This study focused on the computational identification of host proteins targeted by an invading virus, using HIV-1 infection as a case study because extensive study at the molecular level has yielded nearly fifteen hundred experimentally determined HIV-1, human protein interactions, which are catalogued in the HIV-1, Human Protein Interaction Database [1, 2]. Virus and cellular parasite proteins alter host interaction networks by competing with host proteins for binding in the host protein-protein interaction (PPI) network [3–5]. Knowledge of which host proteins interact with virus proteins is important for antiviral drug discovery and treatment optimization using existing drugs [6]. Experimental approaches for finding virus protein binding partners in the human proteome have proved challenging because nearly thirty thousand human proteins must be tested [7]. Computational approaches have helped by reducing the number of host proteins to verify experimentally.

Previous host-pathogen interaction prediction methods focused largely on finding PPIs between human and cellular parasite proteins. One recent method found the probability that two protein domains interact given the human PPI network, and used this probability to find the likelihood that pathogen and human proteins interact given their domain profiles [8]. Another method matched human and pathogen protein pairs to proteins known to form complexes, and then filtered these interaction candidates based on expression data from human and pathogen [9]. Translating these methods to interactions between HIV-1 and human proteins has been difficult because HIV-1 proteins have few domains and their structures are hard to find by comparative modeling. For instance, to find structures for the N-terminal and C-terminal domains of HIV-1 VIF, two different protein structures were required for comparative modeling [10].

In our study, we focus on protein interactions mediated by short eukaryotic linear motifs (ELMs) [11] on HIV-1 proteins and human protein counter domains (CDs) known to interact with these ELMs. We aim to obtain host protein sets enriched with known sets of virus targeted proteins based on ELM and CD associations. The potential functional roles of interactions mediated by ELMs and their CDs in viral infection have been addressed in a number of recent articles [12–14]. The HIV-1 literature contains at least ten examples of HIV-1, human PPIs that are directly associated with motif and domain presence. The motif/domain basis of such PPIs is not restricted to a single HIV-1 protein, but is widely distributed across the HIV-1 proteome, including HIV-1 NEF [7], ENV [15], TAT [16], REV [16], VIF [17], and VPU [18]. This experimental evidence is the motivation for systematically investigating the association of motif/domain pairs with PPIs between virus and host proteins. Although Tastan et al. [19] estimated a relatively weak link between binding motif/domain presence and the actual virus-host PPIs, their work was restricted to predicting direct binding between host and HIV-1 proteins. In this study, we set out to identify host proteins involved in direct interactions as well as those that compete with HIV-1 proteins for binding to their host targets. Moreover, the algorithm presented by Tastan et al. is based on supervised learning and training from known interactions between HIV-1 and human proteins. In their method, each potentially interacting protein pair is associated with a feature vector composed of parameters related to Gene Ontology (GO), global gene expression profiles, the human protein interactome, and protein domains and motifs. Ours is a hypothesis-based approach, and does not require a priori knowledge of virus-host interactions beyond what can be gathered from viral and host protein sequences. As such, it is directly applicable to identifying host protein sets enriched with virus targeted host proteins for a wide scope of infectious diseases. The extremely low p-values we calculated for the overlap between our predictions and experimentally verified HIV-1, host protein interactions indicate the potential value of our approach for deducing a first draft of the molecular vocabulary employed in less studied host-pathogen protein interactions.

The rapid sequencing of viral genomes with next generation sequencing technology [34] makes it possible to link clinical parameters of viral infection to sequence motifs. The task of identifying host proteins targeted by a virus is worthwhile because such proteins may become drug targets to fight infection [6]. Experimental studies for determining virus targeted proteins are expensive and highly challenging [14]. Such efforts, although large-scale, have produced incomplete results for even well studied viruses like HIV-1 [6, 35, 36]. In this study, we used a systems approach to identify host protein subsets enriched by virus targeted proteins. Our method was based on the identification of host motifs on virus sequences. We used the a priori knowledge in the ELM resource to identify the counter domains associated with these motifs and information from the human interactome to focus on host protein interaction pairs with appropriate motif/domain links. KEGG pathways and the GO molecular functions were used to provide biological context to our findings.

The sets of host proteins we predicted as targeted by a given HIV-1 protein in KEGG pathways were highly statistically enriched with host proteins known to interact with the same HIV-1 protein (Figure 4). For example, the match between our predictions and the interactions for HIV-1 NEF in the HIV-1, Human Protein Interaction Database corresponded to a p-value of 4.26 E-21 in KEGG pathways enriched in our predicted set. After combining our predictions for all HIV-1 proteins, we had 607 proteins in HHP enriched KEGG pathways, and of these we matched 241 in the set of 877 experimentally verified proteins with a p-value of 3.11 E-58 (Figure 7). Our predictions were not nearly an exact match for experimental data, but our list was highly enriched with HIV-1 targeted host proteins. Given that HHP in KEGG pathways is about half as large as all HHP, and has a stronger overlap with HHE, experimentalists should begin verification with this set.

In addition to the binding/interaction research compiled in the HIV-1, Human Protein Interaction Database, recent experimental studies based on genome-wide siRNA screens have brought additional light to host-pathogen interactions that facilitate HIV-1 replication [6, 35, 36]. Three studies produced smaller lists of host proteins than the list in the HIV-1, Human Protein Interaction Database. The lower matrix in Figure 7 shows the five-way comparison of HIV-1 targeted protein lists: HHE, HHP, and the three screens. The table indicated the extent of discrepancy between lists, as well as the statistical significance of the matches between them. Our predictions matched HHE with the lowest p-value, and the genome-wide study lists generally matched each other better than the interaction studies. The list of 280 genes presented as host cellular factors required for HIV-1 replication by Brass et al. had 13 genes in common with the list of 295 genes deemed necessary by Konig et al. for regulation of early stage HIV-1 replication, and shared 10 genes with the 311 genes given in the Zhou study. When these proteins were projected into HPRD, the matches led to p-values of 7.35 E-4 and 4.46 E-5. Although the match was significant, there was still a discrepancy between the results. This mismatch may be attributed to the differences in the analysis and experimental methodologies used. Our predictions matched 56 of the 129 HPRD proteins presented by Konig et al. with a p-value of 0.15, 44 of the 91 HPRD proteins in the list by Brass et al. with a p-value of 0.03, and 54 of the 139 HPRD proteins given by Zhou et al. with a p-value of 0.52. These results indicated the challenges faced by experimental studies trying to uncover the grammar of HIV-1, host interactions.

Although our study produced host protein sets statistically enriched with proteins known to be targeted by HIV-1, mismatches between our predictions and experimental data cannot be ignored. It is possible that host-virus interactions employ a grammar that is much more complex than the short linear motif/counter domain interactions assumed in this study. The molecular vocabulary of PPIs is simply not well understood even for proteins belonging to the same species. However, one common mode of interaction is the binding of a linear binding motif on one protein to a domain on another protein [37]. A central hypothesis in the discovery of the linear binding motifs mediating protein interactions has been that proteins with a common interacting partner, such as protein kinases, share a common feature in the form of a motif [38]. Some of the linear binding motifs in the ELM resource have been shown to bind directly to sites at opposing counter domains listed in databases such as PROSITE and Pfam [39]. However, for approximately 30% of the PPI interactions listed in HPRD database, interacting proteins possess none of the already annotated domains. Thus, a model based on known motif/domain interactions would not be able to capture all of the known interactions in the host, let alone those between virus and host.

Another important cause of the discrepancy between our predictions and experimental data might have been the poor annotation of known motifs and counter domains used in this study [40]. Recent studies of domain-motif interactions indicated that the annotation signatures are more specific than those presented in ELM and PROSITE. This was found to be true for the HIV-1 interacting PDZ domain [12], SH3 domain [13] and others [14]. Emerging motif finding tools such as DILIMOT [41], SLIMFinder [42], and D-STAR [43] will help researchers improve the specificity of the motifs that mediate host-virus interactions. Still, the list of host proteins we have provided [see Additional file 5] comprises a candidate set for genome-wide studies of the regulation of HIV-1 replication and infection.

We focused on HIV-1 infection in this study because we desired to assess the effectiveness of our computational approach by comparing our predictions with large-scale experimental data. Our results provided a rationale for applying our method to predict virus-human interactions for sequenced viruses. A systems approach to predicting host-pathogen interactions will at least be partially based on the sequence motifs of interacting genome/proteomes. The present study illustrated the importance of ELMs in the molecular cross talk between host and virus and opened the door for more extensive experimental and computational studies of host-virus interactions.

In this study, we described a bioinformatics model to investigate the crosstalk between the HIV-1 and human proteins. Our method used multiple sequence alignments of HIV-1 proteins, and three datasets related to the host: decoded sequences of the host proteins, a priori knowledge of experimentally observed protein-protein interactions within the host proteome, and interactions between short linear peptide motifs and protein domains. The output of the model was a list of host proteins that may interact with specific HIV-1 proteins using specific sites. This list can be used to draft a connectivity map between virus and host, and to determine a set of protein interaction pathways that are significantly enhanced by host proteins predicted to be targeted by HIV-1.

The model was based on the assumption that virus proteins interact with host proteins though a set of conserved linear sequence motifs present in the host proteome. The conserved spatial organization of these motifs on the rapidly evolving HIV-1 proteome supported the assertion that short linear motifs play critical roles in interactions with the host network. The model's predictions led to host protein sets that are crowded by known HIV-1 targeted proteins. This statistical enrichment was particularly high along cellular pathways modulated by HIV-1. The model's predictions were also consistent with experimental data showing phosphorylation events as key targets of HIV-1 when redirecting cell protein networks toward the goal of virus replication.

The methodology applied here for HIV-1, host protein interactions is applicable to any viruses with multiple sequence alignments and hosts with known interactomes. Therefore, our approach has potential use in the identification of host proteins targeted by emerging and/or understudied viruses. The resulting list will be useful for selecting optimal drug therapies and discovering new antivirus drugs. The systems approach presented here for predicting host-virus protein interactions will benefit from ongoing research on the more specific annotations of short linear motifs and domains involved in protein-protein interactions.