Cluster-based assessment of protein-protein interaction confidence
© Kamburov et al.; licensee BioMed Central Ltd. 2012
Received: 26 March 2012
Accepted: 16 August 2012
Published: 10 October 2012
Skip to main content
© Kamburov et al.; licensee BioMed Central Ltd. 2012
Received: 26 March 2012
Accepted: 16 August 2012
Published: 10 October 2012
Protein-protein interaction networks are key to a systems-level understanding of cellular biology. However, interaction data can contain a considerable fraction of false positives. Several methods have been proposed to assess the confidence of individual interactions. Most of them require the integration of additional data like protein expression and interaction homology information. While being certainly useful, such additional data are not always available and may introduce additional bias and ambiguity.
We propose a novel, network topology based interaction confidence assessment method called CAPPIC (cluster-based assessment of protein-protein interaction confidence). It exploits the network’s inherent modular architecture for assessing the confidence of individual interactions. Our method determines algorithmic parameters intrinsically and does not require any parameter input or reference sets for confidence scoring.
On the basis of five yeast and two human physical interactome maps inferred using different techniques, we show that CAPPIC reliably assesses interaction confidence and its performance compares well to other approaches that are also based on network topology. The confidence score correlates with the agreement in localization and biological process annotations of interacting proteins. Moreover, it corroborates experimental evidence of physical interactions. Our method is not limited to physical interactome maps as we exemplify with a large yeast genetic interaction network. An implementation of CAPPIC is available at http://intscore.molgen.mpg.de.
Accurate interaction networks (interactomes) are fundamental to answering questions about how the biochemical machinery of cells organizes matter, processes information, and carries out transformations to perform specific functions leading to various phenotypes. Toward this goal, a number of experimental  and computational [2–4] techniques have been devised and applied to map the interactions of human proteins [5–8] and those of model organisms such as yeast [9–12]. Despite their incompleteness , current interactome maps already serve as a basis for numerous methods aiming to elucidate biological processes in health and disease [14, 15]. Current interactome maps are contaminated with false positive interactions that can make up a considerable portion of the data [13, 16–20]. These false positive interactions dim the explanatory light of interaction networks and also decrease the predictive value of methods using such data. It is thus of primary importance to derive confidence values for individual interactions, which can serve to refine current interactome maps or can be used as interaction weights. For example, it has been shown recently that the performance of complex detection approaches is better in confidence-weighted protein-protein interaction networks than in non-weighted networks [19, 21].
Several approaches have been proposed for interaction confidence assessment, many of which are reviewed in [19, 22, 23]. Most of these methods integrate additional data like interaction homology , co-expression of genes encoding interacting proteins [17, 24, 25], or a combination of these and other evidence features [26, 27]. The outcome from such methods depends on the additional data sets. Others combine multiple topological features with additional knowledge to achieve better predictions [20, 28]. Methods which are able to use network topology alone to predict interaction veracity [29–32] are the tools of choice for interaction confidence assessment if other types of data are limited or biased.
At various levels (globally as well as locally), the topology of interaction networks encodes biological properties which are largely independent of the biochemical function of the individual members of the network [33, 34]. This has been demonstrated through analysis of global properties exploiting topological features such as node degree  or distance [36, 37]. The biological importance of network topology may be even more clear for local structures, as in the case of specific wiring patterns of interaction partners . Likewise, modularity of interaction networks is currently the most successful concept for addressing the dynamics of cellular processes [8, 38, 39].
Goldberg and Roth  proposed a connectivity based approach for interaction confidence assessment where the number of common neighbors of a pair of predicted interaction partners counts in support of the interaction. They defined interaction confidence as the level of enrichment of common network neighbors of interacting proteins. It is quantified by the hypergeometric distribution P-value given the number of common neighbors and total network neighbors of both interacting proteins. The underlying principle of the approach has been established in seminal studies demonstrating that biological networks are marked with short interaction paths separating random pairs of proteins in the network (small-world property), and densely connected local neighborhoods (neighborhood cohesiveness property) . Real protein-protein interactions are expected to meet the network cohesiveness property more frequently than false positives. More recently, Kuchaiev and co-authors  proposed another method that embeds interaction networks into a low-dimensional Euclidean space based on network metrics (shortest path length) and then calculates confidence of interactions depending on the Euclidean distance between proteins within that space. The basis of the approach is the geometric graph model that was proposed to better reflect biological networks than e.g. the small-world model . Although the biological basis of the geometric graph model remains elusive, the authors show that it measures network distance more reliably. Both of these topology based methods assign confidence as numerical values to protein-protein interactions in a network and are additionally able to predict new interaction candidates by assigning confidence scores to non-interactions. However, both methods have certain shortcomings. The method by Goldberg and Roth is able to assess the confidence of those interactions whose participants have common neighbors only. Often, however, interacting proteins do not share neighbors. The method of Kuchaiev et al. appears limited in that it requires fixing six free parameters. These include algorithm-specific parameters as well as the prior probability for interactions which depends on knowledge about the interactome size.
Here, we propose CAPPIC (cluster-based assessment of protein-protein interaction confidence) – a novel approach that exploits the inherent modular structure of interactomes for confidence assessment of protein-protein interactions. Our method combines the basic principles of the topology based methods described above: high neighborhood interconnectedness of a couple of proteins and short distance between them (the features exploited by Goldberg and Roth and Kuchaiev et al., respectively) are indicators that both proteins participate in the same module. We apply Markov clustering  to the line graph  of an interaction network to dissect it into modules of interactions. As demonstrated in , this strategy can generate interaction clusters that significantly overlap with known biological pathways. Notably, the interaction clusters overlap in their protein constitution. This is biologically more meaningful than clustering the proteins into disjoint modules because pathways and protein machineries are known to overlap [10, 21]. The rationale behind our approach is that proteins that are specific to certain modules are expected to have more interactions with proteins that are specific to the same modules than with other proteins . Intuitively, we assign low confidence to interactions that disagree with the modular structure of biological networks and high confidence to those that comply with it. This rationale has also been used as a basis of approaches for the detection of binary interactions  or protein complexes  from complex purification data or to reveal dynamic interaction patterns during the human spliceosome cycle . While the aim of CAPPIC is to detect false positive interactions, a different approach, which is however also based on the principle of high link density within network modules, has been proposed for identifying false negatives .
Yeast interactome maps used in this study for method evaluation
links with ≥ 3
The value of the fidelity F p,c lies between 0 and 1, with values near or equal to 1 if a protein p is specific to cluster c, i.e. if it has relatively many links in that cluster. For a fixed L p,c it holds that the smaller the cluster (smaller L ·,c ), the greater the fidelity value. Finally, if all the links of two proteins lie within a cluster, the fidelity is greater for the protein with the higher degree.
Interactions get high confidence values if both proteins are specific to the cluster containing the interaction, and low confidence values when one or both of the proteins are not specific to the cluster.
The interaction confidence scores calculated by CAPPIC are dependent on the granularity of the interaction clustering. It has been previously shown that modules in many complex networks, including protein interaction maps, are organized in a hierarchical manner . Accordingly, interaction clustering can yield protein complexes, cellular machineries, pathways, or higher-order biological processes depending on the clustering granularity. To estimate the clustering granularity for a network that will result in the best discrimination between true and false interactions, we first randomly rewire a small part of the links in that network to generate a false interaction set. In the rewiring procedure, pairs of interactions are selected at random and two of the proteins are swapped so that no real interaction is reconstituted and the network stays connected. This way, two false interactions are generated for two real ones while the degree of each protein is preserved. Then, we calculate interaction confidence values of the resulting partially rewired network as described above using different inflation values. The inflation parameter of the Markov clustering algorithm essentially controls clustering granularity . For every inflation value, we quantify the significance of the difference between confidence score distributions of the rewired and the remaining non-rewired links. This is done with the Wilcoxon rank-sum test under the alternative hypothesis that the confidence scores of the non-rewired links are greater than the confidence scores of the rewired links. The inflation value minimizing the Wilcoxon test P-value is considered optimal.
Experiments have shown that randomly rewiring 3% of the links in the granularity estimation procedure described above is a good choice because this yields a false interaction set of reasonable size while keeping most of the network intact. If the set of false interactions obtained through random rewiring is too small, the granularity estimation will lack statistical power, while if too many interactions are rewired, the network’s original modular structure will be altered which will affect the granularity estimate. For all networks CAPPIC was applied on, random rewiring of 1%, 3%, 5%, or 10% of the interactions yielded very similar optimal granularity estimates.
Our granularity estimation strategy builds upon the assumption that the optimal granularity value inferred from a partially rewired network instance (where both false positive and false negative rates are increased compared to the real network) is transferable to the real network. We aimed to scrutinize this reasoning and verified for all reference networks that 1) the estimated optimal granularity was rather independent of the random choice of links for rewiring; and 2) that interaction clusters were similar for the intact and the partially rewired networks clustered with the same inflation value (see Additional file 3: Supplementary Text).
In the case of well-studied organisms such as yeast, data on protein complexes can be used to define the positive interaction sets alternatively to literature evidence as used above. We used two complex-based positive sets from yeast complexes obtained from CYC2008  and from ref. . The performance of CAPPIC (and often of the reference methods) was better with the complex-based compared to the literature-based positive set for almost all networks (Additional file 4: Figure S1). For example, the AUC for CAPPIC increased from 82% to 87-89% for CPDB-yeast and from 66% to 70-72% for Costanzo when the literature-based positive reference set was replaced by a complex-based one; improvements by 1-2% AUC were also observed for the Tarassov-all, Tarassov-hq and Collins networks (Additional file 4: Figure S1 versus Figure 2 in the Main text). However, despite the better performance with complex-based positive reference sets, such sets are not well-suited for measuring the performance on networks obtained by techniques such as yeast-two-hybrid . This could be the reason for the slight decrease in performance (by 1-2% AUC) on the Yu-Ito-Uetz yeast-two-hybrid network compared to a literature-based positive set (Additional file 4: Figure S1 versus Figure 2 in the Main text). Moreover, the complex-based performance estimate may be positively biased since protein complexes in the reference data may have been defined at least partially on the basis of the analysed interaction networks.
Furthermore, if low-confidence interactions are removed from interaction clusters, the latter become more consistent regarding the pathway annotations of the contained proteins (see Additional file 3: Supplementary Text). Our approach can thus be used to obtain more refined functional modules in interaction data sets.
Network topology-based approaches are motivated by the fact that the structure of interaction networks is not random but reflects biological functionality [33, 34]. Modularity is a topological property that is inherent to protein-protein interaction networks [10, 39, 56]. We propose a novel method (CAPPIC) to assess the confidence of individual protein interactions in an interaction network. Our method exploits network modularity alone for estimating the confidence of interactions and does not require any additional knowledge about the interacting proteins or the techniques used to generate the data. We demonstrate the power of CAPPIC in discriminating between true and false interactions on the basis of five physical protein interaction networks and one genetic interaction map from yeast, as well as two distinct interaction data sets from human.
CAPPIC compares well to previous topology-based approaches by Goldberg and Roth and Kuchaiev et al. in assigning continuous confidence scores to all interactions in a given physical interaction network. The method of Goldberg and Roth is dependent on shared network neighbors of interacting proteins; however, many interacting proteins do not share neighbors. As a result, many interactions are scored with a confidence value of zero. However, integrative approaches operating on networks usually take probabilistic rather than binary data as input. Thus, the goal of confidence assessment is often to assign a continuous score to all interactions rather than to filter for a small subset. In particular, all proteins with a single interaction partner are disregarded by Goldberg and Roth’s method, albeit these single protein associations could give important clues about the function of these proteins. Both methods, Kuchaiev et al. and CAPPIC, are able to assign continuous scores also to such interactions. In contrast to the method of Kuchaiev et al., CAPPIC does not require any parameter input. The only parameter that influences the resulting confidence scores – clustering granularity – is optimized internally for each individual input network. Our results have shown that the number of clusters obtained at the optimal granularity tends to be small for all reference networks, ranging from 10 to 50 clusters (see Additional file 5: Figure S2 and Figure ST1 in Additional file 3: Supplementary Text). This alleviated our initial concerns that interactions executing essential crosstalks between related pathways could be assigned low confidence. Because the optimal granularity tends to be very coarse, closely related pathways will probably not be separated but clustered together.
CAPPIC should be applicable for weighting any binary network with an inherent modular structure (for examples, see ). Notably, it does not consider the technique used to generate the network (unlike other approaches that integrate a fixed, subjective judgment on the reliability of different techniques, e.g. ref. ). CAPPIC fails to generate reliable confidence scores in cases where modularity is not pronounced, i.e. if many of the real links within biological modules (complexes, pathways, etc.) are missing. This is probably the case with the Yu-Ito-Uetz and Y2H-human reference networks: here, the topological signal that our method exploits seems to be weaker and it achieves only 60-62% AUC. Absence of modularity in this example is evidenced by the relatively low clustering coefficient  of 0.08 which is nine times lower than that of the Collins network where CAPPIC achieves 94% AUC and six times lower than that of the Mazloom network (90% AUC). Moreover, the Yu-Ito-Uetz data set is the sparsest of all yeast reference networks (Table 1). To conclude, results on all example networks suggest that CAPPIC is well suited to score datasets with moderate to high interaction density.
Unlike the reference methods, CAPPIC is able to accommodate experimental evidence weights of interactions. Interaction detection techniques often associate such weights with predicted interactions, reflecting for example the number of times an interaction is observed in repetitions of a yeast-two-hybrid experiment [7, 9, 13] or the reporter intensity value in the case of a protein-fragment complementation assay . If available, such weights can be exploited by our method in its random walk based interaction clustering step. This can improve the interaction clustering result and consequently increase the performance of confidence assessment. However, since we set out to estimate the performance of CAPPIC in comparison to other methods that cannot accommodate interaction weights, we did not make use of this advantage in this work and considered all interactions equal. Moreover, the ability to incorporate experimental interaction weights helps to avoid interaction data pre-filtering, commonly executed to derive binary interaction networks (where pairs of proteins either interact or not). Such filtering of probabilistic interaction data is inherently associated with data loss. Similarly, it is a common practice to remove interaction hubs in a dataset to improve its quality (e.g., ref. ). As exemplified in Additional file 6: Figure S3 for the yeast hubs PHO85 (a Cyclin-dependent kinase; 467 interactions) and UBC7 (an E2 ubiquitin ligase; 622 interactions) in the CPDB-yeast network, CAPPIC assigns on average lower scores to interactions of hubs. However, a considerable fraction of their interactions scores highly: 29% of the interactions of PHO85 and 25% of the interactions of UBC7 are assigned higher CAPPIC scores than the median score of the complete network. This suggests that a complete removal of hubs from the network could unnecessarily remove high-quality protein-protein interactions and emphasizes the utility of confidence scoring.
Our approach can be combined with other lines of interaction evidence like other topological features, protein co-expression, or interaction homology to achieve even better scoring performance . While the aggregation of different features holds the promise of even more reliable interaction confidence assessment, it depends on reference interaction sets. At present, even for yeast the construction of an appropriate reference set is still a daunting task .
Since biological interaction networks contain false positives, assessing the confidence of individual interactions in order to weight or filter interaction data is a crucial step that should precede network-based inferences. Here we propose a network topology based method called CAPPIC that estimates interaction confidence by exploiting the network’s inherent modularity. CAPPIC requires no reference interaction sets or parameter settings. Based on five large-scale physical interaction networks from yeast, we show that our method compares well to other topology-based approaches. Confidence scores calculated with CAPPIC also correlate well with the Gene Ontology co-annotation of interacting proteins, and corroborate experimental evidence of physical interactions. CAPPIC is limited neither to physical interactome maps nor to yeast networks as it also performs well on a large yeast genetic interaction network and on two human protein-protein interaction data sets.
To cluster a network of interactions, we use the original implementation of the Markov clustering algorithm (version 10-201 downloaded from http://www.micans.org/mcl/sec_software.html). The inflation scan which aims to optimize clustering granularity is carried out in two steps: a coarse scan with step size of 0.1 within a fixed range I∈[1.1,2.0] (where I is the Markov clustering inflation value) is followed by a fine scan with step size of 0.025 around the optimal inflation value resulting from the coarse scan ± 0.1. In general, the inflation parameter takes values from the interval I∈(1.0,30.0] with higher values resulting in finer granularity. In all our experiments, the optimal inflation estimate was far below 2.0 (see Figure ST1 in Additional file 3: Supplementary Text), motivating the choice of this value as an upper boundary of the inflation scan.
To conduct ROC analysis, we constructed true and false interaction sets. The positive set comprised interactions published in at least three papers in total. An exception was made for the Costanzo network because of the scarcity of genetic interaction data: the positive set in this case consisted of interactions that are also reported in . Literature evidences were retrieved with the interaction evidence mining ConsensusPathDB plugin . The negative interaction set was constructed by randomly rewiring 3% of the interactions in the respective network. For each partially rewired network, we ranked interactions according to confidence as calculated with CAPPIC and reference methods and created receiver operating characteristic (ROC) curves. The performance of a given confidence assessment method in ranking positive interactions higher than negative ones was quantified with the area under the ROC curve (AUC). The AUC is around 50% if a method does not perform better than random interaction ranking, and is closer to 100% the better it ranks positive interactions higher than negative ones. Since the constitution of the negative and positive sets involves a random process (that is, the random selection of interactions for rewiring), we repeated the procedure 100 times and averaged ROC results.
We set the number of yeast genes to 6,000 in the method by Goldberg and Roth. The parameters of the method by Kuchaiev et al. (implemented as Matlab scripts downloaded from http://www.kuchaev.com/Denoising) were set as follows: priorEdge=0.002945 (which results when the estimated yeast interactome size of 53,000 interactions  is divided by the number of all possible protein pairs, 6,000 choose 2); priorNonEdge=1-priorEdge; dim=5 (default); d=3 (default); learnSetSize=min(5,000 or half the number of interactions); delta=1.0; and stopEps=0.01 (default). In the case of Costanzo, dim=3 because the program (run on a standard AMD X2 5600+ machine with 8GB of RAM running Matlab version 220.127.116.119 under Linux) did not return results within five days for a higher number of dimensions.
For each network, we obtained the GO semantic similarity of biological process and cellular component annotations of interacting proteins using the method proposed by Resnik  implemented in the software package GOSemSim version 1.8.0 . GO annotations inferred from physical interaction (GO evidence code ‘IPI’) were excluded from the semantic similarity calculation to avoid circularity. For each network, interactions were ranked by increasing confidence score and divided into five equal sized bins. The mean semantic similarity values for interacting proteins within each bin were calculated. Additionally, the mean GO semantic similarity for random pairs of proteins from the respective network was assessed by completely rewiring the networks while preserving each protein’s degree and then calculating the mean GO semantic similarity of links in those randomized networks.
This work was funded by the European Commission under its 7FP grant diXa (grant number 283775) and the German Ministry for Education and Research under its grants MedSys - PREDICT (grant number 0315428A) and NGFNplus (NeuroNet TP3, grant number 01GS08171).
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.