy and acatalytically productive binding mode.
High-throughput-multiple independent MD simulations (HTMI-MD) were used for in silico screening of the generated libraries and to rank the primary designs with an orthogonal tool . Independent initialization of multiple trajectories increases the conformational space sampled by molecular dynamics and decreases the computational cost of the screening step relative to a single long MD run. [46] The reactivity and selectivity of each mutant were predicted by scoring the fraction of snapshots in which the enzyme-substrate complex is in a proRR or proSS near-attack conformation (NAC). The latter are defined by geometric constraints (Figure 1), which should be fulfilled for ar eaction to become feasible. The geometric criteria for proRR and proSS attack conformations were as defined using published quantum mechanical modeling. [72] The ratio of proRR and proSS NAC frequencies were considered to reflect regioselectivity of attack and thus product enantioselectivity according to Equation (1), e:e: in which ee is the predicted enantiomeric excess, [NAC] proRR and [NAC] proSS indicate the fraction of snapshots in which the enzymesubstrate complex is in a proRR or proSS conformation, respectively.P ositive values indicate predicted (R,R)-diol preference, negative values (S,S)selectivity.
The only modification from the existing procedures are listed in this paragraph. More design calculations were done than previously,a nd also more seeds per MD simulation. [32] For the current study,a pproximately 25 thousand design calculations were run per target substrate (Table 1), which is two times more than previously. Like earlier, [32] as tepwise scheme to rank the variants was adopted in which variants were eliminated as soon as they failed ac riterion ( Finally,t he best ranked mutants were visually inspected. For each of the targeted product enantiomers, only variants predicted to have ah igh enantioselectivity were visually inspected, starting with those variants that were predicted to have the highest fractions of NACs. The main reasons for elimination of designs were a too spacious active site cavity or an orientation of the substrate relative to the water that seemed in disagreement with the predicted enantioselectivity (Table S1). Only 15 of the 45 inspected designs were eliminated at the stage of visual inspection. Furthermore, no mutations were added at this stage even though this is common in the field. [73] As ar esult, the visual inspection only took afew hours.
Mutagenesis, expression and purification. For the expression of LEH and variants thereof in E. coli ap BAD based expression vector was used. This vector contained the gene of the thermostable variant LEH-P with an N-terminal hexa-histidine tag. [56] The computationally designed variants of LEH-P were constructed by Quik-Change site-directed mutagenesis using Pfu Ultra Hotstart PCR Mastermix (Agilent), combining multiple mutations in as ingle primer when possible, and omitting sequence verification between individual mutation steps. PCR reactions, transformations, plating and final sequencing were done in microtiter plate format. [74] The obtained plasmids were used to transform chemically competent E. coli To p10 or E. coli NEB10b cells (Thermo Fischer Scientific). For expression, cells were grown overnight in 5mLL uria-Bertani broth at 37 8C. All cultures were supplemented with 50 mgmL À1 ampicillin. The resulting culture was used to inoculate 500 mL Te rrific Broth medium and incubated at 37 8Ca nd 135 rpm. When an OD 600 of 0.6 was reached, expression was induced by adding 0.04 %( w/v)a rabinose and growth was continued at 30 8Ca nd 135 rpm. After 24 ht he cells were harvested by centrifugation at 6700 g and 4 8Cf or 15 min.
For protein isolation, cells were resuspended in 50 mm HEPES buffer,p H8,c ontaining 500 mm NaCl (3 mL per gram of cells) and half of ap rotease inhibitor cocktail tablet to prevent proteolysis (Roche Applied Science). After sonication (60 10 sw ith 20 si ntervals, Labsonic M), the extract was centrifuged at 35 200 g and 4 8C for 1h.T he supernatant was collected and the enzyme of interest was purified by gravity-flow affinity chromatography under native conditions using Ni-NTAa garose resin (Thermo Fischer Scientific). The protein concentration of the collected fractions was determined by aB radford assay,a nd selected fractions were desalted by Econo-Pac 10DG desalting columns (Bio-Rad). The purity of the prepared enzymes (yield 50-150 mg per LT Bm edium) was analyzed by SDS-PAGE. Enzymes were stored at À80 8Cu ntil further use.
Catalytic properties. Chiral chromatography was used to determine the enantioselective hydrolysis of three meso-epoxides (cis-2,3-butene oxide, cyclopentene oxide and cis-stilbene oxide) to chiral diols. In case of cis-2,3-butene oxide and cis-stilbene oxide, 5mgo fp urified enzyme was added to 50 mm substrate (virtual concentration of the suspension) in 50 mm HEPES pH 8( 800 mL total reaction volume). After incubation of the reaction mixture at 30 8Cf or 1h,5 00 mLo f5m NaCl was added and the samples were extracted three times with 600 mLe thyl acetate. The combined extracts containing diols were dried by adding anhydrous sodium sulfate, concentrated under vacuum, and dissolved in 100 mLe thyl acetate. For cyclopentene oxide, reactions were done in as imilar way after which 250 mg K 2 CO 3 was added to the reaction mixture followed by extraction for two times by 600 mLo fn -butanol. The combined extract was dried, concentrated under vacuum and resuspended in 100 mLn -butanol. In case of cis-2,3-butene oxide and cyclopentene oxide, chiral analysis was carried out by injecting 2 mLofthe extracts into an Agilent gas chromatograph equipped with af lame ionization detector and aH ydrodex b-TBDAc column (Aurora Borealis, initial temperature 40 8C, 10 8Cmin À1 to 150 8C, hold 20 min). For cis-stilbene oxide and its diols, samples were analyzed by HPLC on aL uxcellulose-3 column (Phenomenex, Utrecht, the Netherlands) with heptane/2propanol (90/10) as the mobile phase (flow rate 1mLmin À1 ,d etection at 254 nm). Samples from preparative scale reactions with cisstilbene oxide and the diols were also analyzed by HPLC on aC hiralpak AS-H column (Daicel Corp, Illkirch, France) with n-hexane/2propanol (90:10 (v/v)a st he mobile phase (1 mL min À1 ,d etection at 254 nm). Enantiomeric excess (ee)v alues were calculated from concentrations of the (R,R)-and (S,S)-product enantiomers.
To obtain steady-state kinetic parameters, initial velocities at different substrate concentrations were determined and fitted with the Michaelis-Menten equation.
Determination of the apparent melting temperature. The Ther-moFluor assay was used to determine the apparent melting temperatures (T app m )o ft he purified enzyme variants. [75] This method is based on monitoring the change in fluorescence of Sypro Orange dye during the thermal unfolding of ap rotein. The dye binds to the unfolded and exposed hydrophobic protein core, increasing its fluorescence signal. The assays were done as described before. [26] Synthesis of stilbene diols. Reaction mixtures (total volume 1mL) contained 10 mg of cis-stilbene oxide (final concentration 50 mm, suspension) and 6.37 mg of enzyme (final concentration 320 mm)i n 50 mm HEPES, pH 8.0, and were incubated (at 30 or 40 8C, 135 rpm) for 48 h. Substrate and products were extracted three times by 4mLe thyl acetate, dried over MgSO 4 and filtered. The solvent was removed by ar otary evaporator.T he residue was analyzed by 1 HNMR for conversion and by chiral HPLC to determine the enantiomeric excess. Chiral HPLC was used as described above.