Engineering sensor proteins for antibody detection by design and evolution

Engineering sensor proteins for antibody detection by design

and evolution

Citation for published version (APA):

van Rosmalen, M. (2017). Engineering sensor proteins for antibody detection by design and evolution. Technische Universiteit Eindhoven.

Document status and date: Published: 20/11/2017 Document Version:

Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.

• The final author version and the galley proof are versions of the publication after peer review.

• The final published version features the final layout of the paper including the volume, issue and page numbers.

Link to publication

General rights

Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain

• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:

www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:

openaccess@tue.nl

providing details and we will investigate your claim.

Engineering sensor proteins for antibody

detection by design and evolution

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan de Technische Universiteit

Eindhoven, op gezag van de rector magnificus prof.dr.ir. F.P.T. Baaijens,

voor een commissie aangewezen door het College voor Promoties, in het

openbaar te verdedigen op maandag 20 november 2017 om 16:00 uur

door

Martijn van Rosmalen

van de promotiecommissie is als volgt:

voorzitter:

prof.dr. P.A.J. Hilbers

1

_promotor:

_{prof.dr. M. Merkx}

2

_promotor:

_{prof.dr.ir. L. Brunsveld}

leden:

prof.dr. G.J. Poelarends (Rijksuniversiteit Groningen)

dr. S.G.D. Rüdiger (Universiteit Utrecht)

prof.dr.ir. M.W.J. Prins

dr. T.F.A. de Greef

adviseur:

dr. J. Tel

Het onderzoek dat in dit proefschrift wordt beschreven is uitgevoerd in

overeenstemming met de TU/e Gedragscode Wetenschapsbeoefening.

1

Table of contents

Chapter 1

Introduction: Engineering sensor proteins

3

using directed evolution

Chapter 2

Yeast display for engineering FRET sensor

37

proteins by directed evolution

Chapter 3

Affinity maturation of a cyclic peptide

63

handle for therapeutic antibodies using

deep mutational scanning

Chapter 4

Dual color bioluminescent sensor proteins

97

for therapeutic drug monitoring of

anti-tumor antibodies

Chapter 5

Tuning the flexibility of glycine-serine

127

linkers to allow rational design of

multidomain proteins

Chapter 6

Epilogue

147

Summary

156

List of publications

160

Curriculum vitae

161

Dankwoord

162

3

C

HAPTER

1

Introduction: Engineering

sensor proteins using directed

evolution

Van Rosmalen, M. and Merkx, M. Engineering sensor proteins using directed

evolution, (in preparation)

4

Abstract

Engineering fluorescent / luminescent sensor proteins by rational design is a laborious process. Directed evolution offers a compelling alternative to rational design, but currently this strategy is not widely adopted in the field of protein sensors. Here we review the progress that has been made in this regard in recent years and the challenges that must still be addressed. In addition to the lack of high-throughput screening methods for sensor proteins a major impediment to their engineering is that each sensor requires unique target binding domains and therefore relies on a unique sensing mechanism. More generic sensor designs would be advantageous. We review potentially generic sensor formats that have been put forward by various groups and discuss their pros and cons. Finally, we propose a generic platform, FRET-bodies which makes use of insertion of hypervariable loops onto the surface of fluorescent proteins.

Engineering sensor proteins using directed evolution

5

Introduction

This thesis concerns the combination of two widely used protein engineering strategies, rational design and directed evolution for the engineering of sensor proteins. Engineering such sensors purely using rational approaches entails much trial and error and is a slow and tedious process. Directed evolution, generating and screening diverse mutant libraries may be a valuable complementary strategy. The literature review presented in this introduction serves as a reference that places the other chapters into the proper context. Specifically, it addresses the question which directed evolution strategies have been used so far for sensor proteins. In addition, the concept of a generic sensor scaffold is discussed along with several examples of such scaffolds. Chapter 2 presents an attempt to use yeast display for engineering the FRET-sensor protein for antibody detection, AbSense, towards recognition of new target antibodies. This proved to be unsuccessful, because the yeast displayed sensor protein adopted a low-FRET conformation, markedly different to that of the same sensor purified from bacteria. Chapter 3 then describes the affinity maturation of a so-called ‘meditope’ peptide that specifically binds to the therapeutic antibody cetuximab using yeast display and ‘deep mutational scanning’. Subsequently, in chapter 4 the peptides from chapter 3 were incorporated into luminescent sensor proteins for cetuximab detection. Several sensor proteins for detection of different therapeutic antibodies are presented in chapter 4 as well. This two-step approach was aided by a rational thermodynamic understanding of the sensor mechanism. Chapter 5 details the tuning of the length and stiffness of glycine-serine linkers often used in sensors and other multidomain proteins. Finally, the epilogue in chapter 6 provides suggestions for further research on how the antibody sensor proteins presented in this study may be further improved using a two-step approach of both rational protein design and directed evolution.

Genetically encoded luminescent or fluorescent sensor proteins report on the concentrations of target molecules and thus provide a valuable window into the inner workings of the cell. (1) Sensor proteins consist of engineered versions of naturally occurring luciferases and / or Fluorescent Proteins (FPs) as well as ligand binding domains. Upon engaging their intended target molecules, these sensor proteins change their luminescent or fluorescent properties. It is difficult to exhaustively categorize sensors, as many different designs have been reported and summarizing all of them is beyond the scope of this thesis. However, in general terms sensors can be grouped in several different ways by their sensing mechanism. A first distinction can be made between fluorescent and luminescent sensor proteins. Fluorescent sensor proteins require and external excitation light source for fluorescence, while luminescent sensor which employ luciferases require the presence of a substrate in order to function. This distinction has important practical consequences. Because fluorescence is much brighter than luminescence, these sensors are in principle better suited for microscopy applications. Fluorescent sensors are also the preferred choice if fast kinetics (<1 s) are required, such as for monitoring membrane

6

depolarization. This is because luminescence is less bright and therefore often exposure times of up to a second may be required in order to obtain enough signal. Bioluminescent sensors have recently gained in popularity with the availability of very bright luciferases. The main advantage of luminescence is the absence of noise from scattering of the excitation light. Therefore, very low (picomolar) concentrations of these sensors can be used, which makes it possible to accurately determine very low target concentrations. Another distinction can be made between intensiometric and ratiometric sensor proteins. As the names imply, intensiometric sensors respond to their targets with a change in emission intensity, while ratiometric sensors change in the relative ratios of two intensities at two different wavelengths, either in their excitation or emission spectra. An often used class of ratiometric sensors are Förster Resonance Energy Transfer (FRET) sensors, which respond to their targets with a change in FRET, or their bioluminescent counterparts, BRET sensors. The dynamic ranges of intensiometric sensors are generally larger, which makes them more sensitive than ratiometric ones, (although ratiometric sensors can in some cases also have very large dynamic ranges). The main advantage of ratiometric sensors is that their output signal does not depend on the sensor concentration. Ratiometric sensors are therefore preferred if they are used to accurately quantify the concentration of their target, while intensiometric sensors are better suited to visualize subtle changes in target concentrations. Sensor proteins can also be divided into genetically encoded and semisynthetic ones. Fully genetically encoded sensors make use of protein-based luciferases or FPs, which sometimes limits the choice of available proteins. Synthetic dyes can be brighter, more photostable or which may have other desired spectral or chemical properties which are not available in an FP. These dyes can be incorporated using site-specific self-labeling tags, such as SNAP-tags or CLIP-tags. However, for intracellular applications genetically encoded sensors are more straightforward, because the dye does not have to be delivered inside the cell. Finally, sensors can be classified by their targets, which can be as diverse as the constituent molecules of the cell itself and range from general physicochemical properties like ionic strength, redox state and pH (2, 3), to metal ions (4–7) or other messenger molecules (8, 9), metabolites (10, 11), post-translational modifications (12) or enzymatic activities (13, 14). As such, they allow researchers to follow events such as neuronal action potentials in real time in cells or even in vivo (15–17). Besides being used as tools in basic cell biology, luminescent and fluorescent sensor proteins find interesting applications in diagnostics (18) and forensics (19). Luminescent sensor proteins in particular, which require no sophisticated equipment to read out, lend themselves well for Point-of Care (POC) testing. Antibodies are important diagnostic biomarkers Point-of infectious diseases. Since the areas where the most severe infectious diseases occur with high frequencies often lack access to laboratories, POC antibody detection can quickly identify infected individuals, aiding both treatment and the management of epidemics. Another interesting application for sensor proteins is Therapeutic Drug Monitoring. The pharmacokinetics of drugs vary between patients and therefore patients receiving the same dose will have different serum levels, which may compromise the drugs’ efficacy. TDM involves frequently measuring patients’ serum levels to

Engineering sensor proteins using directed evolution

7 deduce individual pharmacokinetic parameters (such as clearance rates). This then allows the drug’s dose to be adjusted. TDM can be done in standard laboratories (e.g. by taking a sample every two weeks), but POC tests may be beneficial when more frequent (e.g. daily) measurements are required, such as in oncology where the time to find the optimal dose is very limited.

Figure 1.1 Schematic representations of engineered sensor proteins. A) Single-FP sensors increase in fluorescence upon binding a ligand, while in FRET-sensors (B) the distance and relative orientation between a donor and acceptor FP changes in response to ligand binding.

The main challenges in sensor engineering are 1) finding (or engineering) a binding domain which engages the ligand with a suitable affinity so that the sensor responds to biologically relevant target concentrations and 2) translating the often subtle conformational change induced by ligand binding into as large a change in output as possible, so as to increase the sensitivity (20, 21). This output signal can be a change in fluorescence intensity, in which case the sensor is called intensiometric (Figure 1A), or a change in the ratio of intensities at different wavelengths. Depending on whether this is a change in the emission ratio between wavelengths in the excitation or emission spectra, such sensors are called excitation- or emission ratiometric. An often-used principle for constructing ratiometric sensors is that of Förster Resonance Energy Transfer (FRET) between a donor and acceptor FP. FRET is highly dependent on the distance between the two chromophores and their relative orientation. If the conformational change in the target binding module results in a change in distance between and relative orientation of the FPs, this can be observed as a change in acceptor / donor emission ratio occurs (Figure 1.1B). One of the first reported FRET sensors, Cameleon, reports on intracellular calcium concentrations (4). It consists of FRET donor and acceptor fluorescent proteins (BFP and GFP or CFP and YFP) connected by the calcium binding domain Calmodulin and the calmodulin binding region of chicken myosin light chain kinase M13. Calmodulin and M13 interact when Calmodulin binds calcium (Figure 1.2A). This interaction decreases the distance between CFP and YFP and result in an increase in FRET (Figure 1.2B). In this example, FRET is increased by target binding, but the reverse is also possible.

8

Figure 1.2 Sensor mechanism of Cameleon FRET sensor for calcium ions. A) Schematic drawing of the Cameleon sensor in which the FPs are drawn as simple cylinders, reflecting their beta barrel structure. Calmodulin and the largely unstructured M13 are drawn as simple lines. Upon calcium binding, calmodulin and M13 form a complex, bringing the FPs close together, although their exact orientations are unknown. B) Emission spectra (excited at 380 nm) of cameleon-1 (with BFP and GFP as FRET pair) at the indicated free Ca2+ _{values at pH 7.47. Figure adapted from reference (4)}

Figure 1.3 eCALWY sensors for intracellular zinc measurements. A) Schematic drawing of the sensor mechanism. The FPs Cerulean and Citrine form a weak intramolecular dimer by ‘sticky’ hydrophobic surface mutations indicated by small circles. Zinc binding to ATOX and WD4 is mutually exclusive with this weak FP dimer. B) Emission spectra of eCALWY-1 before (black line) and after (red line) addition of 0.9 mM Zn2+ in 1 mM EGTA. Figure adapted from reference (5).

Our own group has previously engineered eCALWY sensors for zinc ions, which rely on the principle of mutually exclusive interactions (5). This sensor consists of the FRET donor / acceptor pair Cerulean and Citrine carrying hydrophobic surface mutations which cause them to form a weak dimer. Between the FPs, two zinc-binding domains, ATOX and WD4, connected by a long and flexible linker, were fused. These domains can coordinate a single zinc ion between them. In the zinc bound state, the N-terminus of ATOX and C-terminus of WD4 are positioned too far apart to allow the FPs to stay associated with each other (Figure 1.3A). Therefore, zinc binding results in a sharp decrease in FRET, which can be read out in the fluorescence emission spectrum (Figure 1.3B).

Engineering sensor proteins using directed evolution

9 These two examples (and many others) are modular, because the ligand binding domains can be simply replaced by other domains recognizing a new target. However, because the mechanisms rely on very specific geometries of the ligand binding domains, these new domains will inevitably require extensive optimization of binding domains and linker lengths.

Introducing this ‘switching’ behavior into a sensor protein – i.e. a ligand-induced conformational change in one domain leading to a global reorientation of different protein domains – is still a formidable protein engineering challenge. Most sensors are therefore engineered using a combination of structure-guided rational design and trial and error. This approach is both labor intensive and time consuming and requires structural information about the ligand binding domain to be available. Thus, while many sensors have been developed over time, there are still a lot of target species for which good working sensors are not available. Directed evolution methods would offer a more high-throughput approach to complement classical rational protein design.

Directed evolution strategies applied to sensor proteins

Directed evolution is a protein engineering strategy complementary to rational design, which has been extremely successful in the engineering of antibodies (22–24), antibody mimetics (25) and enzymes (26–28), with techniques such as phage display (29) and ribosome display (30) being instrumental to its success. This high-throughput approach mimics the natural process of protein evolution by (semi)-random mutation and selection based on phenotype in the laboratory. Therefore, it does not require extensive structural / biochemical information on the proteins of interest in order to improve them. A typical directed evolution experiment consists of the construction of a diverse mutation library at the DNA level, which is then expressed in a suitable host. From the library, mutant proteins are selected for improved properties and the mutations which gave rise to the improvement are identified by sequencing. This is an iterative process that can be repeated as many times as needed until no further improvements are obtained. As this strategy requires no knowledge on the effects of introduced mutations beforehand, it is particularly suited when no structural information is available and when the property to be improved is a simple one (such as binding affinity, catalytic turnover, etc.).

The past decade has seen some important examples of sensor proteins engineered using directed evolution methods, although none of these sensors were evolved to bind entirely new targets. The most extensively applied method for directed evolution of sensor proteins is bacterial colony screening (31). Random or focused mutant libraries are expressed in E. coli bacteria and plated on agar. Plates containing colonies can then be imaged in an automated fashion using appropriate excitation and emission filters and the colonies with the desired properties can be picked. This process is reasonable in terms of throughput, allowing libraries of up to 105 _{to be screened. It has been used to develop FPs with exceptionally high quantum yields} by screening on fluorescence lifetime (32–34). It has also been used to engineer sensors

10

indirectly, by evolving parts of sensors, such as FP-pairs (35–37), which were then assembled into full-length sensors in a rational way. This is essentially a suboptimal route, however, since it does not allow one to screen on the sensor’s response or fine-tune its affinity by evolution.

Figure 1.4 Engineering of GECOs by directed evolution. A) Schematic of the system for image-based screening of E. coli colonies. The GCaMP variant, as represented by GCaMP2 (PDB ID 3EVU and 3EVR) (38) has a TorA periplasmic export tag (39). B) Fluorescence excitation (Ex) and emission spectra (Em) of G-GECO1.1 normalized to the Ca2+-free state. C) Normalized excitation spectra of Ca2+-free (dashed line) and Ca2+-bound (solid line) B-GECO1 (blue), G-GECO1 (green), and R-GECO1 (red). D) Emission spectra represented as in (C). E) Absorbance (Abs) and emission spectra for Ca2+-free (dashed line) and Ca2+-bound (solid line) GEM-GECO1. Figure adapted from reference (40).

A challenge in applying this technique to evolve sensor proteins is presenting the target molecule to the sensors in such a way that the same agar plate can be imaged both before and after target addition. Sensors for Ca2+ _{ions have been evolutionarily engineered by Zhang et al. (40). These} Genetically Encoded Calcium indicators for Optical imaging (GECOs) were evolved from GCaMP3 (15, 41), which consists of a circularly permuted EGFP fused to calmodulin at its C-terminus M13 at its N-terminus. Ca2+ _{binding to calmodulin causes it to engage the M13 module, rearranging} the interface with the fluorescent protein and causing an increase in fluorescence. The sensor mechanism has been described in much detail elsewhere (38, 42).

Key to the successful directed evolution of the GECO’s by colony screening was their expression in the periplasm, rather than the cytosol. The periplasm has a naturally high calcium concentration and, in contrast to the cytosol, is directly linked to the extracellular medium via outer membrane pores (Figure 1.4A). Therefore, the calcium concentration in this compartment can be manipulated in colonies by evenly spraying a fine mist of chelator solution on the agar plates. Imaging the plates before and after chelator spraying enabled the authors to select clones with a high intensity in the calcium-bound state and a significant drop in intensity in the unbound state. In this manner, 104 _{– 10}5 _{clones were screened at a time. In this way an improved green} variant called G-GECO1.1 with a twofold increased dynamic range was selected (Figure 1.4B) as

Engineering sensor proteins using directed evolution

11 well as improved blue (B-GECO1), and red (R-GECO1) variants (Figure 1.4C, D). Finally, through several rounds of mutation and selection, even a ratiometric version, GEX-GECO-1 was developed. This variant with an excitation spectrum similar to that of B-GECO-1, shifts from green emission in the absence of calcium to blue emission upon calcium binding (Figure 1.4E). In a follow-up study, the red R-GECO-1 was further evolved into an even more red-shifted version ‘carmine’ CAR-GECO-1 (43).

Figure 1.5 Directed Evolution of Twitch calcium sensors. A) Emission spectra of Twitch-1 at zero calcium and calcium saturation. B) Schematic representation of Twitch-1, showing where random mutations were introduced. C) Directed Evolution workflow. D) Change in emission ratio divided by the starting ratio of Twitch-1, its parent TN-XXL and two improved versions Twitch-2 and 3 in hippocampal neurons exposed to up to 160 short-field stimuli. Error bars are mean ± S.E.M. (n = 60 – 240). Figure adapted from (44).

Other studies have used colony screening to develop calcium sensors. Thestrup et al used a combination of rational protein design and large-scale library screening to engineer so-called ‘Twitch’ FRET sensors (44, 45), using Troponin C (TnC) as a ligand binding domain. Previous sensors with similar design bound multiple ions, making their response highly non-linear (46–48). Therefore, the authors first generated a sensor called ‘Twitch 1’, that consisted of a minimal TnC fragment, binding a single Ca2+_{-ion, flanked by ECFP and a circularly permuted (cp) Citrine (44).} The 68 residue TnC fragment is largely unstructured in the Ca2+_{-free state, but folds upon calcium} binding, bringing the N- and C-termini close together and increasing FRET (Figure 1.5A). To decrease the residual FRET in the off-state of Twitch 1, a library containing mutations at 22 key TnC residues as well as in the linker regions was screened (Figure 1.5B). Because these FRET sensors were not efficiently exported to the periplasm, they were expressed cytosolically. Colonies were blotted onto Whatman paper, imaged, sprayed with calcium and ionophores to permeabilize the cells and imaged again. In this manner, between 104 _{and 10}5 _{colonies were} screened for low initial FRET and a large FRET increase upon calcium binding. The top 1% (1,000 colonies) was selected for small scale purification and spectroscopy in the absence and presence of calcium.

12

This screen was then followed by a lower-throughput (i.e. 120 clones) functional screen in primary neurons exposed to electric field stimulation to determine their in vivo response and subcellular localization (Figure 1.5C). In this way the sensors Twitch-2 and 3 were identified with over two-fold improved dynamic ranges (Figure 1.5D).

These two examples show how sensors for calcium (and probably other metal ions) can be engineered using bacterial colony screens. Engineering sensors for other (larger) targets in this way may be more challenging. While ions can relatively freely enter the periplasm, this is not true for most other molecules. Neither can the two membranes of E. coli be permeabilised to other targets without lysing the bacteria and thereby severing the link between genotype and phenotype.

Synthetic biology can be used to enable colony screening of sensors whose targets cannot be externally applied, but can be genetically encoded, e.g. sensors that report on transcription levels and mRNA localization. The mRNA FRET sensor system VAmPIRe was engineered in this way (49). Its design consists of an aptamer-binding peptide Rsg1.2 (50, 51) fused between ECFP and a circularly permuted citrine. Rsg1.2 binds to the HIV1 Rev Response Element (RRE) RNA stem loop structure with high affinity. In order to translate aptamer binding into an optimal change in FRET, two linkers of four fully randomized codons were inserted between Rsg 1.2 and each FP. This library was expressed in bacteria under an IPTG-inducible promotor, while the RRE aptamer was expressed under an orthogonal tetracycline-inducible promotor. In this way, plates could be imaged, sprayed with tetracycline and then imaged again so as to obtain information on the initial FRET and its change upon RNA binding in each colony. By screening 60,000 colonies, 1,500 of which were selected for expression and purification, the authors identified a superior sensor with a dynamic range of 160% as opposed to 80% at the start of the evolution (49). The same principle can be applied for sensors that respond to enzyme activity. This was first shown by Ibraheem et al (52), who improved the dynamic range of a sensor for Histone 3 lysine 27 trimethylation (H3K27me3) using colony screening. H3K27Me3 is an important epigenetic mark of gene inactivation. The starting sensor (12) contained a H3 peptide (residues 24 – 35) fused via a flexible linker to the H3K27Me3 binding Polycomb (Pc) domain and flanked by CFP and YFP. Enzymatic methylation of the histone peptide caused the Pc domain to interact with it, bringing the fluorescent proteins in close proximity. A library of sensors with different H3K27me3 binding domains and linkers of varying length and composition was then screened in E. coli to improve this sensor’s limited dynamic range of 28%. Assessing each colony in both the unmethylated and trimethylated states was facilitated by co-expressing a 119 residue vSET domain, which specifically catalyzes the methylation of H3K27 (53, 54) under an arabinose-inducible promotor. An attempt to induce vSET expression by spraying arabinose solution on colonies grown in the presence of IPTG proved unsuccessful due to inhomogeneity in spraying density across the plates. The authors then settled for the more laborious approach of manually replica-plating colonies on plates containing either glucose (repressing vSET) or arabinose

Engineering sensor proteins using directed evolution

13 (inducing vSET). In this way, they were able to improve the dynamic range of the sensors over two-fold. Similarly, Belal et al (14) used the same dual-expression vector to optimize the properties of the B Kinase Activity Reporter (BKAR). This sensor consists of a Protein Kinase B (PKB)-specific target peptide connected to a Forkhead-associated domain-2 (FHA) and flanked by CFP and YFP.

Figure 1.6 Strategy to evolve sensors for H3K27 trimethylation. A) Schematic of the dual expression vector allowing orthogonal induction of the library and the methylating enzyme. B) Schematic of the workflow. Ratiometric imaging of replica-plated colonies on agar with either glucose or arabinose to repress or induce expression of the methylating enzyme. Figure adapted from reference (52).

The FHA domain binds the target peptide in its phosphorylated form, reducing the FRET by 30%. A 420 member library of BKAR variants with different FP pairs and linkers was screened for increased FRET-change in the presence of a constitutively active mutant of the bovine PKB1 expressed under an arabinose-inducible promotor by spraying bacterial colonies expressing the library with arabinose. In this manner, a longer and stiffer central linker, (EAAAK)3, was found to increase the change in FRET from 30% to 40%. To show the modularity of their screening approach the authors then exchanged the PKB target peptide to a Cyclin-B/Cyclin dependent Kinase 1 (CDK1) target peptide and exchanged the bovine PKB mutant with the yeast homologs of cyclin 1 and CDK1 and selected a Cyclin B1-CDK1 sensor with a 60% dynamic range.

These examples underpin the feasibility of bacterial colony screening, extending the range of sensor proteins for which the approach can be applied to reporters of enzyme activity. However, a critical requirement is the availability of a constitutively active enzyme which can be expressed in soluble form in bacteria. This is not entirely trivial, as many enzymes (e.g. kinases) are toxic to

E. coli or require eukaryotic protein folding and processing machinery to become functional.

Another complicating factor is that the exact concentration or activity of the target can never be controlled, because of variability in expression levels between cells. Thus, although this approach is proven to be successful in specific cases, this does not necessarily mean that it can be applied as a generic platform for evolving new sensors.

14

Figure 1.7 Improvement of QuasAr voltage sensor by rational library design and three-step screening. A) model of QuasAr–mOrange2, represented by the crystal structures of Arch-2 (PDB ID 2EI4) and mOrange (PDB ID 2H5O). The color scale represents the degree of order in the crystal structure, as reported by the B-factor. B) Design of linker truncation libraries. QuasAr1.2, the best available variant at the time of the screen, was used as the electrochromic quencher. C) Hierarchical screen of truncated linker libraries for eFRET GEVIs. Constructs were first screened in E. coli for mOrange2 brightness and then screened for membrane trafficking in HEK cells. Voltage sensitivity was then tested via field stimulation in HeLa cells co-expressing the eFRET sensor and Kir2.1 (to lower the resting voltage to −60 mV). Figure from (55)

Nonetheless, even when colony screening cannot be used to select for an increased dynamic range it can still aid the evolution of sensors, if it is followed by a functional screen in mammalian cells. Such screens are carried out in 12-well or 24 well-plates, which limits their throughput to libraries of several hundred clones maximum. Therefore, high-throughput bacterial colony screens are used as a pre-selection step to filter out non-fluorescent mutants. The functional screen is then used to select mutants for their improved response in cells. One such challenging target is membrane voltage, an important signaling property in neurons, which transmit action potentials along their membranes. Zou et al (55) developed ‘eFRET’ sensors consisting of a previously described mutant archaeorhodopsin membrane protein, called QuasAr (56) fused to mOrange2. Membrane depolarization leads to protonation of the retinal chromophore within the QuasAr, shifting its absorbance spectrum to overlap with the emission spectrum of mOrange2, thus quenching it. The authors reasoned that optimal quenching in the depolarized state would be obtained by making the junction as tight as possible (Figure 1.7A). However, too short linkers would result in background quenching and impair the folding of mOrange2,

Engineering sensor proteins using directed evolution

15 resulting in diminished fluorescence even in the resting state. Therefore, seven libraries were made by stepwise omitting up to 15 residues from the flexible C-terminus of the QuasAr and up to 17 residues from the N-terminus of mOrange2. In each library, the deleted sequence was replaced by two fully randomized codons (Figure 1.7B). Each library was screened in E. coli for mOrange brightness, selecting the 24 brightest colonies from each library for secondary screening in HEK cells to select for plasma membrane co-localization.

The top 5 clones with the shortest linkers, but still having good brightness and membrane trafficking were then investigated in neurons for their response to action potentials (Figure 1.7C) As expected, the response increased with larger deletions with the optimal sensor having a deletion of 14 residues from the QuasAr and 10 from the FP and a linker consisting of Arg-Leu. This sensor exhibited an increase in dynamic range over its parent of over twofold. Deletion beyond this point resulted in severe reduction of fluorescence in the resting state. The same basic procedure has also been applied for the engineering of a red fluorescent voltage sensor, FlicR (57, 58), which consists of a circularly permuted mApple protein fused to a transmembrane

Ciona intestinalis voltage-sensing domain (CIVSD) (59). In order to more efficiently transmit the

voltage-induced conformational change in the CIVSD to an intensity change in mApple, random mutagenesis and library screening were applied to this construct. Again, selection consisted of a colony screen for red fluorescence and subsequently a functional screen for voltage sensitivity. Because primary neurons are not ideal as a cell model for high-throughput screening, HeLa cells co-expressing the inward-correcting potassium channel Kir2.1 as well as the green voltage sensor ArcLight (as a control) were used. In these cells a robust membrane depolarization can be induced by electric field stimuli (60). Through 8 rounds of random mutation and selection the relative response of FlicR variants over that of ArcLight (a co-transfected green control sensor) gradually increased and the final sensor, FlicR1, carried 12 mutations and had a 3.6-fold higher voltage sensitivity compared to ArcLight, Notably, in both of these studies, the high-throughput bacterial colony screen was merely used to pre-select for brightness. The functional screens used in these studies to select for response magnitude still require manual field stimulation of cultured cells and are therefore very limited in throughput. Expanding the throughput of the functional screen would allow a more streamlined directed evolution workflow. An improved version of ArcLight was engineered in this way (61). Instead of bacterial colony screens, nine single-site saturation mutagenesis libraries of ArcLight (62) were screened by automated imaging in an improved cellular model system of HEK cells expressing the sodium channel Nav1.3 and the potassium channel Kir2.1. These cells do not require field stimulation, but spike spontaneously, much like neurons (63). Compared to other functional screens for membrane voltage, the throughput and amenability to automation of this model is much improved. The authors succeeded in reverting the signal change of ArcLight (61). Rather than decreasing in brightness upon membrane depolarization, their evolved sensor, Marina, increased in brightness, which is a preferable property for sensors.

16

While the examples highlighted above, clearly underscore the benefit of using bacterial colony screening and secondary functional screens in mammalian cells for engineering sensor proteins, as a general rule this method can only be used to improve the properties of existing sensors. Engineering entirely new sensors in this way is not feasible, because of limitations in terms of screening capacity and because it can be challenging to screen sensors in both on and off states for increased dynamic range.

Fluorescence Activated Cell Sorting (FACS) offers a throughput which is several orders of magnitude higher than that of colony screens. Cell sorters can routinely analyze thousands of cells per second, allowing millions of clones to be screened each round. Each of these cells can be interrogated by multiple excitation lasers, each of which is equipped with several emission filters. This allows selection to be based on complex, multidimensional criteria (e.g. correction for differences in expression levels between different cells). FACS-based selections of bacteria expressing libraries of FPs have been extensively reported (64), but relatively little has been done with regard to sensor engineering. A notable exception is the evolution of the CyPet YPet FRET pair (65). The evolutionary trajectory that yielded this improved FRET pair consisted of four rounds of mutagenesis and FACS selection, starting from ECFP and YFP fused via a flexible linker containing a caspase cleavage site. In the first round a library in which YFP was randomly mutated by error prone PCR was screened for increased FRET. The enriched mutations were then recombined by DNA shuffling and fused to a randomly mutated CFP pool to create a second library. Manually adding the mutations from the fast-maturing YFP variant mVenus to the best candidate selected from this library then yielded YPet. Next, the cyan FRET-partner was optimized to improve its brightness, while maintaining a high energy transfer efficiency to YPet. To this end, a random mutation ECFP library was expressed without the acceptor and enriched for increased brightness. The enriched pool was then fused to YPet and further selected for high FRET. A final CFP library was then created by shuffling, thus combining the effects of different individual mutations. This library was again screened first for high CFP brightness and then for high FRET in combination with YPet. The resultant proteins, CyPet and YPet, contained six and seven mutations respectively and exhibited a 7-fold improved FRET compared to the CFP and YFP parents.

This example shows how powerful FACS can be in isolating rare beneficial mutations from large and diverse libraries, but at the same time it underscores how difficult it can be to really screen for a complex photo physical properties. The spectral overlap between CyPet and YPet and the quantum yield of the acceptor did not sufficiently change over the course of their evolution to explain their efficient FRET. Rather, two mutations S208F and V224L present from YFP3 onwards were later found to enhance dimer formation in GFP-derived FPs (5, 66–68).

Engineering sensor proteins using directed evolution

17

Generic sensor scaffolds

In addition to investigating a directed evolution method for sensor proteins, another major question is how to design sensors to entirely new targets in a generic way. Most current sensors utilize existing ligand binding domains, but these are unique for each different target, with different stabilities, N-and C-terminus positions, etc. This makes it difficult to quickly assemble sensors for new targets.

Figure 1.8 Quenchbodies as a generic sensor scaffold. A) Schematic of the incorporation of the TAMRA dye into a scFv using amber suppression and in vitro expression. B) Model of the TAMRA-scFv, built using the WAM antibody modeling server (http://antibody.bath.ac.uk) with CONGEN side chain building and Accessibility profile screen methods. Trp residues, Trp33H, Trp36H, Trp47H, Trp103H, and Trp35L are colored

green. C) Fluorescence emission spectra of TAMRA-scFv with excitation at 550 nm in the absence and presence of BGP-C7 peptide. D) Titration curve of the fluorescence intensity at 580 nm. The intensities are relative values with respect to that in the absence of BGP-C7 peptide. Figure adapted from reference (69).

If a single, diverse library could be used to select sensors for many different targets based on their change in fluorescence, this would greatly accelerate the development of new sensors. Therefore, a generic scaffold from which to construct such a generic sensor library would be a major contribution. A scaffold is a protein which is largely constant, but contains a few dedicated target binding loops which can be readily diversified, because they do not contribute to the overall fold of the scaffold itself. The antibody protein is nature’s own most well-known scaffold

18

protein. Antibodies are naturally able to recognize a wide range of targets with high affinity and specificity. This tremendous versatility is made possible by six hypervariable Complementarity Determining Regions (CDRs) that engage the target. These loops of 5 – 15 amino acids in length protrude from the target-binding face of the antigen binding fragment (Fab) and are unique to each different antibody. Then how to make a sensor scaffold that can similarly recognize a myriad of target molecules and respond to them with a large change in fluorescence or luminescence? A few studies have addressed this question. Antibody fragments themselves have been proposed as a generic sensor platform. (69–73) By incorporating an unnatural amino acid carrying the fluorescent TAMRA dye into the N-terminus of a single-chain variable fragment (scFv) against human osteocalcin (Bone Gla Protein BGP), a ‘quenchbody’ was obtained (69). Through optimizing the amino acid context around the amber codon (74), the unnatural amino acid could be efficiently incorporated by in vitro transcription / translation (Figure 1.8A). In the absence of antigen, the dye was found to fold back onto the antigen-binding face of the fragment, where it was quenched by nearby tryptophan residues (Figure 1.8B) via Photo-induced Electron Transfer (PET) (75–77). Antigen binding disrupted the physical contact between the tryptophan residues and TAMRA and effectively de-quenched the dye. Indeed, upon addition of BGP or its C-terminal 7-residue peptide was found to increase fluorescence over fivefold in a concentration dependent manner (Figure 1.8C, D)

This type of sensor has the potential to be generic, because some relatively conserved tryptophan residues are situated near the antigen binding site. Indeed, the authors introduced the TAMRA dye in the same way into a few other scFvs. The method has also been applied to the more stable Fabs instead of scFvs (72) and it has been shown to work with different dyes, although those must be conjugated in a semi-synthetic fashion (78). Recently, a cannabis sensor was developed using this method (19). A ratiometric version of a quenchbody has also been developed, in which the TAMRA dye acts as a FRET-acceptor of rhodamine green (73). The main disadvantage of this format is that scFvs are not very stable, while Fab fragments require a disulfide bridge, which makes their use in the cytosol problematic.

Other designs which harness the binding versatility of antibodies to make generic sensors have also been put forward. One example is an intermolecular sensor format consisting of single VH and VL domains fused to CyPet and YPet. (79) The authors used a ‘split scFv’ against the C-terminal peptide of BGP. Binding of BGP-C7 to both VH and VL domains brings the FPs into close proximity, resulting in FRET (Figure 1.9A). Emission spectra of this construct indeed confirmed an increase in FRET upon BGP-C7 binding in a concentration-dependent manner (Figure 1.9B) with an EC50 of 38 nM (Figure 1.9C), close to the IC50 value of the antibody that was used. This design may be problematic for large targets such as proteins which may not fit within the relatively small space between the scFv and the FPs. For these targets, a similar, but intramolecular design has also been reported in which the VH and VL domains of an anti-human serum albumin (HAS) antibody were linked by a C-terminal disulfide bridge (80).

Engineering sensor proteins using directed evolution

19

Figure 1.9 Generic intramolecular FRET sensor using FP fusions to VH and VL domains. A) Model of the sensor

mechanism. B) Fluorescence spectra of a mixture of 50 nM CyPet-VH and 50 nM YPet-VL excited at 430 nm

and normalized at 475 nm in the presence of the indicated concentrations of BGP-C7 peptide. C) Titration curve of fluorescence intensity ratio F525/F475 (n = 3). Figure from reference (79).

In this case the N-terminally fused FPs were close together in the absence of antigen, sterically blocking the antigen binding site. HSA binding therefore pushed the FPs apart, decreasing FRET. This ‘open flower’ immunoassay is in principle generic, because the sensor mechanism relies only on the fact that the N-termini of the VH and VL domain are close to the antigen binding site, which is the case for all antibodies. Therefore, unless the target would be so small that it could still access the CDRs without displacing the FPs, the method should be applicable to all antibodies irrespective of their target. However, both of these designs can only be used as purified proteins in vitro (e.g. as a diagnostic test). Their in vivo use is difficult, because of inevitable differences in expression level of the different individual sensor components.

A semi-synthetic bioluminescent design called LUCID was recently reported (81) in which the luciferase NanoLuc and a SNAP-tag were fused to the N-termini of the light chains of three Fab

20

fragments against the small molecule drugs, methotrexate, theophylline and quinine. The SNAP tag is a small engineered DNA-repair enzyme that reacts with benzylguanine and forms an irreversible covalent bond with it. Synthetic tethers containing a spacer, an acceptor fluorophore and weaker-binding derivatives of the target molecules were attached to the SNAP-tag (Figure 1.10). In the absence of target, the weak, tethered ligand binds the antibody in an intramolecular interaction and brings the acceptor fluorophore into close proximity to NanoLuc, resulting in high Bioluminescence Energy Transfer (BRET), while the strongly binding native target outcompetes the tethered one, resulting in weak BRET. This design was shown to be applicable for Point-of-Care therapeutic drug monitoring of the drugs and could be used directly in serum spotted on paper. It is generic insofar that different Fabs can be used to construct LUCIDs for new targets, but engineering them by directed evolution may be problematic. A key step in the engineering of LUCIDs is tuning the affinity of the tethered ligand, so as to obtain sensors which respond in a relevant concentration range. In the reported LUCIDs this optimization was done in a rational way guided by the structure of the antibody-antigen complex. In general, antibody-based sensor designs carry the major disadvantage that they cannot be used for intracellular purposes. Therefore, other proteins may be explored which can be endowed with binding properties similar to those of antibodies, by grafting variable loops onto their surface. Display technology can be used to select non-IgG binders, such as Affibodies, Affimers, DARPins and FN3-monobodies. Such antibody mimics may serve as generic binding modules in sensor proteins. FN3 monobodies are small and highly stable protein domains lacking disulfide bridges. They consist of a stack of two beta sheets each containing three antiparallel strands. The beta strands are connected by flexible loops, three of which (B-C, D-E and F-G) can be extended and fully diversified without influencing the overall structure of the protein. These three loops comprise the same face of the protein and thus create a unique binding surface (25, 82, 83).

Figure 1.10 Schematic depiction of a LUCID antibody sensor. The LUCID consists of a Fab fragment in which a NanoLuc and SNAP tag are fused to the N-terminus of the light chain. To the SNAP-tag a synthetic linker containing a BRET acceptor fluorophore and a competitor ligand are attached. Binding of the target in solution, displaces the competitor and increases the distance between the fluorophore and NanoLuc, thereby decreasing the amount of BRET. Figure from ref. (81)

Engineering sensor proteins using directed evolution

21 Monobodies have been modified into generic sensor scaffolds by the group of Shohei Koide, who assembled so-called affinity clamps (84). This clamp consists of a peptide-binding PDZ domain, fused to a monobody. The PDZ domain contains a groove that can bind linear peptides with low affinity and specificity, while the monobody serves as a ‘lid’ on the clamp structure, closing over the binding groove when the peptide is bound. The peptide-binding interface can be diversified and clamps for different target peptides screened by directed evolution. Because this clamp structure undergoes a large conformational change upon peptide binding (85), it was used to construct a FRET sensor, by genetically fusing CyPet to the N- and YPet to the C-terminus of an affinity clamp binding to the C-terminal peptide motif PQPVDSWV derived from the human RVCF gene (Figure 1.11A) (86).

Figure 1.11 Affinity clamp sensor. A) Diagram showing the mechanism of this generic scaffold. Closing of the affinity clamp is sterically incompatible with a CyPet-YPet dimer. B) Fluorescence emission spectra (excited at 415 nm) of the clamp sensor in PBS in the absence (solid line) or presence (dashed line) of 1.2 μM target peptide. Figure adapted from reference (86).

Emission spectra of this clamp sensor showed a prominent YPet peak very similar to that of CyPet-YPet separated by only a short linker. Addition of the target peptide resulted in a drop in FRET emission ratio of 250% (Figure 1.11B). Titrations of the peptide to this sensor revealed a Kd of 170 nM, 3-fold larger than that of the clamp without the FPs (56 nM), which can be explained by the fact that the FP-dimer must be broken, which costs energy. This design is very generic so long as the target is a linear peptide. Affinity clamps can be engineered for new targets by directed evolution of the interface between the FN3 and PDZ domains. This could therefore be a very suitable sensor scaffold for applications such as sensing post-translational modification status on linear peptide motifs, such as histone tails.

Other affinity reagents have also been used in the design of a generic sensor scaffold. Designed Ankyrin Repeat Proteins (DARPins) were used to construct sensors for large proteins or protein complexes (87). The Ankyrin repeat fold is essentially a linear polymer of repeating units that consist of two antiparallel alpha helices connected by a long and flexible loop. The helices of adjacent units assemble into a linear, but slightly curved stack, the concave side of which is lined by the loops from each successive unit. DARPins are engineered Ankyrin repeats with special capping units to stabilize the start and end of the DARPin.

22

Figure 1.12 Fluorogen Activating Designed Ankyrin Repeats (FADAs) as parts of generic sensors. A) Crystal structure of the FADA-3210 homodimer (cyan and magenta cartoons) binding one MG-2p molecule (green sticks). B) Structural models of homodimers (yellow and brown) of rigid DARPin-DARPin fusions of the GFP-targeting 3G124 (N-terminal) and FADA-3210 (C-terminal) alone (left) or in complex with GFP (right). The models were obtained by assuming complete rigidity of the connecting helix and by superimposition with the (unpublished) crystal structure of 3G124 in complex with GFP. C) FADA-3210 or DD_3G123_13_FADA-3210 (100 nM) was incubated with 100 nM MG-2p in the absence or presence of 200 nM GFP. Data were normalized to FADA-3210 and DD_3G123_13_FADA-3210 in the absence of GFP. Error bars indicate SD (n=3). Figure adapted from reference (87)

Between these capping units, a variable number of internal units can be inserted. Furthermore, the loops of these units can be fully diversified. Because of this versatility and modular design principle, DARPin binders to many different targets can be selected from diverse libraries with randomized loops. This strategy was followed in the engineering of a Fluorogen Activating DARPin (FADA-3210). This DARPin was selected and affinity matured by a combination of ribosome- and yeast display to bind the fluorescent dye Malachite Green (MG) or a cell-permeable derivative MG-2p. This dye exhibits hardly any fluorescence when free in solution, but its intensity increases over 11,100 fold upon interaction with the FADA. Contrary to what its name suggests, the emission maximum of MG is around 650 nm. Mass spectrometry and the crystal structure of the FADA-MG complex revealed that the DARPin formed a homodimer with the flexible loops facing inward, sandwiching the malachite green (Figure 1.12A). Because fluorescence was dependent on dimer formation, this design lent itself well for the construction of a sensor relying on the principle of mutually exclusive interactions. In this sensor design a second ‘targeting’ DARPin was fused to the FADA via a shared rigid 13 residue alpha helix. The rigid fusion was done in such a way that dimerization of the FADA was not sterically impaired in the free state, but that binding of two large protein targets to the two targeting DARPins would

Engineering sensor proteins using directed evolution

23 disrupt FADA dimer formation. As a proof of concept, an anti-GFP DARPin 3G124 was used. Models of this fusion-DARPin suggested that in the dimeric structure, two bound GFP molecules would sterically clash with each other and thus be mutually exclusive with FADA dimerization (Figure 1.12B right). Indeed, experiments confirmed that the Malachite green fluorescence reduced by approximately a third upon GFP binding (Figure 1.12C). This elegant concept may be further optimized. For example, the incomplete loss of fluorescence was likely caused by the limited rigidity of the helix connecting the targeting module to the MG binding module. In addition, the strength of the FADA dimer could be attenuated to reduce background binding. The targeting DARPin’s loops can be diversified to create a sensor library, which can be screened by e.g. yeast display for high fluorescence in the absence of target and subsequently for low fluorescence in its presence. This sensor platform is quite generic, although the target must be sufficiently large for it to disrupt dimer formation. Another major disadvantage is that the fluorescence decreases rather than going up upon target binding.

FRET-bodies

The most direct way to make a generic FRET sensor scaffold would be to graft variable loops onto the fluorescent proteins themselves, creating so-called ‘fluorobodies’. In order to translate target binding to a fluorobody into a detectable response, we envisioned a ‘FRET-body’, essentially consisting of a donor and an acceptor FP, each carrying a target-binding loop, connected by a long semi-flexible linker. If the target molecule would bind to the loops of both FPs, this would decrease the distance between them and result in high FRET (Figure 1.13A). From a FRET-body library, sensors could be selected for entirely new targets using cell display methods, such as yeast display (88), bacterial display (89, 90) or mammalian display (91, 92).

Of these three methods, bacterial display allows screening of the largest libraries, since bacterial transformation efficiency can be up to 1011_{. Yeast display also allows relatively large libraries of} 107 _{members to be screened, although even higher transformation efficiencies of 10}10 _{have been} reported (93). An additional advantage of using a eukaryotic expression host over bacteria is that its secretory machinery ensures proper folding and weeds out any thermodynamically unstable library members (94). Mammalian cell display is useful in particular if human glycosylation is important, but the limited transfection efficiency of mammalian cells caps the throughput of this method at approximately 104_{. For GFP-based biosensors such as FRET bodies, yeast display may} be the most suitable method, combining a high throughput with a eukaryotic secretory pathway. Yeast display has been applied successfully for the directed evolution of antibodies (95), antibody mimics (82), T-cell receptors (96–99) and enzymes (100). GFP and fusions of GFP have also been displayed on yeast (101, 102). The method involves fusion of a FRET-body library to the cell-wall protein Aga2. Cells displaying the library members can be suspended in buffers containing a desired concentration of (fluorescently labeled) target molecules and sorted using FACS, where the FRET ratio can be used as a selection criterion. In order to select FRET bodies on their

24

response magnitude, libraries should first be sorted in the ligand-free state, sorting members for low FRET. In a second selection, the sorted cells should be incubated with the target and then sorted for high FRET.

The beta-barrel topology of Aequorea victoria GFP derived FPs is made up of 11 anti-parallel strands, connected by loops of varying lengths (Figure 1.13B). Extensive research over the past two decades has shown that some of these loops can be extended. The most tolerant insertion positions are between Gln 157 and Lys 158 (loop 8), between Glu 172 and Asp 173 (loop 9), situated at opposite sides of the beta barrel, although insertion between Asp 102 and Asp 103 (loop 5) and between Leu 194 and Leu 195 (loop 10) have also been reported (103–118).

However, one of the main obstacles of this approach is that insertion of loops into GFP derivatives in many cases results in a decreased fluorescence or expression level. The length and composition of the insert can also affect the fluorescence and in some cases, target binding was found to affect the protonation state of the chromophore. This is unwanted, because it makes screening FRET body libraries especially prone to selection artefacts. However, in order to reduce this effect, the GFP framework can be engineered to be more tolerant toward insertions.

Figure 1.13 FRET-bodies as a generic sensor platform. A) Schematic drawing of the proposed FRET-body mechanism. A FRET donor and acceptor FP, each harboring a CDR-like loop, are brought in close proximity by the target binding to both loops, resulting in high FRET.B) Topology diagram of GFP-derived FPs showing the position of loops 5, 8, 9 and 10 with respect to the beta barrel backbone. In 3D beta strands 1 and 6 also border each other to complete the barrel structure.

Engineering sensor proteins using directed evolution

25

Figure 1.14 Evolution of the GFP framework to tolerate double-loop insertions. A) Side and top views of GFP showing Asp 102-Asp 103 (loop 5) in red and Glu 172-Asp 173 (loop 9) in blue. The chromophore is shown in space-fill green. B) FACS dot plots of yeast displayed GFP (GFPM), GFP carrying CDR-L3 between Asp 102 and Asp 103 (GFPM-ASL3), and dual loop-inserted GFP (GFPM-H3L3). PE fluorescence (y axis) is indicative of full-length expression using the c-terminal c-myc epitope for detection and GFP fluorescence (x axis) is indicative of properly folded GFPM or its variants. The negative population lacking both PE and GFP fluorescence located in the lower left quadrant of dot plots is a nondisplaying population characteristic of yeast display. C) Flow cytometric measurements of GFP fluorescence (external GFP), full-length expression (c-myc), and fluorescence per molecule (GFP/c-myc) for three transformants of each evolved scaffold induced at 20 °C and normalized to non loop-inserted GFPM. Clones are named based on the selection they were recovered from (temperature – evolution round – clone number). ND = not detectable. Secretion yields using a baseline expression system are denoted beneath each scaffold in (mg/L). Figure from reference (113).

Pavoor and coworkers (113) performed an extensive directed evolution experiment on enhanced GFP in order to make it a better scaffold for loop insertions and to enable double loop insertions without prohibitive loss of fluorescence. Two loops, if positioned close together, may co-operatively engage their target and enable high-affinity fluorobodies to be isolated. They initially chose to insert two CDRs (CDR-H3 and CDR-L3) of a model antibody against lysozyme between Asp 102 and Asp 103 (loop 5) and between Glu 172 and Asp 173 (loop 9) respectively, because these positions are very close together in the three-dimensional structure of GFP (Figure 1.14A). Then both single insertions and the double insertion were displayed on yeast and found that EGFP carrying an CDR-L3 in loop 9 was relatively well displayed and fluorescent. In contrast,

26

inserting CDR-H3 into loop 5 resulted in a poorly displayed and non-fluorescent mutant (Figure 1.14B). The double insertion was hardly displayed at all and completely non-fluorescent (Figure 1.14B). The authors then attempted to engineer the beta barrel framework in such a way that it would better accommodate the loops. Therefore, random mutagenesis was applied to the dual loop-inserted GFP, restricting the mutations to framework residues and leaving the loops untouched. This library was then screened by yeast display for both enhanced display levels and for increased fluorescence. The beneficial mutations were then shuffled to create a second library, from which further improved variants were isolated. Two more rounds of mutagenesis and sorting were applied and in order to further enhance framework stability, the induction temperature was increased in the later rounds. Throughout the course of this evolution experiment the display level (as assessed by fluorescently staining a c-myc tag fused C-terminal to the fluorobody) and the GFP fluorescence gradually increased (Figure 1.14C). Finally, they obtained a mutant, 37-5-8, with 9 mutations that exhibited a fluorescence per molecule of approximately 60% that of non-inserted GFP and was well displayed. The CDR loops were then exchanged with fully randomized loops and fluorescent binders against a variety of targets (biotin, streptavidin, the extracellular domain of the neurotrophin receptor TrkB and glyceraldehyde-3-phosphate dehydrogenase, GAPDH) were selected. Without any further affinity maturation, low- to high nanomolar binders could be identified, which had fluorescence per molecule ranging from 25–160% of the parent scaffold.

Engineering sensor proteins using directed evolution

27

Aim and scope of this thesis

In the research presented in this thesis we have attempted to combine directed evolution approaches with rational protein design in the construction of fluorescent and luminescent sensor proteins with the ultimate aim to develop the FRET-body platform described in this chapter.

In chapter 2 we have attempted to molecularly evolve a previously reported sensor for antibody detection, AbSense, to respond to new antibodies by using yeast display. As a proof of principle we first displayed the original sensor on the surface of yeast cells, but found the displayed sensor to no longer function correctly. The sensor, which adopts a high FRET state in the absence of its target antibody, was found to have very little FRET when it was displayed on yeast, even though both donor and acceptor fluorophores were present. Thus, we concluded that this direct approach of displaying libraries of full-length sensors on yeast was not feasible. Instead, we then employed a two-step approach, using yeast display to molecularly evolve new antibody binding peptides and subsequently incorporating these into full-length sensors.

In chapter 3, we have used yeast display to improve the affinity of a cyclic ‘meditope’ peptide that binds specifically to the therapeutic antibody cetuximab in a unique pocket within the Fab fragment. We have employed a deep mutational scanning approach, using next-generation sequencing of the library DNA both before and after selection to determine the enrichment of each mutant in the library. We confirmed that our methodology was able to identify even mutations conferring very subtle affinity improvements and found four mutations that together increased the affinity for cetuximab almost tenfold.

In chapter 4 these peptides were then incorporated into LUMinescent AntiBody Sensors (LUMABS) proteins to detect therapeutic antibodies, which may find interesting applications in Therapeutic Drug Monitoring (TDM), individualizing the dose of a drug to take patient-specific pharmacokinetic clearance rates into account. Engineering LUMABS for therapeutic antibodies presented a challenge in that these do not recognize short, linear epitopes, but conformational, and often discontinuous ones. Instead, employing disulfide-bonded cyclic epitopes, mimotopes and meditopes, we engineered sensor proteins for the antibodies trastuzumab, obinutuzumab and cetuximab. The cetuximab sensors, which had the most robust change in BRET and which responded to therapeutically relevant concentrations of the drug, were then characterized in more detail by comparing them to a commercially available ELISA. Rationally incorporating molecularly evolved antibody binding peptides into sensors requires a detailed understanding of the sensor’s thermodynamic behavior. A thermodynamic model of the LUMABS sensor mechanism was able to accurately describe the sensors’ behavior as a function of the monovalent epitope / mimotope / meditope affinity.

28

An important factor influencing the behavior of sensor proteins is the length and stiffness of the linkers that connect its various parts. These linkers control the effective concentrations involved in intramolecular interactions and can therefore have a profound effect on a sensor’s affinity and dynamic range. Often used linkers consist of repeats of glycine and serine in various ratios. These linkers are quite flexible, which makes them suboptimal for crossing large (>10 nm) distances. In chapter 5, we have assessed whether the stiffness of linkers can be increased by reducing their glycine content. Using FRET between ECFP-linker-EYFP fusions to measure the average end-to-end distance of linkers of varying lengths and glycine content, we found that – compared to Gly-Gly-Ser repeat linkers, polyserine linkers are indeed significantly stiffer, but that the effect was substantially less pronounced than expected based on theoretical predictions.

In summary we have shown that combining directed evolution of specific modules within a sensor with rational assembly, assisted by thermodynamic modelling is a viable approach in the engineering of sensors for new targets. In chapter 6 we discuss several interesting directions in which to continue this research, such as improvements of the dynamic range of LUMABS, new interesting targets for developing additional sensor proteins and improved deep mutational scanning strategies that may be applied.

References

1. Bolbat, A., and Schultz, C. (2017) Recent developments of genetically encoded optical sensors for cell biology. Biol. Cell. 109, 1–23

2. Germond, A., Fujita, H., Ichimura, T., and Watanabe, T. M. (2016) Design and development of genetically encoded fluorescent sensors to monitor intracellular chemical and physical parameters.

Biophys. Rev. 8, 121–138

3. Sugiura, K., Nagai, T., Nakano, M., Ichinose, H., Nakabayashi, T., Ohta, N., and Hisabori, T. (2015) Redox sensor proteins for highly sensitive direct imaging of intracellular redox state. Biochem.

Biophys. Res. Commun. 457, 242–248

4. Miyawaki, A., Llopis, J., Heim, R., Michael McCaffery, J., Adams, J. A., Ikura, M., and Tsien, R. Y. (1997) Fluorescent indicators for Ca2+ _{based on green fluorescent proteins and calmodulin. Nature. 388,}

882–887

5. Vinkenborg, J. L., Nicolson, T. J., Bellomo, E. A., Koay, M. S., Rutter, G. A., and Merkx, M. (2009) Genetically encoded FRET sensors to monitor intracellular Zn2+ _{homeostasis. Nature Methods. 6,}

737–740

6. Carter, K. P., Young, A. M., and Palmer, A. E. (2014) Fluorescent sensors for measuring metal ions in living systems. Chem. Rev. 114, 4564–4601

7. Gibhardt, C. S., Zimmermann, K. M., Zhang, X., Belousov, V. V., and Bogeski, I. (2016) Imaging calcium and redox signals using genetically encoded fluorescent indicators. Cell Calcium. 60, 55–64

8. Everett, K. L., and Cooper, D. M. F. (2012) cAMP measurements with FRET-based sensors in excitable cells. Biochem. Soc. Trans. 40, 179–183

9. Fabritius, A., and Griesbeck, O. (2015) Design and directed evolution of genetically encoded cGMP sensors. BMC Pharmacol. Toxicol. 16, A48

10. Cameron, W. D., Bui, C. V., Hutchinson, A., Loppnau, P., Gräslund, S., and Rocheleau, J. V. (2016) Apollo-NADP+: A spectrally tunable family of genetically encoded sensors for NADP+. Nature