Next Generation Techniques for Determination of Protein-Protein Interactions: Beyond the Crystal Structure

Abstract

Purpose of Review:

We discuss recent advancements in structural biology methods for investigating sites of protein-protein interactions. We will inform readers outside the field of structural biology about techniques beyond crystallography, and how these different technologies can be utilized for drug development.

Recent Findings:

Advancements in cryo-electron microscopy (cryoEM) and micro-electron diffraction (microED) may change how we view atomic resolution structural biology, such that well-ordered macrocrystals of protein complexes are not required for interface identification. However, some drug discovery applications, such as lead peptide compound generation, may not require atomic resolution; mass spectrometry techniques can provide an expedited path to generation of lead compounds. New crosslinking compounds, more user-friendly data analysis, and novel protocols such as protein painting can advance drug discovery programs, even in the absence of atomic resolution structural data. Finally, artificial intelligence and machine learning methods, while never truly replacing experimental methods, may provide rational ways to stratify potential druggable regions identified with mass spectrometry into higher and lower priority candidates.

Summary:

Electron diffraction of nanocrystals combines the benefits of both x-ray diffraction and cryoEM, and may prove to be the next generation of atomic resolution protein-protein interface identification. However, in situations such as peptide drug discovery, mass spectrometry techniques supported by advancements in computational methods will likely prove sufficient to support drug discovery efforts. In addition, these methods can be significantly faster than any crystallographic or cryoEM methods for identification of interacting regions.

Introduction

Protein-protein interactions (PPIs) are essential macromolecular interactions in cancer biology. From dysregulated protein signaling networks to aberrant gene regulation, altered protein-protein interactions redirect the flow of information responsible for regulating cell growth. Perturbations from the healthy state, either through mutation or changes in protein expression, can have tumorigenic consequences. Despite the known role of PPIs in cancer biology, there have been relatively few successful drugs targeting PPIs that have survived the trip to the clinic [1]. Small molecule examples include BCL-2 inhibitors, such as venetoclax, which inhibits the pro-survival activity of BCL-2 by mimicking the alpha-helical BCL-2 binding BH3 domain on natural BCL-2 protein ligands [2]. Antibody examples include immuno-oncology drugs targeting the PD-1/PD-L1 protein interface such as anti-PD-1 antibodies pembrolizumab and nivolumab, or anti-PD-L1 antibodies atezolizumab, durvalumab, and avelumab [3]. Other molecules in clinical trials include bromodomain/acetylated histone inhibitors [4,5], MDM2/p53 inhibitors [6], or SMAC/IAP inhibitors [7].

Despite the potential wealth of PPIs as anti-cancer drug targets, the general lack of such agents in the clinic demonstrates significant challenges remain in translating our knowledge of aberrant protein interactions into compounds that can safely disrupt them, particularly intercellularly. A number of reviews cover some of these challenges, primarily focusing on the difficulties designing small molecules that can successfully bind protein-protein interfaces due to the lack of small, well-defined catalytic sites [8–11]. Additional excellent reviews cover the exciting field of peptide and peptidomimetic therapeutic optimization, with a focus on increasing the cellular penetration of peptide therapeutics, decreasing their conformational flexibility, and reducing their degradation by proteolytic enzymes [12–18]. However, before drug optimization can begin, one must identify protein-protein interaction partners and the specific site of interaction to be targeted. Methods for identification of interacting proteins, such as affinity purification coupled to mass spectrometry [19,20] and microarray techniques [21,22], have been thoroughly covered in a number of reviews, and will not be the focus of this review. Here, we will address the structural biology challenges in identifying protein-protein interfaces and how to choose the most appropriate method for interface identification. While the workhorse technique for this purpose has been protein x-ray crystallography, there are a number of exciting advancements in other techniques that we believe will increase their adoption into structural biology workflows. While x-ray crystallography is poised to remain the most widely used and trusted method for determining sites of PPIs, we believe the next generation of macromolecular interaction tools will rely more heavily on alternative techniques, such as mass spectrometry and computational methods, to increase the speed at which sites of interaction can be identified for drug development. While there are many possible tools for identifying protein-protein interactions, we will review several of the most common methods and include a short discussion of the most impactful recent advancements in these fields, with an overview given in Figure 1. This next generation of protein-protein interface identification tools, especially when used in tandem, will facilitate faster development of therapeutic compounds targeting protein-protein interactions and hopefully expedite their translation to the clinic.

Figure 1.

Techniques utilized to determine the sites of protein-protein interactions.

Atomic Resolution Methods

Methods used to solve protein-protein structures to atomic resolution or near atomic resolution are shown in Figure 2 and described below.

Figure 2.

Techniques utilized to identify protein-protein interactions at atomic resolution.

Crystallography

Protein crystallography is by far the most common method of determining protein-protein interfaces. Once protein complexes are formed, crystals are grown when the protein solution is driven towards a supersaturated state as concentrations of solutes (such as salts) increase, and volume of the sample decreases (slow dehydration) [23]. This is most frequently accomplished using methods such as vapor diffusion, in which a protein sample is mixed with a mother liquor (often an aqueous solution of various salts or water-soluble organics) and added to a sealed dish, where is it suspended above a large volume of concentrated mother liquor. Equilibrium diffusion of water allows the drop to slowly increase in solute concentration and protein concentration. If an appropriate mother liquor is chosen, and suitable protein concentration, temperature, and vessel type are utilized, protein crystals may form. Crystallography is completely empirical; there are no predictive methods to determine a priori the ideal conditions to support the growth of a crystal of a given protein complex. As discussed in great detail in [23] and [24], the most common challenge to determining a protein structure is first obtaining a suitable crystal. There are many different methods of promoting supersaturation beyond vapor diffusion that can be investigated, and many parameters that can be rigorously optimized, including concentrations of salts, additional stabilizing additives, modification of pH, increased/decreased protein concentration, microseeding, or temperature. Once suitable crystals are obtained, they are cryocooled and subjected to x-ray radiation. The diffraction pattern (Bragg diffraction) of the x-rays collected while irradiating the crystal can then be used to calculate the electron density of the sample in three dimensions, allowing for the determination of the average position of the atoms that comprise the sample. As protein crystals are on average 50% solvent, they diffract x-rays only weakly; for this reason, crystal size is a vital concern to ensure that significant enough sample is irradiated to reach atomic level diffractions limits [23]. The key bottleneck in a crystallography workflow is obtaining suitable diffraction quality crystals, which can only be addressed through rigorous empirical optimization. Therefore, it quickly becomes apparent that crystallography can be a time-consuming process. This is assuming appropriate crystals can be obtained of the sample at all. Conformationally flexible proteins or membrane-bound proteins with poor solubility in aqueous crystallography solvents prove difficult to crystallize, although research continues in these areas [25,26]. Despite the incredible utility of x-ray crystallography, and the more than 30,000 multimeric protein structures deposited in PDB with resolution better than 2.5 Å [27], there remains a need for additional methods to determine PPI sites.

Cryo-Electron Microscopy

Cryogenic electron microscopy (cryoEM) is a technique that is gaining rapid attention in the field of structural biology, with the developers of cryoEM sharing the 2017 Nobel Prize in Chemistry [28]. The National Cancer Institute at NIH has also founded its own National Cryo-Electron Microscopy Facility dedicated to providing cryoEM access to cancer researchers at NIH and in academia [29]. Unlike crystallography, cryoEM allows for the visualization of single particle protein complexes on a standard transmission electron microscope (TEM) grid. TEM grids loaded with aqueous protein samples are flash frozen in a cryogen, like liquid ethane, such that the protein’s aqueous buffer freezes in an amorphous, non-crystalline state. Grids frozen in such a way can be examined without any additional staining with toxic heavy metals, such as osmium tetroxide, which increase contrast but reduce resolution [30,31]. Moreover, freezing protein samples allows them to tolerate higher electron beam doses before suffering beam damage and protects them from the vacuum of the electron microscope chamber. Even so, electron beam doses on the sample must be kept low to avoid unacceptable beam damage, meaning that single particle images from cryoEM have very poor signal to noise ratios [32]. In order to determine a structure to suitable resolution, thousands of particle images must be averaged from multiple orientations to generate a three-dimensional representation. Some proteins suffer from freezing in “preferred orientations”, limiting 3D model development [32]. Achieving atomic resolution structures depends heavily on the quality of the electron microscope and detectors used. The development of direct electron detectors has allowed for significant improvements in resolution in single particle cryoEM [32]. Most atomic resolution structures have been solved using a Titan Krios 300 kV TEM, the most expensive electron microscope currently on the market [33], putting the technique out of range for those who cannot gain access to such an instrument. While some atomic resolution structures have been solved using more common and affordable 200 kV microscopes, it remains to be seen whether advancements in technique and data processing allow this to become routine [33]. Size remains a significant concern in solving molecular structures via cryoEM. It can be difficult to solve structures of proteins under 100 kDa given that signal to noise ratios drop as particle size drops, although work continues in this area [33]. Regarding protein-protein interactions, cryoEM is generally easier to use with larger complexes. For example, a 3.5 Å structure was solved for the interaction between membrane-bound receptor Patched-1 (PTCH-1) and ligand protein Sonic Hedgehog (SHH), clearing up earlier contradictory reports on the binding site of SHH to PTCH-1 [34]. Hedgehog signaling is specifically involved in development, and excessive signaling has been implicated in certain cancers. The cryoEM structure of PTCH-1 and SHH demonstrate that SHH can bind two PTCH-1 receptors simultaneously, with SHH contacting a different site on each PTCH-1 [34]. Cancer drug development targeting this complex can therefore focus on targeting one or both implicated SHH-PTCH-1 PPIs, one of which is mediated by calcium binding, and the other mediated by palmitate binding [34]. As demonstrated by the example above, cryoEM does not depend on crystal formation and is capable of imaging non-homogenous samples, dynamic samples, or membrane proteins (such as PTCH-1), and is therefore an exciting frontier for solving protein-protein structures [32].

Micro-Electron Diffraction

One additional technique beginning to gain significant attention in the atomic resolution space is micro-electron diffraction (microED). This technique combines some of the advantages of both x-ray crystallography and cryoEM. In microED, accelerating electrons interact with sub-micron-sized protein nanocrystals to generate a diffraction pattern, whereas x-ray diffraction requires protein crystals hundreds of microns in size. The pattern is collected using a CMOS detector mounted in the electron microscope chamber [35]. The wavelength of electrons accelerated at 200 kV in an electron microscope is significantly smaller than those from a conventional x-ray source such as a copper anode (0.0251 Å versus 1.54 Å respectively), allowing for smaller protein crystals to generate suitable diffraction data [36]. This technique thus circumvents one of the most time-consuming parts of x-ray crystallography: optimizing a microcrystal hit until a sizable crystal for x-ray diffraction can be obtained. It also circumvents one rather significant problem in cryoEM, which is access to a state-of-the-art 300 kV TEM. Furthermore, while the need for protein crystals, even microcrystals, automatically eliminates certain proteins from examination, the fact remains that crystallography has established measures of model quality and straightforward methods to refine structures that are currently lacking in cryoEM [37]. The dependence of microED on protein crystals therefore lends more confidence to models generated, even if it imposes constraints on the proteins examined. One important asset of microED is that the charged state of the proteins under investigation can be elucidated because diffraction patterns are generated with charged particles (electrons) rather than x-rays This provides additional exciting opportunities for structure determination where residue charge or bound ions play a significant mechanistic role. MicroED has proven successful for interrogating protein-protein interactions. For example, the structure of the transforming growth factor β (TGF-β monomer) with its receptor (TGRII, extracellular domain) was solved using crystals unsuitable for standard x-ray diffraction [39]. Elevated TGF-β signaling is correlated to poor patient outcomes in some cancers, and represents an intriguing target for cancer therapies [40].

Mass Spectrometry Methods

Methods used to determine amino acid sequences at the site of protein-protein interactions using mass spectrometry are shown in Figure 3 and described below.

Figure 3.

Techniques utilized to identify protein-protein interactions using mass spectrometry.

Chemical Crosslinking

Chemical crosslinking utilizes 7–30 Å chemical crosslinkers that form intramolecular and intermolecular covalent bonds between reactive groups on amino acids. The output of this method provides data on the spatial proximity of two linked amino acids, allowing topology and relative conformation of protein subunits to be determined with a resolution of several amino acids. Crosslinking was originally used to elucidate individual protein structures [41] before improvements in peptide identification by mass spectrometry and enrichment of crosslinked peptides allowed for investigation of protein complexes [42,43]. An important advancement in chemical crosslinking is the development of new crosslinkers. Recently, heterobifunctional crosslinkers that conjugate different types of reactive groups have rapidly expanded. These include photoreactive crosslinkers that react with almost any functional group and crosslinkers for protein-DNA and protein-RNA complexes [44,45]. Perhaps the most exciting trend involving crosslinking is the advent of computational integrative structural biology platforms, such as ROSETTA [46], XLinkDB [47], HADDOCK [48], and I-TASSER [49], which integrate structural data from multiple techniques to model large protein complexes, traditionally a bottleneck in crosslinking mass spectrometry studies [50]. This integrative approach has been used to elucidate the structure of the eIF3 complex and its interactions with the 40S ribosomal subunit [51]. The eIF3 complex is thought to play a role in oncogene expression, with certain subunits aberrantly expressed in various cancers [52].

Hydrogen-Deuterium Exchange

Hydrogen-deuterium exchange (HDX) is a mass spectrometry-based structural technique that monitors the rate of exchange of labile backbone amide hydrogens for deuterium [53]. Because some backbone amides are buried and not equally accessible to excess deuterium oxide (D2O) in the solvent, the rate at which hydrogens exchange depends on their location. Determining deuteration percentage of protein fragments as a function of time can differentiate fragments that are solvent-accessible from those that are not. Back-exchange of hydrogen for deuterium can be a significant issue in downstream manipulations if post-deuteration steps are not performed as quickly as possible. Back-exchange can be severely limited by dropping the pH to 2.5, requiring the use of acid-insensitive proteases such as pepsin for digestion before mass spectrometry analysis. Because pepsin is a nonspecific protease, determining the identity of individual protein spectra while accounting for changes in mass due to deuteration can be somewhat challenging. Proteins are often required to be very highly purified, or “crystallography grade” [53], to help prevent misidentification of spectra. A “top-down” approach for HDX that replaces the digestion step with gas-phase fragmentation in the mass spectrometry workflow can help reduce some of these issues. While data analysis and back-exchange problems can render HDX a challenging technique, the ability to uniquely probe every amide bond for solvent accessibility is certainly attractive. HDX has become useful to drug discovery by identifying protein-protein interactions, most-commonly antibody epitope mapping [54]. For example, the PD-L1 epitope bound by an anti-PD-L1 antibody was determined using HDX before any co-crystal structures of PD-L1 with anti-PD-L1 mAbs were available [55]. Going forward, the utility of top-down approaches (or so called “middle-down” approaches with incomplete pepsin digestion followed by gas-phase fragmentation) and the continued introduction of more user-friendly software, such as the recently published Deuteros [56], will likely increase the use of HDX in PPI site identification [53].

Protein Painting

We have previously reported our own new technology, protein painting, for identification of PPI interfaces [57,58]. This technique uses small molecular dyes that bind non-covalently to solvent-accessible surfaces of pre-formed protein complexes with a resolution of three amino acids. The dyes inhibit the ability of trypsin to access the protein backbone, thus preventing digestion of dye-covered, solvent-accessible regions. Therefore, these regions are not detected by mass spectrometry. Protein-protein interfaces inaccessible to the dye are digested by trypsin and detected by mass spectrometry. Peptide fragments that are absent in samples with the protein partners painted separately, but detected in samples with the painted complex, represent regions of interaction. This technique differs from other mass spectrometry methods in that no covalent modifications are required, allowing for simpler data analysis. This method can be performed in one day without the need for any specialty equipment or expertise beyond basic mass spectrometry. Using this technique, we identified subtle differences in the recruitment of the interleukin-1 receptor accessory protein (ILRAcP) by interleukin-1 (IL-1) versus interleukin-33 (IL-33) to engage their respective receptors IL1-R1 and ST2 [59]. This finding is particularly significant because the crystallographic interfaces for these two complexes are very similar. Both complexes play a role in inflammation and cancer, and the identification of slight differences between them enables the design of drugs specifically targeting only one interaction. For example, IL-33, secreted by tumor cells, has been reported to be crucial for the recruitment of myeloid-derived suppressor cells to the tumor microenvironment ,where they have immunosuppressive properties [60, 61].

Modeling and Computational Methods

Computational methods used to identify sites of protein-protein interactions, from interaction site prediction through hotspot prediction, are listed in Table 1 and described below.

Table 1.

Servers used to either identify protein-protein interfaces or rank hotspots within interfaces.

Server	Use	Algorithm
HDOCK	Interface prediction	Template-based modeling + free docking
LightDock	Interface prediction	AI (Swarm Intelligence Algorithm)
Robetta	Hotspot prediction	Computational alanine scanning
FoldX	Hotspot prediction	Computational alanine scanning

Protein-Protein Interaction Prediction

Predicting protein interaction partners is the first step in identifying druggable protein-protein interaction interfaces. Computational methods greatly expedite this process and can help narrow down the number of binding partners that must be tested experimentally. To predict which proteins interact, machine-based learning and statistical methods can be used. The typical machine-based technique used for this purpose is supervised learning. Supervised machine learning requires establishment of a prediction model from positive and negative training data sets. This prediction model is then applied to a test data set of new proteins to predict protein interaction partners [62]. Alternatively, statistical methods can be used to predict pairwise protein-protein interactions. These methods can be genome-based, such as gene fusion and gene cluster, or gene coexpression-based. They can also be based on additional factors such as protein sequence conservation, coevolution, or interolog search methods [63]. Advancements in network topology methods, which are based on the graphical representation of proteins (nodes) and their associations with other proteins (edges), do not require input of sequence or structure. If the network topology of two proteins is similar, they likely play a role in similar biological processes and interact with each other. Recently, this method has been combined with a heuristic-based algorithm, a form of unsupervised machine learning, to perform graph clustering of networks to better predict protein complexes [64]. The network topology-based method was also combined with gene expression data and miRNA regulatory networks to identify differentially expressed genes in pancreatic cancer. Using this methodology, five prognostic markers related to the Jak-STAT signaling pathway were identified and validated by microarray data [65].

Protein-Protein Interface Prediction

Computational prediction of protein-protein interfaces is particularly useful for drug design given that many druggable interfaces involve transient interactions that are difficult to investigate with classical structural biology techniques. These methods can be categorized into intrinsic-based and template-based approaches [66]. Intrinsic-based approaches rely on sequence or structural data for interface prediction. Sequence-based approaches only consider features of the primary sequence, such as physiochemical properties and hydrophobicity distribution. Structure-based approaches are more accurate than sequence-based methods, but require knowledge of protein structure [67]. Recent advances include the use of deep-learning algorithms, such as a stacked autoencoder, to study sequenced-based interface predictions without the input of structural data [68]. Template-based approaches predict protein-protein interfaces with the input of interface templates from homologous protein complexes. If no templates from homologous proteins exist, structural neighbors can be used instead that have similar secondary structure to the protein of interest. Recent advances include specialized template-based approaches with loop modeling and refinement of the interface [69]. Docking methods can also be used to determine interaction interfaces. Newer docking servers include HDOCK [70] and LightDock [71]. HDOCK employs a hybrid docking algorithm of template-based modeling, both sequence and structure-based, and free docking. LightDock is based on a Swarm Intelligence algorithm, Glowworm Swarm Optimization, and is useful for analyzing flexible protein complexes.

After identification of interaction interfaces, it is important for drug design to identify the specific residues that are crucial for binding. Computational alanine scanning is an energy-based tool that predicts important residues in PPI hotspots by calculating contributions of different residues towards binding energy. Beginning with an input of the complex structure, each residue is mutated to alanine using a side chain repacking algorithm. A numeric energy function is then used to compare the bound and unbound states of the mutated protein versus original protein. Lastly, the change in binding energy (ΔΔG) of each mutation is calculated thermodynamically. Residues are deemed important if binding energy decreases by least 2 kcal/mol after mutation to alanine [72]. Two examples of computational alanine scanning servers are Robetta and FoldX [73,74]. Current advances include the use of molecular dynamics programs to perform alanine scanning. While computational alanine scanning programs require an atomic resolution model, the technique can be used to prioritize regions within a crystallographic interface for drug development.

Conclusions

Perspectives for Next-Generation Protein-Protein Interaction Site Identification

Atomic resolution will continue to be the gold standard for identifying protein-protein interfaces, as information at this resolution can directly aid in small molecule drug design. X-ray crystallography is the most advanced and standard of these techniques, and will likely continue to be the workhorse technique utilized. However, the revolution within electron microscopy for protein structure determination is especially exciting, particularly as it relates to structure determination of membrane-bound proteins. Despite the hype, we believe that the problems facing cryoEM, including standardization of model refinement and lack of well-defined quality metrics as compared to crystallography, necessitate caution in announcing the dawn of the cryoEM age [37]. We look forward to additional work in the field to overcome these noted issues. Additionally, limited access to top-of-the-line 300 kV electron microscopes, necessary for the highest resolution data, restricts the field from becoming as widespread as we would hope. The institution of cryoEM “centers of excellence” may be a way forward to allow access to these instruments for new scientists entering the field [75–77]. While cryoEM has been the subject of much excitement and many reviews, we feel that microED could potentially make a greater and more rapid impact on PPI identification as a 200 kV microscope is sufficient to generate excellent diffraction data. MicroED also requires significantly smaller crystals than conventional x-ray crystallography, and gathers additional information not available in conventional crystallography, such as the charge state of amino acid side chains. These features suggest microED could rapidly decrease the time it takes to acquire adequate crystals, and increase the functional information gathered from co-crystals of interacting proteins.

While atomic resolution data is most desirable, in many cases drug development efforts can begin even in its absence, particularly if the designed drugs are interfering peptides. In this case, simply knowing the sequences found at the interface can allow for preliminary interfering peptide generation. Given that mass spectrometry-based techniques may be more accessible to new users given the ubiquity of mass spectrometers in research institutions over TEMs, we believe mass spectrometry techniques, particularly when coupled to each other, provide the fastest route from target identification to peptide drug lead. The case studies performed in our lab have demonstrated successful generation of interfering peptides based on mass spectrometry data of protein-protein interfaces [57,58], and we have been impressed at the speed and relative simplicity of protein painting compared to a conventional crystallography/small molecule generation strategy. As newer tools come online to integrate mass spectrometry data from crosslinking or HDX experiments with predictive model building, we anticipate that such techniques will become even more widely adopted.

In addition to experimental techniques, computational techniques should not be overlooked. The biggest limitation in AI or machine learning algorithms is the size of the datasets used to “train” the algorithms. As our datasets on protein-protein interactions grow, so will the sophistication of our predictive algorithms. Given the rate at which the AI/machine learning field is changing and expanding due to large interest and funding outside of scientific companies and scientific endeavors, it seems likely that the technology will continue to improve at a dramatic rate [78]. It would be foolish to discount computational techniques, even though drug discovery does have a history of failed promises in this area, such as computer-aided drug design [79,80]. Rather than a sole dependence on computational methods to identify interaction sites, the true power of these methods will be in their combination with experimental data, such as mass spectrometry data. We see the power in these methods as a way to rank multiple experimentally-derived interface regions for druggability or energy contribution, allowing candidate regions to be labelled as low priority or high priority for drug design efforts based on similarity to other druggable protein-protein interface regions.

PPI sites remain an untapped domain of potential drug targets for diseases such as cancer. The next generation of PPI site identification requires an interdisciplinary approach: coupling multiple mass spectrometry techniques to computational models to quickly identify interacting domains, and subsequently solving atomic resolution structures faster using microED or cryoEM. Crystal structures no longer have to represent the primary way to advance a structure-based drug development program. Proteomics groups, computational chemistry groups, and traditional structural biology groups working together can shorten the time to drug target identification from a known protein-protein interaction and increase the probability that new protein-protein interaction targeting drugs can reach the clinic.