Modified Preparation of Bait mRNA Templates

We prepared 68 bait proteins representing 50 human TF-related proteins (Table S1). All 68 cDNA fragments (full length and/or domain portions of the TF-related proteins) were amplified by a four-step PCR with exTaq (Takara Bio) using a Qiagen Biorobot 8000. The PCR was performed as shown in Figure S1 and Tables S1, S2, S3. The mRNA templates were prepared with a RiboMAX Large Scale RNA Production System-SP6 (Promega) and m7G(5′)ppp(5′)G RNA Capping Analog (Invitrogen Corp., Carlsbad, CA, USA)[19]. The mRNAs were detected by routine western blot analysis using the anti-T7 antibody. Ninety-six percent of the bait proteins were expressed in the in vitro translation system using this method (Table S1). See ‘Supporting Methods’ (Text S1) for additional details.

In vitro Parallel Auto-Selection Using IVV

A commercially available human brain cDNA library (the BioChain Institute, Inc.) was prepared for parallel auto-selection to be used as prey in large scale 96-well plate assays carried out by a Qiagen Biorobot 8000, according to a previously described method [19]. As directed by the reported method [20], a PEG Puro spacer was synthesized on a large scale by Invitrogen Japan K.K., Tokyo, Japan and Takara Bio Inc., Otsu, Japan. The human brain cDNA library to be used as prey was prepared according to a randomly primed cDNA library [19]. This approach reduces bias in the cDNA library. Moreover, interference from UTRs was not an issue in this system due to the use of an in vitro translation system. mRNA templates used as bait and prey were co-translated in a wheat germ extract (Zoegene Corporation, now Molecuence Corporation) for 1 h at 26°C in 96-well plates using a Qiagen Biorobot 8000. After six rounds of selection, the obtained sequences were identified by Takara Bio Inc., Otsu, Japan, and Shimadzu Corporation, Kyoto, Japan. A mock experiment was run without bait protein as a negative control to eliminate technical false positives in the IST analysis. See ‘Supporting Methods’ (Text S1) for additional details.

IST Analysis

Determination of the interaction sequence tags (ISTs; Table S4) obtained from a randomly primed prey library was performed by Takara Bio Inc., Otsu, Japan and Shimadzu Corporation, Genomic Research Center, Kyoto, Japan. Using the IVV analysis system (IWAS)[19] developed by Fujitsu Limited, genes corresponding to each prey sequence were assigned by a BLASTN homology search against the coding sequences of the NCBI human RefSeq. Sequences with an E-value ≤1.0E-5 and a match length ≥30 bp were assigned as positive matches. Frame shift mutants were excluded from our analysis for the purpose of clarity. Finally, ISTs were classified into one of the following three categories: Class 1 ISTs were defined as those sequences overlapping with other prey sequence(s) obtained from the same bait protein (without distinguishing between partial and full-length proteins), excluding those that overlap prey sequences obtained from negative control experiments (mock experiments); Class 2 ISTs were defined as those ISTs that did not overlap with other prey sequences obtained from the same bait protein; and Class 3 ISTs were defined as the sequences that overlap with prey sequences obtained from mock experiments. Consequently, as shown in Figure S2, 1,972, 830, and 185, ISTs were obtained for classes 1, 2 and 3, respectively.

Definition of Interacting Regions (IR) and Clusters

In order to identify important protein interaction regions (e.g., interacting domains, recognition motifs), we assessed the overlap among the IST regions of proteins obtained from common bait proteins. In the case of inclusive relationships between different overlapping regions, the innermost region was selected as the minimum length region. The selected regions were at least 3 amino acids in length. We defined a unique cluster as a group of ISTs sharing minimum length regions Moreover, we defined regions within the lateral ends of each cluster as maximum length regions, corresponding to an ‘interacting region’ (IR) containing the interacting domains and/or motifs (Figure S3 and Table S5).

Verification of PPIs (IRs) by Pull-Down and Real-Time PCR Assays

In order to confirm interaction pairs (PPIs and IRs), pull-down experiments were performed as described in previous studies [19], [20], and the precipitates were analyzed by 10-17.5% SDS-PAGE. Real-time PCR was also performed to evaluate the interaction pairs. Briefly, each 20 µl reaction containing 5 ng of DNA template from the prey library obtained during each round of selection, gene-specific primers and SYBR Green PCR Master Mix (Applied Biosystems) was submitted to real-time PCR in a 7300 Real-Time PCR System (Applied Biosystems). Gene-specific primers corresponding to prey sequences were designed by Nihon Gene Research Laboratories.

Pfam Search and Determination of Protein Contact Regions

Domains and motifs were identified through the following three steps: (1) BLAST search of each IVV sequence against the human RefSeq protein database; (2) extraction of the RefSeq protein fragment corresponding to the hit region; and (3) searching for the domains and motifs in each protein fragment. The “hmmpfam” was used to identify known protein domains and motifs in the Pfam database [28] (http://pfam.janelia.org/). Using all class 1 data that overlapped with Pfam domains, we demonstrated good alignment with the corresponding full-length proteins (Figure S5, S6 and Table S6). In order to determine the amino acids responsible for the interaction between two proteins, the distance between the main chain atoms of the two proteins was considered. We defined interacting amino acids as those amino acids possessing atoms within 4.0 Å of each other (Figure 2A). All 3D protein structures presented in this paper were created using PyMol (http://pymol.sourceforge.net).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Validation of IR data obtained following IVV selection.

(A) Left: IST density of IRs on the 3D protein structures of AP-1. ISTs obtained as prey in selections using FOS and JUN were mapped onto the 3D structure of AP-1 (PDB: 1A02, chain F and J; [55]. Right: Contact regions of AP-1. All amino acids of one protein within 4.0 Å of the other protein are colored blue. IST densities are ranked and colored on a scale of 1 to >5 according to the number of ISTs at each amino acid position. (B) Left: MAX interacting regions in FTH1. Twenty-four ISTs derived from FTH1, obtained using MAX as bait, were mapped onto the 3D structure of FTH1 (PDB:1FHA; [56]). Right: Pull-down assay to evaluate the MAX/FTH interaction. ‘IR’ and ‘full’ correspond to the assays performed with the IR (region: 124.176) and full-length FTH, respectively, as bait. Full-length MAX was used as prey. (C) Left: SMAD2 interacting regions in RHOA. An IST derived from RHOA, obtained using SMAD2 as bait, was mapped onto the 3-D structure of RHOA (PDB: 1OW3, chain B; [57]). Right: Pull-down assay to evaluate the SMAD2/RHOA interaction. ‘IR’ and ‘full’ correspond to the assays performed with the IR (region: 38..63) and full-length RHOA, respectively, as bait. The (522..1401) region of SMAD2 was used as prey.

https://doi.org/10.1371/journal.pone.0009289.g002

Mapping of the Transcription Factor (TF) Network at the Protein and IR Levels

The interaction network of human TF-related proteins was constructed by merging our IVV core data set and LC interactions (ftp://ftp.ncbi.nlm.nih.gov/gene/GeneRIF/interactions.gz) for the 50 TF-related proteins that were used as bait in our experiment (Figure S7 and Table S7). We also generated refined PPI networks based on IR data. An IR-level network graph is different from a classical PPI network graph because it contains intermediate nodes (i.e., IR nodes) between each interacting protein pair. All IR nodes have both intra- and inter-protein edges. Intra-protein edges reflect the fact that an IR is part of a protein. On the other hand, inter-protein edges represent interactions between different molecules (e.g., a bait protein and an IR). All of the network graphs were produced using Cytoscape [34]. Cytoscape files for Figures 3A and S12 are available upon request (contact E.M-S[nekoneko@educ.cc.keio.ac.jp]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. A TF network at the IR level developed using IVV data.

(A) Graphic expression of the PPI network at the IR level. Interacting interfaces of the proteins, determined as IRs by IVV experiments, are drawn on the graph as diamond-shape nodes (IR nodes). Broken and solid lines indicate ‘intra-’ and ‘inter-’ protein edges, respectively. The graph contains 1,572 nodes (842 IR nodes and 730 protein nodes) and 842 intra-protein edges. Note that overlapping IRs are merged into a single node in the constructed network. Also see Figure S12. (B) An example of an underlying network graph at the IR level. Graphical expression of the FOS network at the protein level (upper). PPIs are simply expressed by nodes indicating proteins and edges that connect them. Graphical expression of the FOS network at the IR level (lower). A leucine zipper region of the FOS protein exclusively interacts with leucine zipper regions of other proteins (JUN, JUNB, JUND and ATF2). In addition, a region distinct from the leucine zipper in the FOS protein interacts with SMAD2.

https://doi.org/10.1371/journal.pone.0009289.g003

Analyses of IRs with Multiple Interaction Partners

We defined IRs that shared 50% or more of their targets with other IRs obtained from different bait proteins as IRs with multiple partners. This class of overlapping IRs was unified as a single IR node in the network graphs. Proteins interacting with such IRs may compete with each other.

Analyses of Intrinsically Disordered Regions

Intrinsically disordered regions of each human RefSeq were predicted using DISOPRED2 [30]. The default false positive rate (5%) of DISOPRED2 was used as a disorder/order classification threshold. Disordered regions in each IR were identified by comparing the positions of IRs and disordered regions assigned to the corresponding RefSeq. The proportion of disordered (or ordered) regions in each dataset (Table S11) was calculated as the total number of amino acid residues residing in disordered (or ordered) regions divided by the number of total amino acids in the dataset. The statistical significance of the differences between each group was determined using Fisher's exact probability test in R (http://www.R-project.org).

Additional Methods

The core data set was analyzed for correlations with biological attributes, such as network properties (scale-free, etc.), expressional correlations, gene classification according to GO, and tissue-specific correlations. Descriptions of these network analyses are available in the Supporting Methods section (Text S1).

Large-Scale IR Data Sets Obtained for 50 Human TF-Related Proteins by mRNA Display

For this large-scale mRNA display study, 68 bait proteins were prepared either from full-length proteins or from protein domains of 50 human TF-related proteins (Supporting Data II in Text S1). The display technique consisted of an in vitro parallel automated selection of IVV. The cell-free wheat germ translation system utilized in this study exhibited an excellent bait expression rate (96%) for human TF-related proteins. This system was modified from a previously published bait preparation [19] (Figure 1A, ‘Confirmation of in vitro bait expression’; also see Supporting Data II in Text S1, Figure S1 and Table S1). The success rate of the in vitro PPI selection for TFs was 99% (Table S1). These results suggest that this large-scale in vitro system provides highly effective protein expression and selection of TF proteins (Supporting Data II in Text S1). In the course of selection, we obtained 2,987 interaction sequence tags (ISTs), which were subjected to BLASTN searches to identify their corresponding proteins. The sequences were amplified by RT-PCR and sequenced (Figure 1A; Supporting Data III in Text S1). Of the 2,987 analyzed ISTs, 1,127 IRs were identified among the prey proteins (Supporting Data III in Text S1). The 2,987 ISTs (1,125 IRs and 1,098 PPIs) were subdivided into three classes (Figure 1B, bottom; see also Methods and Supporting Data III in Text S1). The IVV core data set (966 IRs; 943 PPIs for 730 proteins) was composed of classes 1 and 2. Class 3 ISTs (potential false-positives) were excluded from the dataset. However, data for classes 1 and 2 are presented (Table S4 and S5). The IVV core data are also available from the Genome Network Platform (http://genomenetwork.nig.ac.jp/index_e.html).

Confirmation of the Core Data and Testing for False Positives

To confirm the reliability of the core data obtained by IVV parallel auto-selection, we carried out pull-down assays of 100 PPIs (IRs) using the C-terminal protein labeling method ([19], [20]; Figure 1B; Supporting Data IV in Text S1, Figure S4A, and Table S5). As expected, class 1 IRs displayed the highest confidence level (80%), likely due to the multiple ISTs for each prey protein. Further, class 3 IRs showed a much lower confidence level (8%), likely due to agreement with ISTs in a technical false positive data set in the negative control experiment (Figure 1B, bottom). Accordingly, we defined the IVV core dataset as all members of classes 1 and 2. The average verification rate for the IVV core data was 70% (Figure 1B, bottom). Because the pull-down assay is not a definitive verification experiment, real-time PCR [19] was used as an alternative assay to confirm the enrichment of prey genes in the IVV libraries. The results of the real-time PCR experiments were generally consistent with those of the pull-down assay (Supporting Data IV in Text S1 and Figure S4B). Together, these results, which are similar to those observed in large-scale Y2H experiments [9], [10], indicate that the large-scale in vitro experimental data are reliable and valid. Although the large mRNA moiety of IVV would likely interfere with protein interactions and in vitro folding would occur differently, we believe that interaction with and folding are less difficult to demonstrate for a protein domain than for a full-length protein. For this reason, the IVV library is composed mostly of parts of full-length proteins as a randomly primed prey library. Further, we previously attempted to assess the reliability of mRNA display both in vitro [19] and in vivo [26]. Experiments using ‘protein’ (not hybrid molecule) pull-down and co-immunoprecipitation assays demonstrated 80% accuracy, even in in vivo verifications. Note that the verification rate of pull-down assays obtained for the IVV method represents a minimum value because the method detects both direct and indirect interactions [19].

Network Graph and False Negatives

The network graph (Supporting Data VI in Text S1 and Figure S7) shows the union of the IVV core data set (943 PPIs, 730 proteins) and the literature-curated (LC) data set for the 50 human TFs (1,240 PPIs, 796 proteins). This network contains 1,410 nodes (proteins) and 2,172 edges (PPIs; Table S7). The IVV core data set contains 82% more nodes than the LC data set (Table S8). It is of particular interest that we detected an average of 14 PPIs per bait protein in this IVV experiment, whereas previous Y2H experiments only detected an average of 3 PPIs per bait protein ([10], [35]; Supporting Data VI in Text S1). This difference suggests that Y2H might generate more false negative data than IVV in large-scale experiments. The difference between IVV and Y2H probably reflects the differences in the nature of the experiments (i.e., the difference between in vitro and in vivo binding behavior). The in vitro IVV selection system can utilize a larger library size and can also evaluate cytotoxic and self-activating TF proteins (Supporting Data I and II in Text S1). We suggest that the systematic collection of IR data from ISTs obtained with IVV, in addition to data collected by other methods such as Y2H, will be valuable for refining our understanding of protein interactions.

Network Analyses of Biological Attribute

We found that the topological properties of the IVV core data network are similar to those of other interactome networks [9], [10]. Specifically, the IVV core data network shows a degree distribution that is approximately power-law degree, as well as a hierarchical organization and a tendency for highly connected (hub) proteins to interact with less highly connected proteins, as assessed by the degree distributions ([36]; Supporting Data VII in Text S1, Figure S8, and Table S9). In addition, we found that the interacting protein pairs identified by analysis of gene expression data for various cells and tissues using SymAtlas (http://wombat.gnf.org/downloads/GNF1Hdata.zip; [37]) were better correlated than would be expected by chance (random pairs; SupportingData VIII in Text S1 and Figure S9). These results suggest that selection of the prey library is very important for the generation of cellular networks. As we employed human TF proteins as bait, the data showed more frequent functional correlations with transcription-related terms in the Gene Ontology (GO) annotations (Supporting Data IX in Text S1 and Figure S10). Because a human brain cDNA library was used as prey, brain-specific PPIs were much more frequently detected in our experimental data than other tissue-specific PPIs (Supporting Data X in Text S1, Figure S11, and Table S10). Despite the fact that the IVV core data were generated in an in vitro experiment, it showed biological network properties (Supporting Data VII in Text S1, Figure S8, and Table S9) and biological expression correlations (Supporting Data VIII in Text S1 and Figure S9) similar to those found in previous large-scale in vivo experimental data sets [9], [10].

Pfam Search and the Accuracy of IR Data As Functional Domains

Further analysis was performed to confirm the reliability and accuracy of IRs in the IVV core data set. A Pfam search ([28]; Supporting Data V in Text S1) was carried out to identify known domains within the IRs defined by IST analyses (Supporting Data III in Text S1). Based on this analysis, we identified 24% of the known domains (Figure 1B bottom, class 1, IRs (known)) within the class 1 data (136 IRs), indicating that Pfam domains are more concentrated in class 1 IRs than in class 2 IRs (Figure 1B bottom; Table S6). To confirm the accuracy of the IR data for FOS/JUN (activating protein-1; AP-1)-interacting domains aligned with ISTs (Figure S6), we compared the denser regions of IRs in the alignment of ISTs of FOS/JUN (Figure 2A, red and orange) with contact regions identified by the evaluation of 3D FOS/JUN structural data (Figure 2A, blue). We confirmed a precise agreement between the denser regions of IRs in the alignment of ISTs and the contact regions in the 3D structure data (Figure 2A, blue). Also, evaluation of all class 1 data that overlap with Pfam domains demonstrated good alignment with corresponding full-length proteins (Figure S6), demonstrating the reliability of the IR data (Figure 2A). Thus, the IR data provide reliable and accurate information about binding interfaces (functional domains) involved in protein interactions (Supporting Data V in Text S1).

Pull-down assays were employed to validate the IRs within Pfam domains. Figs. 2B and C show representative assays. Proteins evaluated by pull-down were selected from both class 1 and class 2. We identified MAX/FTH1 from class 1 and SMAD2/RHOA from class 2 as PPI pairs with well-conserved domains (Table S6). MAX and FTH1 contain helix-loop-helix (HLH) and ferritin domains, respectively (Table S6). SMAD2 and RHOA contain MH2 and Ras domains, respectively (Table S6). We reciprocally confirmed PPIs for both MAX/FTH1 (Figure 2B) and SMAD2/RHOA (Figure 2C) with C-terminal labeling pull-down assays using both protein domains and full-length proteins. The domain-domain interaction between HLH and ferritin has also been identified in Drosophila melanogaster (1 pair [6]). In addition, the interaction between MH2 and Ras has also been observed in other PPIs in humans (47 pairs [38] and an additional SMAD2/RAN pair in the IVV core data; Table S6) and other species (3 pairs [7]). Accordingly, it is likely that IR data will allow for the prediction of PPIs and domain-domain interactions (DDIs) based on domain information [39], [40]. We verified two interaction domains identified by an IVV experiment, including a globular domain in which the interaction region is clearly distinguished from the rest of the protein (Figure 2C).

Network Graph at the IR Level

In order to provide an overview of our experimental results, we drew a refined interaction network graph at the IR level containing 1,572 nodes (842 IR nodes and 730 protein nodes) and 842 intra-protein edges (Figure 3A; Figure S12). Note that the overlapping IRs are merged into a single node in the network diagram. We have derived an example (FOS) of a network graph at the protein level (Figure 3B, left) and the IR level (Figure 3B, right). The underlying network graph of FOS at the IR level shows two different interactions (two different IR nodes), AP-1 (FOS/JUN) and FOS/SMAD2 (Figure 3B, right), which is analogous to the well-known AP-1/Smad3 complex [41], [42], [43]. The interaction between FOS and JUN is also well known [44]. We have confirmed the newly identified interaction between FOS and SMAD2 via a pull-down assay using the C-terminal protein labeling method (data not shown). FOS contains an IR (IR_17) that competitively interacts with multiple partners (JUN, JUNB, JUND, or ATF2), as well as an IR (IR_435) that exclusively interacts with one partner (Figure 3B, right). It is important to distinguish between interfaces with competitive and non-competitive properties to understand the dynamics of cellular networks in detail [27]. Thus, once we obtain an underlying network graph at the IR level (Figure 3B, right) instead of a classical network graph at the protein level (Figure 3B, left), we can identify refined interaction network interfaces and the biological implications of those results for TF complexes.

Network Analysis of Intrinsic Disorder Rate at the IR Level

Following Fischer's lock-and-key proposal, many counterexample proteins have been identified that require a lack of three-dimensional structure in order to function. The importance of the lack of three-dimensional structure (related to disordered regions) in protein interactions can no longer be ignored [29]. We examined the rates of ordered/disordered regions [29], [31] in the IVV core data set (943 PPIs; 966 IRs) using DISOPRED2 [30] in order to obtain a detailed understanding of the types of structural interactions of human TF complexes (Supporting Data XI in Text S1 and Table S11). Figure 4 shows the proportions of intrinsically disordered regions (residues) in various groups of IRs or proteins. The proportion of disordered regions (residues) in the IVV core data was significantly higher than that in the human Refseq for both regions and protein levels (p<2.2e–16 in each comparison); a more distinctive difference was observed for IRs (Figure 4, IVV core (IR)). A similar trend was observed in three previously reported transcription factor data sets [45]. In a detailed comparison within the IVV core data set, we observed more disordered regions (residues) in class 2 IRs than in class 1 IRs (p<2.0e-06). In addition, IRs identified by the Pfam search program showed a higher proportion of ordered structures than any other set of IRs (p<2.0e-06 in every comparison). Prey proteins, which bind to bait proteins with higher affinity, are expected to be in class 1, a group of IRs consisting of multi-targeted prey sequences. In principle, since this experiment is based on affinity selection, stronger binding to bait proteins should correspond to a higher probability of detection. In addition, a wider variety of proteins (genes) are categorized as class 2 (830 proteins, Figure 1B, bottom) than class 1 (136 genes, Figure 1B, bottom) when the 50 human TFs are used as bait. Taken together with the frequent observation of disordered regions in class 2 proteins, these results indicate that many interacting partners of the 50 TFs appear to display unstable interactions mediated through their disordered regions. A limited number of partners were identified with stable interactions involving ordered regions. In the in vivo situation, various combinations of interactions could occur depending on the physiological context (location, time, etc.). Thus, we consider that disordered regions not only of TFs, but also those of the interacting partners of TFs are employed as interaction interfaces to achieve the dynamics necessary for formation of diverse TF complexes. These findings suggest that IVV technology can identify both stable and less stable interactions involving disordered regions. The current understanding regarding the affinities of binding mediated by disordered regions is that these affinities are weak in terms of entropy [46]. In fact, many low-affinity (flexible) interactions are included in the IVV core data set, and utilization of the IVV method almost doubles the size of the identified interactome network (the IVV method produced 82% more nodes than did the LC data set (Table S8)). On the other hand, a relatively low percentage of known domains in the IVV core data set (10%, Figure 1B, bottom) are derived from the quantitative dominance of low-affinity interactions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Analysis of the rates of disordered regions.

The proportions of intrinsically ordered and disordered regions in 13 datasets consisting of IR (7 datasets) and Protein (6 datasets) were analyzed by DISOPRED2 [30] as follows: IR (IR-level data); IVV Core; IVV (class 1); IVV (class 2); Pfam hit (a set of IRs hit by Pfam search); Multiple partners (IRs obtained from multiple bait proteins); Single partners (IRs obtained from a single bait protein) and Refseq (random regions) or Proteins (protein-level data); IVV Core; LC (a set of known interacting partners for 50 bait proteins); ‘Transcription regulator activity’ (a set of proteins for which GO:0030528 is assigned); ‘Transcription cofactor activity’ (a set of proteins for which the GO:0003712 is assigned); ‘Transcription factor activity’ (a set of proteins for which GO:0003700 is assigned); and All RefSeq: all human RefSeqs. The dataset of random regions was created by random selection of protein regions (n = 10000) from the human RefSeq that together correspond to the same length distribution as that of detected IRs. Information about the assignment of GO identifiers for proteins can be obtained from the Gene Ontology Web site (http://www.geneontology.org).

https://doi.org/10.1371/journal.pone.0009289.g004

We further examined the characteristics of IRs displaying multiple interaction partners in the IVV experiment (Figure 3). We speculate that these multi-targeted IRs interact with numerous other partners in a cellular context. Figure 5 shows plots of the number of interaction partners for each prey protein. The mean number of known interaction partners for proteins containing multiple interacting IRs was 11.1, significantly higher than the mean for other prey proteins of 6.8 (Wilcoxon rank-sum test, p = 0.003). In addition, the proportion of disordered regions (residues) in IRs with multiple partners was 59%, significantly higher than that for any other dataset in our analysis (Figure 4; p<0.0001 for every comparison). These findings indicate that disordered IRs can provide the ability to interact with multiple different proteins. This assumption is consistent with the results of several previous studies on the potential of disordered regions frequently observed in so-called hub proteins [32], [47], [48], [49], [50]. In addition, this finding supports the hypothesis that transcriptional regulatory proteins frequently bind to various partners [51]. The IRs with multiple partners detected in this experiment might also function as flexible interfaces that mediate interactions among various compatible partners.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. IR properties and the number of known interaction partners.

Counts of LCI for each prey gene (protein) were plotted for two datasets: prey proteins having IRs obtained from multiple bait proteins (multiple partners), and proteins having IRs obtained from a single bait protein (single partner).

https://doi.org/10.1371/journal.pone.0009289.g005