Hepatitis C virus infection protein network

Construction of an HCV–human interactome map

A comprehensive interactome map between HCV and cellular proteins was generated by Y2H screens. Twenty-seven constructs encoding full-length HCV mature proteins or discrete domains were cloned using a recombination-based cloning system (Walhout et al, 2000) (Supplementary Figure S1). Four independent screens were performed with each HCV bait protein, probing two distinct human cDNA libraries, either by mating (IMAP1 screens) or by transformation (IMAP2 screens; see Materials and methods). Fetal brain and spleen cDNA libraries were used instead of a liver library because the liver is known to overexpress a large number of secreted proteins, which could interfere with the quality of the screens. Comparing EST data from fetal brain and spleen with EST data from liver revealed that 87% of genes expressed in the liver are also expressed in brain or spleen (http://www.ncbi.nlm.nih.gov/sites/entrez?db=unigene). A total of 314 HCV–human PPIs were identified, involving 278 human proteins (Figure 1B, IMAP Y2H data set in Supplementary Table SI). More than 90% of the cellular interactors identified are expressed in the liver. Pairwise interactions between HCV and human proteins were also extracted from the literature by automatic text mining and checked by expert curation (see Materials and methods; IMAP LCI data set in Supplementary Table SI). A total of 135 PPI were extracted from Pubmed and 89 were extracted from BIND database (Bader et al, 2003) (Figure 1B). The resulting HCV–human interactome is thus composed of 481 PPIs with 65% new interactions, involving 11 HCV proteins and 421 distinct human proteins (Figure 1B). IMAP1 and IMAP2 screens share 22 interactions (7% of IMAP Y2H data set). This overlap is in the range of previously reported data using two Y2H high-throughput screening methods (Lim et al, 2006) and suggests that despite screens characteristics, summarized in Supplementary Table SII, saturation has not been reached. Differences in the screening methods, such as sensitivity of the different yeast strains to selective drugs, differential growth rate of colonies and low penetrance of interaction phenotype, could account for this observation. The low redundancy between IMAP Y2H and IMAP LCI data sets may also emphasize a high false-negative rate of the Y2H system, which would be in agreement with recent studies (Rual et al, 2005; Huang et al, 2007). An interesting hypothesis is that different methods of screening may lead to the exploration of different spaces of the HCV–human interactome. As false-positives may also contribute to the weak overlap of IMAP1 and IMAP2, two validation methods were used to assess the confidence of the IMAP Y2H data set. Two-thirds of the data set was retested by direct Y2H between viral protein baits and cellular protein preys identified by our Y2H screens (Y2H pairwise matrices). From the remaining interactions, 26 PPIs (25%) were retested by co-affinity purification and 22 PPIs could be validated (validation rate: 85%; Figure 1C and Supplementary Table SI). This Y2H data set was thus of very high confidence for further analysis at the topological and functional levels. In Table I, the top 21 new interactions validated by GST pull-down experiments and identified in one or two screens are shown. Analysis of the HCV-infection network (V-H_HCV, Figure 1A) showed that NS3, NS5A and CORE are the most connected proteins, with 214, 96 and 76 cellular partners, respectively, highlighting the potential multi-functionality of these proteins during infection (Supplementary Table SI, Figure 1D). Highly interacting proteins are known to be significantly more disordered than low-degree (LD) proteins (Haynes et al, 2006). Interestingly, NS3, NS5A and CORE are the only HCV proteins predicted to contain at least one intrinsic disordered region, according to DISOPRED2 (Ward et al, 2004) (prediction of protein disorder server; data not shown). This correlates well with the high degree (HD) of these proteins. In addition, 45 cellular proteins are targeted by more than one viral protein, suggesting their essentiality for virus biology (Calderwood et al, 2007) (Supplementary Table SIII).

A human PPI network (H–H network; Figure 1A) was reconstructed from eight databases (Gandhi et al, 2006) (see Materials and methods). This network is composed of 44 223 non-redundant PPIs between 9520 different proteins (Figure 2A, complete list of PPIs in Supplementary Table SIV), corresponding to 30% of the human proteome (the remaining proteins have no known cellular partners and can therefore not be included in this network). Interestingly, human proteins targeted by HCV (H_HCV) are clearly over-represented in this H–H network (IMAP Y2H data set: 76%; and IMAP LCI data set: 88%, exact Fisher test, P-value <2.2 × 10⁻¹⁶). This suggests that HCV preferentially targets host proteins already known to be engaged in protein–protein interactions (Rual et al, 2005; Stelzl et al, 2005). For the IMAP LCI data set, the higher percentage of H_HCV integrated in the human interactome may be explained by inspection bias of well-studied proteins and biological pathways. Analysis of H_HCV–H_HCV subnetwork (all connected H_HCV proteins) showed that cellular proteins interacting with HCV are significantly more interconnected than expected for random subnetworks (Figures 1A and 2B, Supplementary methods). Indeed, the 338 H_HCV integrated into the human interactome are distributed into 131 connected components (versus 276 expected by random subnetworks; z-score-based test P-value <10⁻¹⁰, Supplementary Table SV). The largest one is composed of 196 H_HCV (versus 18 expected by random subnetworks; z-test, P-value <10⁻¹⁰) and 127 are disconnected proteins. The three remaining connected components comprised two proteins. Two contained functionally related proteins (CLEC4M and CD209 are lectins involved in viral entry (Lozach et al, 2003); MVP and PARP4 are involved in Vault complex (Kedersha and Rome, 1986)) and one contained proteins not known to be functionally linked (KIAA1549 and CADPS).

Topological analysis of the HCV–human interaction network

To assess how HCV proteins interplay with the cellular protein network, we next focused on the centrality measures of H_HCV proteins integrated into the H–H interactome. Local (degree) and global (shortest path length and betweenness) centrality measures were calculated. Briefly, the degree (k) of a protein in a network corresponds to its number of direct partners and is therefore a measure of local centrality. Betweenness (b) is a global measure of centrality, as it measures the number of shortest paths (the minimum distance between two proteins in the network, l) that pass through a given protein. To provide an unbiased analysis, calculations were done on the basis of the 213 H_HCV from the IMAP Y2H data set integrated in the human interactome. The average degree, betweenness and shortest path length of the H–H network are 9.3, 1.6 × 10⁻⁴ and 4.04, respectively, which is in good agreement with previous reports (Ramirez et al, 2007) (Figure 3A). As the distribution of properties such as node degree and node betweenness in PPI networks appear to follow a power law, summarizing values by their distributions appears more appropriated for comparative analysis (Goh et al, 2002; Joy et al, 2005). The degree distribution of H_HCV and of the human interactome are significantly distinct (U-test P-value <10⁻³), with an average degree of H_HCV higher than the average degree of the human interactome (15.6 versus 9.3). The comparison of degree probability distribution reveals that H_HCV are preferentially represented in all class above the mean degree (Figure 3B, left). This indicates that HCV proteins have a strong tendency to interact with highly connected cellular proteins. However, as degree measures only local connectivity of proteins, global characteristics that could reflect information exchange and propagation in the network were investigated (Hernandez et al, 2007). At a global scale, the betweenness distribution of H_HCV and of the human interactome are significantly distinct (U-test P-value <10⁻³), with an average betweenness of H_HCV higher than the average betweenness of the human interactome (3.8 × 10⁻⁴ versus 1.6 × 10⁻⁴). As for the degree, the comparison of betweenness probability distribution shows an excess of H_HCV in all class above the mean betweenness (Figure 3B, right). In addition, the shortest path length distribution of H_HCV and of the human interactome were found significantly distinct (U-test P-value <10⁻⁵), with an average shortest path length of H_HCV lower than average shortest path length of the human interactome (3.50 versus 4.04) revealing the topological proximity of H_HCV. Both local and global centrality of H_HCV from the IMAP LCI data set were higher than for the IMAP Y2H data set, emphasizing the problem of literature inspection bias and reinforcing the unbiased approach of Y2H screening (Supplementary Table SV). To ensure that the preferential attachment to central H_HCV was not due to inherent bias associated to false positive in the H–H interactome, we performed the same analysis with a high-confidence, but less comprehensive, human interactome (Supplementary Table SV). This trend was maintained with this data set, confirming that H_HCV are highly central within the human interactome, both locally and globally, and appear relatively close to each other in this network. For comparative analysis of HCV and EBV, the centrality measures were also computed for H_EBV (data set from Calderwood et al (2007). Degree, betweenness and shortest path followed the same tendency with H_EBV proteins (Supplementary Table SV and Supplementary Figure S2) and were in good agreement with a previous report (Calderwood et al, 2007). These results indicate that preferential attachment on central proteins may be a general hallmark of viral proteins as recently suggested by analysis of the literature (Dyer et al, 2008). The high centrality of proteins was previously shown to correlate with their functional essentiality for the yeast model organism (Jeong et al, 2001; Ekman et al, 2006). In mammals, lethal and disease-related proteins were found enriched in central proteins (Wachi et al, 2005; Stark et al, 2006; Goh et al, 2007; Hernandez et al, 2007). This suggests that HCV proteins interacted with essential proteins in the cell.

To determine which of the degree or the betweenness most influences the probability of interaction between viral and cellular proteins, we used a generalized linear model to test the separate and additive effects of both measures (Supplementary methods). This analysis revealed that betweenness better explains the probability of interaction between viral and human proteins (ANOVA P-value <10⁻³). Figure 3C shows a partial correlation between k and b centrality measures (R²=56%, P-value <10⁻¹⁶), explained by the high variability of betweenness at LD values. We thus asked whether this high variability observed at LD could explain the preponderant effect of betweenness. For this purpose, the data sets were split in LD and HD protein classes according to the average degree of the human interactome. For cellular proteins included in LD class, HCV interacts preferentially with proteins of high-betweenness independently of their degree property (Figure 3D). Within the HD class, interaction with HCV proteins is dependent on both betweenness and degree of cellular proteins. On the basis of a recent study in yeast (Joy et al, 2005), it can be extrapolated that LD high-betweenness H_HCV proteins could exert an effect as connectors or bottlenecks between cellular modules and may thus be essential for the infection.

Functional analysis of the HCV–human interaction network

To better understand biological functions targeted by HCV, we next tested the enrichment of specific pathways for all interactors of a given viral protein. This was done by analysing the H_HCV proteins with regard to the KEGG functional annotation pathways (Table II, Materials and methods). Although this approach is not totally unbiased because functions have not yet been attributed to all proteins, it remains a powerful way of incorporating conventional biology in system-level data sets. This analysis showed enrichment for three pathways associated with HCV clinical syndromes (insulin, TGFβ and Jak/STAT pathways) and identified focal adhesion as a novel pathway affected by HCV.

Conclusion

In a network approach of HCV infection, the interaction map identifies all connections potentially needed for the virus to replicate and escape host defence. Whether all interactions really occur and have functional consequences is the open question of all interactome studies. The answer to this question necessitates the integration of system-level data sets of different origins that will set the stage for complex systems analysis of the infection. In a complex biological system, function cannot be predicted without understanding the component parts and their interactions and will result from the combination of theoretical knowledge of the cellular network with biological measurement of the interactions. Biological measurement, however, is still in the realm of low-throughput biology and needs major experimental improvement before prediction becomes the rule rather than the exception. Another fascinating challenge of this approach is to identify molecular signatures common to several viruses at the protein network level to develop original large-spectrum anti-viral molecules. A major step towards this goal is the high-throughput screening of a large variety of viruses, which is the aim of I-MAP.

Construction of the HCV ORFeome

All HCV protein sequences were cloned in full length and domains except NS4B, for which no domain has been designed, using the euHCVdb facilities (http://euhcvdb.ibcp.fr; Combet et al, 2007) (Supplementary Figure S1). NS4A–NS3 fusion protein, as well as NS4A–NS3 protease domain were constructed (Kim et al, 1996; Taremi et al, 1998). All 27 ORFs from the HCV genotype 1b, isolate con1 (AJ238799) (Lohmann et al, 1999), were cloned in a Gateway recombinational cloning system (Walhout et al, 2000). Each ORF was PCR-amplified (with KOD polymerase, Novagen) using attB1.1 and attB2.1 recombination sites fused to forward and reverse primers, then cloned into pDONR223 (Rual et al, 2004). All entry clones were sequence-verified.

Yeast Two-hybrid (Y2H) screens

HCV ORFs were transferred from pDONR223 into bait vector (pPC97) to be expressed as Gal4–DB fusions in yeast. Two different screening methods were used (IMAP1 and IMAP2). For IMAP1, bait vectors were introduced in MAV203 yeast strain, and both human spleen and fetal brain AD-cDNA libraries (Invitrogen) were screened by transformation as described (Li et al, 2004). All primary positive clones (selected on SD−W−L−H+3−AT) were tested by further phenotypic assay using two additional reporter genes: LacZ (X-Gal colorimetric assay) and URA3 (growth assay on 5-FOA supplemented medium). Positive clones that displayed at least two out of three positive phenotypes were retested in fresh yeasts: bait vectors were retransformed into MAV203 and each prey cDNA (obtained by colony PCR, see below) were transformed in combination with linearized prey vector (gap repair; Walhout and Vidal, 2001). Clones that did not retest were discarded. AD-cDNA were PCR-amplified and inserts were sequenced to identify interactors. IMAP2 screens were performed by yeast mating, using AH109 and Y187 yeast strains (Clontech; Albers et al, 2005). Bait vectors were transformed into AH109 (bait strain), and human spleen and fetal brain AD-cDNA libraries (Invitrogen) were transformed into Y187 (prey strain). Single bait strains were mated with prey strain, then diploids were plated on SD−W−L−H+3−AT medium. Positive clones were maintained onto this selective medium for 15 days to eliminate any contaminant AD-cDNA plasmid (Vidalain et al, 2004). AD-cDNAs were PCR-amplified and inserts were sequenced.

Text-mining of interactions between HCV and human proteins

Literature-curated interactions (LCI), describing binary interactions between cellular and HCV proteins, were extracted from BIND database and PubMed (publications before August 2007) by using an automatic text-mining pipeline completed by expert curation process. For the text-mining approach, all abstracts related to ‘HCV' and ‘protein interactions' keywords were retrieved, subjected to a sentencizer (sentence partition) and a part-of-speech tagger for gene name (based on NCBI gene name and aliases) and interaction verbs (Rebholz-Schuhmann et al, 2008) (interact, bind, attach and so on). Sentences presenting co-occurrences of at least one human gene name, one viral gene name and one interaction term were prioritized to curation by human expert.

Validation by co-affinity purification

Cellular ORFs (interacting domains found in Y2H screens) were cloned by recombinational cloning from a pool of human cDNA library or the MGC cDNA plasmids using KOD polymerase (Toyobo) into pDONR207 (Invitrogen). After validation by sequencing, these ORFs were transferred into pCi-neo-3 × FLAG gateway-converted. HCV ORFs were transferred into pDEST27 (GST fusion in N-term). A total of 4 × 10⁵ HEK-293T cells were then co-transfected (6 μl JetPei, Polyplus) with 1.5 μg of each pair of plasmid. Controls are GST-alone against 3 × FLAG-tagged prey. Two days after transfection, cells were harvested and lysed (0.5% NP-40, 20 mM Tris–HCl (pH 8.0), 180 mM NaCl, 1 mM EDTA and Roche complete protease inhibitor cocktail). Cell lysates were cleared by centrifugation for 20 min at 13 000 r.p.m. at 4°C and soluble protein complexes were purified by incubating 300 μg of cleared cell lysate with 40 μl glutathione sepharose 4B beads (GE Healthcare). Beads were then washed extensively with lysis buffer and proteins were separated on SDS–PAGE and transferred to nitrocellulose membrane. A total of 50 μg of cleared cell lysate was analysed by western blot to check the amount of 3 × FLAG-tagged cell protein. GST-tagged viral proteins and 3 × FLAG-tagged cellular proteins were detected using standard immunoblotting techniques using anti-GST (Covance) and anti-FLAG M2 (Sigma) monoclonal antibodies.

Integrated human interactome network (H–H network)

Only physical and direct binary protein-protein interactions were retrieved from BIND (Bader et al, 2003), BioGRID (Stark et al, 2006), DIP (Xenarios et al, 2002), GeneRIF (Lu et al, 2007), HPRD (Peri et al, 2004), IntAct (Kerrien et al, 2007), MINT (Chatr-aryamontri et al, 2007) and Reactome (Vastrik et al, 2007). NCBI official gene names were used to unify protein ACC, protein ID, gene name, symbol or alias defined in different genome reference databases (i.e ENSEMBL, UNIPROT, NCBI, INTACT, HPRD and so on) and to eliminate interaction redundancy due to the existence of different protein isoforms for a single gene. Thus, the gene name was used in the text to identify the proteins. Finally, only non-redundant protein–protein interactions were retained for building the human interactome data set.

Topological analysis

The R (http://www.r-project.org/) statistical environment was used to perform statistical analysis and the igraph R package (http://cneurocvs.rmki.kfki.hu/igraph/) to compute network connected components, centrality (degree, betweenness) and shortest path measures.

The Wilcoxon–Mann–Whitney rank sum test (the U-test) was chosen to statistically challenge observed differences. The U-test is a non-parametric alternative to the paired Student's t-test for the case of two related samples or repeated measurements on a single sample. The generalized linear model and ANOVA analysis was used to respectively model and test the separate and additive effects of degree and betweenness on the probability that HCV proteins interact with human proteins.

Functional analysis using KEGG annotations

Cellular pathway data were retrieved from KEGG (Aoki-Kinoshita and Kanehisa, 2007) and the Kyoto Encyclopedia of Genes and Genomes (http://www.genome.jp/kegg/) and were used to annotate NCBI gene functions. For each viral–host protein interactors, the enrichment of specific KEGG pathway was tested by using an exact Fisher test (P-value <5 10⁻²) followed by the Benjamini and Hochberg multiple test correction (Benjamini et al, 2001) to control false discovery rate.

Cell-adhesion assay

Serial dilutions (from 20 to 0.04 μg/ml) of fibronectin or poly-L-lysine in PBS were coated on 96-well microplates overnight at 4°C. Non-specific binding sites were saturated at room temperature with PBS 1% BSA for 1 h. HEK 293T cells were transfected with pCi-neo-3 × FLAG NS2, NS3, NS3/4A or NS5A (JetPei, Polyplus), collected 2 days later with 2 mM EDTA in PBS, spread in triplicate at 1 × 10⁵ cell per well in serum-free medium with 0.1% BSA and incubated for 30 min at 37°C. Non-adherent cells were washed away and adherent cells were fixed with 3.7% paraformaldehyde. Cells were stained with 0.5% crystal violet in 20% methanol for 20 min at room temperature and washed five times in H₂O. Staining was extracted from 50% ethanol in 50 mM sodium citrate, pH 4.5, and the absorbance was read at 590 nm on an ELISA reader (MRX microplate reader, Dynatech Laboratories). Values were normalized to 100% adhesion at 10 μg/ml. The percentage of adhesion was determined for each cell type at each matrix concentration. 50% of maximum adhesions (FA50) were calculated from the curves (Supplementary Figure S3) (adapted from Miao et al, 2000).