1. Biochemistry and Chemical Biology
  2. Computational and Systems Biology
Download icon

Profiling the Escherichia coli membrane protein interactome captured in Peptidisc libraries

Tools and Resources
  • Cited 1
  • Views 2,084
  • Annotations
Cite this article as: eLife 2019;8:e46615 doi: 10.7554/eLife.46615

Abstract

Protein-correlation-profiling (PCP), in combination with quantitative proteomics, has emerged as a high-throughput method for the rapid identification of dynamic protein complexes in native conditions. While PCP has been successfully applied to soluble proteomes, characterization of the membrane interactome has lagged, partly due to the necessary use of detergents to maintain protein solubility. Here, we apply the peptidisc, a ‘one-size fits all’ membrane mimetic, for the capture of the Escherichia coli cell envelope proteome and its high-resolution fractionation in the absence of detergent. Analysis of the SILAC-labeled peptidisc library via PCP allows generation of over 4900 possible binary interactions out of >700,000 random associations. Using well-characterized membrane protein systems such as the SecY translocon, the Bam complex and the MetNI transporter, we demonstrate that our dataset is a useful resource for identifying transient and surprisingly novel protein interactions. For example, we discover a trans-periplasmic supercomplex comprising subunits of the Bam and Sec machineries, including membrane-bound chaperones YfgM and PpiD. We identify RcsF and OmpA as bone fide interactors of BamA, and we show that MetQ association with the ABC transporter MetNI depends on its N-terminal lipid anchor. We also discover NlpA as a novel interactor of MetNI complex. Most of these interactions are largely undetected by standard detergent-based purification. Together, the peptidisc workflow applied to the proteomic field is emerging as a promising novel approach to characterize membrane protein interactions under native expression conditions and without genetic manipulation.

https://doi.org/10.7554/eLife.46615.001

Introduction

Proteins control biological systems in a cell. While many perform their functions independently, the majority of proteins interact with others to achieve their full biological activity. Characterizing protein-protein interaction networks (the interactome) has traditionally been accomplished by methods such as affinity purification coupled to identification by mass spectrometry (AP/MS) (Arifuzzaman et al., 2006; Hu et al., 2009; Babu et al., 2012; Babu et al., 2018), protein fragment complementation assays (Rochette et al., 2015; Tarassov et al., 2008), or yeast two-hybrid screening (Y2H) (Rajagopala et al., 2014). While high-throughput, these methods are quite often limited in their scope by poor scalability because bait proteins must be independently tagged. The addition of these tags can also have uncontrolled effects on proteins such as disrupting binding sites, altering localization, stability, and thereby the accurate prediction of the interactome. Co-fractionation methods, such as protein-correlation-profiling (PCP) in combination with quantitative proteomics methods, such as label-free quantitation (LFQ) or stable isotope labeling of amino acids in cell culture (SILAC), are therefore emerging as an attractive alternative to identify protein complexes under native expression conditions and without genetic manipulation (Kristensen et al., 2012; Scott et al., 2017; Havugimana et al., 2012). Fractionation of a proteome under these native conditions, followed by quantitative proteomic analysis of co-fractionation profiles, allows identification of protein complexes through a principle of ‘guilt-by-association’. This method can generate thousands of potential interactions in a single experiment, and incorporation of SILAC multiplexing allows simultaneous comparison of multiple states of the interactome (Kristensen et al., 2012; Scott et al., 2017).

While co-fractionation has been successfully applied to soluble proteomes, characterization of the membrane proteome has lagged. This is largely due to the hydrophobic nature of membrane proteins and their sequestration in the lipid membrane. To extract this water-insoluble proteome, it is necessary to solubilize the lipid bilayer with the aid of detergents or amphipathic co-polymers such as styrene maleic acid (SMA) (Dörr et al., 2014). When mild detergents are employed, membrane protein complexes can be directly detected following their separation by techniques such as size exclusion chromatography, density gradient centrifugation (McBride et al., 2017) or blue-native gel electrophoresis (Scott et al., 2017; Heide et al., 2012). However, even the mildest detergents tend to decrease protein stability while increasing protein aggregation (Yang et al., 2014). In fact, prolonged exposure to those detergents tends to delipidate proteins and alter their conformation, which can have confounding effects on membrane protein complex stability. As an additional drawback, micelles of detergent must be removed from all samples before analysis by mass spectrometry, which often decreases protein identification (Yeung and Stanley, 2010; Bechara et al., 2015; Bao et al., 2013; Yang et al., 2014). Thus, while a great deal of useful data has been generated using detergent-based proteomics analysis, there is still a pressing need for novel methods that are unencumbered by detergent side-effects.

We recently developed the peptidisc as a novel membrane mimetic scaffold to keep membrane proteins water-soluble (Carlson et al., 2018). The peptidisc is formed when multiple copies of the 4.5 kDa amphipathic scaffold NSPr (also called Peptidisc peptide) wrap around the solubilized membrane proteins. Reconstitution occurs spontaneously upon removal of detergent, incorporating both endogenous lipids and solubilized membrane proteins into detergent-free particles. The number of scaffolds adapts to fit the diameter of the protein target without bias toward large protein complexes. The end result is peptidisc particles that are stable, free of detergent effects, and soluble in aqueous solution (Carlson et al., 2018). Our previous work has shown that the peptidisc is able to stabilize both inner and outer membrane proteins of Escherichia coli.

In this study, we apply the peptidisc to the trapping of the bacterial cell envelope proteome into water-soluble particles. This is performed by reconstituting the heterogeneous membrane protein mixture immediately after its extraction from the cell envelope with mild detergent. This process minimizes protein dissociation and denaturation because it limits exposure to detergent and thereby protein delipidation. The membrane proteome trapped in the peptidisc library is water-soluble and stable during prolonged incubations. This library is then fractionated by high-resolution size exclusion chromatography (SEC) in the total absence of detergent. Application of the PCP workflow, which includes stable isotope labeled amino acids in cell culture (SILAC) and mass spectrometry (LC/MS-MS), allows us to precisely characterize the content of the peptidisc library across the various fractions. When the peptidisc library from the raw E. coli cell envelope is analyzed this way, we identify and quantify 1209 unique proteins, of which 591 are predicted to be directly membrane integrated. From these 1209 proteins, we predict 4911 binary interactions - each characterized by a degree of precision. Our interaction list is hereafter called the peptidisc interactome.

To computationally validate the precision of the peptidisc interactome, we benchmark the dataset against the recently published E. coli cell envelope interactome (‘CE’) (Babu et al., 2018) and two other unpublished interactomes collected for that earlier study (‘validating interactomes’). We also measure the biological plausibility of the peptidisc interactome by determining enrichment for shared gene ontology terms, binding domains, and correlation of growth phenotype (Erickson et al., 2017; Mosca et al., 2014). We also compare the peptidisc-reconstituted membrane proteome against a membrane proteome prepared using the SMA polymer instead of detergent. We find, however, that large membrane protein complexes are better preserved in the peptidisc workflow.

Guided by the peptidisc interactome datalist, we select three well-characterized membrane protein complexes in order to discover novel interactions. With the Sec translocon, we validate association of SecY with the membrane chaperones YfgM and PpiD. This interaction can be isolated in detergent but only when all subunits are simultaneously over-produced. Remarkably, we also discover significant correlation between certain subunits of Sec and Bam complexes, suggesting an astonishing network of protein associations spanning across the entire bacterial cell envelope. We confirm this observation using SILAC AP/MS, thereby providing direct evidence for the Bam-Sec super-complex. Continuing with the Bam complex, we show that all five subunits are captured in peptidisc in addition to two other interactors - RcsF and OmpA. These interactions were previously inferred from genetic and indirect cross-linking experiments, but direct association was not formally demonstrated (Hart et al., 2019; Konovalova et al., 2014). Accordingly, these interactions are much less apparent in detergent. Finally, with the ABC transporter MetNI, we find that binding of the substrate binding protein MetQ depends on its N-terminal lipidation The importance of the MetQ lipid anchor is novel and is unique case among Type I ABC transporters. Moreover, we identify NlpA, also called lipoprotein 28, as a bona fide novel interactor of the MetNI complex.

Altogether, this work validates the peptidisc library workflow as an efficient method for capturing and stabilizing the membrane proteome into soluble particles. The method enables high-throughput detection of detergent-sensitive membrane protein interactions. When combined with rigorous experimental validation, the peptidisc interactome is revealing novel and transient interactions, many of them of fundamental importance to the transport process and biogenesis mechanism of the cell envelope.

Results

Capture of the E. coli membrane proteome in peptidisc

The peptidisc-based SEC-PCP-SILAC workflow is presented in Figure 1. To identify the optimal solubilization of the E. coli cell envelope, we employed three different non-ionic detergents (DDM, LDAO and β-OG) and one ionic detergent (DOC). Each was tested using E. coli crude membranes enriched for an inner membrane protein marker, MsbA. Upon removal of aggregate by ultracentrifugation, the solubilized membrane proteomes were incubated with Peptidisc peptide at the ratio 2:1 (g/g). Formation of peptidisc libraries was initiated by detergent dilution followed by filtration and concentration on a spin column with a cut-off of 100 kDa. The overall content of each library was then compared to the original detergent extract using SDS-PAGE and Coomassie blue staining (Figure 2A). Visual analysis indicated that most, if not all, of the proteins solubilized in detergent were present in the peptidisc library (Figure 2A). However, the best extraction of the marker protein MsbA was seen with DDM; therefore, this detergent was employed in the subsequent extraction studies.

Overview of the peptidisc-based SEC-PCP-SILAC workflow.

(A) Identical E. coli cultures are labeled in SILAC media (i), lysed with french press and crude membrane fraction isolated by ultracentrifugation (ii). Membranes are solubilized in non-ionic detergent (DDM) (iii), transferred into biotinylated Peptidisc peptides (Bio-Peptidisc) solution, and then filtered to remove excess peptide and detergent (iv). (B) The light and heavy peptidisc libraries are separated by high-resolution SEC in detergent-free buffer. The light fractions are pooled and aliquoted into the heavy fractions as an internal quantification standard. (C) Proteins in each fraction are denatured, depleted for Bio-Peptidisc peptides, digested, and analyzed by LC-MS/MS. Maxquant is used to identify peptides and to quantify heavy protein enrichment in each fraction. Binary protein interactions are identified from the co-elution data using the prediction of interactomes bioinformatics pipeline (PrinCE). Binary interactions are subsequently segregated into predicted complexes using the ClusterONE algorithm.

https://doi.org/10.7554/eLife.46615.002
The peptidisc captures detergent solubilized membrane proteins with high efficiency.

(A) SDS-PAGE analysis of detergent solubilized E. coli crude membrane before and after reconstitution into peptidiscs. The crude membrane preparation was solubilized in either 1% n-dodecyl-beta-maltoside (DDM), 3% ß-octyl glucoside (ß-OG), 1% sodium deoxycholate (DOC), or 1% lauryldimethylamine-N-oxide (LDAO), followed by reconstitution into peptidiscs by dilution and buffer exchange. (B) Protein number and overlap after SEC-fractionation of DDM extract and peptidiscs library prepared from DDM extract. A total of 20 fractions were collected, and the fraction containing the highest concentration of protein (fraction 12) analyzed by electrospray mass spectrometry in triplicate. The mass spectrometry data was searched together in MaxQuant. (C) SDS-PAGE analysis of native E. coli membranes incorporated into peptidisc after fractionation by size exclusion chromatography in detergent-free buffer. (D) As in C, with membranes solubilized in DDM and fractionated in buffer supplemented with DDM. (E) Clear native (CN)-PAGE analysis of crude membrane solubilized in DDM (Lane 1) or in peptidiscs (Lane 2). (F) The peptidisc library containing overexpressed MsbA (Lane 1) was bound to Ni-NTA beads, washed in Buffer A (Lane 2), and eluted in Buffer A + 250 mM imidazole (Lane 3). Samples were analysed by SDS-PAGE. (G) CN-PAGE analysis of MsbA purified in DDM (Lane 1) or purified in peptidiscs (Lane 2).

https://doi.org/10.7554/eLife.46615.003

We next assessed the amenability of our peptidisc library preparations to fractionation by size-exclusion chromatography (SEC). We compared the SEC fractionation profiles between the E. coli membrane proteome solubilized in DDM versus the E. coli membrane proteome trapped in peptidisc library (Figure 2B and Figure 2C). The overall protein profiles of each fractionation were comparable, as assessed by SDS-PAGE. The richest protein fraction from each (i.e. fraction #12; Figure 2B and Figure 2C) was analyzed by mass spectrometry. A total of 125 proteins and 162 proteins were identified from the detergent and peptidisc samples with ~85% overlap between the two, respectively (Figure 2D).

To verify that individual membrane proteins and complexes were trapped in discrete peptidisc particles - rather than being non-specifically clustered together - we isolated MsbA from the peptidisc library via a Ni-NTA pulldown. Analysis by SDS-PAGE revealed that his-tagged MsbA is efficiently isolated from the peptidisc library preparation (Figure 2F). Native PAGE analysis revealed that the purified MsbA is, as expected, homogenous (Figure 2G). These results strongly indicate that individual membrane proteins are efficiently trapped in discrete peptidiscs.

Fractionation of the SILAC-labeled peptidisc library

We next prepared light and heavy isotopically labeled membrane proteomes and reconstituted them into peptidisc libraries. Both libraries were separated by high-resolution size exclusion chromatography using two silica-based BioSep4000 columns placed in tandem. The light fractions were pooled and an equivalent volume was added to each heavy fraction before trypsin digestion and electrospray mass spectrometry analysis. For each fraction, proteolytic peptides were identified and SILAC ratios were determined using Maxquant. Two biological replicates were performed, resulting in the identification of 1209 proteins across the 54 fractions (raw data presented in Supplementary file 1). As expected, a large fraction of these proteins (591 proteins) are predicted to be associated with the cell envelope (Table 1). In addition to predicted cell envelope proteins, there were also soluble proteins which are known to associate into macromolecular complexes, such as the ribosome or GroEL complex. Previous reports have shown that these large assemblies are prone to co-sedimenting with the cell membrane fraction during the ultracentrifugation step after cell lysis, thus explaining their presence in our peptidisc library preparations (Papanastasiou et al., 2013; Papanastasiou et al., 2016).

Table 1
List of GO terms used to predict protein association with the E. coli cell envelope.
https://doi.org/10.7554/eLife.46615.004
Gene ontology term (Associated with cell envelope)
Anchored component of membraneAnchored component of external side of membraneAnchored component of periplasmic side of outer membraneExtrinsic component of periplasmic side of plasma membraneGram-negative bacterium cell wallExtrinsic component of plasma membrane
Integral component of membraneCell envelopeCell outer membraneIntegral component of cell outer membraneIntegral component of plasma membraneIntegral component of membrane
membraneCell wallExternal side of cell outer membraneIntrinsic component of cell outer membraneIntrinsic component membraneIntrinsic component of plasma membrane
Plasma membraneExtrinsic component of cell outer membraneExtrinsic component of membraneIntrinsic component of external side of plasma membraneIntrinsic component of periplasmic side of plasma membraneIntrinsic component of periplasmic side of cell outer membrane
Outer-membrane bounded periplasmic spacePeriplasmic spacePlasma membraneIntrinsic component of cytoplasmic side of plasma membraneOuter membrane
Periplasmic side of outer membranePeptidoglycan-based cell wall

To compare the peptidisc library with another detergent-free fractionation method, we solubilized the same E. coli membrane with the styrene maleic acid co-polymer (SMA). The SMA polymer directly solubilizes membrane lipids and captures proteins into styrene maleic acid lipid-protein nanoparticles (SMALPs)(Dörr et al., 2014). Following SEC fractionation of the SMALPs library and MS analysis, we identified 1576 proteins across 54 fractions (raw data presented in Supplementary file 2). There was good reproducibility between replicates and a large part of the identified proteins (705 proteins) was predicted to be associated with the cell envelope. The overlap in protein content between the SMALPs and peptidisc library was excellent, with 1026 proteins shared between the two libraries. Furthermore, the overall distribution of proteins in each library, as classified according to their gene ontology annotations and their originating compartment, was nearly identical (Figure 3A and Figure 3B). Thus, SMALPs and peptidiscs are both suitable for solubilization and detergent-free fractionation of the E. coli cell membrane. Importantly, the similar repertoire of proteins identified in both the SMALPs and peptidiscs libraries indicate that detergent solubilization followed by immediate reconstitution into peptidisc results in comparable capture efficiency of the membrane proteome as direct solubilization in the SMA polymer.

Figure 3 with 1 supplement see all
Proteomic analysis of soluble, SILAC-labeled E. coli membrane proteins in SMALPs or peptidisc libraries.

Gene ontology analysis of identified proteins and annotated cellular compartment of identified proteins in (A) SMALP library, or in (B) peptidisc library. Co-fractionation profiles for quantified subunits of the ATP synthase complex in (C) SMALPs (raw data presented in Supplementary file 2) or (D) peptidisc (raw data presented in Supplementary file 1). Note: total volume of column is 18 mL and void volume is 6 mL. Void volume is represented as the zero on the x-axis of all co-elution graphs.

https://doi.org/10.7554/eLife.46615.005

Large membrane protein complexes are captured in the peptidisc library

To determine if the peptidisc is able to capture membrane protein complexes, we compared the co-fractionation profiles of three well-characterized protein assemblies after encapsulating the membrane library in either SMALPs or peptidiscs (Figure 3C, Figure 3D and Figure 3—figure supplement 1). In all three cases, the complexes appeared more stable in the peptidisc library than in SMALPs. In the SMALPs library, the ATP synthase complex was dissociated, causing its protein subunits to elute separately (Figure 3C). In contrast, the ATP synthase was preserved in peptidiscs (Figure 3D). The Bam complex (BamABCDE) was solubilized with the SMA polymer, but here the elution profiles for the individual subunits showed weak correlation, suggesting at least partial dissociation of the complex (Figure 3—figure supplement 1). In contrast, each Bam subunit presented an almost identical co-fractionation pattern in peptidisc (Figure 3—figure supplement 1B). Similarly, the individual subunits of the respiratory chain complex (NuoABCDGI) showed a higher degree of correlation in their fractionation profiles in peptidisc versus SMALPs (Figure 3—figure supplement 1D and C). These results indicate that while SMA is an effective solubilization agent, it does not stabilize large membrane protein assemblies. The peptidisc is therefore better suited for stabilization of large, multiprotein complexes.

Prediction of binary interactions and the high confidence subset

Binary protein interactions in the peptidisc library were predicted using PrInCE (Predicting Interactomes via Co-Elution), a software designed for analyzing PCP-SILAC datasets (Stacey et al., 2017). PrInCE predicts which protein pairs are interacting or not according to the similarity or dissimilarity of their fractionation profiles (Figure 4A and Figure 4B, respectively). As described in detail in the method section, a naive Bayes classifier was trained (10-fold cross-validation) using multiple pairwise similarity measures based on either the entire co-fractionation profile (Pearson correlation, Euclidean distance) and whether proteins shared an elution peak. To refine prediction, we also incorporated a single measure of pairwise similarity from the M3D expression database (Faith et al., 2008), since this allowed the classifier to distinguish between true interacting protein pairs and pairs whose fractionation profiles were only spuriously similar. Training labels for the classifier were generated from the gold standard complexes, with ‘interacting’ label applied to protein pairs in the same gold standard complex and ‘non-interacting’ label applied to pairs present in the gold standard list but not in the same gold standard complex. Using this approach, we predicted a list of 4911 pairwise interactions (Figure 4C; Supplementary file 3). These interactions were between protein pairs with well-correlated co-fractionation profiles (average R = 0.78 vs R = 0.16, respectively; Pearson correlation). The predicted list also captures the majority of gold standard pairwise interactions (recall = 0.80, Figure 4D).

The peptidisc interactome is composed of 4911 co-fractionating protein pairs.

(A) Typical elution profiles of an interacting protein pair in the peptidisc interactome. (B) Example elution profiles of non-interacting proteins. (C) Precision vs. accumulated number of interactions. (D) Precision-recall curve of the peptidisc interactome.

https://doi.org/10.7554/eLife.46615.007

Using our predicted peptidisc interactome, we then generated a High Confidence subset of interactions based on two orthogonal high-throughput interactomes collected independently from this study (‘validating interactomes’) (Babu et al., 2018). These High Confidence interactions were detected by three independent experiments in two laboratories. From the predicted peptidisc interactome list (4911 interactions), we identified 824 interactions also present in the two validating interactomes (‘High Confidence’ subset, Supplementary file 6). The extent of overlap is significantly greater than the number of overlapping interactions expected by chance (p<0.001, permutation test, Figure 5—figure supplement 1A). Consistent with the fact that high-throughput techniques are often biased toward detecting certain protein complexes over others (Stacey et al., 2018), our predicted peptidisc interactome has greater overlap with the co-fractionation validating interactome than the AP/MS validating interactome, although both overlaps are significant (N = 2382 and 1623, respectively, p<0.001 and p<0.001, Figure 5—figure supplement 1B and C).

Computational validation of binary interactions

As false positives are inherent to high-throughput interactome studies, it is important to validate computationally that the interactome, on aggregate, indeed resembles a collection of true, biological interactors. To do so, we first used our gold standard protein complexes to calculate the ratio of TPs to FPs, measured as precision (TP / (TP + FP)) (Figure 4C and Figure 4D). However, the set of gold standard interactions in E. coli is relatively small compared to the set of gold standard interactions derived from mammalian studies (Rajagopala et al., 2014; Ruepp et al., 2008), meaning that any estimates based entirely on this gold standard set could be susceptible to noise. Therefore, to further estimate the biological plausibility of our peptidisc interactome, we determined whether predicted interacting pairs were more likely than non-interacting pairs to be enriched for three measures of biological association: shared GO terms, positively correlated stress phenotypes, and shared binding domains (Erickson et al., 2017; Mosca et al., 2014). Further, these enrichment values allowed us to benchmark our interactome against the E. coli cell envelope (CE) interactome recently published by Babu et al. (2018).

Using these measures of biological plausibility, we found the peptidisc interactome to be enriched compared to random chance, and as shown in Figure 5A–5C, the enrichment is significant for all three measures. Converting these measures to z-score, that is measuring relative to randomly rewired networks, our peptidisc interactome tended to be more enriched than the CE interactome (GO: z = 15.6 vs z = 14.2, peptidisc vs CE, respectively; Tolerome: z = 24.8 vs z = 6.2; 3did: z = 8.7 vs z = 6.3). We also calculated these three measures for our High Confidence set of interactions (Figure 5, black circles) and found that the High Confidence set was more enriched than the full peptidisc interactome for shared GO terms and positively correlated Tolerome profiles. The opposite was true for shared binding domains (Figure 5C), perhaps because of the sparsity of shared terms: only 24/824 interactions shared a binding domain in the High Confidence set (162/4911 full peptidisc interactome), indicating a noisier measure of biological association. In addition to benchmarking our enrichment values (GO, Tolerome, binding domains) against the CE interactome, we also confirmed that a significant number of interactions were common between the CE and peptidisc interactomes. Of the 4911 peptidisc interactions, 340 are also present in the CE interactome, a significant overlap (p<0.001, permutation test) (Figure 5D, Supplementary file 6). As expected, interactions that overlap with the CE interactome tend to be higher scoring than non-overlapping interactions (average interaction score 0.66 vs 0.62, respectively; p=5e-11, Wilcoxon rank-sum test; Figure 4C).

Figure 5 with 1 supplement see all
Computational validation of peptidisc interactome.

(A) Fraction of interacting pairs sharing a gene ontology (GO) term for the peptidisc (top) and CE interactomes (bottom). Both the full peptidisc interactome (4911 interactions, white) and the High Confidence subset are shown (black). ‘Random’ shows the expected number of shared terms from randomly rewired peptidisc and CE interactomes (1000 iterations, gray bars). (B) Fraction of interacting pairs with positively correlated Tolerome profiles (R > 0, Pearson correlation). (C) Fraction of interacting pairs sharing binding domains. (D) Number of overlapping interactions between peptidisc and CE interactomes compared to random. (E) True (‘data’) and random distributions for M3D co-expression correlation (Pearson) for peptidisc, High Confidence, and CE. Random distributions generated by randomly rewiring networks.

https://doi.org/10.7554/eLife.46615.008

Finally, we confirmed that protein pairs in our peptidisc interactome had better-than-random M3D expression profile correlation (Faith et al., 2008) (Figure 5E). This is to be expected, since M3D expression correlation was used as a feature in our machine learning classifier (see Materials and methods), meaning high M3D correlation was a criterion on which our peptidisc interactions were selected. However, we also note that protein pairs in our peptidisc interactome had higher expression than protein pairs in the E. coli CE interactome (Babu et al., 2018). Therefore, as expected, protein pairs in the peptidisc interactome were well-correlated as measured by M3D, and the level of correlation is higher than the benchmark CE interactome.

Computational assignment and validation of protein complexes

We used a two-stage algorithm to cluster the identified pairwise interactions into complexes (Wan et al., 2015; Drew et al., 2017). A first stage clustering was performed using ClusterONE, an algorithm that allows moonlighting proteins present in multiple protein complexes (Nepusz et al., 2012). However, because ClusterONE tends to collapse biologically distinct protein groups into the same protein complex, we performed a second stage refinement using the MCL algorithm (Enright et al., 2002). The combination of these two algorithms ensured that the same protein can be assigned to multiple complexes. In addition, since both ClusterONE and MCL have tunable parameters, we performed a grid search optimization to find the parameter set which maximizes the matching ratio value between predicted complexes and our set of gold standard complexes. This procedure produced 202 complexes with a median size of five proteins (Supplementary file 4). As for the pairwise interactions above, we employed GO terms as an evidence for biologically meaningful complexes, and we reported that 36 of the 202 complexes were significantly enriched for at least one GO term (hypergeometric test, Benjamini-Hochberg-corrected p<0.05), a significant number (p<0.001, permutation test). Because clustering method removes pairwise interactions that are inconsistent with the predicted complexes, the subset of pairwise interactions clustered into complexes should be scoring higher than not clustered pairwise interactions (Drew et al., 2017). This was indeed the case: the 3490 pairwise interactions clustered in complexes had a significantly higher interaction score than the 1421 un-clustered interactions (mean interaction score 0.64 vs 0.59, p=3e-72, Wilcoxon rank-sum test).

Experimental validation of binary interactions by affinity purification mass spectrometry (AP/MS)

Parallel to the in silico validation described above, we performed a series of in vitro experiments using three different membrane protein systems - the Sec translocon, the Bam complex and the ABC transporter MetNI. The goal was to use AP/MS to confirm and potentially discover novel pairwise associations predicted from interactome datalist. Principally, we aimed to confirm interactions between the core SecYEG complex and the membrane-anchored periplasmic chaperones YfgM and PpiD. These interactions are detected in our datalist at high (>75%) precision. However, this association is difficult to detect in detergent, unless all subunits are simultaneously over-produced in the membrane (Figure 6—figure supplement 1). We were also interested by the astonishing apparent interaction between the Sec and Bam complexes, also given in our datalist at high (>75%) precision. These include interactions between the SecY complex and the BamA, BamC and BamD subunits of the Bam complex. To perform these validation AP/MS experiments, the his-tagged SecYEG complex was expressed in SILAC labeling conditions. The membrane fraction was briefly solubilized with detergent followed by immediate trapping in peptidisc library. The SecY complex was subsequently isolated by Ni-NTA and the co-isolated proteins were identified by LC-MS/MS. To measure protein enrichment and to control for non-specific co-purifying background contaminants, the pulldown experiments were performed in parallel using a detergent extract or peptidisc library prepared from cells transformed with the empty vector.

Experiments with the detergent extract shows that SecY is highly enriched after affinity pulldown (Figure 6A, raw data presented in Supplementary file 5). There is also enrichment of several ribosomal proteins, which is not surprising given the intrinsic affinity of ribosomes for the Sec complex (Rapoport et al., 2017; Park and Rapoport, 2012). However, many known membrane-bound interactors of the SecY complex are not enriched, likely due to their dissociation during the prolonged incubation with detergent. Accordingly, in peptidisc, several ancillary subunits of the Sec translocon, including the periplasmic chaperones PpiD and YfgM, as well as the holo-translocon subunits YidC and YajC are detected. There is also strong enrichment of the porin OmpA in addition to several subunits of the outer membrane Bam complex. The BamB and BamC subunits in particular are significantly enriched (Figure 6B). The dynamics of a Sec-Bam complex interactions awaits further experimentation, but as it is this series of evidences validates the ability of the peptidisc PCP-SILAC and AP/MS workflows to capture novel protein assemblies that are difficult to isolate in detergent.

Figure 6 with 1 supplement see all
Validation of SecYEG interactors by AP/MS.

(A) Enrichment matrix of each quantified protein identified in the SecYEG detergent AP/MS pulldown. The Log(2) peptide intensity for each quantified protein is plotted against the corresponding Log(2) SILAC ratio. Arbitrary enrichment cutoffs were set for both the x and y axes; these are indicated on the plot as dashed lines to aid the eye. Proteins of interest are highlighted in red. The number of unique peptides detected for each protein of interest is given in parentheses. Each black dot is a protein quantified in the pulldown experiments (B) As in A, but for proteins quantified in the SecYEG peptidisc library pulldown. Raw data for both plots are presented in Supplementary file 5.

https://doi.org/10.7554/eLife.46615.010

Identification of Bam complex interactors

We next applied the AP/MS workflow towards the protein BamA - the major subunit of the outer membrane-embedded Bam complex. Our interactome dataset identified the Sec ancillary subunits YidC and YajC as potential Bam interactors with high precision (>75%), as well as the cell surface-exposed lipoprotein RcsF (Supplementary file 3, Figure 7A). To explore the validity of these predicted interactions, we expressed his-tagged BamA in SILAC labeling conditions and analyzed the peptidisc library or detergent extract using the AP/MS workflow described above.

Validation of BamA interactors by AP/MS.

(A) Pair-wise co-elution plots of select BamA interactors as predicted from the peptidisc PCP-SILAC workflow (raw data presented in Supplementary file 1). Pairwise interaction correlation values are shown above each plot. (B) Enrichment matrix of each quantified protein quantified in the BamA detergent AP/MS pulldown. The data was plotted and labeled as in Figure 6. (C) As in A, but for proteins quantified in the BamA peptidisc AP/MS pulldown. Raw data for the plots shown in B and C is presented in Supplementary file 5.

https://doi.org/10.7554/eLife.46615.012

Experiments with the detergent extract shows that BamC and BamD are the only subunits enriched along with BamA (Figure 7B, raw data presented in Supplementary file 5). This finding is in agreement with an earlier study which showed that BamB is prone to dissociating from the rest of the complex in detergent solution (Gu et al., 2016). The only other interactor that is significantly enriched is RcsF. In peptidisc, by contrast, all four other subunits of the Bam complex (subunits B, C, D and E) are captured along with BamA (Figure 7C). Additionally, there is again significant enrichment of the Sec translocon ancillary subunits YidC and YajC, thereby providing additional evidence to support this potentially novel interaction. The lipoprotein RcsF and the porin OmpA are also significantly present (Figure 7C). We note that BamA-OmpA interaction was not reported in our interactome datalist, probably due to the unusually broad SEC elution profile of OmpA, leading to false negative identification due to low score precision. However, a series of recent publications have shown that OmpA is a bona fide interactor of both RcsF and BamA in the cell context (Hart et al., 2019; Konovalova et al., 2014).

Identification of a unique type I ABC transporter complex

The ABC transporter MetNI was chosen as a third validation target because its PCP profile allowed to predict that i) MetQ forms a stable complex with MetNI and ii) NlpA (lipoprotein 28) is a novel interactor of the transporter (Figure 8). The stable interaction of MetQ with MetNI is unexpected because type I ABC transporters are characterized by weak affinity to their substrate binding proteins (SBPs). Accordingly, the transporters LivFGM and HisQP co-elute separately from the SBPs LivJ, LivK, and HisJ, respectively (Figure 8—figure supplement 1. Raw data presented in Supplementary file 1). In contrast, but as expected, the type II transporter FepC co-elutes tightly with FepB (Bao and Duong, 2012; Bao and Duong, 2014; Hvorup et al., 2007; Rice et al., 2014) (Figure 8—figure supplement 1). As an additional distinctive feature, we discovered the presence of lipobox at the N-terminus of MetQ, which is a unique case among other type I SBPs. This lipid modification could explain why MetQ remains associated to MetNI in the peptidisc library, and why this association has not been detected in earlier biochemical studies which employ the soluble, non-lipidated form of MetQ (Nguyen, 2016; Nguyen et al., 2018; Nguyen et al., 2015).

Figure 8 with 2 supplements see all
Validation of MetNI interactors of AP/MS.

(A) Co-elution plot of MetN with the predicted interactors MetQ and NlpA (raw data presented in Supplementary file 1). Pairwise correlation values for each interaction are shown above the plot. (B) Enrichment matrix of each quantified protein quantified in the MetNI detergent AP/MS pulldown. The data was plotted and labeled as in Figure 6. (C) As in A, but for proteins quantified in the MetNI peptidisc AP/MS pulldown. Raw data for the plots shown in B and C is presented in Supplementary file 5.

https://doi.org/10.7554/eLife.46615.013

To validate these predictions, we expressed his-tagged MetNI in SILAC labeling conditions. As above, we isolated the MetNI complex in both detergent and peptidisc, and identified the co-purifying interactors using LC-MS/MS. The results reveal the interaction of MetQ with MetNI in both detergent and peptidisc (Figure 8 - raw data presented in Supplementary file 5). The interaction of NlpA with MetNI is, however, detected only in peptidisc, and not in detergent (Figure 8B and C). We next tested the importance of MetQ lipidation using a mutant carrying a disrupted lipobox, MetQ-C23A. The protein MetQ and MetQ-C23A were co-expressed with MetNI, the membrane fraction was solubilized in detergent and reconstituted in peptidisc. The his-tagged MetNI complex was subsequently isolated by Ni-NTA chromatography. The results indicate the co-elution of lipidated MetQ with MetNI but not MetQ-C23A, underscoring the importance of lipidation for this interaction (Figure 8—figure supplement 2).

Discussion

Despite important progress in the field of proteomics, characterization of membrane interactomes has lagged due to the reliance on surfactants to maintain protein solubility. In this study, we show that the peptidisc can entrap membrane proteins directly out of a crude detergent membrane extract. The resulting protein library is stable, water soluble and is amenable to biochemical fractionation and characterization by mass spectrometry. There are distinctive advantages using peptidisc instead of detergents or other membrane mimetics in this experimental workflow. Because of its adaptability to membrane protein hydrophobic surfaces, the peptidisc is less likely compared to the nanodisc to bias reconstitution toward a certain protein diameter (Bayburt et al., 2006). Addition of exogenous lipids is also not required, which simplifies the reconstitution while reducing protein aggregation caused by extra lipids (Roy et al., 2015; Wilcox et al., 2015; Marty et al., 2013). The self-assembly is rapid and does not require dialysis or detergent adsorbents. This relative short exposure to surfactants is important as it minimizes detergent-mediated complex dissociation and aggregation. We further discuss the advantages and limitations of the peptidisc-proteomic workflow below, along with our initial biological findings using the E. coli membrane proteome as a model.

To compare the peptidisc against another membrane mimetic such as the SMA polymer, we measured the stability of three large membrane protein complexes, the ATP synthase, the Bam assembly and the respiratory chain I complex. Initial results revealed a high degree of overlap between the protein content of the libraries stabilized in peptidisc or in SMALPs (85.5% similarity), indicating that both methods effectively solubilize the membrane proteome (Figure 3). However, while the three large complexes listed above were largely intact in the peptidisc library, all three were significantly dissociated in the SMALPs library (Figure 3, Figure 3—figure supplement 1). Thus, the peptidisc method is demonstrably superior for capturing and stabilizing multi-subunit membrane protein complexes.

To determine the precision of the protein correlation profile (PCP) obtained with the peptidisc library, we had to separate a relatively small number of well-correlated protein pairs from a much larger background of non-interacting proteins. We used PrInCE, a machine learning bioinformatics tool written specifically for analyzing co-fractionation datasets (Stacey et al., 2017). This generated 4911 predicted interactions at 50% precision out of >700,000 potential random interactions. Like in every proteomic-based discovery method, we note the importance to benchmark the predicted interaction dataset against other databases (validating interactomes). As expected, our dataset was significantly enriched in multiple indicators of association, including gene ontology annotations, shared binding domains, and correlation of shared growth phenotypes. These indicators compare favorably with validating interactomes identified by low-throughput AP/MS, which generally have relatively few false positives (Figure 5; Babu et al., 2018). Importantly, a significant subset of interactions in our peptidisc interactome were also found within two other E. coli interactomes collected independently in a separate study (Babu et al., 2018). The integration of our peptidisc interactome with these validating interactomes lead to the ‘High Confidence’ dataset (Supplementary file 6). The number of interactions in this subset is significantly greater than that expected by chance (p<0.001, Figure 5—figure supplement 1), which supports the validity of the peptidisc workflow.

Parallel to this computational validation, we performed experiments to verify the ability of the peptidisc workflow to reveal novel or transient interactions. Our first validation target was the SecYEG complex. The pairwise peptidisc interaction list indicate a number of interactors at a high degree of precision (>75%; Supplementary file 3), such as the membrane-bound chaperones YfgM and PpiD. Interaction of SecY with these chaperones has been reported in the past using low-throughput 2D gel electrophoresis and cumbersome radio-labeling (Götzke et al., 2014). In our study, we show that combining the SecYEG complex trapped in peptidisc with the SILAC AP/MS workflow lead the facile detection of this interaction (Figure 6). In contrast, this association is hardly seen in detergent unless all subunits are simultaneously over-produced in the membrane (Figure 6—figure supplement 1), which suggests that a proper subunit stoichiometry is critical for complex formation. This later observation highlights the importance of characterizing membrane protein networks of association under native expression conditions with minimal genetic manipulation.

Strikingly, the peptidisc interaction datalist also revealed a network of pairwise interactions between subunits of the Sec and Bam complexes. Interactions between SecYEG and the BamA, BamC and BamD subunits in particular are present in our interaction list at high precision. This Sec-Bam interaction has been suggested previously based on coincident detection between membrane fractionation and Western blotting (Wang et al., 2016). In order to confirm these interactions, the SecYEG complex was trapped in peptidisc and analyzed following the SILAC AP/MS workflow (Figure 6B). There is significant enrichment of the BamB and BamC subunits and these interactions are not detected when the same workflow is performed in detergent (Figure 6A). To obtain additional evidence for the Sec-Bam association, we employed BamA trapped in peptidisc as a bait using the same SILAC AP/MS workflow (Figure 7). Here also the results reveal compelling enrichment of the Sec ancillary subunits YidC and YajC, adding another layer of evidence for a Sec-Bam interaction. This interaction is particularly intriguing - and exciting - as it suggests a trans-membrane pathway for the direct transfer of proteins from the inner to outer membrane of the bacterial cell envelope. Further biochemical work is required to characterize this super-complex, but exciting new data already indicates the possibility to isolate the Sec/Bam assembly for structural analysis (http://dx.doi.org/10.1101/589077).

Other BamA interactors identified in the peptidisc-PCP-SILAC workflow were RcsF and OmpA (Figure 7B). The fact that the peptidisc is able to preserve these native interactions highlights the ability of the method to capture interactions that naturally exist in the cell membrane. Of note, the complex between RcsF, OmpA and BamA - which represents a novel mechanism of lipoprotein translocation to the extracellular side of the outer membrane - was initially revealed from genetic and in vivo cross-linking experiments, but never formally demonstrated by biochemical means (Hart et al., 2019; Konovalova et al., 2014). Accordingly, we find these interactions are much less apparent when the experiment was performed in detergent (Figure 7A).

The third validation target in this study is the transporter MetNI. We demonstrate that MetQ forms an unexpectedly stable complex with MetNI due of its N-terminal lipid anchor (Figure 8, Figure 8—figure supplement 2). The importance of MetQ lipidation in mediating this interaction has been overlooked in the literature to date and appears unique among amongst the 48 SBPs present in E. coli. This novel finding therefore calls for a re-evaluation of the role of the MetQ lipid anchor for complex stability and substrate transport. As an additional original discovery based on the peptidisc interactome data list, we identify and demonstrate the interaction of the MetNI complex with the lipoprotein NlpA (Figure 8). Earlier literature has reported that NlpA overproduction can complement a strain depleted for MetQ (Zhang et al., 2003), thereby providing a strong biological rationale to this interaction.

In conclusion, the peptidisc library combined with PCP-SILAC or AP/MS workflow is a promising novel approach for generating and validating high-throughput membrane protein interactomes. As the interaction list published here is intended to be a tool for future research, we provide two ways for researchers to narrow down the interactions to smaller subsets of interactions with fewer false positives. First, each interaction is given with an interaction score (Supplementary file 3), a measure which has been shown to correlate with measures of biological plausibility (Stacey et al., 2017). Second, the High Confidence datalist of interactions provides researchers with interactions that have been orthogonally validated (Supplementary file 6) (Babu et al., 2018). Further improvement of the precision to reduce the number of false positives is feasible by increasing the library fractionation, including for example separation over density gradient or ion exchange resins (McBride et al., 2017; Maddalo et al., 2011). However, this practice can also bias predicted interactions toward protein pairs that are already supported in the literature at the expense of detecting novel interactions. Looking forward, the peptidisc workflow can also be expanded to comparative analysis of membrane interactomes using a third amino acid isotopologue label. This labeling would allow profiling the changes in the global membrane protein interaction landscape under different conditions or in response to drugs and mutations.

Materials and methods

Reagents and plasmids

Request a detailed protocol

Tryptone, yeast extract, Na2HPO4, KH2PO4, NaCl, imidazole, Tris-base, acrylamide 40%, bis-acrylamide 2% and TEMED were obtained from Bioshop Canada. Amino acid isotopologues were purchased from Cambridge Isotope Laboratories. Isopropyl β-D-1-thiogalactopyranoside (IPTG), ampicillin, kanamycin, and arabinose were purchased from GoldBio. Detergents n-dodecyl-β-d-maltoside (DDM) and octyl-β-D-glucoside (β-OG) were from Anatrace. Detergent N,N-dimethyldodecylamine N-oxide (LDAO) was from Sigma. Columns Biosep 4000 GFC/SEC were purchased from Phenomenex. Ni2+-NTA chelating Sepharose was obtained from Qiagen. Peptidisc peptides and biotinylated derivative Bio-Peptidisc (purity >90%) were obtained from Peptidisc Biotech Canada. All other chemicals were obtained from Fisher Scientific Canada. The genes yfgM or ppiD were inserted into pBAD33 encoding for a C-terminal 6x his-tag via the polymerase incomplete primer extension (PIPE) method (Klock and Lesley, 2009) to form pBad33-yfgM and pBad33-ppiD, respectively. To create pBad33-YfgM-PpiD, the sequence encoding for PpiD without a 6x his-tag was amplified and inserted into pBad33-YfgM using the PIPE method. The gene BamA was inserted into pBad22 by the PIPE method and 6x his-tag was subsequently inserted at the N-terminus of the mature protein. The plasmids pET19-hisMetNI and pET21-MetQhiswere gifts from Dr. Janet Yang (University of San Francisco). Those plasmids were employed to construct pBad33-MetQ and pBad22-hisMetNI. The metQ gene was amplified from pET21-MetQhis and inserted into pBAD33. Plasmid pBAD22-hisMetNI was constructed by sequentially inserting the metN and metI genes from pET19-hisMetNI into pBAD22. The MetQ C23A mutation was inserted into pBad33-MetQ by site-directed mutagenesis. All construct sequences were confirmed by Sanger sequencing (Genewiz). The gene msbA was inserted with a sequence encoding for a N-terminal 6x his-tag into the vector pET28 to form the plasmid pET28-msbA. The plasmids pBad22-HA-EYG and pBad22-his-EYG have been previously described (Tam et al., 2005; Maillard et al., 2007; Young and Duong, 2019).

Preparation of SILAC labeled E. coli

Request a detailed protocol

For preparation of heavy and light labeled crude membranes, E. coli strain JW2806 (ΔlysA763::kan) was labeled with Lys4 (2H4-lysine), as previously described (Zhang et al., 2012). Cells were picked from a single colony and grown overnight in 5 mL of LB + 25 µg/mL kanamycin at 37°C. The overnight culture was isolated by low-speed centrifugation (5000 x g, 6 min), and resuspended in an equivalent volume of M9 minimal media. The culture was pelleted and washed two more times to ensure full removal of residual LB media. Unless otherwise stated, the cells were subsequently diluted 1/2000 into two flasks containing 250 mL M9 minimal media + 0.1% glucose + 100 µg/mL thiamine. The flasks were supplemented with either 0.06 mg/mL lysine or 0.06 mg/mL Lys4 to form the light- and heavy-labeled cultures, respectively. A control culture without supplemented amino acid was also inoculated but no growth was detected due to the inability of JW2806 to produce the essential amino acid lysine. Cells were grown at 37°C for 16 hr until OD reached ≅ 0.9–1.1.

Optimization of the peptidisc library method using the model membrane protein MsbA

Request a detailed protocol

Plasmid pET28-hisMsbA was expressed in E. coli BL21(DE3) (New England Biolabs) for 3 hr at 37°C after induction with 0.5 mM IPTG at an OD of 0.4–0.7 in LB medium supplemented with 25 µg/mL kanamycin. Cells were harvested by low speed centrifugation (10,000 x g, 6 min) and resuspended in SEC buffer (50 mM Tris-HCl, pH 7.2; 50 mM Na-acetate; 50 mM K-acetate). Resuspended cells were treated with 1 mM phenylmethylsulfonyl fluoride (PMSF) and lysed using a french press at 10,000 psi. Unbroken cell debris and other aggregates were removed by an additional low-speed centrifugation. The crude membrane fraction containing overexpressed MsbA was subsequently isolated by ultracentrifugation (100,000 x g, 45 min). The crude membrane fraction was resuspended in SEC Buffer to a concentration of ~20 mg/mL. To screen different detergents, aliquots of MsbA-containing crude membranes were solubilized in either 1% DDM, 3% β-OG, 1% DOC or 1% LDAO. Solubilizations were performed at 4°C for 1 hr. Insoluble material was then pelleted by ultracentrifugation (100,000 x g, 15 min). The detergent-solubilized extracts were subsequently trapped in peptidisc libraries as described below. To compare the efficiency of library capture between the different detergents assayed, aliquots of each detergent extract and each resultant peptidisc library were analyzed side by side on 15% SDS-PAGE followed by Coomassie Blue staining. Peptidisc-MsbA was subsequently isolated by Ni2+-chelating chromatography in SEC buffer, washed in 10 column volumes (CV) of Wash Buffer (20 mM Tris-HCl: pH 7.1; 50 mM K-acetate; 50 mM Na-acetate; 15 mM imidazole), and then eluted in ½ CV Elution Buffer (20 mM Tris-HCl: pH 7.1; 50 mM K-acetate; 50 mM Na-acetate; 400 mM imidazole). For purification of MsbA in DDM, the procedure was repeated except there was no addition of Peptidisc peptide to the solubilized crude membrane and a concentration of 0.02% DDM was maintained in all buffers during the dilution and purification steps.

Incorporation of E. coli cell envelope proteins in peptidisc library

Request a detailed protocol

Cells grown to OD ~0.9–1.2 were pelleted by low-speed centrifugation (5000 x g, 6 min) and resuspended in 2 mL Buffer A. Cells were lysed by French press (10,000 psi, two passages) and cell debris removed by an additional low-speed centrifugation step (10,000 x g for 10 min). Crude membrane was isolated by ultracentrifugation (100,000 x g, 45 min), and resuspended in Buffer A to a protein concentration of 20 mg/mL. The crude membrane was solubilized in 0.8% DDM, and isolated by ultracentrifugation. Solubilized crude membrane (100 µL at 10 mg/mL) was mixed with the Bio-Peptidisc peptide (350 µL at 6 mg/mL), and the mixture diluted to 10 mL ([DDM] ≅ 0.008%). The mixture was concentrated over a 100 kDa cut-off polysulfone filter (Sarstedt) to 500 µL, then diluted again to 5 mL in Buffer A ([DDM] ≅ 0.0008%). The library was concentrated to 250 µL ([Total protein] ≅ 6 mg/mL) and left on ice until fractionation. For pull-down experiments, the libraries were concentrated to ≅1 mg/mL and placed on ice until subsequent use.

Incorporation of E. coli cell envelope proteins in SMALP

Request a detailed protocol

The SMA polymer containing 2:1 styrene to maleic acid ratio was prepared following the procedure described in reference (Dörr et al., 2014). In brief, 10% of SMA 2000 (Cray Valley), was refluxed for 3 hr at 80°C in 1M KOH, resulting in complete solubilization of the polymer. Polymer was then precipitated by dropwise addition of 6M HCl accompanied by stirring and pelleted by centrifugation (1500 x g for 5 min). The pellet was then washed 3 times with 50 mL of 25 mM HCl, followed by a third wash in ultrapure water and subsequent lyophilization. SMA (pre- and post-hydrolysis) was analyzed by Fourier Transform-Infrared Spectroscopy (FT-IR) to confirm full hydrolysis of the anhydride group. The hydrolyzed SMA was later re-suspended at 10% wt/vol in 25 mM Tris-HCl, and the pH of the solution adjusted to 8.0 with 1M NaOH. Cells were pelleted by low-speed centrifugation (5000 x g, 6 min), and resuspended in 2 mL SEC Buffer. Cells were lysed by french press (10,000 psi, two pass) and cell debris removed by an additional low-speed centrifugation step (10,000 x g, 10 min). Crude membrane was isolated by ultracentrifugation (100,000 x g, 45 min), and resuspended in SEC Buffer to a protein concentration of 20 mg/mL. Crude membranes were solubilized by addition of 3% SMA2000 for 1 hr at 4°C, clarified by ultracentrifugation (100,000 x g, 15 min, 4°C), then placed on ice until subsequent use.

Fractionation of cell envelope libraries and digestion of protein samples

Request a detailed protocol

Cell envelope protein libraries were fractionated by size exclusion chromatography as previously described (Kristensen et al., 2012; Zhang et al., 2012; Scott et al., 2017). In brief, 200 µL of prepared libraries were separated over two tandem BioSep4000 columns (Phenomenex) pre-equilibrated in SEC buffer at 8°C. At an isocratic flow of 0.5 mL/min, fractions were collected from 20 min to 44 min. After fractionation - where applicable - detergent was first removed from protein samples by acetone precipitation. In brief, protein sample was mixed with 80% ice cold acetone, then left overnight on ice to precipitate. The precipitated proteins were pelleted by low-speed centrifugation (10,000 x g, 10 min, 4°C), washed with an equivalent volume of ice cold, 100% acetone and pelleted again (10,000 x g, 10 min, 4°C). The supernatant was aspirated away and pellet air-dried at 42°C for 10 min before storage at −20°C until digestion. For peptidisc libraries, detergent was removed during peptidisc assembly, so no acetone precipitation was necessary. We used a modified protocol to digest protein samples into tryptic peptides (Scott et al., 2017). In brief, samples were first denatured in 6M urea. When Bio-Peptidisc peptide was present in the sample, the denatured proteins were incubated with streptavidin coated agarose beads (2 µg beads/µl pre-washed in SEC Buffer) for 30 min at 25°C and the supernatant removed to deplete the peptide. Denatured proteins were incubated with 5 mM DTT for 1 hr at 25°C to reduce any cysteines. Free cysteines were alkylated by addition of 20 mM iodoacetamide for 1 hr at 25°C in the dark, the reaction was then quenched by addition of 40 mM DTT. Samples were pre-cleaved by addition of 0.1 µg Lys-C for 1.5 hr at 25°C, followed by dilution to 1 M urea in 50 mM ammonium bicarbonate, pH 8.3. Proteomics grade trypsin (1 µg; Promega) was added to each sample, and the reactions left to digest overnight at 25°C. Digested samples were acidified to <pH 2.5 by addition of 1% trifluoroacetic acid and the resulting peptide supernatant purified using self‐made Stop‐and‐go‐extraction tips (StageTips) composed of C18 Empore material (3M) packed in to 200 μl pipette tips (Ishihama et al., 2006; Rappsilber et al., 2003; Rappsilber et al., 2007). Prior to addition of the peptide solution, StageTips were conditioned with methanol and equilibrated with 0.5% acetic acid (Buffer A3). Peptide supernatants were loaded onto columns and washed with three bed volumes of Buffer A3. Peptide samples were eluted with 80% acetonitrile, 0.5% acetic acid (Buffer B3) into microfuge tubes, dried down using a vacuum concentrator, and stored at −20°C.

Expression of the validation targets SecYEG, MetNI and BamA in SILAC labeling conditions

Request a detailed protocol

Plasmids pBad22, pBad22-hisEYG, pBad22-hisMetNI and pBad22-hisBamA were chemically transformed into E. coli JW2806. Cells were grown overnight in M9 media supplemented with either 0.3 mg/mL Lys4 (for pBad22-hisEYG, pBad22-hisMetNI and pBad22-hisBamA) or 0.3 mg/mL light Lysine (for pBAD22). The next morning, the cultures were diluted 1/100 into fresh M9 media supplemented with either Lys4 or light lysine. Protein expression was induced with 0.02% arabinose once the cells had reached OD ~0.4–0.6. The cultures were then shifted to 25°C and grown overnight. Cells were harvested and resuspended in TSG buffer containing 1 mM PMSF before being lysed on a French Press (8000 psi, three passes). The membrane fraction was collected and resuspended in TSG (50 mM Tris HCl pH 8; 50 mM NaCl; 10% glycerol) buffer, rather than in SEC buffer. Membranes were solubilized in 0.5% DDM for 15 min on ice. Solubilized material was clarified by ultracentrifugation (100,000 x g, 15 min). An aliquot of the detergent-solubilized material was purified as described above, except that all steps contained TSG buffer, with DDM and imidazole where necessary. The remainder of the detergent-soluble supernatant was mixed with a 4:1 excess of Peptidisc peptide and peptidisc libraries were prepared by the dilution and concentration method described above. The resultant library (~1 mL at 1 mg/mL) was purified as described above for MsbA, except that all steps contained TSG buffer (with imidazole when necessary), not SEC buffer. Eluted proteins were analyzed by 15% SDS-PAGE and visualized by Coomassie Blue staining. For mass spectrometry analysis, the detergent-purified ‘heavy’ and ‘light’ elutions fractions were pooled and acetone precipitated before being digested with trypsin and LysC. The samples were then STAGE tipped and analyzed by mass spectrometry. The ‘heavy’ and ‘light’ peptidisc elution fractions were pooled, denatured with 6M urea and digested with trypsin and LysC before STAGE tipping and analysis by mass spectrometry.

Validation of the YfgM-PpiD-SecYEG interaction

Request a detailed protocol

Plasmids pBad33-YfgM, pBad33-PpiD, and pBad33-YfgM PpiD were transformed into chemically competent BL21DE3 cells harboring the plasmid pBad22-HA-EYG. For expression of his-tagged SecYEG complex only, plasmid pBad22-his-EYG was transformed into BL21DE3 cells. Overnight cultures were prepared in LB media supplemented with appropriate antibiotics at the concentrations specified above. After an overnight incubation, the cultures were diluted 1:100 into fresh LB media with antibiotics. Protein expression was induced at OD ~0.4–0.6 by addition of 0.1% arabinose, and cultures were grown for a further 2 hr before harvesting. Cells were resuspended in TSG buffer before being lysed as described above. Membranes were prepared in TSG buffer and solubilized as described above. Solubilized material was clarified by ultracentrifugation (100,000 x g, 15 min) before incubation for 30 min with Ni-NTA affinity resin. Beads were washed in 10 CV TSG buffer + 0.02% DDM, then eluted in TSG buffer + 300 mM imidazole + 0.02% DDM. Eluted proteins were analyzed by 15% SDS-PAGE followed by either Coomassie Blue staining or a western blot using a SecY-specific antibody as previously described (Dalal and Duong, 2009; Dalal et al., 2012).

Validation of the MetNI-Q interaction

Request a detailed protocol

Plasmids pBad33-MetQ and pBad33-MetQ C23A were transformed into chemically competent BL21DE3 cells containing the plasmid pET21-hisMetNI. Overnight cultures were prepared in LB media supplemented with appropriate antibiotics. After an overnight incubation, the cultures were diluted 1:100 into fresh LB media with antibiotics. Protein expression was induced at OD ~0.4–0.6 by addition of 0.1% arabinose and 1 mM IPTG, and cultures were grown for a further 2 hr before harvesting. Cells were resuspended in TSG buffer (50 mM Tris-HCl pH 8; 50 mM NaCl; 10% glycerol), and the membrane fraction was prepared and solubilized as described above. The solubilized material (~1 mL) was mixed with a 4:1 excess of Peptidisc peptide, and peptidisc libraries were prepared by the dilution and concentration method described above. The resultant library (~1 mL at 1 mg/mL) was incubated with Ni-NTA resin for 30 min at 4°C. The resin was washed with 10 CV of TSG buffer, then eluted in TSG buffer + 300 mM Imidazole. Eluted proteins were analyzed by 15% SDS PAGE and visualized by Coomassie Blue staining.

Liquid chromatography and mass spectrometry analysis

Request a detailed protocol

Purified peptides were analyzed using a quadrupole – time of flight mass spectrometer (Impact II; Bruker Daltonics) on-line coupled to an Easy nano LC 1000 HPLC (ThermoFisher Scientific) using a Captive spray nanospray ionization source (Bruker Daltonics) including a 2-cm-long, 100 μm-inner diameter fused silica fritted trap column, 75 μm-inner diameter fused silica analytical column with an integrated spray tip (6–8 μm diameter opening, pulled on a P-2000 laser puller from Sutter Instruments). The trap column is packed with 5 μm Aqua C-18 beads (Phenomenex, www.phenomenex.com) while the analytical column is packed with 1.9 μm-diameter Reprosil-Pur C-18-AQ beads (Dr. Maisch, www.Dr-Maisch.com). Buffer A consisted of 0.1% aqueous formic acid in water, and buffer B consisted of 0.1% formic acid in acetonitrile. Samples were resuspended in buffer A and loaded with the same buffer. Standard 45 min gradients were run from 0% B to 35% B over 90 min, then to 100% B over 2 min, held at 100% B for 15 min. Before each run the trap column was conditioned with 20 μL buffer A, the analytical – with 4 μL of the same buffer and the sample loading was set at 20 μL. When one column system was used the sample loading volume was set at 8 μL + sample volume. The LC thermostat temperature was set at 7°C. The Captive Spray Tip holder was modified similarly to an already described procedure (Beck et al., 2015) – the fused silica spray capillary was removed (together with the tubing which holds it) to reduce the dead volume, and the analytical column tip was fitted in the Bruker spray tip holder using a piece of 1/16’ x 0.015 PEEK tubing (IDEX), an 1/16’ metal two way connector and a 16–004 Vespel ferrule. The sample was loaded on the trap column at 850 Bar and the analysis was performed at 0.25 μL/min flow rate. The Impact II was set to acquire in a data-dependent auto-MS/MS mode with inactive focus fragmenting the 20 most abundant ions (one at the time at 18 Hz) after each full-range scan from m/z 200Th to m/z 2000Th (at 5 Hz rate). The isolation window for MS/MS was 2 to 3Th depending on parent ion mass to charge ratio and the collision energy ranged from 23 to 65 eV depending on ion mass and charge (Beck et al., 2015). Parent ions were then excluded from MS/MS for the next 0.4 min and reconsidered if their intensity increased more than five times. Singly charged ions were excluded since in ESI mode peptides usually carry multiple charges. Strict active exclusion was applied. Mass accuracy: error of mass measurement is typically within five ppm and is not allowed to exceed 10 ppm. The nano ESI source was operated at 1900V capillary voltage, 0.20 Bar CaptiveSpray nanoBooster pressure, 3 L/min drying gas and 150°C drying temperature.

Analysis of Mass Spectrometry Data was performed using MaxQuant 1.5.3.30 (Cox and Mann, 2008; Cox et al., 2014; Tyanova et al., 2014). The search was performed against a database comprised of the protein sequences from the source organism (E. coli K12) plus common contaminants using the following parameters: peptide mass accuracy 40 parts per million; fragment mass accuracy 0.05 Da; trypsin enzyme specificity, fixed modifications - carbamidomethyl, variable modifications - methionine oxidation, deamidated N, Q and N-acetyl peptides. Proteins were quantified from one peptide identification. Only those peptides exceeding the individually calculated 99% confidence limit (as opposed to the average limit for the whole experiment) were considered as accurately identified.

Binary interaction prediction using PrinCE

Request a detailed protocol

Protein-protein interactions were predicted using PrInCE (Stacey et al., 2017), a co-fractionation data analysis pipeline that assigns interactions based on the similarity of co-fractionation profiles. The PrInCE software is available online: https://github.com/fosterlab/PrInCE-Matlab. Since PrInCE employs a naive Bayes classifier, a set of known interacting and non-interacting protein pairs are required to train the classifier, that is a gold standard set. We constructed a gold standard set of protein complexes by combining the 30S ribosome with membrane protein complexes given by the IntAct protein complex database (www.ebi.ac.uk/complexportal/). True positive interactions (TP) are between proteins present in the same gold standard complex, and false positive interactions (FP) are interactions between proteins present in the gold standard set but which are not members of the same complex. PrInCE calculates an interaction score for each protein pair, with higher scores indicating an interaction is more likely to be a true interaction, as measured by the proportion of gold standard TPs. Specifically, a protein pair’s interaction score is equal to the TP-to-FP ratio, measured as precision (TP/(TP +FP)) of all predicted interactions with a classifier score greater than that protein pair. Both the precision of the full list (50%) and the interaction score are directly related to an interaction false discovery rate (FDR), as FDR = 100% precision. For full implementation of PrInCE see Stacey et al. (2017).

Unlike other co-fractionation analyses, which associate protein pairs using external datasets such as gene co-citation (Larance et al., 2016), PrInCE is designed to use only information derived from the experimental dataset. Using external datasets for interaction prediction can bias results to well-known, highly studied interactions (Skinnider et al., 2018). However, since there are still a considerable number of annular lipids retained in peptidiscs, the molecular weight of protein complexes can vary and broaden elution peaks (Carlson et al., 2018) and thereby increase false positives. Therefore, we struck a balance between predicted interactome size and interaction novelty by including a single external dataset, the M3D database (Many Microbes Microarray Database; Faith et al., 2008). For each protein pair observed in our experimental dataset, we calculated the Pearson correlation between expression profiles from M3D. Protein pairs not in the M3D database were imputed as the mean correlation value.

Protein complex assignment via ClusterONE and MCL

Request a detailed protocol

We used a two-stage procedure to cluster pairwise interactions into complexes (Wan et al., 2015; Drew et al., 2017). A first pass clustering was performed using ClusterONE (Nepusz et al., 2012), the results of which were further refined using the Markov Cluster algorithm (MCL) (Enright et al., 2002). This procedure resulted in clustered and unclustered protein interactions. In order to incorporate unclustered but high-scoring interactions (interaction score >0.75), we re-ran the two-stage clustering (ClusterONE + MCL) using unclustered pairwise interactions with score >0.75 as input. The union of these two sets of clusters formed the final set of complexes. Using both ClusterONE and MCL ensured that the same protein can be assigned to multiple complexes, while avoiding collapsing biologically distinct protein groups into the same protein complex, which ClusterONE tends to do (Wan et al., 2015). We measured clustering performance using the maximum matching ratio, a score calculated on the best one-to-one mapping between predicted and gold standard complexes (Nepusz et al., 2012).

Since both ClusterONE and MCL have tunable parameters, we performed a grid search optimization to find the parameter set which maximized the matching ratio value. The optimized parameters were: p, a ClusterONE parameter that models incompleteness in the network by assuming the existence of interactions outside of the given network; dens, a ClusterONE parameter that controls the minimum density of complexes; I, the single MCL parameter, which controls the granularity of MCL output; and S, which is the minimum interaction score of pairwise interactions fed into the clustering algorithm. The optimized values were p=5000, dens = 0.001, I = 20, and S = 0.50 (equal to 50% precision, that is the entire peptidisc interactome).

Computational validation of binary interactions and protein complexes

Request a detailed protocol

We performed multiple validations of both binary interactions and protein complexes using custom Matlab code. To validate the binary interactions, we first calculated three measures of biological association: i) fraction of protein pairs sharing at least one Gene Ontology (GO) term, ii) fraction of protein pairs with a positively correlated stress phenotypes as measured in the Tolerome database (R > 0, Pearson correlation; Erickson et al., 2017), and iii) fraction of protein pairs sharing at least one three-dimensional interacting domains (3did) (Mosca et al., 2014). Only GO and 3did terms that annotated >20 and<1000 proteins were used. Null distributions for each measure were calculated by generating 1000 random peptidisc interactomes, each composed of 4911 random, unique interactions between proteins in the peptidisc interactome. Each of the three measures of biological association were calculated for the 1000 random random interactomes. Z-scores were calculated relative to these null distributions.

We also calculated whether the number of overlapping interactions between our binary interaction list and a recently published E. coli cell envelope (CE) interactome (Babu et al., 2018) was significant. This was calculated by randomly assigning 4911 unique interactions to the set of proteins participating in our interactions. By calculating the overlap between (Babu et al., 2018) and random interactomes, we estimated the probability that the true number of overlapping interactions was due to chance alone. To validate complexes, we calculated the number of complexes enriched for at least one GO term (hypergeometric test, Benjamini-Hochberg correction). To obtain significance for the number of enriched complexes, we repeated this enrichment analysis with 1000 sets of random complexes, where each set was composed of 227 complexes and the size distribution was preserved from the original predicted complexes. Each random complex was generated by randomly sampling from the 526 proteins participating in the predicted complexes. GO terms were filtered such that only terms assigned to >20 and<1000 proteins were used.

As an additional computational validation, we determined the subset of our interactome that is also detected by two independent high-throughput interactomes. These validating interactomes are i) a detergent-solubilized size exclusion co-fractionation interactome (78984 interactions) and ii) an AP/MS dataset (499605 interactions). For methods of preparation, see Babu et al. (2018). These validating interactomes were collected independently from the dataset used to generate our peptidisc interactome and therefore provide orthogonal validation. Together, the set of pairwise interactions that were common to all three datasets (the peptidisc interactome and both validating interactomes) form our ‘High Confidence’ set of interactions.

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
    The synthetic phenotype of ΔbamB ΔbamE double mutants results from a lethal jamming of the bam complex by the lipoprotein RcsF
    1. EM Hart
    2. M Gupta
    3. M Wühr
    4. TJ Silhavy
    (2019)
    mBio, 10, 10.1128/mBio.00662-19, 31113901.
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
  58. 58
  59. 59
  60. 60
  61. 61
  62. 62
  63. 63
    Rapid detergent removal from peptide samples with ethyl acetate for mass spectrometry analysis
    1. YG Yeung
    2. ER Stanley
    (2010)
    Current Protocols in Protein Science, 16, 10.1002/0471140864.ps1612s59, 20155730.
  64. 64
  65. 65
  66. 66

Decision letter

  1. Richard Aldrich
    Senior Editor; The University of Texas at Austin, United States
  2. Nir Ben-Tal
    Reviewing Editor; Tel Aviv University, Israel
  3. Samuel Wagner
    Reviewer

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

[Editors’ note: a previous version of this study was rejected after peer review, but the authors submitted for reconsideration. The first decision letter after peer review is shown below.]

Thank you for submitting your work entitled "Profiling the E. coli Membrane Interactome in Detergent-free Peptidiscs" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and a Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Carol V. Robinson (Reviewer #1).

Our decision has been reached after consultation between the reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work will not be considered further for publication in eLife. Amongst the issues raised by the reviewers, they particularly emphasised the need for better treatment of false positive identifications.

Reviewer #1:

The manuscript describes a new approach for extracting and maintaining interactions in peptidiscs. This a very interesting paper worthy of publication in eLife. The study is complementary, yet distinct from nanodiscs and SMALPS, and appears to have some unique advantages.

Although the membrane proteome is still extracted with detergent prior to reconstitution results are similar to those obtained in SMALPs and nanodiscs. The number of unique insights obtained through the combination of methods applied here was however difficult to tease out. It seems that the most important one was that MetQ co-elutes as a complex with MetNI. MetQ appears to be the only substrate binding protein that is lipid modified at its N-terminus. This I thought particularly interesting.

Of the SecYEG translocon interacting partners it as surprising to me that simultaneous over- production of the SecYEG core translocon with both PpiD and YfgM is necessary for detection of this interaction. It might suggest that original interactions which are thought to be multiple are disrupted by the initial detergent solubilisation step. Perhaps the authors could comment on this.

Reviewer #2:

In this report, Carlson and co-workers combine SILAC labeling, reconstitution of crude membranes using peptidiscs, size fractionation and mass-spec to characterize the interactome of the membrane proteins of E. coli. This is an ambitious study of potential interest that may contribute to future studies of PPI interactions of the membrane proteome. However, there are some serious problems with this study that in my opinion make it unsuitable in its current state for publication in eLife.

From a technological perspective, the novelty is incremental: the methodologies have been used in the past, including the peptidiscs. If I understand correctly, the only novelty is the use of the peptidiscs instead of detergent or amphiphilic copolymers.

From a biological perspective, all of the interactions that were used for validation have been previously described. The paper presents no new novel information.

Examination of Supplementary file 3 reveals that the interactions are heavily biased towards some proteins that are highly represented, while others are completely missing. In addition, many well-characterized interactions are not detected, while other interactions, that seem non-specific false positives are "detected" (see for examples the interactions of NADH dehydrogenase, BtuB, BamC, LptB, CydC).

In addition, below are some specific concerns that question the validity of the findings:

>The FepBCD complex is larger than MetINQ, yet the latter elutes at an earlier time. Along the same lines, the ATP synthase, which is considerably larger than both ABC transporters also elutes at a similar volume, as do BamABCD and NuoABCDEFGI. This is very suspicious. Can we be looking at aggregates? What is the void volume of this column?

The suggestion that MetQ is anchored to the membrane is new and unexpected. It would greatly add to the paper if this suggestion is experimentally validated with the proper controls of an SBP that is not expected to be membrane anchored.

In addition, why does this explain the co-elution with MetIN? This only suggests membrane association, and a stable transporter-SBP association.

The methodology requires detergent extraction, unlike the use of amphiphilic copolymers. This possibly disturbs many of the endogenous interactions. This issue is not discussed.

"The end result of reconstitution is a membrane protein that displays increased stability": unsupported statement. How many cases were examined?

From the width of the elution peaks of the SEC fractionation it appears that many of the interactions (the FPs) may stem from similar migrations rather than associations. This also likely explains the non-specific interactions that were detected.

For removal of DDM, the authors use dilution and re-concentration in a centricon. This is a major flaw since DDM also gets concentrated in this process (this phenomenon is easily detected using sugar detection methods). So, the assumption that the preparation contains no (or little) detergent is probably wrong.

"Upon removal of aggregate": At what stage? what does this mean?

How is the data MsbA data relevant? This study is supposed to provide proteomic information, so how is the behavior of a single protein relevant? One can relate to the other many advantages of DDM, but the behavior of MsbA is irrelevant. If the authors had chosen EmrE as their benchmark, would they extract the membranes using chloroform?

"As expected, a large fraction of these proteins (591 proteins) are predicted, based on gene ontology terms, to be associated with the cell envelope". Considering that this is a membrane preparation, this is not a large fraction. The membrane proteome is supposed to be 1/3 of all proteins so this is a pretty mild enrichment factor. This is actually quite surprising and should be addressed by the authors. Is there some common theme to the non-membrane proteins that were identified? Is there a common theme to the membrane proteins that are missing?

Subsection “Large membrane protein complexes are captured in the peptidisc library”: How many interactions were analyzed to support this statement?

On what basis the three analyzed cases (Figure 3) were chosen? Was this choice arbitrary?

The lack of rigorous analysis of any novel interaction is greatly missing.

The 13 Type-I SBPs that were identified by the authors: if these are not membrane anchored (like MetQ) and do not stably interact with their ABC transporters, how is it that they were identified? In this respect, why were other Type II SBPs (e.g., HmuT, FhuD, BtuF) not identified?

Reviewer #3:

The work by Carlson et al. describes a derivative of protein-correlation-profiling (PCP) for membrane complexomes based on stabilization of detergent-extracted membrane protein complexes by peptidiscs. Overall the paper is well written and good to follow and the experiments suited the needs.

It is clearly laid out that better approaches are needed for proteomics-based identification of protein-protein interactions of membrane protein complexes. The downsides of current technologies are discussed well. The approach presented here may be an improvement over other existing methods but the based on the data presented, the improvement does not seem to be impressive. Saying this, it is surprising to see that the authors apparently claim that a precision of 50% is sufficient to call something an identified interaction. It is not surprising to see then, that only a small proportion of these 50% overlap with previous analyses and that the validating experiments validate only very few of the found interactions. Interestingly, the authors set a threshold of 80% and 75% precision, respectively, for the validation experiments. First of all, this is inconsistent within the validation experiments (75% vs. 80%) and secondly, this is inconsistent with the previously set threshold of 50%, which the authors don't seem to trust themselves. It would be very beneficial for the paper if a false discovery rate could be calculated and applied because the question clearly remains, which of the identified interactions are trustworthy and which ones not.

Along the same lines, it would be good to show where the 340 overlapping interactions between this study and the one by Babu et al., 2017 rank on the graph in Figure 4C.

In the validation experiments using MipA, the authors identify YajC and AtpF as potential interacting proteins. The authors should discuss whether this makes sense in the light of the membrane and complex topology of these proteins.

A major current problem is the stabilization and identification of membrane protein complexes that bridge inner and outer membranes of Gram negative bacteria. The paper would benefit from presenting and discussion the subcellular localization of the identified complexes.

[Editors’ note: what now follows is the decision letter after the authors submitted for further consideration.]

Thank you for resubmitting your work entitled "Profiling the E. coli membrane interactome captured in Peptidisc libraries" for further consideration at eLife. Your revised article has been favorably evaluated by Richard Aldrich as the Senior Editor, a Reviewing Editor, and three reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

Summary of paper:

The authors report on the application of their recently developed peptidisc approach to identify protein-protein interaction of the E. coli membrane proteome. Although other methods are available, the authors make the argument that the peptidisc approach holds advantages since it uses a membrane mimetic that is broadly applicable to membrane proteins of various sizes and to large complexes. In addition, the peptidisc scaffold protein also traps endogenous lipids that may be important for protein interactions and function. The authors benchmark their findings against interactome analysis provided via alternative approaches and demonstrate that their method is comparable, if not superior, especially for multi-component large complexes. The peptidisc interaction analysis validates known interactions but also identifies an unexpected stable interaction between the methionine substrate binding protein (MetQ) and its cognate ABC transporter (MetNI). The findings and the approach lay the foundations for future work examining the newly-identified interactions and the application of the approach to membrane-interactome analysis of other organisms.

Opinion:

The problem of identifying novel membrane protein interactions is important and interesting, and the development of new and simpler approaches is pertinent. However, in its present state, even after the previous round of revisions, it is unclear whether the peptidisc approach presents sufficient novelty to justify publication in eLife. Major revisions are needed to address the outstanding issues listed below.

Major issues:

1) The main claim of the authors is that their approach is a useful tool to identify novel interactions of the membrane proteome. However, of the newly-identified interactions they validate a single one (MetQ-MetIN), which is also not entirely surprising based on two recent publications from the Rees lab. It would greatly benefit the paper, and convince the reader of the usefulness of the approach, if the authors show a few more examples that demonstrate novel and unpredicted interactions.

2) The authors use 0.8-1% detergent (DDM) for 1 hour, yet claim this is a short time. However, most membrane proteins are fully extracted in a shorter time. Along the same lines, the authors show that the SEC and profiles obtained in DDM and in peptidiscs are similar. Since the use of detergent alone is simpler than that of the peptidiscs, it remains unclear how the presented approach is advantageous. To support this claim, the authors need to show a complete and rigorous comparison between DDM and peptidiscs. How many of the interactions identified by the peptidiscs approach would be lost if detergent only is used? The authors are encouraged to equally minimize the detergent exposure time in these experiments.

3) The authors use a detergent (DDM) concentration that at the conditions used (temperature, salt concentration) is likely above the CMC. As a result, the detergent is probably concentrated along with the peptidisc library. The authors need to measure the DDM concentration during, and at the end of, the concentration step to convince the reader that their approach is indeed superior to the use of detergent alone.

4) The authors prepare membrane fractions using ultracentrifugation, rather than density gradients or floatation assays. This leads to sedimentation of large proteins, aggregates, and complexes that are unrelated their analysis and confound the interpretation of the data. The authors must explain why they have chosen this approach and demonstrate that it does not undermine their analysis.

5) Several of the concerns raised by the reviewers during the initial round of submission remain inadequately answered. For example, the identification of 13 Type-I SBPs remains unexplained. The authors need to directly and clearly address this and other concerns raised by the reviewers.

6) The paper suffers from overstatement and inaccurate citations.

7) The supplementary tables should be reformatted. Now they are unintelligible.

https://doi.org/10.7554/eLife.46615.024

Author response

[Editors’ note: the author responses to the first round of peer review follow.]

Reviewer #1:

The manuscript describes a new approach for extracting and maintaining interactions in peptidiscs. This a very interesting paper worthy of publication in eLife. The study is complementary, yet distinct from nanodiscs and SMALPS, and appears to have some unique advantages.

We thank the reviewer for highlighting the novelty of our study. We have made substantial revisions to improve the quality of the manuscript.

Although the membrane proteome is still extracted with detergent prior to reconstitution results are similar to those obtained in SMALPs and nanodiscs. The number of unique insights obtained through the combination of methods applied here was however difficult to tease out.

There are important differences in the results we obtain in our side-by-side comparison between the peptidisc and SMALPs libraries. Although the solubilization efficiencies of the peptidisc and SMALPs are comparable (see Figure 3A and 3B), we observe dramatic differences in terms of the stability of multi-subunit membrane complexes. We compared the coelution profiles for three well-known complexes – the Bam complex, the respiratory chain complex and the ATP synthase complex – in both SMALPs and peptidiscs (see Figure 3C and 3D; Figure 3—figure supplement 1). In every case, the complexes are preserved in the peptidisc but largely dissociated in SMALPs. We have clarified this important point in the manuscript.

The nanodisc was not used in this report. We have, however, recently shown that – unlike the nanodisc – the peptidisc is a universal, “one-size-fits-all” reconstitution scaffold, able to reconstitute membrane proteins of differing sizes and topologies without needing addition of exogenous lipids (Carlson et al., 2018). In addition to streamlining the library reconstitution process, use of the peptidisc rather than the nanodisc also mitigates possible bias during the reconstitution process based on protein size. We have commented on these points in our revised Discussion.

It seems that the most important one was that MetQ co-elutes as a complex with MetNI. MetQ appears to be the only substrate binding protein that is lipid modified at its N-terminus. This I thought particularly interesting.

We agree with the reviewer. The finding that MetQ is lipidated and that it co-elutes with MetNI is an important and novel discovery. In this resubmission, we incorporate new data to rigorously validate these observations. Specifically: i) we validate the stability of the MetNIMetQ interaction by SILAC AP-MS (now included as Figure 8) and ii) we demonstrate the importance of MetQ lipidation by demonstrating that a non-lipidated MetQ mutant does not interact with MetNI (now included as Figure 8—figure supplement 2). We are also including an additional novel observation which we missed in our first submission. Our pairwise interaction list identifies NlpA (Lipoprotein 28) as a potential interactor of MetQ. We found this particularly interesting in light of a previous report which showed through complementation assays that NlpA may be able to deliver Methionine to the MetNI transporter in a MetQ-deleted strain, albeit at low efficiency (Zhang et al., 2003). Notably, a physical interaction between MetNI-Q and NlpA has not been reported to date in the literature. Our new MetNI SILAC AP-MS data now reveal significant enrichment of NlpA along with MetQ (Figure 8). This observation, along with our PCP-SILAC data, provide strong evidence that a physical interaction between MetNI-Q and NlpA does exist in the cell. Altogether, this further validate the peptidisc in untargeted interactomics studies.

Of the SecYEG translocon interacting partners it as surprising to me that simultaneous over- production of the SecYEG core translocon with both PpiD and YfgM is necessary for detection of this interaction. It might suggest that original interactions which are thought to be multiple are disrupted by the initial detergent solubilisation step. Perhaps the authors could comment on this.

We agree with the reviewer. We are not discounting the dissociative effects of detergent on the YfgM-PpiD-SecYEG interaction. We note that Digitonin, rather than DDM, was employed in the original paper which identified this complex (see (Götzke et al., 2014). It is therefore possible that an interaction between YfgM-SecYEG and/or PpiD-SecYEG would be preserved in Digitonin but dissociated in DDM (as done in our study). However, it is equally possible that strong interaction largely depends on correct stoichiometric ratio between components as our data seem to indicate (Figure 6—figure supplement 1). We have modified the manuscript to include a comment on this point.

Reviewer #2:

In this report, Carlson and co-workers combine SILAC labeling, reconstitution of crude membranes using peptidiscs, size fractionation and mass-spec to characterize the interactome of the membrane proteins of E. coli. This is an ambitious study of potential interest that may contribute to future studies of PPI interactions of the membrane proteome. However, there are some serious problems with this study that in my opinion make it unsuitable in its current state for publication in eLife.

We agree with the reviewer that our initial version had certain flaws and accordingly, we have substantially revised the manuscript including additional computational and experimental validations, as well clarification both in the text and our responses above and below. However, as the reviewer also points out, this study potentially represents a major technical advance in the field of membrane interactomics (as such we submitted it as a “Tools and Resources” article).

From a technological perspective, the novelty is incremental: the methodologies have been used in the past, including the peptidiscs. If I understand correctly, the only novelty is the use of the peptidiscs instead of detergent or amphiphilic copolymers.

The principle novelty of our study is the combination of the peptidisc – a novel membrane mimetic system our laboratory developed recently (see (Carlson et al., 2018)- with the PCP-SILAC workflow. We contend that our study provides a significant technological and methodological advance for the following key reasons:

i) The peptidisc solves a crucial problem in the design of detergent free protein libraries because it enables “flexible”, streamlined reconstitution of membrane proteins and complexes of varying topologies and sizes (see (Carlson et al., 2018). This is in marked contrast to the nanodisc system, which may bias reconstitution toward membrane proteins and complexes of a certain size, based on the length of the scaffold protein used.

ii) From our results with SMA as shown in Figure 3 and Figure 3—figure supplement 1, we can say that the peptidisc is a comparatively gentle technique that is better suited for examining native protein interactions. Many membrane protein complexes which are preserved in the peptidisc are dissociated by the SMA polymer. The reviewer is referred to our response to point 2 from reviewer #1 which addresses this topic in detail.

From a biological perspective, all of the interactions that were used for validation have been previously described. The paper presents no new novel information.

We are presenting this study without emphasis on discovery of novel interactions, but rather as a way to rapidly assess a cell membrane interactome. Nevertheless – as stated in our response to Point 3 from reviewer #1 – we are including new data with this resubmission which reveal novel insights into the ABC transporter MetNI-MetQ. Specifically, we reveal that MetQ is a unique SBP in E. coli, because it is tethered to the membrane by an N-terminal lipid anchor. We further show that this lipidation is required for stable association between MetQ and MetNI. Furthermore, we identify an interaction and validate NlpA (Lipoprotein 28) as an interactor of the MetNI-Q complex (data presented in Figure 8 and Supplementary Table 1G). These are novel discoveries with profound implications for our understanding of the mechanism of MetNI-Q mediated Methionine import in bacteria. These data are nicely validating our peptidisc interactomic approach.

Examination of Supplementary file 3 reveals that the interactions are heavily biased towards some proteins that are highly represented, while others are completely missing.

We agree with the reviewer that our interactome contains highly represented proteins. This likely reflects, at least in part, the true biological promiscuity of some proteins. Proteins participating in medium-to-large complexes can have dozens of interacting partners. For example, the most represented protein in our interactome, P0AEH1 (YaeL), which we detect in 114 interactions, is listed as occurring in 106 interactions in the most recent BioGRID database (https://thebiogrid.org/4261736).

In our interactome, the proportion of very highly represented proteins, e.g. the number of proteins in >100 interactions as a fraction of all unique proteins in the interactome, is comparable to other interactomes: we find 1.6% of unique proteins in >100 interactions (9/562), compared to 0.4%, 1%, 0.2%, and 3% (Babu et al., 2018; Hein et al., 2015; Wan et al., 2015; Scott et al., 2017). While some interactions for highly represented proteins will be false positives, we aimed to control for the false positive rate with multiple computational validations. Regarding missing interactions and controlling for false positives, please see our response just below.

In addition, many well-characterized interactions are not detected, while other interactions, that seem non-specific false positives are "detected" (see for examples the interactions of NADH dehydrogenase, BtuB, BamC, LptB, CydC).

We agree with the reviewer that, ideally, we would capture all well-characterized interactions. Unfortunately, this is beyond the current state of interactome studies. For example, taking the CORUM database as the “full” human interactome of >80,000 interactions, well-cited interactome studies (Hein et al., 2015) and (Wan et al., 2015) capture only 3% and 7% of pairwise CORUM interactions, respectively. Similarly for E. coli, (Babu et al., 2018) captures 40 of the 993 known protein interactions we use as reference in our study.

Some proteins (hence interactions) are missing because they are low abundance or otherwise hard to detect with mass spectrometry (e.g. poor ionization). Importantly, though, many of the interactions we do detect are indeed well-characterized interactions, such as the ATP synthase complex, the Bam complex, and the respiratory chain complex (see Figure 3 and

Figure 3—figure supplement 1).

Because it is difficult to capture known interactions, a common strategy of interactome studies (and one we take here) is to accept non-specific false positives as the cost of detecting true interactors. It is therefore very important to estimate and report the proportion of false positives. We took pains to do that here by i) calculating the TP-to-FP ratio as measured against our reference E. coli complexes; ii) calculating how many interactions share GO terms, binding domains, and have correlated expression levels, which are measures of true, biologically interacting proteins; and iii) cross-validating our detected interactions against independent datasets, namely the cell envelope (CE) interactome (Babu et al., 2018) and the two validating interactomes.

In addition, below are some specific concerns that question the validity of the findings:

The FepBCD complex is larger than MetINQ, yet the latter elutes at an earlier time. Along the same lines, the ATP synthase, which is considerably larger than both ABC transporters also elutes at a similar volume, as do BamABCD and NuoABCDEFGI. This is very suspicious. Can we be looking at aggregates? What is the void volume of this column?

Size exclusion chromatography was performed using two BioSep4000 columns (Phenomenex) connected in tandem. The total volume of the column was 18 mL and the void volume was 6 mL. The void volume is represented as the zero on the x-axis of all co-elution graphs in the manuscript. Thus, any proteins eluting past the zero point in our dataset are fully soluble. We have clarified this last point in legend of Figure 3 as it appears first.

We were also puzzled by the elution behaviour of some protein complexes on the silica resin. Our main hypothesis to explain this behaviour is that certain membrane protein complexes may retain significant amounts of phospholipids, which increase the apparent molecular weight of the complex. These additional lipids may also affect the chromatographic behaviour of the proteins during size exclusion chromatography on silica – specifically, by broadening elution peaks and thus decreasing the resolution of the column. Given the broadness of the peaks we observed, we were concerned about detecting “false positive” interactions caused by co-elution of noninteracting proteins. To filter out false positive interactions caused by spurious co-elution, we incorporated information from a co-expression dataset (M3D) (see Faith et al., 2008).

The suggestion that MetQ is anchored to the membrane is new and unexpected. It would greatly add to the paper if this suggestion is experimentally validated with the proper controls of an SBP that is not expected to be membrane anchored.

Reviewer #1 also asked for confirmation and validation of this important finding. Please see our response to Point 3 from reviewer #1 above. Briefly, we co-expressed MetNI with MetQ and validated the stability of the interaction by showing that MetQ co-purifies with his-tagged MetNI. In this experiment, we noticed that MetQ is localized to the membrane fraction, an observation which supports our suggestion that MetQ is lipidated.

To explore the importance of MetQ lipidation for interaction with MetNI, we constructed a mutant – MetQ C23A – in which the Cysteine residue required for lipidation is replaced with an Alanine. Although this mutant is localized to the membrane fraction, it does not interact with MetNI, thus demonstrating the importance of MetQ lipidation for formation of the MetNI-Q complex. This series of experiments reveals the ability of the peptidisc-PCP-SILAC workflow for capturing and identifying novel, unexplored aspects of membrane protein complexes that are not evident using other experimental methods.

In addition, why does this explain the co-elution with MetIN? This only suggests membrane association, and a stable transporter-SBP association.

Our experiments using the non-lipidated MetQ C23A mutant reveal the importance of MetQ lipidation for its interaction with MetNI – please see the point above.

The methodology requires detergent extraction, unlike the use of amphiphilic copolymers. This possibly disturbs many of the endogenous interactions. This issue is not discussed.

We use minimal non-denaturing detergent, and importantly we rapidly remove it by replacing it with peptidisc, so that dissociation of complexes caused by detergent exposure is meant to be minimized. However, it is true that some complexes will be missed or destroyed. We are explaining this unavoidable caveat in the Discussion of the manuscript. A possible way around to this problem would to construct peptidisc libraries using three different detergents with partially overlapping solubilization solubilities (as has been done in Babu et al., 2018).

Regarding the amphipatic SMA co-polymer, we present data that rather indicate that SMA is not a “soft” solubilization agent. As shown Figure 3, some large complexes which can be isolated in the peptidisc are instead bracked apart when the membrane is solubilized with the amphiphilic SMA polymer. This drawback clearly limits the applicability of SMA as a solubilization agent for interactomics studies.

"The end result of reconstitution is a membrane protein that displays increased stability": unsupported statement. How many cases were examined?

This statement is supported by our previous publication, as well as by unpublished results from our laboratory (see (Carlson et al., 2018), in which we show a major increase in membrane protein stability upon reconstitution in peptidisc.

From the width of the elution peaks of the SEC fractionation it appears that many of the interactions (the FPs) may stem from similar migrations rather than associations. This also likely explains the non-specific interactions that were detected.

This is a caveat inherent to any PCP technique, and we agree that the elution peaks in this study are wider than we would have liked. However, the consequences of the wide peaks are testable: if they’re too wide, then true interacting proteins will be swamped out by spurious co-fractionating proteins, and we will be unable to predict true positives better than false positives. We did not find that this was the case. That is, we were able to predict many interactions that, through computational validation, are of comparable quality to other interactomes. Additionally to mitigate this effect, we use two columns arranged in tandem and we incorporate expression data to avoid spurious interaction identification. That is, interactions are predicted using our peptidisc co-fractionation data and the M3D expression database. In this resubmission, we have also incorporated interaction data from AP-MS studies to provide a high-confidence list of interactions. Altogether, while we agree that the wide elution peaks are in issue, we believe we’ve sufficiently accounted for them, given that our interactome looks relatively “high quality” (computational validation compares favourably to other published interactomes, interactions validated in by wet lab experiments, etc.).

For removal of DDM, the authors use dilution and re-concentration in a centricon. This is a major flaw since DDM also gets concentrated in this process (this phenomenon is easily detected using sugar detection methods). So, the assumption that the preparation contains no (or little) detergent is probably wrong.

The detergent is diluted below its CMC, so that monomeric detergent molecules can pass through the centricon filter. We cannot exclude the possibility that small remnant, yet non-solubilizing amounts of detergent are trapped in the peptidiscs.

"Upon removal of aggregate": At what stage? what does this mean?

Removal of aggregated material occurs during ultracentrifugation immediately following detergent solubilization of the membrane. We have clarified this in the manuscript.

How is the data MsbA data relevant? This study is supposed to provide proteomic information, so how is the behavior of a single protein relevant? One can relate to the other many advantages of DDM, but the behavior of MsbA is irrelevant. If the authors had chosen EmrE as their benchmark, would they extract the membranes using chloroform?

The MsbA data is important because it shows that our peptidisc library method traps individual membrane proteins and complexes into discrete peptidisc particles, rather than non-specifically clustering them together. To show this, we overproduced a model protein (i.e. MsbA) and subsequently purified it from the library. From this, we observed efficient isolation of MsbA homo-dimer – free from contaminants – and perfectly soluble, as evaluated on SDS-PAGE and CN-PAGE respectively. This data confirms that proteins are not randomly captured into peptidiscs. We have clarified this important point in the manuscript.

> "As expected, a large fraction of these proteins (591 proteins) are predicted, based on gene ontology terms, to be associated with the cell envelope". Considering that this is a membrane preparation, this is not a large fraction. The membrane proteome is supposed to be 1/3 of all proteins so this is a pretty mild enrichment factor. This is actually quite surprising and should be addressed by the authors. Is there some common theme to the non-membrane proteins that were identified? Is there a common theme to the membrane proteins that are missing?

The peptidisc libraries employed in this study were prepared from crude cell membranes. The only enrichment performed to prepare the membranes was an ultracentrifugation step following cell lysis. While ultracentrifugation leads to sedimentation of cell membranes, it is well documented that large, soluble cellular complexes – such as GroEL and ribosomes – also pellet along membranes (Papanastasiou et al., 2013, 2015). We have clarified this point in the manuscript.

Furthermore, it is well known that membrane proteins suffer from radically decreased identification rates in mass spectrometry experiments, which would further decrease the number of identified membrane proteins. The common theme among the non-membrane proteins we identify were large, soluble complexes that co-sediment with the membrane fraction (such GroEL, ribosome, AceP, etc.).

Subsection “Large membrane protein complexes are captured in the peptidisc library”: How many interactions were analyzed to support this statement?

We looked at all pairwise comparisons between complex members of the three complexes (ATP synthase, Bam, and respiratory complex). This gave 48 pairwise comparisons in the SMALP condition (21 + 6 + 21 comparisons, respectively) and 55 comparisons in the peptidisc condition (28 + 6 + 21 comparisons, respectively).

On what basis the three analyzed cases (Figure 3) were chosen? Was this choice arbitrary?

The validation targets were picked in large part due to their relevance with existing projects in our laboratory (Sec and ABC). The proteins YfgM and PpiD are relatively uncharacterized ancillary subunits of the bacterial Sec translocon, which is a major focus of study in our lab. As a result, many of the materials were already available to conduct these experiments (i.e. a stable SecYEG expression construct). The complex MetNI was chosen as a validation target due to the surprising observation that it co-elutes with the SBP MetQ (see above points for details). We included data on the scaffold protein MipA in our earlier submission. We now feel that inclusion of this data detracts from the main focus of our study, and have therefore removed it in this resubmission.

> The lack of rigorous analysis of any novel interaction is greatly missing.

Since our first submission, we have validated the MetNI-Q interaction, and we have demonstrated the importance of MetQ lipidation in this complex. This is an important and novel biological observation. Additionally, our newly included SILAC AP-MS experiment reveals the lipoprotein NlpA as an interactor of MetNI-Q. The MetNI-Q-NlpA association has only been characterized to a very limited extent in the current literature. We have now validated this novel interaction using AP-MS, which further validate the validity of the peptidisc-PCP approach.

However, we would like to stress out that our major aim was to demonstrate that our method is suitable for predicting interactomes from membrane proteins. Also, PCP-SILAC, while useful for identifying novel interactions, is a powerful tool for comparative interactomics. While we did not expand this study to a third label, the work still remains a valid proof of concept for subsequent analysis based in detergent-free fractionation of a membrane preparations.

The 13 Type-I SBPs that were identified by the authors: if these are not membrane anchored (like MetQ) and do not stably interact with their ABC transporters, how is it that they were identified? In this respect, why were other Type II SBPs (e.g., HmuT, FhuD, BtuF) not identified?

Type I SBPs have weak affinity for their cognate transporters, which may result in their co-sedimentation with the membrane fraction during ultracentrifugation following cell lysis. However, as reported by other biochemists including ourselves using the maltose transporter, this affinity is too weak to resist the dissociative effects of detergent during membrane solubilization. This interaction is also not restored until the transporter in stabilized in a very specific conformational state requiring non-hydrolyzable ATP analogs. This likely explains why most soluble Type 1 SBPs are eluted at the end of the fractionations and are fully dissociated from their cognate transporters.

Reviewer #3:

The work by Carlson et al. describes a derivative of protein-correlation-profiling (PCP) for membrane complexomes based on stabilization of detergent-extracted membrane protein complexes by peptidiscs. Overall the paper is well written and good to follow and the experiments suited the needs.

It is clearly laid out that better approaches are needed for proteomics-based identification of protein-protein interactions of membrane protein complexes. The downsides of current technologies are discussed well. The approach presented here may be an improvement over other existing methods but the based on the data presented, the improvement does not seem to be impressive. Saying this, it is surprising to see that the authors apparently claim that a precision of 50% is sufficient to call something an identified interaction. It is not surprising to see then, that only a small proportion of these 50% overlap with previous analyses and that the validating experiments validate only very few of the found interactions.

We agree with the reviewer that high level of false positives is an unwanted feature of interactomes. However, as explained above in response to reviewer #2, we view false positives as the cost of detecting true interactors in a high-throughput study. We believe the number of false positives needs to be estimated and controlled for, and we took care to do both of these. Further, one of the major aims in this study was to provide a larger list that researchers can reduce as needed if a higher confidence set of interactions is required. We provide two ways for researchers to refine the full published list (by interaction score and the “high confidence” subset) and we have revised our Discussion to better explain this to the reader less familiar with global interactome analysis. We believe this is the preferred method since it gives researchers more choice. That is, it is better to publish lower-scoring interactions and give researchers the option to use them or not, as opposed to not publishing them at all.

We agree that 50% precision is low, but we note that it is in the same range as other interactome studies. Defining true and false positives as we do in our study, and using the CORUM database as the true interactome, other well-cited studies (Hein et al., 2015) and (Wan et al., 2015) have precisions of 31% and 65%, respectively. Using our E. coli reference complexes as the true interactome, (Babu et al., 2018) has a precision of 32%. Therefore, we note that the proportion of false positives in our study is not unusually high compared to other interactome studies.

We agree that the overlap between our interactome and other interactomes is small in magnitude. However, we believe this does not invalidate our interactome for two reasons. First, because the full interactome of a species is estimated to be large (e.g. hundreds of thousands of interactions in the human interactome (Stumpf et al., 2008) and interactome studies sample a small, random subset of it, overlap between interactome studies tends to be small. For example, two well-cited, high-throughput human interactome studies with 31,000 and 14,000 interactions, respectively, only have 408 interactions in common (Wan et al., 2015; Hein et al., 2015), a much smaller relative overlap than our interactome and (Babu et al., 2018). Second, although small, the overlap between our interactome and other E. coli interactomes is significantly larger than expected by random chance (Figure 5D, Figure 5—figure supplement 1). Therefore, we believe the overlap between our interactome and the E. coli interactome in (Babu et al., 2018) is meaningful and supports the validity of our conclusions.

Interestingly, the authors set a threshold of 80% and 75% precision, respectively, for the validation experiments. First of all, this is inconsistent within the validation experiments (75% vs. 80%) and secondly, this is inconsistent with the previously set threshold of 50%, which the authors don't seem to trust themselves.

The reviewer raises a good point that, by design, using a threshold of 50% estimated precision produces an interactome with a large number of false positives. We chose this threshold because we wanted to provide a comprehensive list with fewer false negatives, at the at the cost of more false positives. Importantly and as mentioned above, we now explicitly say this in the Discussion to inform the reader. Further, we trust these interactions in proportion to their estimated precision level. Therefore we wanted to validate interactions with a higher estimated precision and to provide interpretable graphs, we used a higher cut-off for the validation experiments. Finally, from the immediately previous comment, 50% precision is in the range of well-cited interactome papers (31%, 32%, and 65% precision in (Wan et al., 2015; Babu et al., 2018; Hein et al., 2015), respectively).

It would be very beneficial for the paper if a false discovery rate could be calculated and applied because the question clearly remains, which of the identified interactions are trustworthy and which ones not.

The estimated false discovery rate for the peptidisc interactome can be calculated from the precision (FDR = 100% – precision). We agree this is an important point, and to make this connection explicit we now include the FDR calculation in the text.

Along the same lines, it would be good to show where the 340 overlapping interactions between this study and the one by Babu et al., 2017 rank on the graph in Figure 4C.

Interactions that overlap between our study and the interactome published in Babu et al., 2018 tend to be higher scoring than non-overlapping interactions. We added this information to Figure 4C, and we now discuss this in the text.

In the validation experiments using MipA, the authors identify YajC and AtpF as potential interacting proteins. The authors should discuss whether this makes sense in the light of the membrane and complex topology of these proteins.

MipA is a lipoprotein anchored on the periplasmic side of the outer membrane. Both YajC (part of the Sec complex) and AtpF (part of the F1F0 ATPase) are inner membrane proteins with exposed loops in the periplasm. Therefore, these interactions do make sense in light of the topologies of these proteins. Please note that – as stated above in response to reviewer #2 – we have removed the MipA validation experiment in this resubmission. In light of the additional in-depth validation data we are presenting on the MetNI-Q complex, we feel that inclusion of the MipA data detracts from the main focus of the manuscript.

A major current problem is the stabilization and identification of membrane protein complexes that bridge inner and outer membranes of Gram negative bacteria. The paper would benefit from presenting and discussion the subcellular localization of the identified complexes.

We identify a number of interactions that bridge between the inner and outer membranes. We observe interactions between the lipoprotein Pal (outer membrane), the periplasmic protein TolB and the inner membrane proteins TolQ and TolR. More in-depth characterization of these interactions is currently underway in our laboratory. As suggested by reviewer #1, we have amended Figure 1 so that the reader readily understand our study deals with a complex network of interactions taking place across a double cell membrane.

[Editors' note: the author responses to the re-review follow.]

[…] The problem of identifying novel membrane protein interactions is important and interesting, and the development of new and simpler approaches is pertinent. However, in its present state, even after the previous round of revisions, it is unclear whether the peptidisc approach presents sufficient novelty to justify publication in eLife. Major revisions are needed to address the outstanding issues listed below.

We thank the editor and reviewers for highlighting the importance of our study. We strongly feel that our work is sufficiently novel for publication in eLife, especiallygiven the scarcity of proteomic methods that effectively enable analysis of membrane protein interactions in aqueous solution. In this response, we are including additional new data to unequivocally demonstrate the superiority of the peptidisc approach relative to detergent for stabilizing and detecting novel interactions. We are showing the following:

i) We identify a novel interactor of the MetNI transporter, the lipoprotein NlpA. Using SILAC AP/MS, we demonstrate that this interaction is preserved in peptidisc and undetected in detergent (Figure 8).

ii) We discover that MetQ is a lipidated SBP and we demonstrate the importance of this membrane anchor for association to the MetNI transporter (Figure 8—figure supplement 2). The role of MetQ lipidation has been overlooked in earlier studies (1,2).

iii) We show that YfgM and PpiD are two membrane-embedded chaperones that interact strongly with the Sec translocon and that a proper stoichiometry is important for detection of this association (Figure 6—figure supplement 1). Similar to what we observe with MetNI-NlpA, these interactions are difficult to capture in detergent (Figure 8), which may explain why these interactions have been considered transient and detected only indirectly in the past (3,4).

iv) Continuing with the Sec translocon, our pairwise interaction dataset reveals an astonishing correlation with the Bam complex. We confirm this observation using SILAC AP/MS (Figure 6). There significant enrichment of the BamB and BamC subunits and here also, these interactions are not detected when the same workflow is performed in detergent. Further work will be necessary to understand the interface of association, but the identification of this trans-membrane super-complex has profound implications for our understanding of outer membrane protein biogenesis. In further support, while our manuscript was under revision, a study from the Collinson group was released in bioRvix.org. Their work reports the initial isolation and a low resolution EM structure of a Bam-Sec translocon (https://doi.org/10.1101/589077). This second independent study from a leader in the field of Sec-mediated protein translocation further underscores the importance the Bam-Sec interaction we have identified.

v) Continuing with the Bam complex, we show that all 5 subunits are fully captured in peptidisc in addition to two other interactors – RcsF and OmpA (Figure 7). These interactors were previously inferred from genetic and in vivo cross-linking experiments but never biochemically isolated (5,6). Accordingly, we show that the BamA-RcsF-OmpA interaction is less apparent in detergent than in peptidisc (Figure 7).

We further explain these findings in our point-by-point response below. We also provide a “compare manuscript” document to show the extend of the changes we have made.

Major issues:

1) The main claim of the authors is that their approach is a useful tool to identify novel interactions of the membrane proteome. However, of the newly-identified interactions they validate a single one (MetQ-MetIN), which is also not entirely surprising based on two recent publications from the Rees lab. It would greatly benefit the paper, and convince the reader of the usefulness of the approach, if the authors show a few more examples that demonstrate novel and unpredicted interactions.

Our findings on MetNI-Q are novel because they reveal the importance of MetQ lipidation for association to the MetNI transporter. The role of MetQ lipidation for mediating this interaction has never been reported before, and this lipidation turns out to be a unique case among the Type I family of ABC transporters in E. coli. The recent papers from the Rees lab (Biol Chem 2015 and PNAS 2018) report on the interaction of MetQ-MetNI and high-resolution crystal structure of the complex. However, their work employs a truncated non-lipidated MetQ mutant, which was engineered to artificially increase its affinity to MetNI (1,2). Thus, our findings are novel and furthermore, they call for a re-evaluation of the role of MetQ lipid anchor for complex stability and substrate transport. In our Results, we present data showing the importance of MetQ lipidation. We co-express his-tagged MetNI with either wildtype MetQ or mutant MetQ-C23A with a disrupted lipobox. We find that only wild-type MetQ co-purifies with MetNI, thus demonstrating the importance of MetQ lipidation for the formation of the MetNI-Q complex.

As an additional response to the concern of novelty – and to highlight the advantage of the peptidisc over detergent – we report for the first time an interaction of the MetNI transporter with the lipoprotein NlpA (also termed Lipoprotein 28). This interaction was identified in our pairwise interaction dataset with a high degree of confidence (>80%). Using SILAC AP/MS, we provide direct evidence that MetNI is interacting with NlpA. Interestingly, the same experiment in detergent fails to detect this interaction. Therefore, we confirm using two different methods (AP/MS and PCP-SILAC) that our Peptidisc workflow is able to reveal novel protein interactions. Interestingly, an earlier study back in 2003 has suggested that NlpA can deliver methionine to the MetNI transporter in a MetQ-deleted strain (7). This biological finding supports the relevance of the physical interaction between NlpA and MetNI.

To provide additional examples that interactions are preserved in the peptidisc, we employ the Sec translocon. Our peptidisc interaction list reveal a number of known and unknown interactors of the Sec translocon at high degree of precision (>75%). These include the membrane-bound chaperones YfgM and PpiD. To validate these interactions, we express the SecY complex in SILAC conditions and performed AP/MS in both detergent and peptidisc. The interactions between the SecY complex and the membrane-bound YfgM and PpiD are well preserved in peptidisc but are largely disrupted in detergent.

Our peptidisc data list also reveals surprising correlation between the Sec translocon subunits and the outer membrane Bam complex. Interactions between SecYEG and the BamA, BamC and BamD subunits in particular are present in our interaction list at high precision. Our SILAC AP/MS data now provides direct supporting additional evidence for this remarkable interaction. There is also strong enrichment of the porin OmpA along BamB and BamC subunits. As above with YfgM and PpiD, the interaction between the Sec translocon and the Bam complex is not evident when the same experiments are performed in detergent.

The BamABCDE complex is also employed as a third validation target. We expressed BamA in SILAC labeling conditions and performed AP/MS in both detergent and peptidisc. As expected, we observe enrichment of all subunits of the Bam complex in peptidisc, which is not the case in detergent. We also observe enrichment of the inner membrane proteins YidC and YajC, both of which are ancillary subunits of the Sec translocon. This later result corroborates our earlier AP/MS results with the Sec translocon.

Analysis of the Bam pairwise interaction dataset indicates two other interactors, RcsF and OmpA. An interaction of the Bam complex-RcsF has been previously reported using indirect methods such as genetic complementation and in vivo photo-crosslinking, but has not been biochemically isolated (5,6). Consistent with these previous studies, both OmpA and RcsF are significantly enriched along with BamA in our peptidisc SILAC AP/MS. When the same BamA pulldown experiment is performed in detergent, only the BamC and BamD subunits appear enriched – suggesting that the Bam complex is somewhat labile in detergent as previously reported (8). Additionally, the only other interactor that is significantly enriched is RcsF but not OmpA. These observations highlight again the advantages of the peptidisc method for preserving and identifying transient interactions.

2) The authors use 0.8-1% detergent (DDM) for 1 hour, yet claim this is a short time. However, most membrane proteins are fully extracted in a shorter time. Along the same lines, the authors show that the SEC and profiles obtained in DDM and in peptidiscs are similar. Since the use of detergent alone is simpler than that of the peptidiscs, it remains unclear how the presented approach is advantageous.

The time exposure (0.8-1% DDM for 1 hour) correspond to the initial solubilization step, which is performed at high protein and lipid content. The subsequent fractionation step is performed without detergent. This treatment is therefore short compared to the time exposure that occurs during fractionation in detergent conditions (i.e. several hours with continuous input of new micelles during elution, thereby increasing protein delipidation and subunits dissociation). We have revised the manuscript to clarify this point.

We show in Figure 2 that the protein profiles in detergent DDM and in peptidisc are visually similar, as assessed by SDS-PAGE and Coomassie staining. However, mass spectrometry analysis of the detergent and peptidisc samples reveals significant differences, that is 125 IDs in detergent versus 162 IDs in peptidisc for the most protein-rich fraction. This lesser identification number in detergent is likely due to the acetone precipitation step that is required before MS analysis. Since peptidisc does not require removal of detergent before MS, the acetone step in omitted, and the method yield higher protein ID number which is especially important when dealing with low abundance or hard to detect membrane proteins.

To support this claim, the authors need to show a complete and rigorous comparison between DDM and peptidiscs. How many of the interactions identified by the peptidiscs approach would be lost if detergent only is used? The authors are encouraged to equally minimize the detergent exposure time in these experiments.

As mentioned above, we have included side-by-side comparisons to support our claim that the peptidisc is superior to detergent. Specifically, we have included SILAC-based AP/MS experiments to compare in a quantitative manner the detergent and peptidisc methods. From these additional datasets, it is evident that the peptidisc preserves interactions that are dissociated or simply lost due to the detergent treatment.

3) The authors use a detergent (DDM) concentration that at the conditions used (temperature, salt concentration) is likely above the CMC. As a result, the detergent is probably concentrated along with the peptidisc library. The authors need to measure the DDM concentration during, and at the end of, the concentration step to convince the reader that their approach is indeed superior to the use of detergent alone.

The peptidisc library is water-soluble since its membrane protein content can be fractionated by size-exclusion chromatography in the complete absence of detergent (Figure 2). We also show that the peptidisc library can be visualized on native-PAGE, in contrast to the detergent sample that forms large aggregate at the top of the gel during migration (Figure 2E, compare lane 1 to lane 2). This experimental evidence allows to conclude that detergent has been removed and replaced by the peptidisc during the reconstitution step using filtration. Evidently, the pore of the filtration device is large enough to allow detergent monomers to pass through. The dynamic behavior of detergent micelles has well known and the auto-assembly process that occur upon removal of detergent during filtration is well documented with the nanodisc.

4) The authors prepare membrane fractions using ultracentrifugation, rather than density gradients or floatation assays. This leads to sedimentation of large proteins, aggregates, and complexes that are unrelated their analysis and confound the interpretation of the data. The authors must explain why they have chosen this approach and demonstrate that it does not undermine their analysis.

We agree, purification of the E. coli inner membrane traditionally employs fractionation using density gradient centrifugation after initial ultracentrifugation (3,9,10). These additional steps serve to remove outer membranes, peripherally associated proteins and large complexes such as ribosomes (10). We did not perform these additional fractionation steps because we wanted to cover and preserve as many cell envelope protein-protein interactions as possible; that is, we wanted to preserve peripheral and outer membrane proteins along with inner membrane proteins in our peptidisc libraries. We note that the peptidisc library preparations were fractionated by size exclusion chromatography to remove protein aggregates.

We did not find evidence that ribosomes undermine our analysis. In fact, large complexes such as the ribosomes are important markers for the downstream bioinformatic analysis, since these complexes are employed as gold standard reference. Our machine learning approach uses these gold standard references as training labels, and more training data provides a better-calibrated classifier and thereby improves our ability to predict interacting proteins. Filtering these nonmembrane complexes from the data would prevent their use as training data, thus reducing our ability to give confidence on our interacting membrane protein dataset.

5) Several of the concerns raised by the reviewers during the initial round of submission remain inadequately answered. For example, the identification of 13 Type-I SBPs remains unexplained. The authors need to directly and clearly address this and other concerns raised by the reviewers.

The abundant Type I SBPs have weak affinity for their cognate transporters, yet this affinity (and abundance) is sufficiently high to allow their co-sedimentation with the membrane fraction during initial ultracentrifugation. This explains their presence in our crude membranes stock which as mentioned above, are not purified by density gradient ultracentrifugation. The Type I SBP-transporter complex are unable to survive the downstream detergent solubilization and reconstitution steps, except when a lipid anchor exists – as is the case with MetQ. Thus, the vast majority of Type I SBPs are fully dissociated from their cognate transporters and therefore elute in the late fractions of the size exclusion chromatography.

6) The paper suffers from overstatement and inaccurate citations.

We have edited the manuscript to limit over-statements and we have asked expert co-authors on this manuscript to verify accuracy of the citations.

7) The supplementary tables should be reformatted. Now they are unintelligible.

This manuscript went thought two rounds of revisions which has led to inclusion of many additional tables. We have re-formatted some of these supplementary tables to improve clarity.

References:

1) Nguyen, P. T., Li, Q. W., Kadaba, N. S., Lai, J. Y., Yang, J. G., and Rees, D. C. (2015) The contribution of methionine to the stability of the Escherichia coli MetNIQ ABC transporter-substrate binding protein complex. Biol Chem 396, 1127-1134,

2) Nguyen, P. T., Lai, J. Y., Lee, A. T., Kaiser, J. T., and Rees, D. C. (2018) Noncanonical role for the binding protein in substrate uptake by the MetNI methionine ATP Binding Cassette (ABC) transporter. Proc Natl Acad Sci U S A 115, E10596-E10604.

3) Maddalo, G., Stenberg-Bruzell, F., Götzke, H., Toddo, S., Björkholm, P., Eriksson, H., Chovanec, P., Genevaux, P., Lehtiö, J., Ilag, L. L., and Daley, D. O. (2011) Systematic analysis of native membrane protein complexes in Escherichia coli. J Proteome Res 10, 1848-1859.

4) Götzke, H., Palombo, I., Muheim, C., Perrody, E., Genevaux, P., Kudva, R., Müller, M., and Daley, D. O. (2014) YfgM is an ancillary subunit of the SecYEG translocon in Escherichia coli. J Biol Chem 289, 19089-19097.

5) Hart, E. M., Gupta, M., Wühr, M., and Silhavy, T. J. (2019) The Synthetic Phenotype of Δ. MBio 10.

6) Konovalova, A., Perlman, D. H., Cowles, C. E., and Silhavy, T. J. (2014)

Transmembrane domain of surface-exposed outer membrane lipoprotein RcsF is threaded through the lumen of β-barrel proteins. Proc Natl Acad Sci U S A 111, E4350-4358.

7) Zhang, Z., Feige, J. N., Chang, A. B., Anderson, I. J., Brodianski, V. M., Vitreschak, A. G., Gelfand, M. S., and Saier, M. H. (2003) A transporter of Escherichia coli specific for L- and D-methionine is the prototype for a new family within the ABC superfamily. Arch Microbiol 180, 88-100,

8) Gu, Y., Li, H., Dong, H., Zeng, Y., Zhang, Z., Paterson, N. G., Stansfeld, P. J., Wang, Z., Zhang, Y., Wang, W., and Dong, C. (2016) Structural basis of outer membrane protein insertion by the BAM complex. Nature 531, 64-69,

9) Stenberg, F., Chovanec, P., Maslen, S. L., Robinson, C. V., Ilag, L. L., von Heijne, G., and Daley, D. O. (2005) Protein complexes of the Escherichia coli cell envelope. J Biol Chem 280, 34409-34419.

10) Papanastasiou, M., Orfanoudaki, G., Koukaki, M., Kountourakis, N., Sardis, M. F., Aivaliotis, M., Karamanou, S., and Economou, A. (2013) The Escherichia coli peripheral inner membrane proteome. Mol Cell Proteomics 12, 599-610,

https://doi.org/10.7554/eLife.46615.025

Article and author information

Author details

  1. Michael Luke Carlson

    Life Sciences Institute, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft
    Contributed equally with
    R Greg Stacey and John William Young
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3807-6516
  2. R Greg Stacey

    Michael Smith Laboratory, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Contributed equally with
    Michael Luke Carlson and John William Young
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-4496-8131
  3. John William Young

    Life Sciences Institute, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Contributed equally with
    Michael Luke Carlson and R Greg Stacey
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-3541-509X
  4. Irvinder Singh Wason

    Life Sciences Institute, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Data curation, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
  5. Zhiyu Zhao

    Life Sciences Institute, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Data curation, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
  6. David G Rattray

    Michael Smith Laboratory, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Data curation, Investigation
    Competing interests
    No competing interests declared
  7. Nichollas Scott

    Michael Smith Laboratory, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Data curation, Investigation, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2556-8316
  8. Craig H Kerr

    Michael Smith Laboratory, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Data curation, Formal analysis, Investigation, Methodology
    Competing interests
    No competing interests declared
  9. Mohan Babu

    Department of Biochemistry, Faculty of Science, University of Regina, Regina, Canada
    Contribution
    Data curation, Formal analysis, Writing—review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4118-6406
  10. Leonard J Foster

    Michael Smith Laboratory, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8551-4817
  11. Franck Duong Van Hoa

    Life Sciences Institute, Department of Biochemistry and Molecular Biology, Faculty of Medicine, University of British Columbia, Vancouver, Canada
    Contribution
    Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    fduong@mail.ubc.ca
    Competing interests
    has a website which sells the peptide used in this study
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-7328-6124

Funding

Canadian Institutes of Health Research

  • Mohan Babu
  • Leonard J Foster
  • Franck Duong Van Hoa

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Senior Editor

  1. Richard Aldrich, The University of Texas at Austin, United States

Reviewing Editor

  1. Nir Ben-Tal, Tel Aviv University, Israel

Reviewer

  1. Samuel Wagner

Publication history

  1. Received: March 7, 2019
  2. Accepted: July 30, 2019
  3. Accepted Manuscript published: July 31, 2019 (version 1)
  4. Version of Record published: August 16, 2019 (version 2)

Copyright

© 2019, Carlson et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 2,084
    Page views
  • 432
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Biochemistry and Chemical Biology
    2. Cell Biology
    Leslie A McCauliff et al.
    Research Article Updated
    1. Biochemistry and Chemical Biology
    2. Microbiology and Infectious Disease
    James Lee et al.
    Research Article