In order to investigate the extent of eukaryotic targeting signal usage by the viral proteins considered, we scanned the human, vaccinia virus and cytomegalovirus proteomes using various bioinformatics predictors that identify these signals. To avoid redundancy in the datasets, we considered all proteins available in UniProt 24 from one representative strain of each virus (we chose the AD169 strain for the cytomegalovirus and the Copenhagen strain for the vaccinia virus). As shown in Table 1, despite differences in genome size of several orders of magnitude, several targeting signals are found to a similar extent in both viral and human proteomes. In particular, large numbers of these viral proteins contain N-terminal signal peptides and anchors, consistent with the knowledge that many glycoproteins encoded in these large viruses require entry into the secretory pathway and have evolved to modulate ER quality control mechanisms to ensure that large quantities of viral proteins can be correctly produced and assembled into infectious particles 25. Similarly, a high proportion of viral proteins are predicted to contain at least one transmembrane domain. This reflects the high degree of interaction of these enveloped viruses with cellular membranes for functions that include assembly of viral particles and budding of the virus 26, and thus the need for insertion of a large proportion of their proteins in membranes, to participate in and modulate these processes. The vaccinia virus and cytomegalovirus proteomes also contain proteins that are predicted to harbor mitochondrial targeting peptides. Both cytomegalovirus and vaccinia virus are known to encode at least one protein that is localized to mitochondria, where they play a role in the inhibition of apoptosis 7. GPI anchors, which allow the attachment of proteins to the extracellular leaflet of the plasma membrane, are also predicted to be used by these viral proteins, to a similar extent as by human proteins. This might constitute a significant viral localization mechanism. In contrast to the relatively large proportion of viral proteins harbouring a C-terminal GPI-attachment region, very few of these viral proteins are predicted to be prenylated, which might reflect a greater need for extracellular rather than intracellular anchoring of these viral proteins in the plasma membrane.

Nuclear localization signals (NLSs) can also be detected in the viral proteomes. A larger proportion of cytomegalovirus proteins are predicted to contain NLSs than those encoded by the vaccinia virus genome. This is consistent with the fact that the cytomegalovirus genome replication as well as its viral core and capsid assembly occur in the nucleus whereas the vaccinia virus coordinates these processes in the cytoplasm.

We also detected the presence of short targeting signals in these proteomes. The N-terminal KDEL-like endoplasmic reticulum (ER) retrieval motif that is present in approximately 20% of human ER lumenal proteins 27 does not seem to be used by these viral proteins but the highly non-specific peroxisomal-targeting signal is present to the same extent in these viral and human proteins.

The most prevalent functional eukaryotic domains present in these viral proteins are also shown in Table 1, as predicted by InterPro 28. The immunoglobulin-like domain is the most widely used eukaryotic domain in both cytomegalovirus and vaccinia virus, which are well known to extensively modulate the immune response of the host (reviewed in 2930). The galactose oxidase and proteinase inhibitor I4 domains are over-represented in vaccinia virus but absent in cytomegalovirus suggesting that these domains are not used as part of a viral strategy common to these two viruses but are rather specific to vaccinia virus. Similarly, the rhodopsin-like GPCR superfamily is prevalent in cytomegalovirus proteins but absent from vaccinia virus. Cytomegalovirus is known to encode at least four G-protein coupled receptors, which could allow it to modulate and antagonize host signalling pathways 31.

Interestingly, protein-protein interaction domains such as SH2, SH3, WW and t-snare domains are conspicuously absent from these viral proteomes (data not shown), indicating that mimicry and modulation of this type of cellular communication mechanism might not be part of the survival strategy of these viruses.

The very high proportion of viral proteins containing one or several eukaryotic targeting motifs and functional domains shows the extensive usage of cellular localization mechanisms and machinery by these viruses. This provides a good indication that eukaryotic protein subcellular localization predictors might perform well on these viruses.

We used the PSLT human subcellular localization predictor 22 to predict the localization of cytomegalovirus and vaccinia virus proteins and to investigate whether principles of eukaryotic protein localization prediction can be applied to viral proteins. PSLT is a Bayesian network type tool, trained on human sequences, that predicts the subcellular localization of proteins based on the co-occurrence of protein domains, motifs and targeting signals. Table 2 shows the predictions of vaccinia virus proteins whose cellular localization has already been studied experimentally and is available in the literature (the full prediction dataset is available as supplementary material, please see Additional file 1). As shown in Table 2, the localization of most vaccinia virus proteins is well-predicted. The accuracy of PSLT on this dataset can be estimated to be 78% (proteins localized to more than one compartment are considered to be accurately predicted if at least one predicted compartment agrees with the previous literature annotation). A large proportion (36%) of these proteins are predicted to be secreted or expressed on the cell surface as integral membrane proteins or membrane anchored proteins. For the most part, this prediction is confirmed in the literature, but it should be considered a conservative estimate, since experimental studies cannot always sample the kinetics of viral protein synthesis and trafficking in all systems under all conditions. This estimate of extracellular and cell surface viral proteins is higher than our estimate of 22% for human cellular proteins 22, and likely reflects important viral functions that require using the host secretory pathway. Indeed, several of these proteins are known to be involved in modulating the host immune response including secreted proteins that bind chemokines, interferons and interleukin family members 3032. Other such proteins are incorporated in the viral envelope. Few or no vaccinia proteins are predicted to localize to the peroxisome, lysosome, ER or Golgi apparatus.

Table 3 shows the PSLT predictions for cytomegalovirus proteins whose cellular localization has already been studied experimentally and is available in the literature (the full prediction dataset is available as supplementary material, please see Additional file 2). The prediction accuracy of PSLT on this dataset is estimated to be 60% according to the literature. Almost all proteins classified as wrongly predicted according to the literature are annotated as localized in the ER or Golgi apparatus but predicted by PSLT as being on the cell surface. Several of these proteins display characteristics of cell surface or secreted proteins such as the capability to bind MHC class I and class II antigens. However, instead of being secreted, these cytomegalovirus proteins localize to the ER where they bind the MHC antigens, effectively targeting them for degradation and leading to the protection of cytomegalovirus-infected cells from CD8+ and CD4+ T lymphocytes 33. Many other cytomegalovirus proteins are well-predicted including cell surface receptors, several of which mimic host receptors 34 as well as secreted proteins such as viral chemokine and IL-10 homologues 3536.

We investigated whether the prediction accuracy of PSLT was correlated with the degree of similarity between the viral proteins and their closest human homologue. The two rightmost columns of Tables 2 and 3 show the closest human homologue from the NCBI RefSeq 37 database for each viral protein, as determined by BLAST 38. In general, viral proteins that have close human homologues (BLAST e-value <= 1e-10) are accurately predicted by PSLT. The prediction accuracy for these proteins is 100% for the cytomegalovirus and 91% for the vaccinia virus. Some viral proteins that do not have close human homologues (BLAST e-value > 1e-10) are well-predicted but the overall prediction accuracy of PSLT for these proteins is lower (43% for cytomegalovirus proteins and 67% for vaccinia virus proteins). This is consistent with previous analyses which allowed us to show that the prediction accuracy of PSLT is greater when predicting proteins from organisms that are evolutionarily close to those used to train the predictor 22.