From pathways to networks: connecting dots by establishing protein-protein interaction networks in signaling pathways using affinity purification and mass spectrometry

Abstract

Signal transductions are the basis of biological activities in all living organisms. Studying the signaling pathways, especially under physiological conditions, has become one of the most important facets of modern biological research. During the last decade, mass spectrometry has been used extensively in biological research and is proven to be effective in addressing important biological questions. Here, we review the current progress in the understanding of signaling networks using mass spectrometry approaches. We will focus on studies of protein-protein interactions that use affinity purification followed by mass spectrometry approach. We discuss obstacles to affinity purification, data processing, functional validation, and identification of transient interactions and provide potential solutions for pathway-specific proteomics analysis, which we hope one day will lead to a comprehensive understanding of signaling networks in humans.

Application of mass spectrometry in the studies of signaling pathways

Sophisticated signaling networks are required for development and survival, and a minor disruption of any of these pathways may cause severe diseases such as cancer in humans. A typical signaling pathway, such as transforming growth factor–beta (TGF-β) signaling, usually starts with an extracellular ligand (e.g., TGF-β) binding to a receptor (e.g., TβR) on the cell membrane. This ligand-receptor interaction could trigger a sophisticated phosphorylation cascade, followed by translocation of major signaling molecules (e.g., Smads) to the nucleus and binding to specific DNA sequences to regulate gene expression and thus ultimately modulate cellular activities (reviewed in [1]). Over the years, considerable efforts have been devoted to elucidating signaling transduction pathways in human cells. A number of pathways have been identified, in which protein-protein interactions (PPIs) and post-translational modifications (PTMs) are often involved. As a quantitative and unbiased approach, mass spectrometry (MS) has been proven to be a powerful tool in revealing protein abundance as well as PPIs and PTMs under near-physiological conditions (Fig. 1) (reviewed in [2, 3]). There are many decision points when using MS to study signaling events, which include sample preparation, choice of MS applications, pre-MS strategies, the MS itself, and post-MS data analysis (Fig. 1).

Figure 1. An integrated view of the MS application and strategies in signaling pathway studies.

Step 1: Sample preparation. Basal state cells could be used directly for MS studies, or treated with ligand, drug, or inhibitors to obtain different states in signaling events. Cells may be subjected to fractionation or other approaches to separate certain fractions/organelles, or cross-linked with chemicals, to monitor the proteome/interactome changes of a part of the cells. Step 2: Choose MS applications. MS could be used to monitor the overall protein abundance, post-translational modifications, and protein-protein interactions. Step 3: Pre-MS strategies. Protein complexes could directly subjected to MS (“shotgun” methods) or isotope/isobaric labeling prior to MS. For affinity purifications, antibodies against endogenous proteins could be used to bring down the endogenous protein complexes. One-step or tandem affinity purifications (TAP) could be used to bring down epitope-tagged bait proteins. Step 4: MS and data analysis. Different MS equipment could be used to generate peptide readouts. Processed MS/MS spectra are searched against protein sequence databases thereby assigning MS/MS spectra to peptides, which can be subsequently mapped to proteins. Additional algorithms and software are commonly used in the follow-up data analysis to generate biologically relevant readouts.

Signaling events are by definition associated with changes, for example the change of protein level inside or outside of the signaling cascade, which determine the final outcome. Although many approaches, such as microarray, may be used to detect mRNA level changes, protein levels are frequently regulated at post-transcriptional and post-translational levels and therefore do not always correlate with mRNA levels. Monitoring protein abundance, especially in a complex environment, provides irreplaceable insights into signaling events. Using a “shotgun” strategy, proteins can be directly identified from a complex mixture by enzymatic digestion followed by reverse-phase chromatography and tandem mass spectrometry (MS/MS) [4]. Because of its simplicity, monitoring signal-induced protein level changes, such as those that occur following the binding of an upstream ligand or treatment with an inhibitor, has become one of the most widely employed signaling-related applications of MS. To allow measurement of the absolute or relative protein abundance among different samples, tremendous efforts have been made to ensure that MS results are more reliable and accurate from a quantitative point of view. These efforts began with the traditional semi-quantitative MALDI-TOF and liquid chromatography (LC)-MS/MS (MS2) approach (reviewed in [2, 5]), to recently developed isotope-coded affinity tags (ICAT) [6], stable isotope labeling by amino acids in cell culture (SILAC) [7], isobaric tags for relative and absolute quantification (iTRAQ) [8], tandem mass tags (TMTs) [9, 10] and triple-stage mass spectrometry (MS3) [11]. With these technological advancements, measurements of protein relative and absolute abundance become more robust, precise and accurate. For example, the TMT technology allows as many as 54 samples to be tagged with different combination of isobaric tags and analyzed in a single MS run, thereby providing relative protein abundance in each sample [12–14]. iTRAQ and TMT do not require growing cells in isotope-containing culture medium, therefore they could be used for comparing protein abundance in multiple tissue samples [15]. However, it has been frequently observed that the accuracy and precision of quantitative data from iTRAQ and TMT suffer because the contaminating near-isobaric ions are isolated and fragmented together with the target ions [16, 17]. Efforts have been made to solve this problem, which include the use of MS3 and improving data analysis [11, 16]. Indeed, the technological developments often require accompanying advances in bioinformatics methodologies. A handful of software packages, such as OpenMS [18, 19] and MaxQuant [20, 21], were designed to analyze the label-free or isotope-labeled MS data. Specific software packages, such as virtual expert mass spectrometrist (VEMS), were also designed specifically for iTRAQ or TMT data analysis, to achieve more accurate peptide ratios [22, 23].

These new MS-based technologies provide great advantages in detecting differences in protein abundance in signaling events between different states, such as basal versus activated or inhibited, or healthy versus diseased. Using these technologies, researchers have been able to compare protein abundance in complex environments, such as haploid versus diploid yeast [24] or healthy tissues versus tumor samples [25]. With these quantitative methods in hand, ligand-induced receptor signaling pathways, which can be easily manipulated by external ligand stimulation or treatment with inhibitors, are now being studied. For example, using a γ-secretase inhibitor to block Notch signaling, Dai et al. profiled dose-dependent protein expression in glioblastoma cancer stem cells [26]. Treating Madin-Darby Canine Kidney Epithelial cells with TGF-β ligand, Shen et al. profiled the plasma membrane proteins and found Wnt-5a involved in H-Ras/TGF-mediated epithelial-mesenchymal transition [27]. With a similar strategy, Kim et al. profiled the total protein changes on lipid rafts upon TGF-β1 treatment [28]. Using an iTRAQ-based quantitative screen of Smad4-negative and stably rescued SW480 cells, Ali et al. found a series of protein level differences between the two cell lines, which indicates that they could be potential downstream targets of the TGF-β signaling pathway [29].

MS-based approaches also provide insight into PTMs that regulates protein activity and stability. Proteome-scale studies have been performed to assess the global change of PTMs, such as phosphorylation (reviewed in [30–32]), ubiquitination and Ub-like modifications (reviewed in [33, 34]), acetylation (reviewed in [35]), methylation (reviewed in [36]), and glycosylation (reviewed in [37]). Among many PTMs, phosphorylation is one of the most important PTMs in signaling transduction, and phosphoproteomics is the most studied PTM using MS approaches. Functionally focused large-scale proteomics studies have also been conducted, such as studies of DNA damage–induced phosphorylation by Ataxia telangiectasia mutated (ATM) and Ataxia telangiectasia and Rad3 related (ATR) [38] and of cell cycle regulations [39]. With proper enrichment strategies and quantitative MS approaches, several signaling pathways have now been characterized using MS-based global phosphoproteomic profiling (reviewed in [32]), such as quantitative phosphoproteomics of TGF-β signaling [40], Wnt signaling [41], insulin signaling [42] and proto-oncogene tyrosine-protein kinase Src signaling [43]. Affinity purification (AP)/MS also has been used frequently in enriching and identifying PTMs of important proteins in signaling pathways. Willert et al. purified active Wnt molecules and found that they are lipid modified and act as stem cell growth factors [44]. Morén et al. identified K507 on Smad4 as a ubiquitination site and participates in R-Smad C-terminal phosphoserine recognition [45]. Matsuura et al. found O-GlcNAc modification occurs at multiple sites on Notch epidermal growth factor repeats [46]. Together with the extensive experimental studies, the bioinformatics tools for analyzing protein PTMs have also been rapidly developed. The MS results can be motif analyzed [47], Gene Ontology (GO) and pathway annotated [48, 49], and PPI database incorporated [50] and eventually provide a comprehensive understanding of cellular events under physiological and pathological conditions (reviewed in [51]). These studies provide insights into the regulation of signaling pathways and are valuable resources for basic and clinical research.

Affinity purification followed by mass spectrometry – a powerful tool in identifying protein-protein interactions in signaling pathways

Studies of PPIs have provided immense insights into protein functions. Many biochemical tools have been invented to explore PPIs, including in vivo/in vitro pull-down assays, co-immunoprecipitation (co-IP) assays using epitope tags or antibodies against endogenous proteins, crosslinking protein interaction assays and/or label transfer protein interaction analysis (e.g., biotinylation labeling), and yeast two-hybrid analysis (reviewed in [52]). As a relatively new and unbiased approach, AP/MS offers tremendous advantages over other methods in identifying PPIs under near-physiological conditions and identifying protein complexes instead of binary interactions [2]. By performing AP of a protein of interest (the “bait”), and following it with LC-MS/MS, the partner proteins (the “preys”) that form complexes with the bait can be identified [53]. We have used this approach to study the DNA damage signaling pathways (for examples, see [54–58]; for a review, [59]), since many stable and functionally relevant complexes are formed before and after DNA damage or regulated in a cell cycle–dependent manner. It helped us tremendously, since many insights into the regulation of DNA damage–responsive pathways were initiated from the AP/MS results of individual DNA damage–responsive proteins.

Using AP/MS to study individual proteins in other signaling pathways, such as the TGF-β signaling [60, 61] and Wnt signaling [62] pathways, has also led to many important findings in these fields. With three different MS strategies, Luo et al. profiled Smad2-interacting proteins under different TGF-β stimulation conditions [61]. Several other core components in TGF-β signaling, such as Smad3 [63], Smad4 [60], TβRI [64], and TGFβ1 [65], have also been studied by AP/MS. Using a streptavidin-based tandem AP (TAP)/MS approach, Angers et al. and Major et al. obtained the interactomes of seven core components of the Wnt signaling pathway, including Dishevelled-1/2/3 (DVL1/2/3) [66], β-catenin, AXIN1, adenomatous polyposis coli (APC), and beta-transducin repeat containing E3 ubiquitin protein ligase 1/2 (β-TRCP1/2) [62]. These studies uncovered several novel and important components of Wnt regulators, such as Cullin-3 complex and Wilms tumor gene on X chromosome (WTX) [62, 66]. Using a triple SILAC technology, Hilger et al. also investigated APC, Axin-1, DVL2, and C-terminal binding protein 2 (CtBP2) interactomes under various stimuli and detected several stimulus-dependent interactions such as the binding of APC with Girdin and Axin-1 [67].

Most studies, however, focused only on one or few components in a given pathway. For example, SMAD4 and β-catenin, the respective core components of the TGF-β and Wnt signaling pathways, have been affinity purified by many groups using a variety of methods, while many other important components, such as other Smads in TGF-β pathway, the T-cell factors/lymphoid enhancer-binding factor 1 (TCFs/LEF1), LDL receptor–related protein (LRPs) and Frizzled (FZDs) in Wnt pathways have never or rarely been studied. Moreover, the data generated by different groups using diverse AP/MS strategies and filtration criteria are not comparable. Therefore, it remains challenging to integrate different protein AP/MS results into a unified network for the entire signaling pathway. Systematic investigation using a unified approach is urgently needed to provide a network view of PPIs involved in these signaling pathways.

Using AP/MS in large-scale proteomics studies – methodology, data analysis, and validation

Several large-scale AP/MS-based studies, consisting of hundreds to thousands of protein purifications, have been conducted in yeast [68–70], drosophila [71] and humans [72] and uncovered thousands of PPIs. These studies, as well as yeast two-hybrid screenings performed with baits from many model organisms [73, 74], have been used to build huge interactomes containing millions of PPIs. Since protein-protein interactions imply functional connections between these proteins, these PPI data have become invaluable resources for further investigating functions of individual proteins. Moreover, these large-scale studies provide a broad view of the whole genome complexomes. However, to get a better understanding of certain biological process, more sophisticated and focused study is a must, since it allows in-depth functional validation. Several function-related medium-to-large-scale AP/MS studies in human cells have been conducted in recent years, which focus on a group of proteins with similar functions, such as chromatin remodeling complexes, deubiquitinating enzymes (DUBs), autophagy systems, and histone deacetylase (HDAC) families [75–78]. Some of these studies were coupled with siRNA screenings and functional validations [78]. These focused proteomics studies have uncovered many novel PPIs and protein complexes and provided numerous leads for further functional studies of these cellular processes.

Methodology

There are four major steps in AP/MS: bait presentation, AP, MS, and data analysis [2]. First, to perform the AP, there must be a way to bring down the bait from the cell lysates. Ideally, antibodies targeting endogenous proteins are the best choice because they cause minimal disturbance to the system. Besides their high cost, however, another major drawback of using antibodies against endogenous proteins in large-scale MS is that they often bring down nonspecific cross-reactive binding proteins, which are impossible to eliminate through bioinformatics analysis and are detrimental to the quality of large-scale datasets. Despite these technical difficulties, an exciting attempt using over 3,000 antibodies against endogenous proteins in human cells has led to the discovery of many novel protein complexes [72].

As an alternative approach, epitope tags are frequently used for AP in large-scale proteomics studies [2, 52, 53]. In yeast, efficient homology recombination made it possible to introduce tandem epitope tags to the endogenous gene locus [68, 70]. In mammalian cells, however, this approach is technically challenging and not suitable for large-scale studies [5]. Instead, the tagged proteins are usually introduced into cells through transfection or viral infection. Transiently transfected cells can be directly subjected to AP/MS, but that usually leads to the bait being expressed at relative high levels. Since overexpressed proteins often bind to proteins that they normally do not interact with under physiological conditions, the transient transfection followed by AP/MS approach may lead to considerable false-positive findings. A better approach may be to establish stable pools or cell lines with an expression level of bait protein similar to that of the endogenous protein [79].

One-step AP/MS using different epitope tags such as FLAG peptide tag (FLAG), human influenza hemagglutinin peptide tag (HA), and Green fluorescent protein (GFP) is the most frequently used approach in medium-to-large-scale PPI studies in mammalian cells [76, 78, 80]. The biggest advantages are obvious: it is economically and technically convenient and is able to generate a long list of interacting proteins [79]. Another major advantage is, since it has been used extensively in large-scale studies of mammalian cells and produced many publicly available datasets, the background generated by certain tags, such as FLAG-tag, is well known and is relatively easy to eliminate. With the quick development of MS equipment, advanced mathematical modeling, and data-driven elimination of common contaminants, one-step AP now provides more reliable results than it once did. In the recent Hippo pathway studies, AP/MS analysis using FLAG or 3XFLAG tag were able to generate dependable results with very few previously known contaminants such as keratin and chaperone proteins [80, 81].

Tandem affinity purification usually consists of two or more AP steps against different epitope tags fused to the bait protein. The original version of TAP tags contained a tobacco etch virus (TEV) protease cleavage site between two different tags, which could be cleaved by TEV protease and the protein complex released from the beads and subjected to the second step of AP [82]. Designed for protein complex characterization and proteome exploration [82] and first used in large-scale yeast proteomics studies [68, 70], TAP now has over twenty different modified versions with different combination of tags and has become the standard approach in several model organisms [83, 84], but not yet in humans. The major reasons are that it is relatively costly and technically challenging at the cleavage step during AP, hindering its adaptation for large-scale proteomics studies in human systems.

However, high-throughput MS/MS is also time consuming and very expensive. Significant efforts have been wasted on repeatedly re-identifying the same abundant peptides or peptides from “sticky proteins” that bind to many baits in AP/MS analysis [85]. The presence of these abundant and common contaminants may lead to the low coverage of peptides from specific prey proteins. Thus, the real signals are often buried in the noisy background, such that several technical repeats and sophisticated equipment are needed to achieve a high percentage of coverage for specific preys [86]. Since TAP or its modified versions eliminate a large fraction of abundant proteins during the two-step AP, it is potentially a more effective way of detecting “regulated” PPIs. Because these regulated interactions are critical for studies of signal transduction pathways, the TAP/MS approach may provide some level of advantage in medium-to-large-scale studies focusing on a particular signaling pathway. However, the most technically challenging part of TAP/MS is the cleavage/elution step. In order to establish PPI networks that are useful for directing in-depth functional studies in mammalian systems, a more user-friendly TAP method similar to those used in yeast proteomic studies is critically needed. Moreover, a data-driven computational method specifically designed for the TAP-MS approach would also help to uncover many bona fide interacting proteins, but at the same time eliminate contaminants and abundant proteins that are often associated with the baits in human systems.

In addition to traditional AP/MS or TAP/MS, quantitative AP/MS is another way to capture signal-induced changes in interactomes. Similar to its advantage in monitoring signal-induced whole proteome changes and PTM changes, labeled quantitative AP/MS approaches such as SILAC or iTRAQ also offer the ability to quantitatively compare between different stages in signaling events. SILAC based AP/MS has been used in several small-medium scale studies [67, 87]. For example, using SILAC to differentially label proteins in epidermal growth factor (EGF)-stimulated versus unstimulated cells identified proteins regulated by EGF signaling and their associated partners [88, 89]. The quantitative AP/MS approach holds lots of promise for future studies, but it is still technically challenging and available only in a limited capacity.

Data analysis, visualization and validation

Considerable efforts were devoted to MS data analysis (reviewed in [3, 90]). Raw spectra need to be processed first. The processing steps include noise filtering, background subtraction, isotope and charge deconvolution, peak detection, mass calibration and retention time alignment. Processed MS/MS spectra are searched against protein sequence databases thereby assigning MS/MS spectra to peptides, which can subsequently be mapped to proteins (reviewed in [3]). Target-decoy search strategy is usually used for increasing confidence and estimating false discovery rate (FDR) [91, 92]. The frequently used data processing software include SEQUEST [93], MASCOT [94] and X!Tandem [95].

Then the assigned PPIs are needed to be further analyzed to determine the bona fide interactions from background noises. Currently there are at least three well-accepted data filtration methods for label-free shotgun MS data analysis, all of which are based on the spectra counts from AP/MS results. These are the normalized spectral abundance factor (NSAF) method [75] and its data-driven expansion based on the BCM-CCI algorithm [72]; the Comparative Proteomic Analysis Software Suite (CompPASS) [76]; and the Significance Analysis of INTeractome (SAINT) method [96]. The NSAF algorithm was one of the first computational methods for this application, which normalizes the raw spectra counts with prey protein length and total spectra counts. This normalization idea has been adopted for other methods. The probability of interactions within and between complexes is computed solely on the basis of NSAFs using the Bayes approach [75]. Malovannya et al. established a data-driven version based on this method, which utilizes their large-scale proteomics data and is able to assign scores to each of the interactions [72, 97].

CompPASS was originally designed for analyzing the large-scale DUB PPI dataset [76]. Starting with two scores, Z and D, which are, respectively, the same as the Z-score used in statistics and the score representing prey specificity/data reproducibility, they introduced another score, WD, in their follow-up studies [78], which weighted the D-score to distinguish the “background proteins that were known to be true interactors” for specific baits from “background proteins that were never true interactors.” It largely relies on data reproducibility and has been proven to be effective with different datasets [98].

The SAINT method was originally designed for analyzing the yeast kinase and phosphatase interactomes obtained from yeast two-hybrid data [99]. This method has now been proven to be effective in the analysis of AP/MS data [80, 100, 101]. The SAINT method also uses advanced mathematical models to estimate the abundance and specificity parameters and therefore is capable of generating probability scores for each of the interactions. Several follow-up studies have been published, including SAINT-MS1, dealing with MS1 intensity data [102]; SAINT express, with a simpler statistical model [103]; and contaminant repository for affinity purification–mass spectrometry data (CRAPome), dealing with the common contaminants [104].

Several other modified data-driven algorithms have been used in individual studies [71, 105]. These computational algorithms could also be used together, or even combined with hand curating, to generate the list of high-confidence candidate interacting proteins (HCIPs) [81, 105].

Once the HCIPs have been identified and the interaction networks built up, Cytoscape software is usually used to visualize the interacting networks [106, 107]. With many plug-ins (i.e., apps), Cytoscape provides even more powerful features, such as topological analysis, clustering, functional enrichments, network comparison, and merging [108]. Other “R” based visualization methods are also available. The R framework and Bioconductor [109], are also frequently used for data analysis, clustering, and visualization. GO and pathway annotation [48, 49] are usually carried out using the HCIPs to achieve an understanding of pathway interactions.

For data validation, co-IP assay using epitope tags or antibodies against endogenous proteins is the most frequently used approach [78, 80, 81, 100, 101]. Other approaches include two-hybrid– and flowcytometry– based interaction validations, such as Fluorescence resonance energy transfer (FRET) and Bimolecular fluorescence complementation (BiFC) assays [110–113]. Reciprocal AP/MS is occasionally used for a few preys of interest, with a goal of recovering the original bait from AP/MS using its prey as the bait [77, 100]. Another part of the validation is to use the interaction data and search them in various PPI databases, such as STRING [114], BioGrid [115], Intact [116], BIND [117], DIP [118], and HPRD [119], to show that the newly identified interactomes at least partially overlap with these knowledge databases. The HCIPs could also be searched in the NetPath database [120], a resource of functional protein complexes in many signal transduction pathways, to get an idea of coverage of known proteins in the given pathway. In-depth functional validation, sometimes coupled with RNAi screening, could be conducted for several proteins of interest to prove that the dataset can lead to biological meaningful results [100, 101]. A few examples of data validation could be found in the following section of Hippo pathway studies. The whole strategy of AP/MS has been summarized and illustrated in Fig. 2.

Figure 2. An example of current workflow in AP/MS studies of a signaling pathway.

Bait proteins are usually epitope tagged and stably expressed in the cell of interest. Western blot analysis and immunofluorescence are usually carried out to ensure the proper localization and expression level of bait proteins. Ideally, the expression should be similar to the endogenous level of the protein. Cell lysate could run on sodium dodecyl sulfate–polyacrylamide gel electrophoresis and digested by enzymatic reaction (i.e., trypsin) to short peptides. Samples could be labeled by isotope/isobaric tags before the one-step or tandem affinity purification. Control purifications of mock or unrelated proteins are usually performed with the same strategies to eliminate the common contaminants and abundant proteins in AP. MS is usually performed with biological and technical replicates, and followed by data analysis, to generate a list of high-confidence interactions, which could be used to build up the interactomes. But to ensure the quality of the dataset, data validation including knowledge database overlapping, co-IP, and reciprocal AP/MS validation of a subset of the interactomes, is needed. In-depth functional validation could be conducted for several proteins of interest to prove the dataset can lead to biological meaningful results. The interactomes could be used to predict bait and prey functions and their diseases correlations. The example snapshots were taken from Couzens et al., Kwon et al., Wang et al., and Li et al.

Systematic proteomics studies of Hippo signaling pathway – an example of using AP/MS in the study of signal transduction pathways

There are many advantages to systematic analysis of a signaling pathway using medium-to-large-scale AP/MS, which purifies the core components or all of the known proteins involved in a given pathway. New components or regulators could be identified through AP/MS, which can be further functionally validated [2]. Comparing the purification results under different conditions or following treatment with inhibitors or drugs may also gain additional insights into protein functions and regulations. Several medium-to-large-scale proteomics studies focused on certain signaling pathways have been conducted; for example, soon after the large-scale TAP/MS studies in yeast [68], Bouwmeester et al. performed TAP/MS for 32 known and candidate proteins in the human TNF-α/NF-κB pathways and identified 80 novel interactors [121]. Glatter et al. studied the InR/TOR pathway in Drosophila using AP/MS and identified 58 new network components from 16 known baits in the pathway [122]. These studies provided new insights into these pathways and identified potential novel components of these pathways, which significantly benefited the research in these fields.

Recent studies by four different groups using AP/MS methods for investigating the Hippo-YAP signaling pathway, an important pathway that controls organ size and tumorigenesis, provide an interesting example of the variation in methods for using AP/MS approach in studying signal transduction pathways [80, 81, 100, 101]. Early components of the Hippo pathway, including warts (wts) [123, 124], salvador (sav) [125, 126], hippo (hpo) [127, 128], mob as tumor suppressor (mats) [129], yorkie (yki) [130], merlin (mer), and expanded (ex) [131], all of which are involved in organ size control, was identified from genetic mosaic screens in Drosophila (reviewed in [132]). Later, the mammalian homologs of these proteins, including Lats1/2, WW45, Mst1/2, Mob1, YAP/TAZ, NF2, and FRMD6, were identified and shown to function in a similar signaling pathway (Fig. 3A) (reviewed in [133]). Although these known components have been extensively studied during the last decade, the identification of new components in the Hippo pathway has been relatively slow. Over 30 components of the human Hippo pathway have been identified during the last decade. However, compared with other well-defined signaling pathways such as Wnt and MAPK that are also known to be involved in tumorigenesis, very few mutations have been discovered in known Hippo pathway genes in common human cancers. Neurofibromin 2 (NF2) is the only Hippo pathway gene known to be frequently mutated and inactivated in cancers [134]. This suggests that there might be undefined Hippo pathway components that are mutated and responsible for tumorigenesis in some cases. Alternatively, the regulation of the Hippo pathway, presumably through PTMs and PPIs, may be more relevant to human carcinomas than the somatic mutations of genes involved in this pathway. In either case, identifying new components and new regulatory mechanisms of this pathway is essential for further study of its role in organ size control and human cancers.

Figure 3. An overview of the human Hippo signaling pathway before and after proteomics studies.

(A) The core knowledge of the Hippo signaling pathway before proteomics studies. Red bubbles indicate positive regulators and blue bubbles indicate negative regulators of Hippo signaling. (B) The Hippo signaling pathway refined with newly identified components. The three experimentally validated regulators CIT, CCDC85C, and leash are highlighted. Other novel HCIPs shown in at least two datasets are shown with different subcellular localizations (green: on membrane; tan: nuclear; grey: cytoplasmic and others). The list was generated by Moya et al. using the datasets from Couzens et al., Kwon et al., and Wang et al.

Using a modified GS-TAP system designed for protein purification in Drosophila [135, 136], Kwon et al. performed TAP/MS for 12 core components of the Hippo pathway in Drosophila and identified a high-confidence PPI network consisting of 153 HCIPs [100]. They conducted co-IP validation for 26 interactions using epitope-tagged proteins, and 23 of them were positive. They further carried out an RNAi screening using 255 dsRNAs targeting the 153 HCIPs and monitored the effect on this signaling pathway using a Yki-reporter assay. Seventy-seven negative and 23 positive regulators were identified that showed >1.5-fold change of Yki-reporter activity. They subsequently performed reciprocal TAP/MS using Yki-interacting protein CG4674 (leash) and recovered Yki, which indicates that at least some of these interactions occur under physiological conditions. Finally, they functionally validated this protein leash as a Yki regulator in Drosophila, since overexpression of leash led to significant reduction in wing size [100]. This work provides a standard workflow for future proteomics studies of signaling pathways, which start with AP/MS, followed by interactome validation and further confirmed by functional validation.

Similarly, Wang et al. performed streptavidin-FLAG-S (SFB) triple-tagged TAP/MS for 32 Hippo pathway components in human cells. They identified 343 HCIPs and functionally validated WWC1-interacting protein CIT and YAP1-interacting protein CCDC85C as new Hippo pathway components that regulate YAP1 subcellular localization and activities [101]. Couzens et al. performed 3xFLAG-tagged AP/MS coupled with proximity-dependent biotin identification (BioID) AP/MS [137] of 27 Hippo pathway components treated or not treated with serine/threonine phosphatase inhibitor okadaic acid and identified 400 HCIPs [80]. Since phosphorylation cascades are often involved in signal transductions, the interactomes of the Hippo PPI networks of treated and untreated cells could be compared to suggest putative mechanisms for the regulation of this pathway and further functional implications [80]. In addition, Hauri et al. performed AP/MS for 34 bait proteins related to the human Hippo pathway (8 core components, 26 secondary baits) and identified 270 HCIPs [81]. Taken together, the findings from these four groups uncovered nearly one thousand HCIPs, which should be considered new potential components of the Hippo pathway (Fig. 3B). It is likely that these newly identified components will be the starting points for further study of the regulation of the Hippo-YAP signaling pathway and how it may be connected with other cellular processes and signaling pathways. Moreover, some of these new components may serve as potential biomarkers or drug targets for cancer diagnosis and treatment in the future.

Current obstacles for the AP/MS-based large-scale studies of signal transduction pathways

Unfortunately, there are still some conceptual and technical challenges to proteomics studies of signaling pathways. While current AP/MS methods are effective in identifying stable protein complexes, they are inadequate in recognizing regulated interactions, which are essential for understanding the complex signaling networks in the cell. This shortcoming is particularly noticeable in large-scale proteomics studies, since the appearance of abundant associated proteins in the interactomes drastically reduces the sensitivity for detecting low levels of regulated but biologically significant interactions. Doing AP/MS in different organism, conceptual biases, and differences in AP approaches all lead to very distinct results. Since different organisms are genomically different, the results from different organisms are usually not comparable [138]. However, proteins encoded by genes mutated in inherited genetic disorders are likely to interact with proteins known to cause similar disorders, which suggest the existence of disease PPI subnetworks [138]. This “guilty by association” concept may also apply to signaling pathways, since many signal transduction pathways, such as Wnt and Hippo, are highly conserved across species and their corresponding PPI networks should also display a high degree of similarity.

Another technical challenge is the diversity of AP methods that are being used in the field, which makes it difficult to compare results from different investigators. As already discussed, using epitope tags is still the most cost-effective approach for large-scale proteomics studies. This is unlikely to change before unified and quality-controlled collection of antibodies against endogenous proteins becomes available [2, 52, 53]. It was noticed from the very beginning that one-step purification [69] and tandem purifications [68] in yeast generate very different results even with the same baits [139], which highlights the importance of having unified or comparable technical platforms. In yeast, TEV-based TAP has become a standard approach, and many large-scale studies have been performed on this platform [68, 70]. In human studies, however, there are at least thirty different epitope tags or combinations that have been used for AP. In the recent Hippo signaling studies, for example, five different AP methods were used by four groups, of which three were one-step APs and two were modified TAPs. The datasets from the first three groups mentioned above, one in Drosophila and two in human, are hardly comparable: less than 10% of the HCIPs were present in all three datasets and only ~20% in the two human datasets [140]. Since these three groups used very similar bioinformatics approaches for data filtration [96], the discrepancies likely come from the AP approaches they used.

It is difficult to conclude which AP method is the best for large-scale proteomics studies until a systematic comparison among different AP methods and painstaking data validation have been conducted. Nevertheless, if it is possible, a unified AP/MS method(s) in human proteomics studies, just like the TEV-based TAP/MS method used in yeast, would provide great advantages in understanding signaling events in humans, since it would permit meaningful comparisons of results generated by many investigators. Interlaboratory collaborations have been made to achieve high reproducibility of large-scale human protein-complex analysis by using a standardized TAP/MS approach with same reagent and data analysis [105]. The hope is that one day the MS data generated by different research groups could become comparable and integrated into one large database, as happened for microarray data. This will greatly advance the field and profoundly change how AP/MS and other proteomics data can be used for functional studies.

On the other hand, one could also argue that technological diversity may in some cases be beneficial and required, since using one unified method could in theory introduce strong technical biases. Tandem affinity purification usually provides “cleaner” results with fewer common contaminants, but it is more technically demanding than one-step AP and may cause loss of transient or weak interactions during the cleavage/elution step [2]. One-step AP, such as the use of FLAG-tag, usually produces a list of potential interacting proteins that is longer than the one obtained from TAP, and it has been used in many milestone studies on a large scale. However, one-step AP has its own technical limitations, such as in detecting transient or regulated interacting proteins in the presence of highly abundant, non-specific associated proteins [141, 142]. Several other AP techniques have been developed recently to capture the weak/transient interactions, such as BioID [137]. The BioID method is able to produce meaningful data that reflect “local protein partners” information, which could not be generated by the traditional AP method [80]. However, it remains to be determined whether this method is able to capture weak or transient interactions, such as enzyme-substrate interactions. Ideally, at least two unified AP methods should become standards in the field for signaling studies: a more stringent method to characterize stable or regulated complexes, and a more sensitive method that captures transient or weak interactions that are functionally important for studies of signal transduction.

Another challenge in the field is data filtration. With the advances in MS technology and bioinformatics, we are gradually overcoming the obstacles of generating raw spectra counts and mapping them to correct proteins [143]. We are now increasingly facing the complication of our own success: how to make sense of a list of hundreds, if not thousands, of putative associated proteins? Before the systematic bioinformatics data filtration approaches were developed, the data were usually manually curated and filtered by experience [2]. That may be feasible for studies of individual proteins, since the number of proteins in one MS list, no matter how long it may be, is still limited, and follow-up confirmation and functional studies are usually carried out to make sure that the right PPIs are presented. For large datasets such as whole pathway proteomes, however, automated data filtration algorithms are needed.

The current data filtration methods work efficiently for their originally designated applications [72, 75, 76, 96]. Methods such as SAINT, which use semi-supervised mixture models and Makov chain Monte Carlo (MCMC) sampling, is mathematically advanced and works effectively with a group of unrelated proteins [96]. Such an assumption may not be true in biology, however, since in many cases a connection between unrelated proteins may lead to new scientific discoveries. In signaling pathway studies, data filtration becomes even more challenging, since proteins often interact with other proteins in the same pathway. Together with their interacting proteins, they frequently appear in other purification results across experiments in the same pathway. This severely interferes with the estimated parameters for abundance and specificity and could lead to both false-positive and false-negative identifications. Such a problem may be difficult to solve solely by mathematical modeling, but customized data-driven mathematical algorithms may offer a solution. Combining high-throughput datasets with mathematical modeling will help greatly in identification of interactions with more confidence and fewer false positives [79]. Several attempts have been already made, such as establishment of the Drosophila Protein Interaction Mapping project (DPIM) [71] and contaminant repository for AP/MS data (CRAPome) [104]. More investigations in AP/MS data processing will undoubtedly advance this quickly growing field.

Future Perspectives

The technological advancement in the field of AP/MS analysis for proteomics studies is far exceeding our ability to handle the information properly. Sometimes even simple questions face a dilemma. For example, AP/MS is often performed with biological replicates and technical replicates; should the not-well-reproducible data (e.g., identified twice in four replicates) or the interactions with low peptide counts/interaction scores be automatically excluded? For studies of individual proteins, the answer should surely be no. Since one AP/MS covers only a limited portion of the total peptide population [86], and some proteins are relatively difficult to recognize by MS because of their abundance or sequence/structural features [53], the peptide readings per se may not reflect the real biological importance of these protein-protein interactions, especially if these interactions are regulated in a signaling pathway. In this case, the functional validation and relevance is the most important aspect and these leads should be pursued with a biological question in mind.

For large-scale proteomics studies, the answer to the same question may be ambiguous. Large-scale datasets focusing on signaling pathways should not only provide a valid resource for the research community but also should generate new insights from an overall view of the biological system, such as functional correlation with different biological processes or other signaling pathways, or even provide new disease correlations based on the interactions identified. Without proper data filtration, the GO process and functional implication may not be meaningful, largely because of the presence of abundant contaminants in the datasets. In these cases, the criteria for data filtration should be stringent to reveal the true biological significance.

On the other hand, large-scale MS studies are very expensive and time consuming. We should utilize any information gained from these studies as fully as possible. Luckily, other large-scale approaches, such as RNAi screening, microarrays, metabolomics, next-generation sequencing, and even patient mutation/disease correlation databases, are available and continually growing. The integration of data generated via different technologies will greatly increase the data accuracy and provide much needed new insights into mechanisms and disease relevance. A good example of combining multiple technologies is a genome-wide siRNA screen searching for Wnt pathway regulators in a colorectal cancer cell line, which was conducted together with AP/MS studies of several proteins in this pathway [144]. Several relatively weak hits from AP/MS have been functionally confirmed by siRNA screening [144].

Moreover, the field of mathematical modeling is also growing fast. The new Markov network methodology has been proven effective in building neuron networks and incorporating different distributions such as Gaussian, Poisson, and Binomial in one model [145]. We hope that one day data produced by all different platforms will be integrated into one database and generate a super-interactive network for signaling pathways, which truly reflect the complex biological systems we are working on.

Abbreviations

TGF-β

transforming growth factor–beta

PPI

protein-protein interaction

PTM

post-translational modification

MS

mass spectrometry

MS/MS

tandem mass spectrometry

LC

liquid chromatography

ICAT

isotope-coded affinity tags

SILAC

stable isotope labelling by amino acids in cell culture

iTRAQ

isobaric tags for relative and absolute quantification

TMTs

tandem mass tags

MS3

triple-stage mass spectrometry

Ub

ubiquitin

AP

affinity purification

TAP

tandem affinity purification

GO

gene ontology

co-IP

co-immunoprecipitation

FLAG

FLAG peptide tag

HA

human influenza hemagglutinin peptide tag

GFP

green fluorescent protein

TEV

TAP tags containing a tobacco etch virus protease cleavage site

EGF

epidermal growth factor

FDR

false discovery rate

NSAF

normalized spectral abundance factor

CompPASS

comparative proteomic analysis software suite

SAINT

significance analysis of interactomes

DUB

deubiquitinating enzyme

CRAPome

contaminant repository for affinity purification-mass spectrometry data

HCIP

high-confidence candidate interacting proteins

SFB

streptavidin-FLAG-S triple tags

BioID

proximity-dependent biotin identification

DPIM

the drosophila protein interaction mapping project