- Research article
- Open Access
Context-specific interactions in literature-curated protein interaction databases
BMC Genomics volume 19, Article number: 758 (2018)
Databases of literature-curated protein-protein interactions (PPIs) are often used to interpret high-throughput interactome mapping studies and estimate error rates. These databases combine interactions across thousands of published studies and experimental techniques. Because the tendency for two proteins to interact depends on the local conditions, this heterogeneity of conditions means that only a subset of database PPIs are interacting during any given experiment. A typical use of these databases as gold standards in interactome mapping projects, however, assumes that PPIs included in the database are indeed interacting under the experimental conditions of the study.
Using raw data from 20 co-fractionation experiments and six published interactomes, we demonstrate that this assumption is often false, with up to 55% of purported gold standard interactions showing no evidence of interaction, on average. We identify a subset of CORUM database complexes that do show consistent evidence of interaction in co-fractionation studies, and we use this subset as gold standards to dramatically improve interactome mapping as judged by the number of predicted interactions at a given error rate.
We recommend using this CORUM subset as the gold standard set in future co-fractionation studies. More generally, we recommend using the subset of literature-curated PPIs that are specific to the experimental context whenever possible.
Proteins perform the majority of cellular functions necessary for life. Nearly all individual proteins are modular components of larger macromolecular structures, i.e. protein complexes, and the exact role of a protein within a cell is controlled by its interactions with co-complex members. Uncovering which proteins interact, i.e. the interactome, is therefore central to understanding the molecular mechanisms of life.
This task is complicated by a combinatorial explosion, however: a proteome containing 20000 proteins has nearly 200 million potential pairwise interactions and many more higher order complexes. High-throughput techniques that analyze thousands of proteins simultaneously with minimal bias offer a solution to this problem . For example, PCP-SILAC (protein correlation profiling–stable isotope labeling of amino acids in cell culture), a co-fractionation (CF) technique, separates protein complexes into fractions according to their size (rotational cross-section), and associates proteins whose amounts are correlated between fractions. As each fraction is analyzed with mass spectrometry, PCP-SILAC and other CF techniques can detect tens of thousands of interacting proteins [2,3,4,5,6,7,8]. In order to separate signal from noise, it is common for high-throughput protein interactome studies to consult databases of known, unequivocal interactions (“gold standards”) [2, 3, 9, 10]. For example, co-fractionation studies often use gold standard interactions as training labels in a machine learning classifier [2, 3, 11]. Gold standards are also used to define false positive/negative and true positive/negative interactions in order to calculate common statistics such as precision, recall, and sensitivity [5,6,7, 11, 12].
Gold standard databases are assembled from different experiments and techniques, each with a unique set of biases. Since protein-protein interactions (PPIs) can be context-specific and transient, single datasets, which are typically generated by a single technique, can disagree with gold standards. This variability partly reflects true biological differences. For example, the majority of in vivo yeast PPIs were observed to depend on environmental and chemical conditions . Some assays also impose technical biases that limit detectable PPIs, such as a bias of some high-throughput techniques toward highly expressed or well studied protein pairs, or a bias against PPIs involving transmembrane proteins . Therefore, gold standard databases that include all interactions that can occur will fail to describe the subset of interactions that are either not occurring due to current experimental conditions, or that are unlikely to be detected due to technical limitations.
Therefore, a distinction should be made between the large, curated compilations of interactions across many studies, and the gold standard sets used as a reference for a single dataset. Our own focus has been on interactome mapping using co-fractionation, so here we quantify the proportion of gold standard interactions that fail to display any evidence for interaction in 20 co-fractionation datasets. Using a conservative measure of protein interaction, we find that between 19 and 55% of gold standard PPIs display no evidence of interacting. Across co-fractionation experiments, there is evidence that a subset of literature-curated complexes consistently co-fractionates, suggesting this subset would be a more appropriate gold standard reference set. Indeed, the number of predicted interactions at a given stringency increases dramatically when using this subset as a gold standard set. We recommend using this subset as the gold standard reference in future co-fractionation studies and, more generally, using experiment- and condition-specific gold standards whenever possible.
Discrepancies exist between gold standards and individual datasets
Using the CORUM database of protein complexes , we first examined the degree to which literature-curated PPIs were unsupported by data from single co-fractionation datasets. Many database PPIs show clear evidence of interaction, as shown by their tendency to co-fractionate for the entire chromatogram (Fig. 1a) or a portion of the chromatogram (Fig. 1b). However, other protein pairs from within a single CORUM complex show little evidence of interaction in certain experiments. For example, two chaperone proteins, HSP-90a (UniProt ID P07900) and BiP (P11021) are known to interact as part of a larger chaperone multiprotein complex  (CORUM complex “HCF-1”), yet there is little evidence that the two proteins co-fractionate in our data (Fig. 1c).
More broadly, across 20 PCP-SILAC co-fractionation datasets, the majority of random protein pairs do not co-fractionate, as quantified by anti-correlated fractionation profiles, a conservative measure of which protein pairs are non-interacting (Fig. 1d). While the majority of gold standard protein pairs have positively correlated co-fractionation profiles (black), 23% (34442/149477) are negatively correlated. All 20 datasets include a similar proportion of negatively correlated gold standard pairs (23 +/− 5%, mean +/− st.d.). This pattern is similar when co-fractionation is measured with Euclidean distance, another standard measure (Fig. 1e).
While the full set of CORUM complexes is a widely used gold standard [6, 7, 9,10,11], there are many other literature-curated interaction databases. In addition to CORUM, we examined nine databases of protein interactions [16,17,18,19,20,21,22,23,24] and two subsets of CORUM used previously as gold standards [2, 3]. These range from databases that include interactions from high-throughput experiments to manually curated databases composed exclusively of low-throughput experiments. All had anti-correlated protein pairs in our co-fractionation datasets (Fig. 2). As a baseline, 62% of all protein pairs, the large majority of which can be assumed to be non-interacting, were anti-correlated (Fig. 2, left). The proportion of anti-correlated pairs in gold standard sets ranged from 55% (HPRD) to 19% (CORUM). Restricting gold standard PPIs to those supported by two or more publications limits but does not eliminate uncorrelated protein pairs (Additional file 1: Figure S1). Therefore all interaction databases investigated here contain protein pairs that are not supported by our co-fractionation data, and comparisons to CORUM give a conservative estimate of the discrepancy between our data and interaction databases.
Since some PPI databases include evidence codes, we also analyzed whether the proportion of anti-correlated protein pairs differed as a function of evidence code (Additional file 1: Figure S2). Indeed, database PPIs with evidence codes relating to co-fractionation-like experiments (e.g. “molecular sieving”, “density sedimentation”) tended to have fewer anti-correlated pairs in our co-fractionation data than PPIs derived from other experiment types (e.g. “LUMIER”). This suggests that there is a subset of database PPIs that more accurately describes co-fractionation datasets, and that this subset tends to be derived from co-fractionation-like experiments.
Discrepancies are consistent within and between high-throughput techniques
A certain level of discrepancy between raw co-fractionation profiles and interaction databases should be expected, as all experimental samples undergo some degradation owing to the constraints imposed by each assay. But do individual datasets differ randomly or systematically from interaction databases? For certain gold standard complexes we see a strong tendency for co-complex members to co-fractionate, and, conversely, a strong tendency for other complexes to fail to co-fractionate (Fig. 3a). Across 20 co-fractionation datasets, we made 39846 pairwise comparisons between fractionation profiles of cytoplasmic ribosomal proteins, and 17616 pairwise comparisons of proteins in the C complex spliceosome. The collection of ribosomal gold standard interactions are significantly better correlated than chance (R = 0.64, chance R = 0.48, p = 0.005, permutation test; Fig. 3b), while the collection of spliceosome gold standard interactions are significantly worse correlated (R = 0.27, p < 0.001; Fig. 3c). Calculating significance for all 1253 observed CORUM complexes (permutation test, Benjamini-Hochberg correction), 419/1253 correlate significantly higher than average, while 294/1253 are significantly lower. This suggests that some gold standard complexes are enriched for interacting protein pairs, while others are enriched for non-interacting protein pairs, where non-interacting pairs likely represent interactions disrupted by the particular assay.
Other high-throughput techniques display consistent over- and under-enrichment of specific gold standard complexes. Figure 3d shows gold standard complexes that were consistently predicted in one of three high-throughput techniques - CF, affinity purification mass spectrometry (AP-MS), or yeast two-hybrid (Y2H) - and largely absent from the others. Eighty gold standard complexes were predicted (≥1 interaction per complex) in 4/6 co-fractionation interactomes, while being predicted in no more than a single AP-MS or Y2H interactome (chance = 14 complexes, p = 0.005, bootstrap). Similarly, 61 gold standard complexes are predicted in at least 2/3 Y2H interactomes, while being predicted in at most one co-fractionation or AP-MS interactome (p < 0.001). Only 22 AP-MS-specific complexes are selected in this way (p = 0.49) possibly due to the low CORUM coverage of interactome AP3 . Technique-specific consistency is also seen at the level of pairwise interactions (Fig. 3e).
In addition to being truly non-interacting, the absence of some gold standard complexes from published interactomes (Fig. 3d) could result from low expression of interacting partners (rendering them difficult to quantify) or from none of the co-complex members being included as baits. To control for this, we additionally looked at the subset of gold standard complexes where at least one interaction could potentially be predicted in each study, as defined by having quantified proteins and baits (see Methods). The same pattern seen in Fig. 3d persisted (Additional file 1: Table S1), indicating that gold standard complexes seen by one method but not others cannot be explained by lack of co-expression or choice of bait, and therefore likely reflect the fact that the physical association of gold standard complexes is not guaranteed and depends on the local conditions of the experiment.
We note that the context-specific nature of gold standard complexes is not limited to the type of high-throughput experiment (CF, AP-MS, or Y2H). For example, using co-fractionation data where proteins were fractionated using a variety of techniques  (Additional file 1: Table S1), the 60S ribosome gold standard complex consistently co-fractionated via sucrose fractionation (Additional file 1: Figure S3A) but consistently failed to co-fractionate via heparin dual ion exchange (Additional file 1: Figure S3B).
Universal gold standards improve interactome mapping
If a subset of database PPIs consistently fails to resemble interacting proteins for a given assay, performance should improve when these PPIs are removed from the gold standard set. We confirmed this was the case. We generated co-fractionation-specific gold standard subsets by selecting those complexes that were significantly enriched for interactions in interactomes CF4, CF5, and CF6 [2,3,4]. Evaluating significance at four p-value thresholds (p < 1, 10− 2, 10− 6, 10− 10) produced four subsets of CORUM complexes that contain 302, 122, 95, and 80 complexes, respectively (Additional file 2). To avoid training and testing on the same data, we defined the co-fractionation-specific gold standard subsets using interactomes published by other groups (CF4, CF5, CF6), and these gold standard subsets were then used to predict interactomes using co-fractionation data generated by our group.
These co-fractionation-specific CORUM subsets correspond significantly to housekeeping protein complexes. Using the Gini coefficient, a measure of inequality, we calculated consistency of mRNA expression (Fig. 4a)  and protein expression (Fig. 4b)  across tissues and cell types. As quantified by lower Gini coefficients, the expression levels of co-fractionation-specific complexes are significantly more consistent across tissues than other CORUM complexes (mRNA: Gini = 0.28 vs Gini = 0.40, p = 2.2 × 10− 16, Welch two-sample t-test; protein: Gini = 0.40 vs Gini = 0.49, p = 2.2 × 10− 10). This agrees with an orthogonal analysis of a mouse co-fractionation dataset collected by our group, which analyzed protein co-fractionation across seven tissue types . Quantifying co-fractionation via Pearson correlation, 15 CORUM complexes were found to be housekeeping complexes, as defined by average pairwise correlation significantly greater than chance in all seven tissues (p < 0.05, permutation test; Fig. 4c). Of these 15 complexes, 8 overlap with the 122 co-fractionation-specific CORUM complexes, a significant overlap (p = 7.6 × 10− 8, hypergeometric test; overlapping complexes marked by asterisk * in Fig. 4c). Additionally, we performed a functional enrichment analysis of the p < 10− 2 co-fractionation-specific CORUM subset vs the subset of CORUM with observable proteins in co-fractionation studies (p < 101 subset) (Gene Ontology BP, CC, and MF; hypergeometric test). Consistent with the two previous analyses, this enrichment analysis confirms that the co-fractionation-specific subset of CORUM is enriched for housekeeping processes such as translation and transcription (Additional file 3: Table S2).
Using gold standard subsets generated in this way drastically alters the predicted interactomes (Fig. 5). Controlling interactome quality via the ratio of true positives (TP) and false positives (FP), calculated as precision (TP/(TP + FP)), well-chosen gold standard subsets increased the size of the predicted interactome by up to an order of magnitude over randomly-chosen subsets (Fig. 5a-c). Since FPs are defined as inter-complex protein pairs, they grow with the square of the gold standard set size. TPs, intra-complex pairs, grow linearly. Therefore there is a tendency for precision estimates to increase artificially as the gold standard set shrinks. For this reason we compared all co-fractionation-specific subsets (Fig. 5 black) to random subsets of the same size (red). Precision-recall curves, which visualize the tradeoff between quality and quantity of the interactomes, are also improved over random for increasingly stringent co-fractionation-specific gold standard subsets (Fig. 5d-g).
Here we have estimated the discrepancies between interactome data generated by co-fractionation and curated gold standard interactions from the CORUM database. Across 20 datasets, 37% (54859/149477) of gold standard protein pairs display at most weak evidence for interaction (R < 0.25, Pearson correlation), and 23% (34442/149477) show no evidence of interaction (R < 0) (Figs. 1d, 2). Other databases have a larger proportion of anti-correlated interactions, with up to 55% of database PPIs showing no evidence for interaction (Fig. 2). Protein interaction networks have been compared elsewhere. For example, comparing the power of five PPI networks to predict cancer genes , benchmarking 21 networks for their ability to predict disease genes , and investigating their impact on recovering novel PPIs from high-throughput data . However, to our knowledge our study is the first to specifically address the context-specific nature of PPI entries in these databases.
Since CORUM is manually curated from low-throughput experiments, we do not interpret these anti-correlated pairs as errors in the database. Rather, we attribute any discrepancy between our raw data and the databases to the context-specific nature of protein interactions and the fact that databases compile evidence from many different experiments and conditions. Indeed, under certain conditions, 60S ribosomal proteins, which have been extensively studied and shown to interact, display poor evidence of interaction (Additional file 1: Figure S1).
Therefore studies should take care not to conflate interaction databases, which attempt to list all interactions that can interact, with the subset of interactions that are in fact interacting in a given experiment. Doing so limits high-throughput interactome mapping studies. First, it artificially raises all estimates of error rates, since by definition a portion of the reference positive set is indistinguishable from the negative set. Second, when gold standard interactions are used as training labels in a classifier [2, 3, 7, 11], classification accuracy will be reduced and fewer interactions and/or more noisy interactions will be predicted.
One solution is to use technique-specific gold standard subsets. We show that subsets of gold standard databases that have consistent, independent evidence taken from similar conditions to those under which the raw data was produced can increase the size of interactomes judged at the same precision level (Fig. 4). We include this set of CORUM gold standard complexes and recommend it for future co-fractionation studies.
Motivated by the fact that local conditions can modulate PPIs , this study investigated the degree to which PPI databases (“gold standards”), which compile PPIs and complexes across many experiments, agree or disagree with data from a single experiment. Using a conservative measure of protein interaction (negatively correlated co-fractionation profiles) we find that up to 55% (39 +/− 11%, mean +/− s.d.) of gold standard PPIs show no evidence of interacting during a single experiment. Further, for some gold standard complexes the discrepancy between databases and single experiments is systematic. For example, across CF experiments, proteins in Chaperonin Containing TCP-1 complex consistently resemble interacting proteins, while proteins in the Spliceosome consistently do not (Fig. 3). When data is gathered from different assays (AP-MS and Y2H) we find similar consistency but for different gold standard complexes. We interpret this to mean that there is a subset of database PPIs that is inherently more detectable by CF datasets than by AP-MS and Y2H, and vice versa – i.e. experiment-specific gold standard subsets. Indeed, using a CF-specific subset of gold standards dramatically improves interactome mapping for CF data (Fig. 4). Therefore, we recommend using this subset as a gold standard for future CF studies. More generally, our analysis highlights the error inherent in conflating the entire set of PPIs listed in PPI databases with the subset of PPIs that are truly interacting in any given single study.
Gold standards databases
We primarily used CORUM core complexes as an LC database of known protein complexes (Comprehensive Resource of Mammalian protein complexes, released February 2017) . CORUM is based entirely on experimentally verified interactions, all of which must have extensive low-throughput supporting data. To provide a broad sample of databases, we also analyzed interactions from an additional ten LC interaction databases: HPRD (release 9, last modified April 13, 2010) , MINT (downloaded June 8, 2017) , MENTHA (release June 5, 2017) , BIND (release 1.0, last modified May 20, 2014) , HIPPIE (release 2.0, last modified June 24, 2016) , IID (release April 2017) , BioGrid (release 3.4.149, accessed June 8, 2017) , PINA (version 2, last updated May 21, 2014) , HINT (version 4, downloaded June 8, 2017) , and DIP (release February 5, 2017) . We analyzed a subset of the full BioGrid database for which interactions were supported by at least two publications (Nfull = 254886 interactions, Nsubset = 39524). For databases such as CORUM that list complexes rather than pairwise interactions, gold standard interactions were defined as all protein pairs that are co-members of at least one gold standard complex. Only non-redundant, i.e. unique protein interactions were analyzed.
Co-fractionation profile datasets and mass spectrometry
The majority of co-fractionation data analyzed in this study was collected by our group and constitutes a broad sampling of SILAC-labelled co-fractionation datasets. Data was collected for four independent experiments, each mapping interactome rearrangements to an experimental treatment. Datasets in this study were separated by condition and replicate, such that an experiment with two conditions and three replicates would yield six datasets analyzed here. A total of 20 datasets are included in this study. Three experiments are previously published: two that map the reorganization of the HeLa interactome in response to stimulation with EGF  and Salmonella enterica infection , and one that maps the response of Jurkat T cells to Fas-mediated apoptosis . All fractionation was achieved by size exclusion chromatography except  which used blue-native page. Both methods separate protein complexes by molecular weight. The third co-fractionation dataset is available in this manuscript (Additional file 4: Table S3). All co-fractionation data was quantified using SILAC ratios over successive fractions of a separation gradient, i.e. a chromatogram. Only protein chromatograms with quantification in five or more fractions were analyzed. In order to compare interactions seen by different fractionation techniques, we also analyzed previously published co-fractionation data generated by extensive biochemical fractionation . All co-fractionation profile datasets are composed of co-fractionation profiles, which are protein amount measured over successive fractions, quantified by mass spectrometry. There is one profile per protein or protein group for each combination of replicate and condition.
Published PPI interactomes
In addition to raw co-fractionation data, we analyzed twelve published human protein interactomes: three derived from co-fractionation data published by our lab (CF1 , CF2 , CF3 ), three derived from co-fractionation data not published by us (CF4 , CF5 , CF6 ), three AP-MS derived interactomes (AP1 , AP2 , AP3 ), and three Y2H interactomes (Y2H1 , Y2H2 , Y2H3 ). All interactomes were high-throughput and represent a broad sampling of the full human interactome.
Evaluating gold standard complexes
Raw co-fractionation profiles: To evaluate the degree to which gold standard PPIs are supported by co-fractionation data, we calculated the Pearson correlation coefficient and Euclidean distance between each pair of chromatograms in a dataset. For both measures, missing values in the chromatograms were replaced by zeros. When calculating Euclidean distance, all chromatograms were normalized to have a minimum value of 0 and a maximum value of 1. High correlation or low Euclidean distance was taken as evidence that the gold standard interaction was interacting in the sample.
Published interactomes: We mapped published pairwise protein-protein interactions to gold standard CORUM complexes. For each published interactome, we tested whether gold standard complexes were enriched for published interactions, meaning they contained significantly more pairwise interactions between complex members than the average rate (hypergeometric test). We took significant enrichment as evidence that the published interactome supported the gold standard complex. To standardize interactomes with each other and CORUM, all isoform tags were removed from protein IDs.
To control for expression and different baits, we also defined the subset of gold standard complexes in each study in which at least one interaction could be predicted. For CF interactomes this was defined as a complex in which at least two co-complex members are present in the raw data (raw data for  downloaded here http://metazoa.med.utoronto.ca/;  and  raw data taken from publication). For AP-MS interactomes we assumed the matrix model, meaning that a bait protein need not be present in a gold standard complex for an interaction to be predicted in the gold standard complex, as long as a gold standard complex member is associated with a bait protein. Therefore if at least two members of the gold standard complex were present in the AP-MS interactome, we defined that gold standard complex as a complex that could be predicted by the study. Finally, for Y2H interactomes we defined a gold standard complex as able to be predicted if at least one complex member was a bait protein.
Co-fractionation-specific gold standard subsets
In this study, we used subsets of the gold standard complexes that are consistently supported by co-fractionation interactomes. For these co-fractionation subsets, we used all CORUM complexes that were significantly enriched for interactions (hypergeometric test) in CF4, CF5, and CF6. Significance was assessed at a range of p-value thresholds: 1, 10− 2, 10− 6 and 10− 10. A threshold of p = 1 produced the subset of CORUM complexes with at least one interaction in CF4, CF5, and CF6. The number of CORUM complexes (interactions) in each subset were 302 (33378), 122 (10953), 95 (6326) and 80 (4861), respectively.
To control for the effects of simply reducing the size of the gold standards, we generated random subsets of gold standard PPIs with the same size as the selected subsets. Of the full 46413 unique. PPIs in the core CORUM complexes, we randomly sampled 33378, 10953, 6326, and 4861 PPIs.
without replacement. Random sampling was repeated 100 times for each of the subset sizes, and interactomes were predicted using each random subset with the PrInCE software package.
For this study, interactomes were predicted using the PrInCE software (Predicting Interactomes via Co-Elution), a software package developed by our lab for the analysis of co-fractionation datasets . PrInCE measures the similarity between every pair of co-fractionation profiles using a variety of similarity measures, such as Pearson correlation and Euclidean distance. Gold standard interactions are used as true positive labels (TP) in a Naive Bayes classifier. False positive interactions (FP) are defined as interactions between a pair of proteins that both occur in the gold standard database, but are not members of the same gold standard complex, e.g. an interaction between a ribosomal protein and a proteasomal protein. PrInCE assesses the quality of the predicted interactome using precision, where precision = TP/(TP + FP).
Affinity purification mass spectrometry
Chaperonin containing TCP-1 (protein complex)
Comprehensive resource of mammalian protein complexes
Messenger ribonucleic acid
Protein correlation profiling – stable isotope labeling of amino acids in cell culture
Kristensen AR, Foster LJ. High throughput strategies for probing the different organizational levels of protein interaction networks. Mol BioSyst. 2013;9(9):2201–12.
Havugimana PC, Hart GT, Nepusz T, Yang H, Turinsky AL, Li Z, et al. A census of human soluble protein complexes. Cell. 2012;150(5):1068–81.
Wan C, Borgeson B, Phanse S, Tu F, Drew K, Clark G, et al. Panorama of ancient metazoan macromolecular complexes. Nature. 2015;525(7569):339–44.
Kirkwood KJ, Ahmad Y, Larance M, Lamond AI. Characterization of native protein complexes and protein isoform variation using size-fractionation-based quantitative proteomics. Mol Cell Proteomics. 2013;12(12):3851–73.
Kristensen AR, Gsponer J, Foster LJ. A high-throughput approach for measuring temporal changes in the interactome. Nat Methods. 2012;9(9):907–9.
Scott NE, Brown LM, Kristensen AR, Foster LJ. Development of a computational framework for the analysis of protein correlation profiling and spatial proteomics experiments. J Proteome. 2015;118:112–29.
Scott NE, Rogers LD, Prudova A, Brown NF, Fortelny N, Overall CM, et al. Interactome disassembly during apoptosis occurs independent of caspase cleavage. Mol Syst Biol. 2017;13(1):906.
Heide H, Bleier L, Steger M, Ackermann J, Dröse S, Schwamb B, et al. Complexome profiling identifies TMEM126B as a component of the mitochondrial complex I assembly complex. Cell Metab. 2012;16(4):538–49.
Huttlin EL, Ting L, Bruckner RJ, Gebreab F, Gygi MP, Szpyt J, et al. The BioPlex network: a systematic exploration of the human Interactome. Cell. 2015;162(2):425–40.
Hein MY, Hubner NC, Poser I, Cox J, Nagaraj N, Toyoda Y, et al. A human interactome in three quantitative dimensions organized by stoichiometries and abundances. Cell. 2015;163(3):712–23.
Stacey RG, Skinnider MA, Scott NE, Foster LJ. A rapid and accurate approach for prediction of interactomes from co-elution data (PrInCE). BMC Bioinformatics. 2017;18(1):457.
Rolland T, Taşan M, Charloteaux B, Pevzner SJ, Zhong Q, Sahni N, et al. A proteome-scale map of the human interactome network. Cell. 2014;159(5):1212–26.
Celaj A, Schlecht U, Smith JD, Xu W, Suresh S, Miranda M, et al. Quantitative analysis of protein interaction network dynamics in yeast. Mol Syst Biol. 2017;13(7):934.
Ruepp A, Brauner B, Dunger-Kaltenbach I, Frishman G, Montrone C, Stransky M, et al. CORUM: the comprehensive resource of mammalian protein complexes. Nucleic Acids Res. 2008;36(Database issue):D646–50.
Wysocka J, Myers MP, Laherty CD, Eisenman RN, Herr W. Human Sin3 deacetylase and trithorax-related Set1/Ash2 histone H3-K4 methyltransferase are tethered together selectively by the cell-proliferation factor HCF-1. Genes Dev. 2003;17(7):896–911.
Peri S, Navarro JD, Amanchy R, Kristiansen TZ, Jonnalagadda CK, Surendranath V, et al. Development of human protein reference database as an initial platform for approaching systems biology in humans. Genome Res. 2003;13(10):2363–71.
Chatr-aryamontri A, Ceol A, Palazzi LM, Nardelli G, Schneider MV, Castagnoli L, et al. MINT: the molecular INTeraction database. Nucleic Acids Res. 2007;35(Database):D572–4.
Calderone A, Castagnoli L, Cesareni G. Mentha: a resource for browsing integrated protein-interaction networks. Nat Methods. 2013;10(8):690–1.
Bader GD, Betel D, Hogue CWV. BIND: the biomolecular interaction network database. Nucleic Acids Res. 2003;31(1):248–50.
Alanis-Lobato G, Andrade-Navarro MA, Schaefer MH. HIPPIE v2.0: enhancing meaningfulness and reliability of protein–protein interaction networks. Nucleic Acids Res. 2016;45(D1):D408–14.
Kotlyar M, Pastrello C, Sheahan N, Jurisica I. Integrated interactions database: tissue-specific view of the human and model organism interactomes. Nucleic Acids Res. 2016;44(D1):D536–41.
Cowley MJ, Pinese M, Kassahn KS, Waddell N, Pearson JV, Grimmond SM, et al. PINA v2.0: mining interactome modules. Nucleic Acids Res. 2011;40(D1):D862–5.
Das J, Yu H. HINT: high-quality protein interactomes and their applications in understanding human disease. BMC Syst Biol. 2012;6:92.
Xenarios I. DIP: the database of interacting proteins. Nucleic Acids Res. 2000;28(1):289–91.
Ewing RM, Chu P, Elisma F, Li H, Taylor P, Climie S, et al. Large-scale mapping of human protein-protein interactions by mass spectrometry. Mol Syst Biol. 2007;3:89.
GTEx Consortium. The Genotype-Tissue Expression (GTEx) project. Nat Genet. 2013;45(6):580–5.
Kim M-S, Pinto SM, Getnet D, Nirujogi RS, Manda SS, Chaerkady R, et al. A draft map of the human proteome. Nature. 2014;509(7502):575–81.
Skinnider MA, Scott NE, Prudova A, Stoynov N, Stacey RG, Gsponer J, Foster LJ. An atlas of protein-protein interactions across mammalian tissues. bioRxiv. https://doi.org/10.1101/351247.
Li Y, Sahni N, Yi S. Comparative analysis of protein interactome networks prioritizes candidate genes with cancer signatures. Oncotarget. 2016;7(48):78841–9.
Huang JK, Carlin DE, Yu MK, Zhang W, Kreisberg JF, Tamayo P, et al. Systematic Evaluation of Molecular Networks for Discovery of Disease Genes. Cell Syst. 2018;6(4):484–95.e5.
Skinnider MA, Stacey RG, Foster LJ. Genomic data integration systematically biases interactome mapping. https://doi.org/10.1371/journal.pcbi.1006474.
Ruepp A, Waegele B, Lechner M, Brauner B, Dunger-Kaltenbach I, Fobo G, et al. CORUM: the comprehensive resource of mammalian protein complexes—2009. Nucleic Acids Res. 2009;38(suppl_1):D497–501.
Chatr-Aryamontri A, Breitkreutz B-J, Oughtred R, Boucher L, Heinicke S, Chen D, et al. The BioGRID interaction database: 2015 update. Nucleic Acids Res. 2015;43(Database issue):D470–8.
Wang J, Huo K, Ma L, Tang L, Li D, Huang X, et al. Toward an understanding of the protein interaction network of the human liver. Mol Syst Biol. 2011;7:536.
Rual J-F, Venkatesan K, Hao T, Hirozane-Kishikawa T, Dricot A, Li N, et al. Towards a proteome-scale map of the human protein-protein interaction network. Nature. 2005;437(7062):1173–8.
We thank N. Brown, C. Kerr, A. McAfee, and A. Nastasa for critical suggestions. M.A.S. is supported by a CIHR Frederick Banting and Charles Best Canada Graduate Scholarship, a UBC Four Year Fellowship, and a Vancouver Coastal Health-CIHR-UBC MD/PhD Studentship Award.
This work was supported by Genome Canada (https://www.genomecanada.ca/) and Genome British Columbia (https://www.genomebc.ca/) (214PRO, awarded to LJF); and the Canadian Institutes of Health Research (http://www.cihr-irsc.gc.ca/e/193.html) (MOP77688, awarded to LJF). Funding bodies played no role in the design of the study, the collection, analysis, and interpretation of data, or the writing of the manuscript.
Availability of data and materials
The 11 protein-protein interaction databases supporting the conclusions of this article are available at: HPRD (http://www.hprd.org/), MINT (https://mint.bio.uniroma2.it/), MENTHA (https://mentha.uniroma2.it/), BIND (https://www.bindingdb.org/bind/index.jsp), HIPPIE (http://cbdm-01.zdv.uni-mainz.de/~mschaefer/hippie/), IID (http://iid.ophid.utoronto.ca/iid), BioGrid (https://thebiogrid.org/), PINA (http://omics.bjcancer.org/pina/), HINT (http://hint.yulab.org/), DIP (http://dip.mbi.ucla.edu/dip/), and CORUM (http://mips.helmholtz-muenchen.de/corum/).
The 12 published, publicly accessible datasets supporting the conclusions of this article are available at: Kristensen et al., 2012 (https://doi.org/10.1038/nmeth.2131); Scott et al., 2015 (https://doi.org/10.1016/j.jprot.2014.10.024); Scott et al., 2017 (https://doi.org/10.15252/msb.20167067); Havugimana et al., 2012 (https://doi.org/10.1016/j.cell.2012.08.011); Wan et al., 2015 (https://doi.org/10.1038/nature14877); Kirkwood et al., 2013 (https://doi.org/10.1074/mcp.M113.032367); Huttlin et al., 2015 (https://doi.org/10.1016/j.cell.2015.06.043); Hein et al., 2015 (https://doi.org/10.1016/j.cell.2015.09.053); Ewing et al., 2007 (https://doi.org/10.1038/msb4100134); Rolland et al., 2014 (https://doi.org/10.1016/j.cell.2014.10.050); Wang et al., 2011 (https://doi.org/10.1038/msb.2011.67); Rual et al., 2005 (https://doi.org/10.1038/nature04209).
Finally, one dataset supporting the conclusions of this article is included within the article (and its additional files).
Ethics approval and consent to participate
Consent for publication
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Figure S1. Restricting gold standard PPIs to those supported by two or more publications does not eliminate uncorrelated protein pairs, as measured by Pearson correlation R < 0. Each point is one dataset. Horizontal lines show medians. Red: all non-gold standard protein pairs. Black: non-redundant gold standard pairs. “All pairs” and “BioGrid” correspond to Fig. 1. Figure S2. Co-fractionation datasets are more consistent with database PPIs with “co-fractionation-like” evidence codes, e.g. “density sedimentation” and “molecular sieving”. Top: Percent fraction of anti-correlated pairs across co-fractionation datasets, sorted by average percent. Evidence codes are given on the x-axis. Bottom: Average Pearson correlation co-efficient. This analysis is similar to Fig. 2. Each dot represents on co-fractionation dataset. For each evidence code, only datasets with at least 100 database PPIs are shown. All evidence codes with at least one such dataset are shown. Table S1. Some CORUM complexes are predicted by a single high-throughput technique, as measured by average complex coverage. Complex coverage = number of pairwise interactions in a published interactome / total pairwise connections within a complex. Parentheses show number of complexes averaged. CF-specific complexes correspond to numbers 1–80 in Fig. 3d, AP/MS-specific to 81–102, and Y2H-specific to 103–163. To control for expression and bait selection, only complexes that could be be predicted in each interactome are included (see Methods). Figure S3. 60S ribosome co-fractionates via sucrose fractionation (A) but not via heparin dual ion exchange (B). Pearson R. Plots show replicates. Missing (black) are protein pairs where neither protein was detected. (DOCX 561 kb)
Table S4. Subset of CORUM core complexes that consistently co-fractionate (Feb 2017 CORUM release). Complexes were chosen if they were significantly enriched for pairwise interactions in three published co-fractionation interactomes (Wan et al. 2015, Havugimana et al. 2012, and Kirkwood et al. 2013). Enrichment was calculated with a hypergeometric test, and significance was evaluated at four thresholds: p < 1, p < 1e-2, p < 1e-6, and p < 1e-10. All data aside from columns “p < 1”, “p < 1e-2”, “p < 1e-6”, and “p < 1e-10” are taken from the CORUM core data file. (XLSX 202 kb)
Table S2. Gene Ontology (GO) enrichment of proteins in co-fractionation-specific subset of CORUM. Hypergeometric test. “Hits” are proteins from the co-fractionation-specific subset (p < 10–2) and “background”/“universal” are all proteins from CORUM complexes that were detected, but not necessarily co-fractionating (p < 1) (see Methods for CORUM subset definition). (XLSX 43 kb)
Table S3. Third co-fractionation dataset. (CSV 782 kb)
About this article
Cite this article
Stacey, R.G., Skinnider, M.A., Chik, J.H.L. et al. Context-specific interactions in literature-curated protein interaction databases. BMC Genomics 19, 758 (2018). https://doi.org/10.1186/s12864-018-5139-2
- Protein-protein interaction
- Literature curated database