Reuse of structural domain–domain interactions in protein networks
© Schuster-Böckler and Bateman. 2007
Received: 23 February 2007
Accepted: 18 July 2007
Published: 18 July 2007
Skip to main content
© Schuster-Böckler and Bateman. 2007
Received: 23 February 2007
Accepted: 18 July 2007
Published: 18 July 2007
Protein interactions are thought to be largely mediated by interactions between structural domains. Databases such as iPfam relate interactions in protein structures to known domain families. Here, we investigate how the domain interactions from the iPfam database are distributed in protein interactions taken from the HPRD, MPact, BioGRID, DIP and IntAct databases.
We find that known structural domain interactions can only explain a subset of 4–19% of the available protein interactions, nevertheless this fraction is still significantly bigger than expected by chance. There is a correlation between the frequency of a domain interaction and the connectivity of the proteins it occurs in. Furthermore, a large proportion of protein interactions can be attributed to a small number of domain interactions. We conclude that many, but not all, domain interactions constitute reusable modules of molecular recognition. A substantial proportion of domain interactions are conserved between E. coli, S. cerevisiae and H. sapiens. These domains are related to essential cellular functions, suggesting that many domain interactions were already present in the last universal common ancestor.
Our results support the concept of domain interactions as reusable, conserved building blocks of protein interactions, but also highlight the limitations currently imposed by the small number of available protein structures.
One way to understand a protein's function is to look at its composition of conserved domains. Such families of related sequence regions, collected in the Pfam database , usually constitute structurally and functionally conserved modules. It is assumed that binding interfaces, too, are conserved evolutionary modules that are reused between proteins of different functions and retained during evolution [2, 3].
Therefore, domain–domain interactions are often regarded as the currency of protein–protein interactions. Based on this assumption, Ng et al. described an approach to predict domain–domain interactions using literature curation, evolutionary history and the distribution of domains in protein interactions . Wuchty et al. compared the relationship between this set of predicted interacting domain pairs to the domain coocurrence network . More recently, other groups have come up with sophisticated statistical methods to estimate putatively interacting domain pairs, based on the assumption of domain reusability [6–10]. However, none of these approaches offers structural evidence that the predicted domain pairs are able to form an interaction.
For complexes with known structure, it has been shown that domains can mediate interactions [11, 12]. Such interactions between pairs of domains are stored in the iPfam database . The structural evidence lends strong support to the inferred domain pair, resulting in a high confidence set of domain pairs.
Unfortunately, the selection of complexes in the PDB database of protein structures  is rather small and biased . There is often only a single structure that shows a certain protein pair to interact, while other complexes like haemoglobin have been crystalized dozens of times. This makes it difficult to assess whether some domain pairs act as reusable modules in protein interactions from PDB data alone.
High-throughput experiments [16–18] and extensive literature curation efforts  have yielded large databases of protein interactions [20–24]. Despite the continuing growth of protein interaction databases, even the best studied protein interaction network of S. cerevisiae is thought to be incomplete and inaccurate [25–27]. Given that this network already comprises around 60000 interactions, questions arise as to how such networks have evolved and how they are organised. Furthermore, methods for assessing the quality of high-throughput experimental results are in high demand due to the error prone nature of the methods used.
In this study, we investigate how pairs of protein families taken from iPfam are distributed in experimental protein interactions from five major model species. This allows us to address a number of questions: what proportion of each organism's protein interaction network, its interactome, can be attributed to a known domain–domain interaction? How conserved are domain–domain pairs between species, and how many interacting domain pairs are still unknown?
We analysed the distribution of Pfam families known to interact from a PDB structure (iPfam domain pairs) in experimentally derived protein interactions (experimental interactions). The experimental interactions were filtered to only include interactions with exactly two partners (see Methods). The fraction of experimental interactions that contain at least one iPfam domain pair is referred to as the iPfam coverage. Accordingly, the fraction of experimental interactions that contains any pair of Pfam domains (excluding the iPfam domain pairs) is called the Pfam coverage.
iPfam domain pair coverage on protein interactions
Proteins in proteome
% proteome in interaction set
Protein pairs in interaction set
Protein pairs with iPfam domain pair
Protein pairs with iPfam domain pair (randomised mean)
The relatively low iPfam coverage is by itself a disappointing finding. However, the fact that only a small fraction of protein interactions contain known domain pairs could be a result of the scarcity of available structures of protein complexes. Therefore, we asked whether the observed iPfam coverage is larger than would be expected by chance. To test this, we created 1000 random networks per species using the algorithm described in Methods. We then calculated the iPfam coverage on the protein interactions in each randomised network. Mean and standard deviations of the randomisation experiments are shown in Table 1. No P value (see Methods) was greater than 1.84 · 10-06. This proves that the observed iPfam coverage is significantly higher than expected and iPfam domain pairs are enriched in real experimental protein interactions.
All three distributions in Figure 2 resemble a power law distribution, according to the good fit of log-linear functions (log(f(x)) = k log x + log a) shown as dotted lines. The slopes k of the H. sapiens and S. cerevisiae distributions are very similar (-1.53 and -1.60, respectively), while E. coli has a markedly smaller slope (-2.78). This suggests that the ratio of specific to promiscuous iPfam domain pairs is very similar in S. cerevisiae and H. sapiens, whereas E. coli features fewer multiply reoccurring iPfam domain pairs.
The power law distribution of iPfam frequencies implies that the majority of covered protein interactions can be attributed to a minority of iPfam domain pairs. 51.7% of the iPfam domain pairs in S. cerevisiae and 45.3% in H. sapiens are seen in just one experimental interaction. Conversely, 92.4% of H. sapiens and 85.4% of S. cerevisiae covered experimental interactions contain an iPfam domain pair that occurs more than once. Even more, half of the covered experimental interactions in H. sapiens contain an iPfam domain pair that occurs in more than 16 different experimental interactions (5 for S. cerevisiae).
We reasoned that if there are iPfam domain pairs that act as reusable modules in protein interactions, then highly connected proteins should also be more likely to contain promiscuous iPfam domain pairs and vice-versa.
We find that for proteins from a degree of 1 to 50, there is strong correlation in both H. sapiens and S. cerevisiae (Spearman correlation coefficients of 0.68 and 0.71, respectively) between degree and number of iPfam domain pairs on adjacent edges. For the 1.2% of proteins in H. sapiens and 6.4% in S. cerevisiae which have a degree higher than 50, the correlation gradually diminishes.
Additional file 1 contains a list of all iPfam domain pairs and their frequencies in the experimental protein interactions, while Additional file 4 lists the frequencies of the single domains. Interactions between protein kinase domains (Pkinase, Pfam acc. PF00069 and Pkinase_Tyr, Pfam acc. PF07714) are the most frequent iPfam domain pairs, as well as interactions involving recognition domains such as SH2 or SH3. In S. cerevisiae, the Proteasome family (Pfam acc. PF00227, a family of peptidases) and WD40 (Pfam acc. PF00400, a repeat involved in multimer assembly) are also amongst the five most frequent iPfam domain pairs. As expected, more frequent domains are also more likely to be found as pairs in interacting proteins. It should be noted however that in the PDB structures, some of the observed domain pairs (Pkinase_Tyr ↔ SH3_1, Pkinase_C ↔ Pkinase and others) are only seen to interact within one protein (intrachain interactions) as opposed to interactions between two distinct proteins (interchain interaction). The table in Additional file 5 lists the number of PDB structures for each iPfam domain pair, distinguishing between intrachain and interchain interactions. Looking for example at the covered experimental interactions in H. sapiens(Additional file 1), only 8 out of the 100 most frequent iPfam domain pairs are seen in intrachain interactions exclusively, while 61 are exclusive to interchain interactions and 31 are seen in both.
A possible explanation for the occurrence of purely intrachain iPfam domain pairs in the covered experimental interactions is that they frequently cooccur together on the same protein with other iPfam domain pairs. A list of all combinations of iPfam domains (the domain architecture) on interacting proteins is given in Additional file 2. It reveals that certain iPfam domains such as SH2, SH3_1 or Pkinase_tyr frequently occur in the same architecture. Without further experiments, we cannot assign the correct interacting domains with certainty.
This highlights a basic assumption of this study that could be a source of error. We assume that interacting proteins that contain an iPfam domain pair interact through these domains. This, of course, is not necessarily the case. Although it has been shown that sequence similarity is linked to the mode of interaction , not every protein interaction that contains an iPfam domain pair is necessarily mediated by exactly this domain pair. To gain a rough estimate of the false positive rate due to this assumption, we counted how many protein pairs in the PDB contain an iPfam domain pair that does not mediate an interaction in one complex structure but does so in another. 3671 out of a total of 5380 interacting protein pairs from the PDB contain an iPfam domain pair that does not interact in one complex structure but does so in another. This means that for more than 32% of the protein interactions in the PDB, the iPfam domain pair assignment is correct. For the remaining 68%, the iPfam domain pair assignments are wrong in one case but correct in another. The real false positive rate is likely to be smaller, because some iPfam domain pairs might still independently mediate an interaction with a different, possibly unknown, partner protein.
We tested whether iPfam domain pairs are enriched in known protein complexes from S. cerevisiae. This is interesting firstly because domain–domain interactions are thought to be more common in obligate interactions. Secondly, the described modularity of known S. cerevisiae complexes lends support to the assumption that the underlying iPfam domain pairs are modular. In fact, we find a two-fold enrichment for iPfam domain pairs in the complexes described by Gavin et al. . From the 294 binary protein interactions in this data set, 24 contained an iPfam domain pair, which corresponds to a coverage of 8.16% (P value 2.7 · 10-47).
We also analysed the full dataset of protein complexes. From 491 complexes described by Gavin et al., 157 contained at least one pair of proteins with an iPfam domain pair (31.9%). In total we found 617 pairs of proteins that contained an iPfam domain pair. Interestingly, we find that the distribution of iPfam domain pairs on complexes is uneven. When we drew 617 protein pairs randomly from all possible protein pairs in the complexes, we covered 192 complexes on average, with a standard deviation of 7.22. The probability of covering only 157 complexes is just 6.24 · 10-07. Thus, some complexes contain a greater number of iPfam domain pairs, while other complexes do not contain any at all. This suggests that some sets of domain pairs are specific to certain complexes or pathways. Typical examples are the RNA polymerase II complex (IntAct id: EBI-815049) or the U1 snRNP complex which contain numerous iPfam domain pairs that are specific to these complexes.
Matrix of mutual shared iPfam domain pairs
iPfam domain pairs in total
We also compared the iPfam domain pair frequencies between H. sapiens and S. cerevisiae directly. We derive a Spearman correlation coefficient of 0.50 for the frequencies of all 524 iPfam domain pairs that are conserved between S. cerevisiae and H. sapiens. To test whether the correlation is an artefact of the distribution of the values, we recalculated the correlation 1000 times, each time shuffling one distribution randomly. From these random results, we derive a P value of 3.6 · 10-30 that the observed correlation is random. This suggests that iPfam domain pairs with a large number of occurrences in one species tend also to be more frequent in the other.
Our analysis allow us to estimate how many iPfam domain pairs would eventually cover all protein interactions. This corresponds to the predictions made by Aloy and Russel . Similar to their approach, we make a linear estimation with the following factors:
χ S The number of iPfam domain pairs observed in species S
θ S The number of observed interactions in species S that contain an iPfam domain pair
Θ S The total number of observed interactions in species S
ψ S The number of proteins from species S that are seen in an interaction screen
Ψ S The proteome size for species S
ξ S The number of Pfam domains observed in all protein of species S
Ξ The total number of known Pfam domains
This means we scale the observed number of iPfam domain pairs to cover all observed interactions. We then use the relative proteome coverage to estimate the total number of iPfam domain pairs in all proteins.
Prediction of total number of iPfam domain pairs
The coverage of iPfam on experimentally derived protein interactions is low. For S. cerevisiae, the species with the best mapped interactome, only 4.47% of the protein interactions contain an iPfam domain pair. Even in H. sapiens, where we suspect a positive bias due to the overrepresentation of disease-related proteins in both the PDB and protein interaction databases, 81% of protein interactions do not contain an iPfam domain pair. This reveals the limits of our understanding of the molecular structure of protein interactions.
Figure 1 also shows that a majority of protein interactions contains at least one pair of Pfam domains. While there is no structural information about putative interactions between these pairs, this fraction can already be analysed using statistical methods to identify putative domain interactions [7, 9, 10]. This in turn creates new targets for future structural genomics projects . Prioritising these targets according to the number of covered experimental interactions could increase the coverage of databases like iPfam quickly.
We find, however, that iPfam domain pairs occur significantly more often in experimental interactions than would be expected by chance. This requires that at least a subset of the iPfam domain pairs are reused in several experimental interactions.
Despite the low overall coverage, iPfam domain pairs are found in more protein interactions than would be expected by chance (see Table 1). This statistical overrepresentation suggests that certain iPfam domain pairs constitute modules of molecular recognition which are reused in different protein interactions . In fact, we find a characteristic power law distribution when we plot the histogram of experimental interactions per iPfam domain pair, see Figure 2. This underlines that a few promiscuous iPfam domain pairs are responsible for the majority of the iPfam coverage. These iPfam domain pairs are most likely to be reusable modules. In fact, we find the most frequent iPfam domain pairs to be recognition domains in signal transduction. Conversely, a large number of iPfam domain pairs are specific to a small number of protein interactions. This implies that recognition specificity amongst proteins is often achieved by maintaining an exclusive interacting domain pair. This could pose a problem for purely statistical approaches to infer domain interactions: if for many interfaces the real interacting domain pair will only occur once in an interactome, it will be hard to elucidate this on a statistical basis.
The concept of modularity of interacting domain pairs is furthermore supported by the positive correlation between the number of protein interactions an iPfam domain pair is seen in and the connectivity of the interacting proteins. We hypothesise that if during the course of evolution a protein is duplicated, it is likely to retain connections with other proteins which contain the same domain interaction modules. It is clear, however, that even though recognition domains are reused in various proteins, their specificity is bound to be controlled.
We tried to estimate how many iPfam domain pairs exists in all interactomes. Our predictions lie almost an order of magnitude higher than the 10000 domain interaction types proposed by Aloy and Russel . While all such estimates should be taken with caution, our results show that at best 10% of all structural domain pairs are represented in iPfam. The statistical approaches described in the introduction can only cover a small fraction of this domaininteraction space. Riley et al. for example report only 3005 interacting domain pairs which could be inferred from protein interactions . Even under the assumption that many interactions involve short linear motifs, it seems likely that a large number of domain interactions remain to be resolved.
iPfam domain pairs are not only recurrent within the protein interaction network of one species. They also appear to be conserved between species. In a small set of protein structures from S. cerevisiae, it has been shown that interacting domain pairs are more conserved than non-interacting domain pairs . Here, we call an iPfam domain pair conserved if there are protein interactions in two species which contain the same iPfam domain pair.
In a recent study , Gandhi et al. have assessed the conservation of protein interactions as the co-occurrence of orthologous interacting proteins. They found only 16 orthologous interacting protein pairs that were conserved in S. cerevisiae, C. elegans, D. melanogaster and H. sapiens. Conversely, we find that 71 iPfam domain pairs are conserved in the experimental interactions of these species. Even between a prokaryote like E. coli and the two eukaryotes S. cerevisiae and H. sapiens there is a considerable proportion of conserved iPfam domain pairs, to the extent that 53% of the iPfam domain pairs from E. coli are also observed in a eukaryote (Table 2). 107 domain pairs are shared between E. coli, S. cerevisiae and H. sapiens. These domains are predominantly related to transcription, translation and other basic essential cellular activities, which is in congruence with the findings of Gandhi et al..
Although the low overall iPfam coverage hampers the interpretation of our results, it looks as if there has been a diversification of domain interactions from E. coli to H. sapiens. While more than half of the iPfam domain pairs in E. coli have been retained throughout evolution, numerous new ones seem to have emerged in eukaryotic development. The significant positive correlation in the frequency of iPfam domain pairs conserved between S. cerevisiae and H. sapiens also suggests that the binding interfaces are more often kept or even reused rather than lost in the course of evolution. Conversely, this also raises the question of whether one could establish a comprehensive set of domain interactions that were present in the last universal common ancestor.
In this study, we addressed the utility of current knowledge about structural domain interactions in order to interpret experimental protein interactions. Disappointingly, only a small fraction of all experimental interactions can be attributed to a known domain interaction. Within this subset of interactions, we nevertheless made several reassuring observations: structural domain pairs are enriched in experimental protein interactions. Some of the domain pairs seem to mediate a large number of protein interactions, thus acting as reusable connectors. This property is also conserved between species. Taken as a whole, this further underlines that solving structures of protein complexes should be an important focus for future structural genomics projects. Targeting the most frequent domain pairs would increase the coverage of databases such as iPfam, shedding more light onto the molecular mechanisms underpinning cellular networks.
The complete interaction sets from BioGRID , DIP , HPRD , IntAct  and MPact  were downloaded. A wide range of databases were used to cover as many distinct experimental data sets as possible. BioGRID for example contains a large manually curated set of protein interactions for S. cerevisiae . Similarily, HPRD hosts a set of manually curated protein interactions for H. sapiens. IntAct on the other hand contains results from high-throughput screens and integrates data from other protein interaction databases as part of the IMEx collaboration. The MPact database combines the manually curated S. cerevisiae protein complexes data set formerly known as the MIPS complexes with other high-throughput interaction experiments data. Taken together, these databases represent most of the protein interactions currently stored in machine-accessible form.
Despite great efforts to unify access to protein interaction data , acquiring large data sets from diverse sources is still far from trivial and error prone. The PSI-MI XML data exchange format provided by the aforementioned databases was used to generate a local relational database of protein interactions. All entries were mapped to UniProt  by either relying on existing annotations from the source databases or by pair-wise sequence alignment to all UniProt proteins from the same species as the query protein. The direct sequence comparison was performed using pmatch, a very fast pairwise alignment algorithm developed by Richard Durbin (unpublished, source code available ).
To allow cross-species comparisons, the data were split into five distinct species sets: E. coli, S. cerevisiae, C. elegans, D. melanogaster and H. sapiens. It should be noted that the proportion of proteins for which an interaction is known varies greatly between the species, see Table 1. This might affect the results if there is a systematic bias on the composition of a protein interaction set.
To prevent bias from multiple alternative versions of the same protein, all interacting proteins were mapped to reference proteomes as defined by Integr8 , again using pmatch. An average of ≈ 16% of interaction entries were lost in the mapping process, either if no sequence was provided with the original entry or if no significant matching sequence could be found in Integr8. The total number of missing proteins will be lower, as several entries from different databases refer to the same sequence.
There are many types of experiments used to derive protein interactions, with different properties and error rates. For this analysis, solely the properties of physically interacting proteins is of interest. Therefore, only interactions between exactly two proteins per experiment were considered. That means all protein complex data that were derived by co-purification methods were removed, unless a particular experiment had identified exactly two binding partners. All genetic interactions were also removed.
Randomised protein interaction networks with identical degree distributions were generated from the original filtered experimental interaction data for each species. In each randomisation step, a mapping is created that assigns every node a randomly chosen replacement node. In this way the edges of the network remain in place, while the nodes are shuffled randomly. It should be noted that the degree distribution per node is not maintained. Instead, this behaviour simulates a network with a high false positive rate.
P values for observations x were calculated as P(X ≥ x) = f(x; μ, σ), where f(x; μ, σ) is the probability density function of the normal distribution with mean μ and standard deviation σ. μ and σ are estimated through the randomisation experiments. The density function thus provides the probability that a value less than or equal to x is observed by chance, given the distribution estimated by a random resampling method. Where appropriate, the inverse probability P(X > x) = 1 - f(x; μ, σ) was applied.
The authors thank Robert Finn for valuable comments and suggestions. We are grateful for the numerous detailed comments and suggestions by the anonymous reviewers. BSB and AB are funded by the Wellcome Trust.
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.