The complete ortholog matrix from 18 eukaryotic species is shown in Additional File 1: Table S1. For brevity, only the orthologs between H. sapiens and five common model organisms are presented (Table 1). These orthologs were based on the HomoloGene database. Interologs were determined from the model organisms M. musculus (mouse), R. norvegicus (rat), D. melanogaster (fruit fly), C. elegans (worm) and S. cerevisiae (yeast).

Based on ortholog information, the conservation of genes and ortholog groups among 18 eukaryotic species were identified. We found 81 genes that were conserved in all 18 species presented in HomoloGene (Additional File 2: Table S2), suggesting that these genes are fundamental and/or vital to eukaryotes. Interestingly, 243 genes are missing in P. falciparum, but found in the other 17 species, including members of the proteosome, various ATP synthases and many mitochondria-related genes. While most species in the HomoloGene database share a high proportion of orthologs with other species (ranging from 48.3% in O. sativa to 87.4% in H. sapiens), less than 20% of the 5,266 genes in P. falciparum can be grouped with genes from other species. This suggests that the lifestyle and biological processes of this parasite deviate from those of other organisms.

PPIs from the various model organisms were used to infer PPIs (interologs) in higher order organisms such as H. sapiens. Because experimental PPIs from the target organisms are needed to verify these inferred PPIs, collections of PPIs are essential for an ortholog-based approach. The POINT database has collected most of the available public PPI data for a range of organisms (Table 2). It contains more than 44,000 human PPIs with available ortholog information. In addition, more than 70,000 yeast interactions are available, suggesting that a considerable number of human interologs can be inferred. Most of these interactions were obtained from high-throughput techniques such as yeast two-hybrid screening, which is prone to a high rate of false positives and other errors. Within the high-confidence dataset, where only PPIs verified by two or more methods or reported in the literature two or more times are included, there are 28,559 human PPIs and 25,612 yeast PPIs with available ortholog information.

While the use of high-confidence PPIs eliminates many potential PPIs that are present in the available datasets, this trimming process reduces the false positive rate. Among the organisms listed in Table 2, S. cerevisiae shows the most dramatic drop in the number of PPIs when only high-confidence PPIs are selected. The reason for this is obvious: this species is a single cell organism. Most of the PPI datasets were obtained using high-throughput approaches, and have not been verified by other methods or reported independently in the literature. For H. sapiens, the number of high-confidence PPIs exceeds even those of yeast. However, some species in the HomoloGene database do not have PPI data available. For example, P. troglodytes (Chimpanzee) and C. familiaris (dog) have no inferred human interologs despite the large number of orthologs they share with H. sapiens.

Given n objects in an undirected network (graph), there will be n(n-1)/2 relationships among these n objects and n*n relationships for a directed network. Since there are 19,491 human ortholog groups (Table 1), we therefore can assume that there are 19,491*(19,491-1)/2 pairwise interactions among these gene products. Certainly, a complete graph is not reasonable or biologically feasible. However, we can assume that each interaction can be associated with a probability and that the probability for a non-interacting pair will be 0. At this stage, we do not have enough information to assign a probability for each theoretical interaction. However, we can expand all 189,939,795 interactions among these 19,491 orthologous groups.

The interologs were inferred from ortholog information. Using the orthologous groups shared by humans and other species, we can obtain the maximum number of interologs from currently available interactomes. Only two orthologous groups shared by more than two species can be used to infer interologs. For example, if orthologous group A is shared by humans and mice, and orthologous group B is also shared by humans and mice, there will be a potential interolog A-B between humans and mice, although the probabilities associated with these two interactions (one in human and one in mouse) are not known.

Based on this assumption, we analyzed a number of orthologous group pairs and identified a number of species sharing these orthologous groups for H. sapiens (Additional File 3: Table S3). Among the 189,939,795 interactions, 180,191,177 interologs were inferred from ortholog information. This translates to 94.86% coverage of interologs (IC^HSA). Although the theoretical interolog coverage is high (nearly 95%), the interolog coverage on currently available PPIs is not significant. For all available human PPIs, only 3,859/44,251 interactions (8.72%) can be inferred from known interactions in other model organisms. Using the trimmed set of high-confidence PPIs, this coverage drops to 4.61% (1,316/28,559). There is an obvious gap between the theoretical upper boundary and the experimentally observed data.

To investigate the origin of this gap, we further analyzed the interolog coverage of each model organism. Five common model organisms were selected. The number of inferable interologs, experimental PPI derived interologs and their interolog coverage were calculated (Table 3 and Table S3). It is interesting that the most commonly used model organism, S. cerevisiae (yeast), has a theoretical interolog coverage of only 0.67% (interologs inferred from yeast divided by all human interactions), whereas the IC^HSA of M. musculus (mouse) and R. norvegicus (rat) are larger by two-orders of magnitude. However, for experimental human PPIs, the IC^HSA of mouse is only 2-fold higher than that of yeast, and the IC^HSA of rat is lower than that of yeast. The species contributions, C^Sp, shown in this table are also informative. While mouse contributes 43.07% of the known interologs, yeast contributes only 19.85%. This trend was mostly unchanged for high-confidence PPIs, except the contribution of yeast was boosted to 32.29%.

The mapping of all orthologous group pairs permits interolog prediction for various eukaryotic species. For example, in the POINeT web service http://poinet.bioinformatics.tw/, interologs can be inferred for seven eukaryotic species (H. sapiens, M. musculus, D. melangaster, C. elegans, S. cerevisiae, A. thaliana, and P. falciparum). Currently, more than 32,000 human interologs can be inferred. Among them, 3,859 have been confirmed by experimental evidence. The continual growth of interactomes in every eukaryotic species will continue to improve the ability to predict interologs.

P. falciparum is a parasite with a complex life cycle, and this malarial parasite threatens millions of lives worldwide. Based on the HomoloGene database, P. falciparum has the least similar genome in comparison to other species. Only roughly 20% (990/5,266) of the genes share orthologous groups with other organisms. This suggests that many cellular processes vital to other eukaryotes may be missing or replaced in P. falciparum, and the interplay between the parasite and its two hosts may compensate for the functions missing in the P. falciparum genome. The interactome of P. falciparum has been determined experimentally 18 and modeled genome-wide 19. This allows comparisons to be done between the genomes and interactomes of P. falciparum and its two hosts, H. sapiens and A. gambiae (the African malaria mosquito).

Using the experimental PPIs and interologs, 3,090 inter-species interactions between P. falciparum and H. sapiens (and not intra-P. falciparum interactions) were found (Additional File 4: Table S4). The Gene Ontology annotations of the P. falciparum and H. sapiens genes were identified. These inter-species PPIs have been grouped based on the ontology of their biological processes. The resulting network is illustrated in Figure 1. The nodes in Figure 1 are biological processes from P. falciparum and H. sapiens. Links between P. falciparum and H. sapiens biological processes were derived from interactions linking two genes that participate in the respective biological processes in the two species. Darker lines indicate the involvement of more interactions, allowing more interplay between the two biological processes. The P. falciparum biological processes are shaded using different levels of grey. Darker nodes indicate that more genes are involved in the process. In Figure 1, the metabolic processes and cellular processes of P. falciparum are most abundant in the host-parasite interaction network. This is understandable, since P. falciparum is a parasite and needs to acquire nutrients from the host erythrocyte. In the genomic-wide model of the P. falciparum interactome, only a small fraction of intra-P. falciparum interactions contributed to metabolic processes 19, which supports the notion that P. falciparum metabolic processeses may be dependant on human metabolic and cellular processes. There are also other interesting interactions between P. falciparum and the antimicrobial, antibacterial, cell killing and immune system processes of H. sapiens.

Although more than 3,000 H. sapiens-P. falciparum PPIs were inferred, not all of these interactions are likely to take place under physiological conditions due to spatiotemporal constraints. Filtering using gene ontology annotations resulted in 918 host-pathogen interactions. Further filtering of P. falciparum sequences using the presence/absence of translocational signals led to 95 PPIs (Figure 2). Only 15 P. falciparum proteins participate in these 95 PPIs (Table 4). One of the P. falciparum proteins, calmodulin (PF14_0323), interacts with 50 human proteins. It is well known that P. falciparum requires an environment with high Ca²⁺ levels 20, and the abundence of calmodulin-based interactions may help P. falciparum maintain this high Ca²⁺ concentration 21. Among the 50 human proteins interacting with PF14_0323, 13 also interact with human calmodulin (CALM3). This suggests that P. falciparum calmoduin shares some of the targets of human calmodulin, and may hijack these PPIs for its own purpose. The protein with the second highest number of interactions was N-myristoyltransferase (PF14_0127). Many proteins interacting with calmodulin require myristoylation in N-terminal 222324, further supports the functioning of the calmodulin-centric network.

Previously, Dyer et al. 16 have inferred host-pathogen interactions using Bayesian statistics. H. sapiens-P. falciparum PPIs predicted by the Bayesian approach are mainly enriched in 'blood coagulation' and 'membrane integration' protein interactions. This may partly be due to the gene ontology terms used to filter the PPIs. It is difficult to compare the two works, since the datasets and methodology used are different. However, the intersection of the two datasets reveals 3 interactions between PF14_0359 and the TNF receptor associated factor family (TRAF1, TRAF2 and TRAF6). PF14_0359 is a hypothetical protein. Inspection of the HomoloGene database reveals that PF14_0359 may be a homolog of DNAJA1 (HSP40). The functional implications of these three interactions require further investigation. However, TNF associated factor family are known to be involved in host immune response, suggesting that P. falciparum may interfere with this defence mechanism in H. sapiens. All in all, the diversity of different host-pathogen interaction inference methods suggests that these and other approaches may complement each other. And further development of the ability to predict host-pathogen interactions may benefit from the combination of multiple diverse approaches.

One of the limitations inherent in the analysis of interologs is the assignment of the orthologs, which is achieved using various BLAST algorithms together with several additional criteria 69112627 or from the NCBI HomoloGene and other protein/gene cluster databases. In this work, the ortholog information for each human gene was identified using the NCBI HomoloGene Release 54 28. The NCBI HomoloGene database contains homologous information for 18 eukaryotic organisms and has been augmented with homology and phenotype information drawn from various sources, e.g., MGI 29 and Fly base 30.

The new version of POINT integrated several publicly accessible PPI datasets (Additional File 5: Table S5). These data sources have diverse entry formats, disparate ID systems and different protein symbols. The diversity of these datasets made the task of performing cross-site browsing or iterative querying very tedious and challenging. We systematically re-organized these datasets to improve and standardize the publicly accessible PPIs. High-throughput PPI datasets are prone to false positives and errors. Therefore, we also generate a relatively high-confidence PPI subset, which refers to a PPI subset where the PPIs have been verified by two or more experimental methods or published in the literature two or more times.

The interolog coverage is quantifiable from an estimation of the ortholog-based PPI prediction power. The definition of interolog coverage is as follows:

I C H S A = N T H S A × 100 % MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemysaKKaem4qam0aaWbaaSqabeaacqWGibascqWGtbWucqWGbbqqaaGccqGH9aqpjuaGdaWcaaqaaiabd6eaobqaaiabdsfaunaaCaaabeqaaiabdIeaijabdofatjabdgeabbaaaaGccqGHxdaTcqaIXaqmcqaIWaamcqaIWaamcqGGLaqjaaa@3F05@

where T^HSA is the total number of human (H. sapiens) interactions (whether theoretical, experimental, or highly confident), N is the number of interologs, and IC^HSA is the interolog coverage for the human interactome. Another measure is the contribution of a given model organism to the human interologs and this is defined as

C S p = I S p T I H S A × 100 % MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qam0aaWbaaSqabeaacqWGtbWucqWGWbaCaaGccqGH9aqpjuaGdaWcaaqaaiabdMeajnaaCaaabeqaaiabdofatjabdchaWbaaaeaacqWGubavcqWGjbqsdaahaaqabeaacqWGibascqWGtbWucqWGbbqqaaaaaOGaey41aqRaeGymaeJaeGimaaJaeGimaaJaeiyjaucaaa@40FA@

where TI^HSA is the total number of human interologs, I^Sp is the number of interologs inferred from species Sp, and C^Sp is the contribution of species Sp to human interologs.

With the expanded orthologous pairs, intra- and inter-species PPIs can be inferred with ease. The inference of H. sapiens-P. falciparum interactions are based on orthologous pairs with one-side orthology to P. falciparum. For example, given a PPI between M_a and M_b in species M, if M_a has an ortholog in P. falciparum (P_a), and M_b has an ortholog H_b in H. sapiens (but not in P. falciparum), an interaction between P_a and H_b is inferred.

However, interologs inferred from orthologous pairs may not occur in vivo, especially in the case of inter-species interactions. P. falciparum inhabits a parasitophorous vacuole after its entry into the red blood cell. A translocational signal peptide (RELXE/Q) is required to translocate P. falciparum proteins into red blood cell cytoplasm for host-cell manipulation 313233. Also, proteins localized in the nucleus (both H. sapiens and P. falciparum) are not likely to participate in inter-species PPIs. Two filters have been applied to reduce such unlikely cases. The first filter utilizes gene ontology annotations. Human proteins with the following annotations were removed: mitochondria, nucleus, ribosome, cell process, helicase activity, complex, nuclease activity, nucleic acid binding, nucleotide binding or proteolysis. The second filter utilizes the translocation signal RELXE/Q, where X refers to any amino acids. P. falciparum sequences matching this pattern within the first 25% of its length are kept, since translocation signals are likely to appear at the N-terminal.