Protein sequences and profiles

Sequence profiles of secretion systems components were generated by selecting orthologues of each component in order to capture the diversity of the T3SS, T4SS and T6SS. Following the approach described by Abby and Rocha [20], we selected protein sequences corresponding to the components of the flagellar and non-flagellar T3SS (NF-T3SS). Meanwhile, to build profiles that represent the variety of T4SS, protein sequences of the components of 18 archetypal T4SS [25] were also selected. In both cases, sequences of each component were extracted from a set of model organisms representative of the diversity of all these types of systems. We based the construction of T6SS components profile on a previously reported list of coding sequences belonging to several bacterial genomes; sequences that were found to be orthologues of the components of the first described T6SS, namely, V. cholerae, Pseudomonas aeruginosa and Burkholderia mallei [26]. The sequences of these orthologues were also included in our set of protein families. All these sequences were downloaded from the NCBI website and selected based on their RefSeq genome annotation. Then, sequences corresponding to each component were aligned with Muscle [27] and manually adjusted with Seaview [28]. Finally, protein profiles were built with HMMER3 [24]. When fewer than five representative model organisms were found to code for a given component, no profile was built; instead, files were generated in multi-FASTA format.

We extended the above set by collecting protein sequences from AtlasT4SS [29] and proceeding in the same way as above to generate profiles of orthologue clusters described in that database. Secretion systems loci identified in this study were manually screened, and additional profiles were incorporated based on the RefSeq genome annotation. Consequently, our final dataset comprises sequence information for a total of 364 components of the T3SS, T4SS and T6SS (65, 449 and 20, respectively). Further information about each of the component profiles can be found in S1 Table.

Identification of secretion systems clusters

T346Hunter performs BLASTp [23] and HMMER3 [24] searches of the protein sequences and profiles described above against user-supplied genomic sequences. Regions containing homologous genes (Evalue < = 0.0005) of at least 4 different core components of T3SS, T4SS or T6SS and spanning up to 70 kb are retained and included in the output report. We consider core components of T3SS, T4SS and T6SS as described by Abby and Rocha [20], Bi et al. [25] and Shrivastava and Mande [26], respectively (see S1 Table for details).

Implementation

T346Hunter runs on a Linux platform with an Apache web server. The web interface was implemented using HTML and CSS, and data pipelines were developed using PHP, Perl, R and shell scripts. Circular genome images are generated using circos [30], and gene maps are produced using the R package genoPlotR [31]. Open reading frame predictions are generated with Glimmer v3.02 [32].

Using T346Hunter

T346Hunter provides a simple and user-friendly web interface (Fig. 1A) for the search of T3SS, T4SS and T6SS in user-supplied genomic sequences. A user can upload the sequence of interest at http://bacterial-virulence-factors.cbgp.upm.es/T346Hunter in two ways. First, the user can provide raw DNA sequences in FASTA format, in which case the system makes use of Glimmer v3.02 [32] to predict coding regions. When more than one sequence is submitted, T346Hunter interprets the set of sequences as a draft genome (S1 Fig.). Alternatively, the user can upload an NCBI-formatted sequence. In such a case, three types of data have to be uploaded to the server, including DNA sequence, protein sequences and the location of protein-coding genes. The format of the data should be similar to that used by the NCBI Genome FTP Server (fna, faa and ptt files). Users are encouraged to submit their sequences in this format, since T346Hunter does not take phylogeny into account to predict coding regions, leading to a potential decrease in the quality of gene calls. Besides the input sequence, several parameters are provided for the user to configure the search, including HMMER3 [24] and BLASTp [23] E-value thresholds, secretion systems to be predicted and sequence shape (circular or linear). Once the prediction is completed, secretion systems loci are displayed in an intuitive HTML document containing tabulated and graphical information of the regions, along with a whole-genome graphical view. A tab-delimited summary of the gene-by-gene search, which can be easily incorporated into data pipelines, is also provided. These results are stored on the T346Hunter server for a week.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Example of execution of T346Hunter using the sequence of B. pseudomallei 668 chromosome 2 as input.

A. Web interface of T346Hunter. A sequence file together with a hypothetical email address are shown as selected to be uploaded into the server. B. Genome-wide graphical view showing the predicted secretion systems of B. pseudomallei 668 chromosome 2. C. Genomic representation of one of the three NF-T3SS clusters identified, including a graphical gene map and a tabulated gene list with detailed information of each component. Hyperlinks to NCBI for direct execution of BLASTn and BLASTp against the non-redundant nucleotide and protein databases are provided for each gene within the loci. Some other relevant information is also included, such as the percentage of core components found in the cluster and PubMed hyperlinks to the studies we have based our methods on to build the component profiles found in such a cluster.

https://doi.org/10.1371/journal.pone.0119317.g001

T346Hunter output

Here, the result of the execution of T346Hunter on the sequence of B. pseudomallei strain 668 is shown as an example (Fig. 1). The genome of B. pseudomallei is typically comprised of two circular chromosomes, which encode several secretion systems [33,34,35]. B. pseudomallei is an aerobic, Gram-negative bacterium that infects humans, animals and even plants. It is the causative agent of melioidosis, an often-fatal disease that is endemic to Southeast Asia and Northern Australia, and whose infection can take place by ingestion, inhalation and skin abrasion. There is no vaccine available to protect against this pathogen, which is also highly resistant to antibiotics. All this makes it a potential organism to be used as a bioterrorism agent [36], and a growing interest on its virulence mechanisms has recently emerged. Consequently, hundreds of genome sequences are nowadays available for B. pseudomallei [37], and the role of T3SS and T6SS in the virulence of this species has been previously reported [34,35]. Fig. 1B shows the whole-genome graphical overview generated by T346Hunter displaying the predicted secretion systems clusters for B. pseudomallei 668 chromosome 2 (NCBI Refseq NC_009075). The system localises three clusters of NF-T3SS, one cluster of flagellar T3SS and five clusters of T6SS. These predictions are consistent with previously reported in silico analyses [20,38]. The bsa NF-T3SS of B. pseudomallei has been shown to be an important part of the virulence armoury of this strain [39]. Fig. 1C shows the output generated by T346Hunter containing detailed information of such cluster, including a tabulated output and a gene map graphic representing its genomic context.

Core components

The sets of core components used for T3SS, T4SS and T6SS were as described by Abby and Rocha [20], Bi et al. [25] and Shrivastava and Mande [26], respectively. However, there is no consensus on the definition of core in terms of secretion systems components. As long as we understand, “core” is the minimum set of components experimentally proven to be necessary for a secretion system to be functional. That appears to be the meaning used by Abby and Rocha [20] and Shrivastava and Mande [26] when they suggest a set of T3SS core components and T6SS components of major requirement, respectively. On the other hand, Bi et al. [25] do not explicitly describe minimum required sets of components, but suggest a list of core components for each of the 18 T4SS they collect in their database. It is not clear though whether such proteins are indispensable for these T4SS to be functional. For instance, the trb T4SS encoded by A. tumefaciens C58 [40] lacks TrbN, which belongs to the above core list. Given this controversy, T346Hunter makes no discrimination regarding the different uses of the core set, leaving to the user the role of interpreting the results. We chose 4 as the minimum threshold of core components after trying different values and manually screening the predicted secretion systems, since it offers a trade-off between false positives and false negatives. Again, it is the user who has to sift through the predictions.

Identification of T3SS, T4SS and T6SS gene clusters in sequenced bacterial genomes

Complete bacterial genomic sequences of 2997 chromosomes and 2164 plasmids sequenced available as of 14 February 2014 were downloaded from GenBank Refseq. T346Hunter was executed on these sequences and localised clusters enriched in either NF-T3SS, T4SS or T6SS components. In total, 2,814 clusters were identified (512 NF-T3SS, 1,466 T4SS and 836 T6SS) across 1,121 organisms. Predicted clusters are summarised in S2 Table and can be queried at http://bacterial-virulence-factors.cbgp.upm.es/T346Hunter#predicted_ref. Sequences with negative predictions are listed in S3 Table.

Comparison with currently available tools

In order to validate the performance of T346Hunter, systematic comparisons with other available applications for secretion systems prediction would certainly be the best choice. By crossing our predictions with loci identified by other servers one could have a measure of the relative accuracy of our tool. But, in practise, this is not straightforward to carry out. On the one hand, predicted data are not always available in a format that are ready to be systematically analysed. Servers usually provide their data in a way that either they have to be queried using some kind of criteria (e.g. strain name) or they are just embedded in the webpage. On the other hand, few servers are available to predict secretion systems clusters as such. To our knowledge, only SecReT4 [25] provides a specific tool for genomic localisation of T4SS clusters. Despite these difficulties, and given the need for assessing the accuracy of our predictions relative to others' work, we attempted to accomplish comparisons either by systematic processing, when possible, or by manual inspection, when not.

First, we aimed to compare our predictions of T3SS with those of T3SSscan-FLAGscan [20]. Such server does not localise T3SS clusters in user-submitted sequences, but does keep a repository of predicted NF-T3SS loci that can be accessed using different features. Therefore, we randomly queried clusters for 100 genomic sequences and manually compared them with our predictions. Since negative predictions are not provided in the server, this selection was restricted to positive predictions. We found that T346Hunter predicted any NF-T3SS cluster in the 100 sequences examined. More precisely, T346Hunter and T3SSscan-FLAGscan identified the same number of clusters in 95 sequences (S4 Table). For each of the resting five sequences, the number of predicted clusters just differs in one. This difference is probably explained by the different searching criteria used by both methods, particularly in defining contiguous genes within a cluster and setting the minimum required number of core components. In total, 125 clusters were identified by T3SSscan-FLAGscan and 128 by T346Hunter.

To test how T346Hunter performs in predicting T4SS, we compared it with SecReT4 [25]. This server provides a summary of T4SS predictions on a number of sequences by means of a table in the webpage. We could, then, proceed to systematically compare the two methods. Again, such a comparison was necessarily restricted to positive predictions. In total, 387 sequences were examined and all of which were found to contain at least one T4SS cluster. Of 387 such replicons, 324 (∼84%) were predicted by both methods to contain the same number of T4SS (S5 Table). In this case, differences in searching criteria may have had a stronger impact in the predictions. SecReT4, for instance, does not restrict T4SS genes to be located in a specific cluster. In contrast, T346Hunter identifies a cluster whenever orthologues of 4 core components are found in a window of up to 70 kb. Despite such divergent parametrisation, 53 out of the 63 sequences with discordant predictions (∼84%) only differ in one cluster. Accounting for the 387 sequences analysed, SecReT4 and T346Hunter identified 522 and 595 T4SS, respectively.

We went ahead to inspect the accuracy of our T6SS predictions. As a specific server for the prediction of T6SS is not available, tool-by-tool comparison was precluded in this case. However, systematic localisation of T6SS loci has been previously performed [26,38,41], reporting data we could use to compare our predictions with. We focused on Boyer et al. [38], which completed the widest analysis. Since no readily processable summary of the predictions was provided, comparisons had to be manually performed. Among the 100 sequences reported with positive predictions, T346Hunter identified at least one T6SS in 98 of them. Furthermore, both approaches predicted the same number of T6SS clusters in 93 sequences (S6 Table). Such a subtle difference is explained by the requirement of 4 core components in a cluster imposed by T346Hunter, which was not applied by Boyer et al. [38]. Nonetheless, predictions on the resting 7 sequences of both approaches differed in only one cluster. Summing up all identified clusters in the 100 sequences examined, T346Hunter and Boyer et al. [38] predicted 170 and 175, respectively.

Regarding general annotation engines, some of the most widely used tools are the NCBI Prokaryotic Genome Annotation Pipeline [21] and the RAST server system [22]. In the last few years, RAST has become particularly popular and is now frequently used to rapidly annotate bacterial genomes against its comprehensively curated subsystem database. Due to its constant growth, RAST automated annotations are nowadays of a great quality, having reached a high degree of specificity. Indeed, T3SS, T4SS and T6SS are included among the subsystems collection of RAST, and thorough reports of related genes are provided within general annotations. Such reports include visual and tabular information of the corresponding genomic clusters, thus offering an exhaustive output. However, bacterial strains that has not been incorporated into RAST database are reported with no subsystems, and users need to manually inspect individual features to infer the existence of secretion systems clusters. This makes RAST subsystems search dependent on its database of bacterial isolates, and makes it particularly not suitable for the analysis of newly characterised bacteria. Furthermore, when subsystems are reported, the number of secretion systems clusters are not directly shown in the output and it rather needs to be derived from reported tables. Besides, no information regarding core components is provided, and some conjugal T4SS are not categorised as such. Therefore, even though RAST performs quite well in detecting genes encoding secretion systems when compared to other general annotation tools, its annotations lack some relevant information on T3SS, T4SS and T6SS, and do not directly offer the whole picture of the underlying genomic clusters.