Evolutionary pressures shaping protein interaction networks are practically unexplored, although the importance of such studies has been recognized [1]. The earlier studies addressing this question analyzed the conservation of network motives [2, 3, 4, 5], or the rate of link dynamics in interaction networks [6, 7]. Although the importance of different protein properties to the evolution rates of proteins has been questioned [8, 9, 10], proteins interacting with multiple partners simultaneously have been shown to have a slower pace of evolution [11]. Thus, the properties of interaction networks can be thought to influence the evolution of organisms. More general knowledge about the effectors shaping protein interaction networks would give insights on several aspects of the formation of complex biological systems.

For detailed analysis of the evolution of a protein-protein interaction (PPI) network it is necessary to have knowledge about the emergence of all the proteins in the network and their interactions. Some investigations in this line have been presented [12, 13]. The problem is that the steps which lead to the current situation cannot be directly addressed. To be able to follow the development in biological systems, information about biological evolution should be applied. One attempt in this direction was to use so-called isotemporal categories to express the appearance of a gene or protein in organisms [14].

DNA and protein sequences hold their past in their structure, which can be reconstructed with the standard tools of phylogenetics. Here we performed to our knowledge the first PPI network analysis utilizing phylogenetic information. One of the reasons for the lack of this kind of study is the problem of obtaining accurate and validated data for a system that covers the biological network, its gene and protein constituents and their evolutionary history and homologs. We have collected such data for the human immune system.

The human immune system, which is one of the most complex biological machineries, has been widely investigated at the molecular, cellular and organ level in its normal state and during disease. It is a very complex system built up from several different tissues, cell types, molecules and processes. The evolutionary history of the human immune system is widely studied (see e.g. [15]), but we still need new data to complete the picture.

The Essential Human Immunome is a reference collection of genes and proteins involved in human immunity which is distributed in the Immunome database [16, 17]. Evolutionary data for all these proteins is collected in the ImmTree database [18]. Experimentally verified protein-protein interaction data for the immunome proteins was collected from the Human Protein Reference Database (HPRD) [19] and used to reconstruct the protein interaction network.

Recently, several studies have been published about natural networks [20], ranging from social interactions [21, 22] via protein-protein interactions [23, 24] to the spreading of epidemics [25] as well as human made networks like telecommunication networks [26, 27] and the Internet [28, 29]. It was initially surprising that the different networks share several common characteristics, which suggests common organizing principles for their emergence [30].

Preferential attachment [31, 32, 33] suggests that a new protein is more likely to be connected to a highly connected protein than to a protein with fewer interactions. Networks built up according to this principle, have been shown to have scale free characteristics [34, 35]. Scale free networks have been detected from several different sources, but thus far it has been very rare to observe the preferential attachment rule at work, especially in the case of protein interaction networks [36, 37].

The evolution and development of networks has been widely studied [38, 39]. Many earlier reports on the evolution of protein interaction networks have focused on yeast high throughput datasets [14, 40, 41]. Evolutionary comparisons have been limited to a few reference genomes. The new dataset on the human immunome [16] related proteins and their evolution allowed us to assign evolutionary levels to the proteins [18]. These levels can be interpreted as indicating the steps of emergence of proteins. Reliable protein interaction data can be assigned from HPRD, which contains experimentally verified interactions from literature [19]. We studied the emergence of the immunome PPI network and elucidated which of the network parameters are shaped by evolutionary pressures and how these parameters have changed during time.

Further, interactions for low level subnetworks, levels 6–9, were collected for Drosophila melanogaster [42, 43] and Caenorhabditis elegans [42]. This was done to reconstruct early interactions between immunome proteins which might have been lost during the evolutionary processes. Only 13 new interactions were identified in these subnetworks and thus they did not affect the overall trends. It would have been beneficial to have PPI data for model organisms on all the 10 evolutionary levels. This is not currently possible due to lacking proteome wide PPI networks.

For the analysis, we made some assumptions. We used PPIs identified in the human proteome, and assumed that if two proteins which interact in human, they also interact in any organism in which the two proteins coexist. We also assumed that this interaction existed during the evolution since both of the proteins emerged in a linage. The model is a simplification of all possible cases. This assumption is also in use in interaction predictions [44, 45], and although not exclusive, it is still true for the majority of interactions. Our results for interactions in fruitfly and worm confirm this idea, because hardly no new interactions were found in these low level subnetworks. The model simplifies the evolution of the entire interaction network to the evolution of its nodes. This is necessary since we do not have a method to track the evolutionary past of interactions, while the phylogenetic analysis of the proteins has well established and accepted procedures.

The human immunome PPI network contains 1349 interactions for 584 proteins (Table 1). Since the network includes only experimentally proven interactions from the HPRD database, we can assume it represents a real, albeit incomplete, network model of protein interactions in the human immune system. Not all the immunome proteins are included because the data is not complete and does not cover all proteins in all cell types and conditions.

Of the investigated proteins, protein-tyrosine kinase FYN has the highest number of interactions. The subnetwork of FYN and its first neighborhood includes 47 proteins and 93 interactions, which account for about 8% of all the immunome nodes and 6.9% of interactions. Another Src-family member, lymphocyte-specific protein tyrosine kinase (LCK), is the second most linked protein with 37 interactions. 64 of the 584 proteins have more than 10 interactions. Many of these are mediators of signal transduction pathways, for example the Janus kinases (JAKs), the signal transducer and activator of transcription (STAT) family members, and the TNF receptor-associated factors (TRAFs) (see Additional file 1).

For further analysis, subnetworks were created for eight evolutionary levels. Levels 8 and 9 were excluded from the analysis because these levels contain only a few proteins and interactions. All subnetworks contain the nodes from the examined and earlier levels, and the interactions between them (Table 1, Fig 1, Additional file 2). They thus represent the interaction network that existed at different steps during evolution. The lowest, level 7, subnetwork is small, with 112 proteins and 133 interactions which existed in the ancestors of Homo sapiens when the taxon Bilateria was formed. In each evolutionary step the number of proteins and the interactions between them grows substantially. The degree distribution of the subnetworks follows power law distribution with the power law exponent between 2 and 3 (Fig 2.). The log-likelihood ratio (-2logΛ), which marks the likelihood of the power law degree distribution of the dataset, is much higher in the higher level networks (Table 1).

When analyzing the relationship between the evolutionary levels and the degrees (the number of interactions of the proteins) we expected proteins which appeared early on to have more interactions [37], and thus nodes with higher evolutionary levels should have higher degrees. However, a simple comparison of node level numbers and their degree in the level 0 network does not show this phenomenon (Fig 3). When we tested whether the new nodes introduced in evolutionary steps tend to attach to nodes with higher numbers of connections, a statistically significant preferential attachment is clear (Fig 4). We compared the degree distribution of all the nodes and of those nodes which get new connections in the next evolutionary level. The nodes with new connections have higher degrees in each step. This implies that when a new node is introduced into the immunome network, it most likely attaches to a node with a higher degree, so there is a bias toward attachment to higher degree nodes.

We used scale free models although the immunome protein interaction network, like other PPI networks, does not contain enough nodes to fulfill the statistical criteria for scale freeness. Therefore we mostly used general descriptive measures of networks, like efficiency, and avoided in our conclusions the scale free network specific aspects.

An important feature of the scale free protein interaction networks is that highly connected nodes tend to be essential and therefore more conserved [40, 46]. We used the average entropy [47] of the proteins to measure how conserved, and thus how essential they are. Entropy was used to measure the variability of the sites in a multiple protein sequence alignment instead of comparing a human sequence to an ortholog in a reference genome. We thus take variability into account from many sequences instead of a sequence pair. Proteins with high connectivity never have high entropy; yet on the other hand, some of the proteins with just a few connections have very high average entropy, which means that they are not conserved (Fig 5). We detect this phenomenon on all the evolutionary levels. These levels also contain enough nodes to allow binning of the data (Fig 6). More conserved proteins are more connected during the evolution of the immunome protein interaction network.

We further studied the effects on several network characteristics during the evolution of the network to find out what kind of selective forces affect its development. Global efficiency quantifies the efficiency of the network in sending information between nodes [48]. According to earlier studies, the efficiency of a scale free network is expected to decrease when the size of the network is growing [36, 48]. Surprisingly, the efficiency of the immunome network grows through the evolutionary steps, from an initial value of around 0.24 to a final value of around 0.41 although the number of nodes and edges is also growing (Fig 7A). This means that despite the number of nodes increasing from 112 to 584, the average number of steps necessary to reach one random node from another decreases. Since this is against the expected behavior (Fig 7A), we assume that a selection pressure exists which shapes the immunome PPI networks to became more efficient during evolution.

The critical components of a network can be searched by looking for the most vulnerable nodes [49]. Vulnerability is defined as the drop in efficiency when a node and all its edges are removed from the network. The maximal value of the vulnerability is the overall vulnerability of the whole network. The maximal vulnerability of the immunome network constantly decreases during the evolutionary steps from the initial value of 0.28 (Fig 7B). At the level of Homo sapiens (level 0) the value is 0.003, which means that maximal drop in the efficiency of the network is 0.3% if one of the nodes is deleted from the network. Scale free networks are known to be tolerant of errors in randomly chosen nodes [28], a feature that is also important for biological interaction networks.

Our findings of the increasing efficiency and the decreasing vulnerability of the evolving network also raise new aspects in evolution. Apparently, it is advantageous for an interaction network to become more efficient and less vulnerable to random errors. More robust networks can possibly increase the fitness of the organism. It is likely that similar behavior will be found from some other biological, and possibly other types, of networks. Our analysis revealed new concepts for evolutionary biology, as well as new insights on the emergence of scale free networks. Large scale experimental studies are needed to test and verify the observed network properties and their effect during evolution.