Optimization and ZSPORE Analysis of Affinity Purification Coupled with Tandem Mass Spectrometry in Mammalian Cells

The availability of high affinity monoclonal antibodies against the HA and FLAG epitope tags has led to their frequent utilization for affinity purification of protein complexes. These are the most commonly used epitopes for mapping the proteome. To compare the efficacy of these epitopes for affinity purification, FLAG and HA fusion proteins were purified in parallel. The bait proteins used for optimizing affinity purification included MDA5 and other well-known components of the innate immunity network ^{8, 16, 17}. HEK293 cells stably expressing FLAG and HA double tagged MDA5 N-terminal card domain (MDA5-N) were used to compare efficiencies. A total of 4x107 cells were collected and the cell lysate was incubated with either 20 µl anti-FLAG or anti-HA conjugated beads for 12 hr. After washing, the MDA5-N complex was eluted with FLAG or HA peptides. Both strategies yielded purified MDA5-N protein complexes as shown by silver staining (Figure 1A). Quantitative analysis of band intensity demonstrated that purification from anti-FLAG conjugated beads yielded 6.8 fold more protein than anti-HA beads (Figure 1A). Furthermore, purification with anti-FLAG beads pulled down several additional bands (Figure 1A). One band was identified as MAVS, a known MDA5 interactor. In contrast, purification with anti-HA beads did not pull down detectable quantities of MAVS. Taken together, purification with anti-FLAG conjugated beads yield about 7 fold more protein than anti-HA beads, suggesting the FLAG tag is more efficient for AP-MS.

Figure 1.Optimization of AP-MS. (A) Comparison of affinity tag purifications with anti-FLAG and anti-HA antibodies. HEK293 cells stably expressing FLAG and HA tagged MDA5 N-terminal card domain (MDA5-N) are compared. Left panel shows silver staining of MDA5-N complexes purified with HA and FLAG tag. Right panel indicates lane intensity quantitated by densitometry. NSBP stands for non specific binding protein. (B) Comparison of one-step and tandem purifications. Protein numbers identified in TBK1, NAP1, IRF3, and SINTBAD complexes are depicted. F stands for FLAG affinity purification and FH indicates FLAG and HA tandem purification.

Single tag AP-MS and tandem affinity purification (TAP) are both broadly applied methods for protein purification. TAP is a two-step procedure requiring sequential purification using two different affinity tags. FLAG and HA double tags are most commonly applied for tandem purification of protein complexes. To compare the effect of tandem tag vs. single tag purification on the yield of total prey and HCIP, we compared protein complexes purified by single purification with FLAG vs. a two-step purification with FLAG followed by HA. To compensate for the lower binding capacity of anti-HA beads we used 4 times more anti-HA beads than anti-FLAG beads for immunoprecipitation. The number of HCIP associated with the kinase, TBK1, was determined by the algorithm ZSPORE detailed below. MS analysis revealed that TAP purified TBK1 complexes lacked several known interactors, including optineurin (data not shown). As with other screening methods, AP-MS is unable to detect all interactors. For example, neither single-step nor tandem purification of TBK1 pulled down A20, a TBK1 known interactor ¹⁸. We next examined the efficacy of tandem purification using three additional proteins associated with the innate immune signaling pathway (NAP1, IRF3, and SINTBAD). The number of total interactors was dramatically reduced in all protein complexes isolated by TAP purification. However, the ratio of HCIP to total prey did not increase. Consistently, more HCIP were detected by single step affinity purification (Figure 1B). In brief, tandem purification reduces the NSBP at the price of HCIP loss.

As with many screening methods, unfiltered AP-MS data contain many non-specific binding proteins caused by binding to the antibody coated beads, epitope tag, aggregation, or carryover from prior MS runs. Several computational tools have been developed for processing AP-MS data to eliminate NSBP and identify HCIP ^{7, 10, 15}.

We aimed to create a simplified method for analysis of AP-MS data. Three main parameters (protein abundance, the frequency of observed protein in the database, and reproducibility) were combined to generate an algorithm. Total spectral counts (TSC) have gained acceptance as a practical, label-free, semi-quantitative measure of protein abundance for proteomic studies ^{19, 20}. We adopted the Z-score statistic to compare protein abundance because conventional Z-score determines the number of standard deviations away from the mean. However, Z-score does not reflect reproducibility. In our protocol bait protein complexes are retrieved in two independent experiments, and then each eluate is divided into two aliquots which are analyzed by MS on different days. As each protein complex is tested in at least 4 MS runs, reproducibility can be readily factored into the analysis. Z-score also does not analyze information about prey occurrence (i.e. prey uniqueness). To explore the likelihood that an interaction is specific, we empirically set a value of prey occurrence at < 7%. We now propose a simple 3 stage scoring system to identify HCIP. This algorithm combines Z-score plus prey occurrence and reproducibility (ZSPORE). In the ZSPORE scoring system each interaction must pass all 3 criteria to merit classification as HCIP (Figure 2A). This scoring method is flexible as each parameter can be adjusted to vary the stringency of analysis. The flowchart of ZSPORE is shown as in Figure 2A and a detailed description is provided in Materials and Methods.

Figure 2.ZSPORE analysis and validation. (A) Flowchart of ZSPORE analysis and algorithm. (B) Comparison of HCIP identification by ANOVA and ZSPORE using the innate immune database. (C) Validation of HCIP by co-IP and comparison with BioGRID and STRING databases.

We first evaluated the performance of ZSPORE using our published database of Human Innate Immunity Interactome for type I Interferon (HI5) ⁸. The HI5 dataset consists of 58 innate immunity genes and 5 negative controls. A total of 9432 pairwise interactions were detected between 1218 preys and 58 baits. We calculated Z-scores for each Xi,j (where X is the maximum TSC associated with prey i which interacts with bait j) based on 4 MS runs. For this analysis we set the Z-score cutoff at 2. The threshold of reproducibility was established at 2 of 4 runs. To minimize NSBP in this small set of baits, prey occurrence was set at <7%. Empirically adjusting the percent of prey occurrence is used to effectively discriminate known interactors in this signaling pathway which features a high rate of interconnectivity among baits. Established bait-to-bait interactions are listed (Table S1). In total 279 HCIP forming 415 total interactions were identified using ZSPORE compared with 425 interactions detected by ANOVA (Table S1). The high overlap of 390 interactions between the ANOVA and ZSPORE analyses validates the efficiency of ZSPORE (Figure 2B). We examined the overlap between ZSPORE and experimental findings by comparing the interactions in public databases and interactions verified by co-immunoprecipitation (co-IP). As shown in Figure 2C, ZSPORE verified 88 out of 92 known interactions reported in the BioGRID and/or STRING databases. ZSPORE also confirmed 107 out of 110 interactions detected by co-IP and reported in our previous study ⁸. In total, > 96% (195 of 202) verified and known interactions detected with the ZSPORE strategy were validated. This is an improvement over the previous ANOVA-based computational approach ⁶. The stringent statistical criteria for selecting HCIP may have masked relevant false negatives. Of 345 known interactions 20 (6% false negative) were filtered out as NSBP in the innate immune database according to ZSPORE. Conversely, 3 out of 110 interactions found by ZSPORE could not be validated by co-IP. Therefore, the false positive rate of ZSPORE on the innate immune database is < 3%. Taken together, the ZSPORE method combines three metrics (Z score based on TSC, prey occurrence, and reproducibility). These parameters are embedded in most AP-MS datasets. Thus, ZSPORE is a simple, efficient and robust way to analyze AP-MS data. Recalculation of the innate immune interactome with the ZSPORE algorithm (Table S2) provided an expanded innate immune interactome with 415 interactions.

Our AP-MS database and the simplified computational strategy represent a valuable resource for investigators using the outlined purification procedures, especially those analyzing a small number of baits. Combining datasets will enhance the resolution of HCIP. Dissemination of the database to interested members of the research community can be arranged by contacting the authors. The ZSPORE algorithm may also be applied to large datasets involving other species and cell types.

To evaluate the ability of the ZSPORE strategy to identify novel HCIP, we applied these tools to a well-characterized adaptor protein. Growth factor receptor-bound protein 2 (GRB2) is ubiquitously expressed and plays a critical role in receptor tyrosine kinase signaling pathways ^{21, 22, 23}. Thus, GRB2 was chosen to validate our AP-MS pipeline (Figure 3). In brief, FLAG-tagged GRB2 was cloned into a CMV-driven vector with hygromycin resistance and transfected into HEK293 cells. After selection with hygromycin the stable cell line was cultured in five 150 cm2 dishes. After affinity purification of the GRB2 complex sample was split into halves and each was run onto a NuPage gel. Short electrophoresis migration (~1 cm) was adopted and the whole lane was excised for mass spectrometry. In our hands MS analysis of individual bands separated by full length electrophoresis failed to show significant improvement of HCIP identification when compared with short electrophoresis. Furthermore, collection of the entire sample in one fraction avoids the complication of analyzing MS data merged from different MS runs of individual bands.

Figure 3.Overview of Optimized Pipeline of AP-MS.

Data from 4 MS runs (replicates from 2 independent purifications) were collected and in total 166 interactors were detected (Table S3). Z-scores were calculated based on the maximum TSC of each prey among 4 independent MS runs and analyzed against our current database of 211 protein complexes (Table S2 and unpublished data). Increasing the size of the core database allows one to apply greater flexibility or stringency in selecting HCIP. 73 preys showed Z-scores higher than 2 (p< 0.05). Next we investigated prey occurrence; one prey was filtered out based on its occurrence in >7% in our database. Reproducibility of GRB2 data revealed 54 preys (79%) appeared at least twice out of 4 MS runs. Finally, 4 proteins with only one peptide hit were removed. Taken together, we identified 50 HCIPs associated with GRB2 (Table 1).

Cytoscape software was used to visualize the interconnectivity of the GRB2 complex and combine the interactions into one map (Figure 4). The Z-score and TSC values are represented by line color and thickness, respectively (Figure 4). Comparisons with the BioGRID and STRING public protein interaction databases identified 24 known GRB2 binding proteins (23 in BioGRID, 13 in STRING, and 12 in both), including some well-known partners, SOS1/2, GAB1/2, PI3K and CBL (Figure 5A). The data were consolidated into a GRB2 interaction map, prey colors represent known interactions found in the BioGRID and/or STRING databases (Figure 4). Overview of the GRB2 network reveals clusters of related proteins (including members of the PI3K, WIPF, DNM and KIF4 families), which belong to structurally and functionally related gene families. GRB2 associates with 18 known GRB2 interactors highlighted by dashed lines forming a sub-network with high interconnectivity.

Figure 4.GRB2 complex purification and analysis. Protein interaction network of GRB2 complex using Cytoscape. Z-score and TSC are represented by line color and thickness, respectively. The fill color indicates novel or known GRB2 interactors as designated. Dashed lines represent known interactions between GRB2 interactors.

Figure 5.Analysis of GRB2 interactome. (A) Comparison of GRB2 complex with BioGRID and STRING public protein interaction databases. (B) Functional categorization of GRB2 interactors. (C) Pfam domain distribution among GRB2 interactors.

Successful identification of known interactions established the efficiency and robustness of our AP-MS pipeline. We also identified 26 new interactors associated with GRB2. For example, protein diaphanous homolog 1 (DIAPH1) is a new GRB2 partner with 11 peptide hits. DIAPH1 is involved in MEMO1-RHOA-DIAPH1 signaling pathway, which plays an important role in ERBB2-dependent stabilization of microtubules at the cell cortex ²⁴. Since GRB2 is an adaptor of EGFR receptors, GRB2 may bridge DIAPH1 to ERBB2. Functional categorization of GRB2 interactors suggests that GRB2 not only plays a role in growth receptor pathways like ERBB and VEGF, but also crosstalks with the JAK-STAT, mTOR, and immune cell receptor signaling pathways (Figure 5B). Another interesting novel GRB2 binding partner is ARAP1 (Arf-GAP with Rho-GAP domain, ANK repeat and PH domain-containing protein 1). ARAP1 is strongly associated with GRB2 (61 peptide counts). ARAP1 is a PIP3-dependent Arf GAP that regulates Arf-, Rho-, and Cdc42-dependent cell activities ²⁵. Furthermore, analysis of the domains contained among GRB2 interactors found several novel GRB2 interactors share domains with previously established interactors (Figure 5C). For example, SNX18 and SH3PXD2B have the PX and SH3 domains characteristic of PI3K family members.

In summary, the methods described here represent a broadly applicable pipeline from affinity purification of protein complexes to statistical analysis of AP-MS data. We validated the utility of this strategy by defining an enlarged innate immune network. To further demonstrate the advantage of these tools on a well-studied protein we analyzed the GRB2 complex. In addition to 24 known interacting proteins, we also identified 26 new binding partners. The robustness of our AP-MS pipeline supports its widespread application in the characterization of protein interaction networks for various signaling pathways in different cell types.

Many processes in a cell depend on protein-protein interactions and perturbations of these interactions can lead to pathophysiology. Comprehensive knowledge of protein interaction networks will identify novel components and yield new insights on how cells respond to different environments. Ultimately such knowledge may provide new targets for therapeutic application. In conclusion, this approach to AP-MS is an invaluable tool for identification of new protein-protein interactions and mapping protein interaction networks.

HEK293 cells were purchased from ATCC. Human GRB2 cDNA was ordered from PlasmID (Dana-Farber, Boston). Antibodies specific for FLAG and HA were purchased from Sigma Chemical Co (St. Louis, MO).

Public protein interaction databases include the STRING database (protein.links.v7.1.txt.gz, found at http://string.embl.de/) and the BioGRID database (http://www.thebiogrid.org/downloads.php). The protein interaction network was generated in Cytoscape (http://www.cytoscape.org).

The methods and criteria used to remove non-specific binding proteins (NSBP) and identify high confidence interacting proteins (HCIP) include Figure 2A. (2)each Xi,j (prey i interacts with bait j) based on the maximum total spectral counts (TSC) of4 MS runs.