“NAD‐display”: Ultrahigh‐Throughput in Vitro Screening of NAD(H) Dehydrogenases Using Bead Display and Flow Cytometry

Abstract NAD(H)‐utiliing enzymes have been the subject of directed evolution campaigns to improve their function. To enable access to a larger swath of sequence space, we demonstrate the utility of a cell‐free, ultrahigh‐throughput directed evolution platform for dehydrogenases. Microbeads (1.5 million per sample) carrying both variant DNA and an immobilised analogue of NAD+ were compartmentalised in water‐in‐oil emulsion droplets, together with cell‐free expression mixture and enzyme substrate, resulting in the recording of the phenotype on each bead. The beads’ phenotype could be read out and sorted for on a flow cytometer by using a highly sensitive fluorescent protein‐based sensor of the NAD+:NADH ratio. Integration of this “NAD‐display” approach with our previously described Split & Mix (SpliMLiB) method for generating large site‐saturation libraries allowed straightforward screening of fully balanced site saturation libraries of formate dehydrogenase, with diversities of 2×104. Based on modular design principles of synthetic biology NAD‐display offers access to sophisticated in vitro selections, avoiding complex technology platforms.


Introduction
Enzymes play an increasingly important role in the industrial preparation of chemicals,f uelled by sophisticated approaches to improve their properties through mutagenesis and selection of desired variants through screening. [1] The basic steps that make up ad irected evolution campaign require anumber of choices to be made that can be critically important in determining the successful outcome;D NA library generation (random or (multi)site-directed), library propagation (in vivo,transformation or in vitro,for example, PCR amplification), protein expression (in cellula or cellfree), catalysis (in cellula or cell-free), screening for product formation (directly by chromatography or MS,i ndirectly using af luorescent leaving group or af luorescent sensor of product). Arguably,m ethods for sensitive and selective product detection are key to enabling high throughput screening (e.g.inmicrofluidic droplet screening systems with capacities > 10 6 per day). [2,3] Ty pically,m odel substrates with fluorogenic leaving groups [4,5] are used for easy optical detection, but their structures (with large,h ydrophobic moieties such as fluorescein) often are too different from the substrate of interest, meaning the desired specificity and activity may not be selected for. Themore direct the detection of product is,t he better will the assay readout propel the screening and selection campaign in the desired direction in sequence space ("you get what you select for"). Fluorescent protein-based sensors,engineered to directly output arobust fluorescent signal in response to small molecule binding, would allow sensitive and specific detection, with ac onvenient, direct, optical readout. Although such sensors have been developed almost exclusively for the measurement of cell signalling,metabolism and homeostasis, [6] we argue that they present exquisitely sensitive tools with which to carry out high throughput screening.
We exemplify this approach with pyrimidine dinucleotideutilising redox enzymes that encompass one sixth of all characterised enzymes [7] and constitute an important target in industrial biocatalysis,f or example for the production of optically pure secondary alcohols. [8,9] Directed evolution has helped to bring about improved stability,a ctivity and (enantio)selectivity in NAD(P)H-utilising enzymes. [10][11][12][13] Focusing on the redox state of the co-factor,N AD(H), which many of these enzymes share,w eu sed ab est-in-class fluorescent protein-based sensor of NADH, called SoNar, featuring robust, excitation-ratiometric detection and alarge, 1500 %dynamic range [14] to develop an ultrahigh throughput screen. To avoid interference of NAD(H) redox state detection with cellular processes other than the desired transformation, we completely avoid cells by using in vitro transcription &t ranslation (IVTT) to effect expression of NADH dehydrogenase variants,e nabling interrogation of large libraries of enzymes without endogenous biological interference.I nl ieu of the genotype-phenotype linkage normally provided by cells, [15] we link the redox-phenotype to the DNA-encoded genotype on micrometre-sized paramagnetic beads.W edemonstrate our approach, which we call NAD-display,byultrahigh throughput selection of aformate dehydrogenase site saturation library,g enerated using our recently developed SpliMLiB method. [16] NAD-display is straightforward to implement in molecular biology settings, relying on am odular assembly of functional molecular components.T his approach enables ready access to ultra-high-throughput screening in droplets, [2,3] obviating the need to install microfluidic technologies to create compartments.

Results and Discussion
On-Bead Fluorescent Detection of Immobilised NAD(H) Redox State NAD-display uses water-in-oil emulsion droplets to generate aphenotype-genotype linkage by compartmentalization of single genes.I ne ach droplet compartment one protein variant was expressed (by IVTT) from agene library member attached to aSpliMLiB bead [16] and catalysis took place with an immobilised co-factor analogue as co-substrate.O nce the substrate is oxidised, the redox state of the co-factor indicates whether catalysis is complete,w hich is monitored using an immobilised fluorescent protein-based sensor.Then beads are screened for product cofactor state,a tu ltrahigh-throughput, and selection was achieved using flow cytometric sorting ( Figure 1).
Thek ey design features of the multifunctionalised bead centrepiece of NAD-display and their practical implementation are: (1) Cofactor immobilization. To immobilise the NAD(H) co-factor, abiotinylated analogue of NAD + (the commercially available biotin-17-NAD + ,F igure 2A)w as linked to the surface of beads covered with ab iotin-binder.A sab iotinbinding protein Tamavidin-2-HOTwas chosen and covalently attached to the bead as afusion protein with the SpyTag [17] in readiness to form an additional attachment point described later in (3). Recombinant expression and purification of Figure 1. Overview of NAD-display,using SpliMLiB beads [16] as the library input. (i)F irst paramagnetic, micrometre-sized beads, prepared with immobilisedSpyTag, as well as functionalities for DNA and biotin binding, are subjected to SpliMLiB, [16] resulting in beads densely coated in "monoclonal" DNA making up asite saturation library of the NAD(H)-dependent enzyme to be evolved. (ii)After SpliMLiB,b eads are furnished with an NAD + co-factor analogue (biotin-17-NAD + ). [For clarity only asingle dsDNA DNA molecule is drawn, although > 10 6 identical DNA molecules exist per bead. [16] ]( iii)B eads are singly encapsulatedinwater-in-oil emulsion droplets, together with in vitro transcription & translation mixture (IVTT) and the enzymatic substrate. Upon expression of functionald ehydrogenase enzyme, catalytict urnover is permanentlyr ecorded in the form of reduction of the immobilisedco-factor to NADH. (iv) The emulsion is broken and af luorescentprotein-based sensor of NAD + :NADH is attached to the beads, mediated by isopeptide bond formation between the SpyTag and aSpyCatcher-sensor protein fusion.( v) Beads are then sorted by flow cytometry,where the specific fluorescentcomponents of the NAD + :NADH sensor reveal the functionality of the dehydrogenase encoded by the DNA immobilisedo nthe bead.  Figure S4). B) The in vitro functioningo fthree different fluorescentprotein-based sensors of NAD(H)r edox state was tested with biotin-17-NAD + and its reduced form, both in the presence and absence of Tamavidin-2-HOT.C)Reduction and re-oxidation of biotin-11-NAD + immobilised on bead. Immobilised co-factor reduction was brought about by treating beads with LDH and sodium lactate, while the reverse was achieved with LDH and sodium pyruvate. D) Monitoring of bead redox state using SoNar-SpyCatcher and flow cytometry.
(2) Choice of sensor. Af luorescent protein-based sensor should report on the NAD + :NADH ratio to reliably report the redox-state of individual beads carrying immobilised cofactor with an optical signal. Although an umber of NADH sensors are on record in the literature, [19][20][21] the recently described "SoNar" NADH sensor produces ah ighly robust signal with ab est-in-class dynamic range [14] (DR, defined as the maximal change in excitation ratio divided by the minimum excitation ratio [22] ). We required the probe to bind the NAD(H)-end of biotin-17-NAD + (Figure 2A), even when the biotin-end of this analogue was bound to Tamavidin-2-HOT.W henw et ested sensor response in presence and absence of this biotin binder,SoNar was the only probe with high DR for NAD + /NADH detection in in the presence of Tamavidin-2-HOT( 1111 %). FREX [19] suffered ad ramatic decline in DR in the presence of Tamavidin-2-HOT (8 %), and Peredox, [20] while unaffected by the biotin binder, had aD Rt hat was low in both conditions (71 %), Figure 2B). SoNar was thus selected for further work.
(3) SoNar immobilization. To avoid non-quantitative labelling of beads due to reversible SoNar:NAD + binding (K d % 5 mM), [14] the SpyCatcher-SpyTag system, [17] which allows in situ formation of ap ost-translational, covalent isopeptide bond between two protein components,w as introduced as as econdary linkage.S pyCatcher was fused to the SoNar C-terminus,a llowing immobilization to the Tamavidin-2-HOT-SpyTag fusion described above ( Figure S3). In this way the sensor became covalently attached to the beadimmobilised biotin-binder ( Figure 2C).
Together these features should allow each bead carrying one library member gene to be screened for redox reaction turnover by flow cytometric monitoring of cofactor redox state.T he ability to reliably distinguish beads carrying either immobilised-NAD + or its reduced counterpart, was tested with fully assembled beads in af low cytometric experiment. Beads bearing immobilised NAD + were left either untreated or exposed to lactate dehydrogenase (LDH) and sodium lactate ( Figure 2C). After washing, the beads were labelled with SoNar-SpyCatcher and profiled by flow cytometry.A clear separation in ratiometric excitation signal, corresponding to their expected redox state,i se vident, suggesting that thetwo samplescan be clearlydistinguished with excellentDR (277 %, Figure 2D). To show that thescreen also functioned in theopposited irection,potentially allowing valuable reactions such as ther eduction of ketonesi ntoc hirala lcoholst obe monitored, ther eversibility of thes olid-phase redoxr eaction wasd emonstrated (Figure2D).B eads were firste xposedt o LDHa nd sodium lactate( butn ot yetS oNar-SpyCatcher) to achieveo xidation in thef orward reaction,w eret henw ashed ande xposed to LDHa nd sodium pyruvate( to triggert he reverser eaction, oxidation).T he redoxs tate of theN AD + / NADH cofactorw as monitoredb yl abelling thet hreeb ead samplesw ithS oNar-SpyCatcher.T he ratiometrice xcitation signal of ther e-oxidised beads( Figure2D, bottom panel) reflects thes tartingv alue (Figure2D, topp anel)d emonstratingt he solid-phaser edox reaction wasc ompletelyr eversible.

NAD-Display and Catalytic Assays for Formate Dehydrogenase (FDH)
With the NAD-display assay in hand, we next tested its functioning in the context of compartmentalised, in vitro protein expression. We chose to focus on the enzyme formate dehydrogenase from Candida boidinii; [23] CaBoFDH is used for biotransformations,w here NAD + is typically recycled back to the reduced NADH form using sodium formate as as acrificial substrate. [24] We confirmed that CaBoFDH was able to accept the unnatural NAD + analogue,consistent with previous reports that many dehydrogenases can accept modifications of NAD(H) at the adenine end of this substrate. [25][26][27] However,i ts hould be noted that in the presence of Tamavidin-2-HOT, there was amarked reduction in CaBoFDH activity with biotin-17-NAD + ,likely aresult of steric hindrance ( Figure S6). Nevertheless,a ctivity was still readily detectable in the latter situation and future implementations of NAD-display may be able to address this lowered activity by synthesizing analogues with longer linkers between the biotin and adenine moiety of biotin-17-NAD + . DNAe ncoding wild-type CaBoFDH and DNAe ncoding an inactive mutant of CaBoFDH (R258A) were separately loaded on Tamavidin-2-HOT-SpyTag beads,w ith DNA labelled with two different distinguishable fluorescent dyes ( Figure S7B), before biotinylated co-factor was bound to these beads.B eads were singly encapsulated with IVTT mixture in the droplets of aw ater-in-oil emulsion, together with 25 mM of sodium formate substrate ( Figure 3A). The emulsions were then incubated for several time periods, before they were chemically broken. To reveal the NAD + :NADH ratio of each bead, these were loaded with the SoNar-SpyCatcher sensor and subjected to flow cytometry.U pon addition of the enzymess ubstrate,as hift in the excitation ratio (from 0.36 to 0.81, Figure 3B i) and ii), respectively) indicated successful reduction of bead-bound NAD + ,albeit without aphenotypic linkage to genotype,due to the bulk conditions employed during catalysis for sample ii. When beads had been singly encapsulated in aw ater-in-oil emulsion during the expression &c atalysis phase,aclear separation of phenotypes could be discerned ( Figure 3B,iii)vi)). This separation thus also indicates that beads were largely singly encapsulated (consistent with the assumption of aP oisson distribution) as otherwise the phenotype of the beads that had been functionalised with DNAe ncoding the inactive CaBoFDH mutant would have merged with the phenotype of beads functionalised with DNAe ncoding active,w ildtype CaBoFDH (i.e., as in Figure 3B,i i)). Incubation of the emulsion sample for 16 hours led to an unexpected reduction in excitation ratio ( Figure 3C). This observation can be explained by aN ADH-oxidising background activity in the IVTT system, becoming apparent only once the rate of reduction slowed down to as ufficiently low level, due to depletion of the sodium formate substrate. Indeed, when ac ontrol experiment was carried out to probe directly for background oxidation by IVTT of NADH, this effect was clearly demonstrated ( Figure S5). To avoid this reverse background reaction from interfering with the assay and to find the optimal sorting gate,wefurther exploited the data to calculate the optimal sorting gate that would lead to the highest possible enrichment of wildtype from inactive mutant. As we were able to flow cytometricallyestablish the genotype of each bead, we could determine that the sorting gate as indicated ( Figure 3D), would lead to an enrichment of 312-fold. In addition, the fact that all possible events could be accounted for, including true positives and false positives, enabled the positive predictive value [28] of the screen to be calculated:when applying this sorting gate,itwas found to be 99.1 %. Even accounting for potential mis-sorting events caused by as orting flow cytometer,t his established an excellent maximal enrichment. Although this stringent gate does come at the inevitable cost of ar elatively high false negative rate (Q1 in Figure S7C), the ability to greatly oversample the library size (see library screen below) will allow many hits to be selected.

Combining NAD-Display with SpliMLiB to Test 20 000 CaBoFDH Mutants
FDH is prized in industry for its acceptance of arelatively cheap substrate with which to recycle NAD + back to NADH, as well for its catalysed reaction, in which the escape of the gaseous reaction product (CO 2 )h elps to drive the reaction equilibrium forward. [24] However,t he CaBoFDH and its homologues suffer from arelatively poor stability [29] and k cat , while their specificity for NADH has limited their application for recyclingo fN ADPH. [30,31] Work by Arnold et al. [32] showed that targeting sites for mutagenesis close to the binding site of the adenine end of NADH led to marked improvements in k cat in ad iverse set of dehydrogenases. [32] Inspired by this work, we set out to screen al ibrary of four sites simultaneously saturated, using NAD-display.T he sites, Y194, Y196, A229 and G234 were chosen based on their distance (maximally 5 )t oa deninese xocyclicn itrogen ( Figure 4A).
To generate beads densely coated with DNAe ncoding single variants of FDH (i.e.m onoclonal beads), SpliMLiB libraries were generated. In this recent library approach [16] beads are split into separate tubes,aDNA fragment carrying adifferent codon at atargeted site is immobilised, and beads are pooled and split again. Solid phase ligation is used to further extend the DNAand, at the same time,for introducing am utation at another targeted site,with the split &p ool process repeated. [16] We limited the number of residues per site to 12, keeping the library to am oderate size of 12 4 = 20 736 variants,b yl imiting replacement in saturation mutagenesis to those amino acids with side chains possessing similar physical and/or chemical properties to the wild-type amino acidsside chain (Table S5). Adetailed, sequence-level overview of the library construction design is provided (Supplementary Figure S2).
TheC aBoFDH SpliMLiB beads (1.5 million) were subjected to selection, using a4 -hour incubation at 25 8 8Cf or the compartmentalised expression &c atalysis phase.T he incubation time was deliberately set to introduce as tringent selection, as we had previously established apparent saturation at the 8-hour mark ( Figure 3C). DNAf rom the sorted beads ( Figure S8) and-as an egative control for SpliMLiB-NAD-display-mediated enrichment of functional variantsfrom beads not subjected to NAD-display selection was Figure 3. Optimization of NAD-displayincubation time and monitoring of potential for enrichment. A) Schematic representation of IVTTexpressed wildtype FDH catalysing the reduction of bead-immobilised biotin-17-NAD + using sodium formate as substrate and producing CO 2 . DNA encoding wildtype CaBoFDH was labelled with Cy5dye, while DNA encoding R258A FDH was labelled with TexasRed (TxR), allowing flow cytometry-based discrimination of both bead types. Both bead types were mixed and individually encapsulatedi nto water-in-oil emulsion droplets in the presence of IVTT and sodium formate. This step resulted in the compartmentalization of single beads in aqueousd roplets,t ogether with potential genotype-dependent reduction of the immobilisedco-factor.B )Excitation ratiometric histograms (normalised to the highest peak) of bead samples measured by flow cytometry.Beads were gated on the basis of TexasRed (grey trace) or Cy5f luorescence (black trace). Identical aliquots of 1:1bead mixtures (500 000 beads per sample)w ere subjected to various conditions during the IVTT and catalysis phase:i )beads exposed to IVTT but not sodium formate;i i) beads exposed to IVTT and sodium formate but without emulsification;iii-vi)beads exposed to IVTT and sodium formate, within an emulsion that was incubated at 25 8 8Cf or iii)2hours;iv) 3hours;v)8hours;vi) 16 hours. Emulsions (i.e. samples iii-vi) were chemically broken after the specified time period, all samples were labelled with SoNar-SpyCatcher and subjected to flow cytometric analysis. The diagram in vi)shows aclear distinctiono fpositive and negative hits that was gated for selection as shown in panel D. C) Median excitationr atio (405 nm/488 nm, emitting at 520 nm) of beads identified as carrying wildtype CaBoFDH DNA (due to the presence of Cy5) as af unction of emulsion IVTT incubationt ime. D) Determination of the optimal sorting gate (shaded in red) for sample iv) in panel B. The potential n-fold enrichment was calculated by dividing the ratio of beads functionalised with DNA encoding wildtype CaBoFDH (i.e.,positives, black trace) to beads with DNA encoding inactive mutant R258A enzyme (i.e.,negatives, grey trace) within the sortingg ate (406) by the same ratio within the entire sample (1.3). recovered by PCR, cloned into an acceptor vector and bacterially expressed. Enzyme activity was measured in bacterial lysate in ap late reader ( Figure 4B &T able S8 & S9). We can now use the sequence output of this selection in ahotspot analysis that assesses which amino acid substitutions remain functional. Thes equences of selected sorted hits (Table S9) displayed al imited number of different amino acids at each position, with most hits displaying activity close to wild-type.I nterestingly,p osition Y196 was almost completely conserved as the wild-type amino acid residue,w ith the very small number of alternative residues found at this position attributable to low-activity false positives (Table S9). By contrast, position G234 was found not to have been under selective pressure in our screen and we thus conclude that it does not appear to play as ignificant role,d espite being located within 5 of the adenine end of NAD + .A comparison with the input beads showed that these displayed low activity ( Figure 4B &T able S8) and an ear-random distribution of amino acids at each of the target positions, indicating that the SpliMLiB library was indeed randomised in each position ( Figure 4D). Ther ecovery of clones with wild-type activity,b ut different sequences suggests that selections for an activity that is higher than the rest of the library pool are possible,v alidating the NAD-display selection workflow.

Conclusion
We have established anew UHT assay principle,inwhich beads are equipped with multiple functions-linking genotype and phenotype,c arrying the product detection sensor and ar eaction cofactor. NAD-display supports ultrahigh throughput selections of NADH dehydrogenases through flow cytometric sorting of beads displaying immobilised NAD(H) and an NADH fluorescent sensor attached to those same beads.A ss uch, NAD-display falls under the "display" category of high throughput screening of enzymes,which are more commonly embodied by cell display technologies such as yeast [33] and bacterial display, [34] as well as phage display [35][36][37] and retroviral display. [38] In certain cases,i th as proven possible to select enzyme activity where both the enzyme and its product remain immobilised on the same surface,a llowing genotype-phenotype linkage without the need for artificial compartmentalization. [33,36,37] However,t o select enzymes with small-molecule products,i ti st ypically necessary to create artificial compartments from the start of catalysis up to the point of screening. [34] Thea dvantage that NAD-display provides is ap hysical decoupling between the catalysis phase-occurring within artificial compartmentsand the screening phase,which is carried out on the bulk, noncompartmentalised bead population. NAD-display is thus au ser-friendly approach, as it allows ap ause point to be introduced between the catalysis phase and the screening and sorting phase.This is an important advantage,asthe flexibility with regard to incubation time is crucial for the level of stringency( by setting al imit on the number of turnovers for variants with poor k cat ), ak ey parameter in ad irected evolution experiment.
To explore sequence space of NAD(P)H dehydrogenases beyond the limitations of plate-based campaigns,s everal groups have developed ultrahigh throughput screens.Inredox balance screens,t he bacterial host cellsn ative NADHconsuming pathway is genetically perturbed, such that cells carrying an NADH-dehydrogenase library variant successfully turning over NADH enjoy ag rowth advantage. [39] This approach, which has since been adjusted to allow for NADPH-utilising dehydrogenases, [40] benefits from straightforward selection schemes,b ut is limited to cell membranepermeable and cell growth-compatible substrates.S imilarly, as election scheme in which ar edox-sensitive transcription factor (TF) produced GFP in response to NADPH-oxidation by enzyme variants of interest suffered from cell-to-cell heterogeneity. [41] Thec entral positioning of NAD(H) and its associated enzymes in basic biochemical pathways renders these live,w hole cell-based screens troublesome.A na lternative approach was demonstrated recently where am icrofluidic device was employed to optically select droplets of aw ater-in-oil emulsion containing single lysed E. coli cells, substrate,N AD + co-factor and ac hemical sensor of NADH. [11] However,t he required expertise in microfluidic device creation and sorting remains as ignificant hurdle to Figure 4. CaBoFDH SpliMLiB library design and screening by NAD-display. A) On the left, the structure of CaBoFDH (PDB 5DN9) is shown with the targeted sites, Y194, Y196, A229 and G234, depicted in green and the 5 radius around the exocyclicn itrogen of adenine shown as ablue transparentsphere. On the right, the targeted residues and adenine are depicted in greater detail. Each of these residues was varied with codons encoding 12 different amino acids (Table S5), resulting in SpliMLiB library size of 20 736. B) Secondary screening of bacterial lysate before (input) and after (sorted output) flow cytometric sorting of beads. Progress curve slopes were normalised to wild-type CaBoFDH lysate activity measured with 10 mM NAD + and 10 mM sodium formate in abuffer consisting of 20 mM Tris-HCl (pH 8) and 100 mM NaCl. C) Sequence logo depicting the frequency of amino acids encounteredateach position for atotal of 36 clones sequenced each for the beads before and after sorting. take-up.N AD-display,b yc ontrast, simply relies on on-bead assembly of components that are either commercially available or easily produced as recombinant proteins.I nstead of am icrofluidic device,asimple filter brings about emulsion droplets.NAD-display substitutes on-chip sorting with aflow cytometric sorter,t hat is,t emporary access to aw idely used instrument is the only requirement for sorting.N AD-display thus helps "democratise" UHTS by enabling access to this important technique within biocatalysis for non-microfluidics-specialised laboratories,f or which there is ap ressing need. [42] Unlike most other examples listed above,NAD-display is an entirely in vitro platform, using cell-free expression, meaning it is i) time-saving, as bacterial transformation steps to generate the screening library (typically taking days) are skipped;i i) insensitive to potential cell-toxicity problems associated with either enzymes themselves or their substrates &products.Furthermore,weshowed here how NAD-display could be directly integrated with SpliMLiB,a llowing fully non-degenerate,m ulti-site saturation. NAD-display could therefore also fit within directed evolution approaches such as ISM and CASTing, [43] as well as approaches where sequence fragments from homologous proteins are shuffled. [44] More generally,t he design of the display construct is based on the ambition of synthetic biology to create bespoke complex functionality by assembly of simple building blocks. Them odular design principles of NAD-display may become ab lueprint for ag eneral assay principle based on two requirements,t hat is,t hat enzymatic products can be readily immobilised and af luorescent sensor is available.F or applications beyond NADH dehydrogenases,N ADPH-dependent enzymes may be monitored by iNAP [45] or the redox cofactor may be conceptually replaced by ATP, using sensors such as ATeam [46] or iATPSnFRs [47] providing access to ultrahigh throughput screening of ATPases/ATP synthetases or other enzymes coupled to this cofactor. As more sensors become available,avariety of functional manifolds in analogy to NAD-display can be constructed, enabling UHT campaigns to identify and improve enzymes catalysing am uch wider range of reaction than currently possible.