Computational design and experimental optimization of protein binders with prospects for biomedical applications

Abstract

Protein-based binders have become increasingly more attractive candidates for drug and imaging agent development. Such binders could be evolved from a number of different scaffolds, including antibodies, natural protein effectors and unrelated small protein domains of different geometries. While both computational and experimental approaches could be utilized for protein binder engineering, in this review we focus on various computational approaches for protein binder design and demonstrate how experimental selection could be applied to subsequently optimize computationally-designed molecules. Recent studies report a number of designed protein binders with pM affinities and high specificities for their targets. These binders usually characterized with high stability, solubility, and low production cost. Such attractive molecules are bound to become more common in various biotechnological and biomedical applications in the near future.

Introduction

Protein–protein interactions (PPI) are crucial to virtually all cellular processes including signal transduction, cell growth and division, transcription/translation, multicomponent protein assemblies and others. Mutations in interacting proteins could change binding affinity between the two partners sometimes disrupting and sometimes strengthening a particular PPI. Such changes affect functionality of the whole PPI network, frequently resulting in disease such as cancer. PPIs are also important in viral and bacterial diseases, where pathogenic proteins bind to a variety of human targets, promoting pathogen propagation and disturbing human PPI networks. In neurogenerative diseases, assembly of the same protein units into long amyloid fibers is the driving force for the disease progression. Thus, design of proteins that bind to one of the targets in the disease-associated PPI and disrupt the complex is an attractive strategy for drug design. Furthermore, proteins engineered to bind to targets overexpressed in a particular disease could be used in diagnostics and imaging. Engineered binders could also serve as tools in protein purification as they could pull down their target protein from a protein mixture. Furthermore, protein agonists could be designed to stimulate receptor activation while binders to specific receptors could promote intracellular protein delivery through receptor-mediated endocytosis. Many important applications of PPI engineering and design brought about the increased scientific interest in the field in recent years.

Three classes of protein scaffolds have been utilized for engineering protein binders (Figure 1). The most traditional scaffolds for binder engineering are antibodies and smaller constructs derived from antibodies (Beckman et al., 2007). Antibodies are used by our immune system to recognize and neutralize foreign antigens by binding to them with their Complementary Determining Regions (CDRs) constituting six often flexible loops in each antibody arm. The initial low-affinity antibodies are able to go through a hypermutation process accumulating various mutations in CDRs, acquiring high binding affinity to their target (Di Noia and Neuberger, 2007). Dozens of antibodies have been already approved as drugs and many more are in clinical trials (Lu et al., 2020). While flexible CDR loops facilitate recognition of multiple proteins, antibodies possess no unique structural and physico-chemical features that make them superior binders. In fact, many small protein domains could be evolved to bind to target proteins with high affinity and specificity (Sha et al., 2017, Skrlec et al., 2015). Among such proteins are natural protein effectors that already associate with their target and could be redesigned for enhanced potency and specificity and novel binding domains that do not bind a desired protein but could acquire binding function through a number of mutations. A variety of novel binding domains of different sizes and geometries have been reported (Vazquez-Lombardi et al., 2015). Novel binding domains possess high structural robustness, high solubility and high expression yield, characteristics that make them easier to produce and characterize. Both natural effectors and novel binding domains have been engineered and employed in biochemical assays, separation technologies, diagnostics and therapeutics (Konning and Kolmar, 2018, Richards, 2018, Skerra, 2007).

Fig. 1.

Three types of binders that target the receptor binding domain (RBD) of the SARS-CoV-2 spike protein (blue). All scaffolds target the same epitope on RBD but vary in size, secondary and tertiary structure (A) Antibody-based (Fab) binder (cyan) (PDB ID 7jx3). Six CDR loops are shown in red. (B) natural interaction partner, ACE2 receptor, optimized for binding to RBD (Glasgow et al., 2020) and (C) a novel binding domain designed and optimized for targeting RBD40. All the three scaffold proteins bind to the same epitope on RBD (Cao L. et al., 2020).

Two complimentary approaches for PPI engineering have been applied to all three classes of scaffolds. In a ‘rational’ computational protein design approach, a structure of the protein–protein complex is used as an input for computational optimization. Starting from the complex structure, the sequence of one or both binders is computationally redesigned according to the energy function that models binding interactions between the two proteins. This approach relies on our ability to reproduce with high precision intermolecular interactions that govern protein–protein binding and to predict mutations that improve affinity/specificity. While fast and inexpensive, computational protein design is limited by the inaccuracies in the energy function and by inability to sample a large number of protein conformations. An alternative ‘irrational’ approach for PPI engineering is based on experimental selection of protein binders from a large combinatorial library of mutants using phage display, cell surface display, ribosome display, or messenger RNA display technologies (Galan et al., 2016). In yeast surface display (YSD) technology (Gai and Wittrup, 2007) for example, the fluorescently-tagged library of the binder protein is expressed on the surface of yeast cells and incubated with the fluorescently-labeled target protein. Subsequently, Fluorescence-activated Cell Sorting (FACS) is used to select cells displaying proteins that show high expression and high binding affinity to the target. After several rounds of selection, high-affinity mutants are sequenced and the corresponding protein is expressed and characterized. The experimental protein engineering approaches proved to work exceptionally well for a variety of scaffolds and targets. Yet, not all protein targets could be displayed on the cell surface, the number of explored binder mutants is limited (to 10¹⁰ for cell surface and phage displays and 10¹⁵ for ribosome display) and there is no guarantee that the selected binder protein interacts with a particular binding epitope on the target. Recent studies demonstrated that combining computational and experimental approaches for PPI engineering is particularly rewarding as limitations of one approach could be overcome by application of the second one (Rosenfeld et al., 2016). In general, computational design should be used prior to experimental selection to decrease the combinatorial complexity of protein libraries explored by experiment (Guntas et al., 2010) and to direct the binder to the desired binding epitope, which is crucial to disease inhibition (Whitehead et al., 2012). Figure 2 summarizes the most common steps in computational binder design and optimization. In this review, we aim to describe recent advances in computational methodology for PPI engineering and experimental optimization, providing examples of the most interesting applications of this methodology.

Fig. 2.

Schematics for computational design and optimization of protein binders (1) Scaffold(s) for binder design are selected. Scaffolds could be antibodies, natural effectors, or novel binding domains. (2) scaffold(s) are docked into the desired epitope of the target protein and the residues on the binder are computationally optimized. A number of binder sequences or a focused library of binders are designed (3) The gene constructs corresponding to particular binder designs or a focused library of binders are cloned. High-throughput selection (such as yeast surface display) is used to select binders that show interaction with the target. In an optional step, several cycles of optimization of the binder either through random mutagenesis or through computational design are performed to further improve affinity (backward arrow). (4) Binders selected by a high-throughput assay are expressed in a suitable host and purified. (5) Exact binding affinities are measured for purified binders interacting with their targets. (6) X-ray crystallography or CryoEM are used to solve the structure of the binder in complex with the target. The initial model of the complex shown in (2) is compared to that of the actual determined structure.

Natural effector redesign

Natural protein effectors already bind to their targets and are directed toward the desired epitope, thus constituting attractive scaffolds for design of inhibitors, imaging agents and drug candidates. Moreover, as human proteins, natural effectors are non-immunogenic and non-toxic, which is crucial when developing new drugs. However, natural effectors frequently interact with multiple homologous proteins in the cell, possessing low binding specificity. Binding affinities of natural effectors for their targets are frequently weaker than what they could possibly be since the ability of effectors to dissociate from their targets is often important for effector function (Cohen et al., 2017). Binding affinities and specificities of natural effectors toward one particular target protein could be improved through protein design and/or engineering (Fromer et al., 2010). Affinity and specificity optimization could be performed through computational saturation mutagenesis that scans all effector binding interface positions with twenty amino acids and predicts affinity- and specificity-enhancing mutations (Aizner et al., 2014, Rosenfeld et al., 2015, Sharabi et al., 2014). Computational predictions for single mutations or non-optimized cold-spot positions (Shirian et al., 2016) could be used to design focused libraries of mutants, thereby significantly improving our chances of selecting high-affinity binders via experimental techniques (Guntas, et al., 2010). Using such an approach, we converted a broad metalloproteinase (MMP) family inhibitor TIMP-2 into specific picomolar inhibitor of one family member (MMP-14 or MMP-9) (Arkadash et al., 2017, Shirian et al., 2018). Glasgow et al engineered a high-affinity binder of SARS-CoV2 receptor binding domain (RBD) starting from the human angiotensin-converting enzyme 2 (ACE2) (Glasgow et al., 2020).

Another powerful approach to optimize natural effectors lies in experimental saturation mutagenesis, frequently referred to as deep mutational scanning (Fowler and Fields, 2014). In such an approach, a library of the binder protein is constructed containing all possible single mutations and YSD is used to select high-affinity binders. The selected clones are deep sequenced and the frequency of each mutation in the high-affinity pool relative to its frequency in the unsorted population is used to calculate the enrichment ratio for each mutation (Wrenbeck et al., 2017). Mutations that show high enrichment value with high probability result in affinity enhancement measured on purified proteins. Single affinity-enhancing mutations could be further combined into combinatorial libraries that are used for further affinity maturation. Such an approach was recently utilized to design very potent inhibitors of interaction between ACE2 and RBD of SARS-CoV-2 (Chan et al., 2020). These inhibitors, based on optimized ACE2 dimers inhibited the SARS-CoV2/ACE2 interaction with 0.6 pM K_D.

While deep mutational scanning has been repeatedly used to identify affinity-enhancing mutations, the methodology could be further improved to measure exact values of changes in binding free energy (ΔΔG_bind) for thousands of binder protein mutants in one experiment (Adams et al., 2016, Heyne et al., 2020a, Heyne et al., 2020b, Reich et al., 2016). For this purpose, during the YSD selection, the mutant clones are collected in several gates with various affinity ranges, the clones from each gate are deep sequenced and the enrichment values are calculated for each mutant in each gate. A small set of experimental data points for ΔΔG_bind values are measured on pure proteins and are used to obtain a normalization formula that converts enrichment values into ΔΔG_bind. Our group has used this approach to map binding landscapes for several homologous serine protease/inhibitor complexes and to establish striking differences in such landscapes depending on the PPI evolutionary optimality (Heyne et al., 2020b). A similar approach has been explored by the Keating group to design highly specific alpha helical binders to just one member of the Bcl-2 family of proteins: Bcl-x_L, Mcl-1 and Bfl-1(Jenson et al., 2018).

Design of novel binding domain

Design of novel binding domains presents a challenging problem for computational design since it requires as an input a structure for the novel protein–protein complex. Producing such a structure through conventional docking techniques is not possible since the sequences of the binding partner(s) are not known a priori and are subject to subsequent redesign. To overcome this difficulty, several strategies have been explored and implemented. In the first strategy, the novel binding interface is designed by extending a frequently observed secondary structure element across the binding interface and subsequently redesigning the binding interface residues. Using such a strategy, homodimers were successfully created through an intermolecular beta-sheet, providing initial binding energy through intermolecular backbone-backbone hydrogen bonds (Stranges et al., 2011).

In a second approach, termed hot-spot centered approach (Fleishman et al., 2011a and b), novel PPI is created by first designing a few energetically favorable interactions between the target protein and a cluster of disconnected amino acids. Such a cluster is later grafted onto various highly stable scaffolds, and the residues surrounding the hot-spots on the binder protein are redesigned. Such an approach yielded a successful de novo binder to the conserved region of influenza hemagglutinin (Fleishman et al., 2011b) and several other novel PPIs (Procko et al., 2013, Strauch et al., 2014). In the third approach, termed epitope grafting (Azoitei et al., 2012, Azoitei et al., 2011, Capelli et al., 2017), a continuous structural unit such as a loop or an α-helix known to interact with a specific target is grafted onto an unrelated scaffold protein, preserving the geometry and the intermolecular interactions at the binding interface. Next, the backbone structure of the chimeric protein is optimized, and the residues surrounding the grafted region are computationally optimized to enhance novel binding domain stability and affinity. The fourth approach, termed Rotamer Interaction Field (RIF) docking (Dou et al., 2018), first generates an ensemble of billions of amino acid conformations that confer favorable hydrogen bonds and hydrophobic interactions with the target and records the positions of their Cα atoms. The algorithm then docks various scaffolds into the interacting ensemble using a grid-based hierarchical search algorithm. Epitope grafting and RIF docking approach have been recently used to design novel binding domains to SARS-COVID-2 spike RBD with the aim to disrupt its interaction with the human ACE2 receptor (Cao et al., 2020).

In spite of the development of several approaches for novel binder design described above, the success rate for purely computational design of such binders remains relatively low. Thus, it is crucial to be able to quickly assess tens to thousands of computational designs prior to time-consuming protein expression, purification and affinity measurements. Most frequently, YSD has been employed as a method for initial selection of successful computationally designed novel binding domains. Most recently, an attractive approach has been proposed where large pools of various novel binding domain designs can be encoded in one YSD library and selected by FACS to identify initial binders (Cao, et al., 2020). In one example, the Baker group synthesized oligo pools corresponding to 22660 computational designs of small protein domains to target influenza haemagglutinin and botulinum neurotoxin B and identified ~ 2000 initial binders through YSD (Chevalier et al., 2017). The weak affinity of the initial novel binding domains could be improved either through computational optimization or through experimental selection. Among experimental approaches for affinity enhancement, deep mutational scanning (Fowler and Fields, 2014) of the binder protein is especially attractive since it allows to explore the effect of each single mutation on binding affinity and to quickly identify affinity-enhancing mutations. Combining several such mutations in one design was shown to further increase affinity, resulting in several examples of novel binding domains with low nM to pM affinities. In summary, novel binding domains could be designed starting from multiple scaffolds with different structures, providing a large variety of highly stable and specific binders for development of drugs and imaging agents.

Antibody design

Full-length antibodies and shorter antibody-based constructs are the most common protein scaffolds for drug design. As such, they have been pursued widely in pharmaceutical industry and experimental approaches for their engineering have been well established (Saeed et al., 2017). Yet, only recently, strategies for computational design of antibodies have been explored. Such strategies could be divided into two general classes: those that are based on structural modeling of antibody–antigen complexes (structure-based) and those that use experimental binding data on particular antibodies to predict high-affinity binding sequences (sequence-based).

In structure-based approaches, structural model for the antibody-target complex has to be first created, requiring modeling of interactions between the antibody CDRs and the target protein. Such modeling is challenging due to high flexibility and length variability of the antibody CDR loops. CDRs could be modeled de novo; however, such modeling requires considering a huge number of loop conformations and their interaction with the target protein, sometimes making such approach intractable (Huang et al., 2011). Alternatively, compatible loops from proteins in the PDB could be grafted onto the antibody, producing a new CDR conformation. Loop grafting is relatively fast but is limited by the loop structures already available in the PDB. An early work by the Kuhlman group demonstrated that a loop could be computationally grafted into a single protein with high structural precision (Hu et al., 2007), thus serving proof of principle that such grafting could be applied to antibody design. Fleishman and colleagues developed an antibody design software AbDesign (Lapidoth et al., 2015) that is based on CDR grafting from numerous antibody structures in the PDB. Initially, the authors tried to graft all six CDRs separately from different structures, but computational analysis revealed some drawbacks of the resulting designs such as poor residue packing. Subsequently, the authors followed natural antibody V(D) J recombination strategy, thus grafting four larger segments that combine CDR1 and CDR2 from the same antibody and the two CDR3s from a different antibody. The antibody models were then docked onto the desired target and their sequences were redesigned to reflect natural amino acid variability in CDRs. The AbDesign methodology was applied to design antibodies against human insulin and Mycobacterium tuberculosis acyl-carrier protein-2 (Baran et al., 2017). The authors experimentally screened 130 designs and found variants with mid-nanomolar affinities, which were further improved through random mutagenesis and selection to low nM affinity. Dunbrack and colleagues implemented a similar strategy called RosettaAntibodyDesign, which is based on CDR grafting from a widely accepted set of the canonical clusters of CDRs (North et al., 2011). This approach was verified by improving affinity of two existing antigen–antibody complexes by 10- and 50-fold (Adolf-Bryfogle et al., 2018).

Computational design was also applied to another important problem, that is design of bi-specific antibodies as such antibodies become increasingly attractive drug candidates in pharmaceutical industry. To design bispecific antibodies, two antibody arms with different specificities need to be combined into one antibody. To produce such chimera antibodies, interface between the two C_H3 constant regions of the two different antibodies is redesigned to optimize intermolecular interactions (Leaver-Fay et al., 2016). Using such an approach, Leaver-Fay et al alternated between design and docking of antibody halves, while using the multistate design approach (Leaver-Fay et al., 2011) to simultaneously stabilize antibody heterodimers and destabilize homodimers. In this study, computational design supplemented by a few manually selected mutations resulted in bispecific antibodies that were more than 90% pure.

In another structure-based approach called re-epitoping (Nimrod et al., 2018), machine learning was utilized to learn favorable residue-residue contacts from multiple structures of antibody–antigen complexes and to predict mutations that increase affinity of the scaffold antibody to the desired epitope. The approach was tested by docking 501 antibodies against a modelled antigen, cytokine interleukin-17 (IL-17A), selecting five best antibody scaffolds and designing five combinatorial libraries of antibodies by randomizing cold-spot positions on each antibody. The libraries were then sorted by YSD, and the best antibody was selected with a K_D value of 80 nM after humanization.

The second class of antibody design approaches do not model antibody–antigen structures but utilize the information on antibody sequences that bind and do not bind particular targets. Most recently, deep learning approaches have been introduced that use neural networks to learn from the available sequence/function relationship and to predict antibody mutations that improve binding affinity. To optimize an existing therapeutic antibody, trastuzumab, Mason et al. first performed deep mutational scanning of the antibody when it is bound to its target, extracellular domain of the HER2 receptor (Mason et al., 2019). Experimental results were used to construct a small combinatorial library of 5 x 10⁴ antibody variants that was sorted into two groups of binders and non-binders. Deep sequencing of the two groups resulted in two large sequence datasets that were used to train neural networks to predict high-affinity antibody sequences. With the trained network, the authors scored 7 x 10⁷ antibody sequences and predicted 3 x 10³ binders, 30 of which were tested experimentally and exhibited binding affinities ranging from of 0.1–10 nM. Variants predicted to be binders were then subjected to developability filters such as low immunogenicity and high solubility, resulting in thousands of highly-optimized antibodies. In a different study, Lui et al used sequence data coming from phage display panning experiments to train neural networks to predict sequences of binding antibodies (Liu et al., 2020). Several experimental datasets were utilized for training of six neural networks. Finally, a combination of all these neural networks with a gradient accent algorithm were applied to predict antibody sequences that possess high binding specificity and high affinity for their target. Among 19 experimentally assessed designed antibodies, all exhibited similar to WT or better affinities to their targets, demonstrating great potential of machine learning methods in antibody optimization. While a number of additional deep learning approaches have been proposed for binder design and optimization (Graves et al., 2020), the success of such protocols has yet to be experimentally verified.

Challenges in clinical development of protein binders

So far none of the computationally designed binders have been approved as drugs. Nevertheless, efforts to develop protein therapeutics from such molecules are underway, with several companies working on commercializing their inventions (Cao, et al., 2020, Chevalier, et al., 2017, Dang et al., 2019). To convert the designed binder proteins into drugs, one has to address several potential problems such as their possible immunogenicity and drug delivery. To minimize immunogenicity, scaffolds for binder design could be derived from human proteins (e.g. adnectins, anticalins, avimers, Fynomers and Kunitz domains) or from proteins that have shown low immunogenic potential (e.g. DARPins and knottins) (Simeon and Chen, 2018). Interestingly, several characterized small de novo–designed proteins generate little or no immune response (Chevalier, et al., 2017, Silva et al., 2019). Furthermore, the immunogenic potential of a protein could be computationally minimized by choosing protein sequences that are predicted to possess reduced affinity for the major histocompatibility complex (MHC) (Doneva et al., 2021). Even so, the immunogenicity of the engineered protein needs to be determined as early as possible in the development process, especially for previously untested protein scaffolds. The second issue that needs to be addressed when converting protein binders into drugs is drug delivery. While engineered proteins could easily target extracellular proteins, targeting intracellular proteins with macromolecules becomes a challenge since such macromolecules do not easily cross cellular membranes. While the intracellular delivery problem has not been completely solved, many strategies for protein intracellular delivery have been recently explored including utilization of cell-penetrating peptides, nanoparticles, liposomes, supercharging and bacterial toxins (Lee et al., 2019, Verdurmen et al., 2017). Thus, intracellular delivery should not preclude the efforts for designing protein binders for intracellular targets.

Conclusions and future directions

Design of protein binding interactions is an attractive research direction with many biomedical applications. Recently, a number of new computational methods for binder design has been proposed and tested, resulting in several molecules that demonstrate high therapeutic potential. The advantage of computational methodology for binder design over purely experimental approaches lies in the ability of computational methods to select the most compatible scaffold for a particular target and to direct the binder to the correct target binding epitope. Yet, computational design frequently produces low-affinity initial binders, which need to be further optimized experimentally. New experimental protocols for whole gene randomization (Kowalsky et al., 2015) and deep mutational scanning (Heyne et al., 2020a, Jenson, et al., 2018) greatly facilitate binder optimization. In addition, recent improvements of the YSD platform for higher protein expression, higher fluorescent signal (Zahradník et al., 2020) and faster binder selection (Cohen-Khait and Schreiber, 2018) are bound to further increase capabilities of new designs. New improvements in computational methodology for binder design are still needed and might come from deep learning approaches that rely on quickly growing number of protein complex structures and binding affinity measurements. With the new developments, designed protein binders are bound to become more common in various biomedical applications in the near future.

Funding

This work was supported by the Israel Science Foundation 3486/20; Israel-US Binational Science foundation 2017207; HUJI/University of Toronto Research alliance and National Institutes of Health R01CA258274.

Conflict of interest

The authors declare no conflict of interest.

Edited by: Dr. Arne Skerra