Allele frequency

Figure 1 shows the MHC allele histogram of the predicted epitopes of all 553 Mtb H37Rv exported proteins. The horizontal axis represents the alleles available for prediction through the NetMHC software (version 3.0), and the vertical axis represents the absolute number of epitopes predicted by each allele of all exported proteins. The MHC alleles are ordered according to their decreasing number of predicted epitopes. The first five MHC alleles are human and represent 52.32% of all predicted epitopes, the first 15 represent 80.83%, and the last 24 MHC alleles only represent 2.58% of the overall NetMHC epitope prediction.

Control datasets

In the Figure 2, the control groups were divided in panels exhibiting protein quantity, percentage regarding this quantity and the average MED score. The horizontal axis of all three panels states the predicted sub-cellular location (cytoplasmic, membrane bound, PSE or secreted) for three groups of proteins: the Doytchinova et al. (2007) control groups (positive and negative control groups represented by Dplus and Dminus, respectively) and an Mtb positive control group (Mtbplus) taken from the AntigenDB. The vertical axis displays the data (from top to bottom): number of proteins, the percentage represented by the number of proteins and the average MED score for each control group. The number of proteins (top panel) and percentage (middle panel) predicted as cytoplasmic represent the majority for both Dminus and Mtbplus groups, while the Dplus group has more predicted exported proteins. Curiously, the Mtbplus group has the majority of cytoplasmic predicted proteins, which is surprising as it was expected that the majority of antigenic proteins would be exported to the extracellular milieu, as observed in the Dplus group that contains several pathogenic organisms.

Two results should be noted in the bottom panel. Firstly, the average MED scores were very similar among the three control groups, showing that MED is not necessarily a binary statistic classifier for targets but also a continuous statistic measure capable of defining the preferable targets; however, when significant differences between MED scores are shown, it can be used just like a binary classifier. This procedure was assessed in the evidence dataset shown in the next section. Secondly, the average MED score for proteins predicted as membrane-integral were shown to be twice as great as in the other sub-cellular compartments. This result agrees with other work in which signal peptides and trans-membrane domains were found to have exceptionally high CD8+ T cell epitope densities [17].

Evidence dataset

Figure 3 shows a histogram representing the distribution of MED scores for all 553 Mtb exported proteins. As seen in Table 1, MED scores range from 15.67 to 27.00 nM/mer, with the highest MED score data set represented on the far right side of Figure 3. These values strongly contrast with MED scores of Table 2, which are between 0.00 and 3.19 nM/mer, with the lowest MED score dataset represented on the far left side of Figure 3. As mentioned in the previous section, the MED score is not a binary classifier but is also capable of analyzing proteins scored within these extremely different ranges, allowing us to develop evidence for the general importance of MED scores.

MED score limitations

Figure 4 is useful to understand the main limitation of MED scores. It shows two pair of box plots, each pair representing a numerator (predicted epitopes) and a denominator (possibilities or chances for epitopes) that are used in Equation 1. The first pair of boxes show data from the numerator and denominator from the 30 lowest MED scored proteins from the Mtb exported proteins, shown at the far left side of Figure 3; the second pair of boxes show data from the 30 highest MED scored proteins from the Mtb exported proteins, shown at the far right side of Figure 3. The numerators and denominators were investigated to determine how protein length can influence the MED score. The number of epitopes predicted in the highest-scored subset is more than twice as high as the lowest-scored subset. This result was expected because there is evidence that the highest-scored subset is composed of proteins related to antigenicity or contributing to the bacterial pathogenicity while the majority of the lowest-scored subset is not. The number of possibilities for linear epitopes in the lowest-scored subset is almost three times higher when compared to the highest-scored subset. This numerical difference in the denominators is the major limitation for the MED score strategy, especially for data above the average. Quartiles Q3 and Q4, among those with lowest chances, include half (7/14) the evidence, in contrast to our hypothesis of an existing relation between MED and promising reverse vaccinology targets. These quartiles include denominators between 537 and 1,860 (just one greater than 1,498). Thus, according to the data, MED scores tend to indicate false positives when there is a difference factor of at least five between the number of predictions and the number of epitope possibilities located in the mature amino acid sequence portion. No false positives were observed when this factor was less than two. An interesting result is that the two biggest control groups from Figure 2, Dplus and Dminus, had average factors (fold) of 3.22 and 2.82, respectively.

MED score sensitivity

Among the 30 proteins that were lowest scored by MED, 14 showed contrary evidence and just four favorable evidence to the MED score concept. Among the 30 highest scored proteins, there was favorable evidence for 21 proteins based on the MED score and no protein with contrary evidence. Among the lowest and highest scored remainders, none showed favorable or contrary evidence related to MED scores. These results were used to create Figure 5 with a ROC curve graph that calculated sensitivities of 84% for MED scores with 7% false positives.

Novel probable putative Mtbantigens

The Mtb H37rv proteins Rv0235c, Rv0492A and Rv1004c were predicted to have some of the highest MED scores: 17.78, 20.31 and 18.58 nM/mer, respectively. The former two were predicted to be potentially exposed on the bacterial surface, and the latter was predicted to be secreted. Respectively, there are 78, 43 and 228 predicted epitopes against 138, 73 and 386 epitope chances for these proteins. This is the first published indication of their roles in bacterial antigenicity; MED scoring results suggest these proteins as useful putative targets for future investigations.

Allele frequency

The available methods for MHC epitope prediction take into account allele frequency in the selection of potential candidates [18, 19]. Some alleles are extremely rare; others are specific to some population or widespread [20]. The tools applied here to search for epitopes are not novel, but the way the results are read from standard software tools can be considered a novelty. We proposed to interpret not only epitope prediction from some specific MHC alleles, but from all available alleles. This proposition has a rationale: the idea of assessing the immunogenic potential of a protein, independent of alleles, helps to avoid excluding a protein from a list of in silico candidates just because the suitable allele for a specific population was not selected. For example, there are pathogenic organisms that cause different diseases in different hosts, including humans, caprines, ovines, equines, bovines and buffaloes [21–29]. In such cases, it is not reasonable to exclude a single allele from the current limited number available in software tools.

Control datasets

Even within the Dminus group, the average MED scores were similar to those from the Dplus and Mtbplus groups. Because of this, we focused on predicted exported proteins to create a priority list of targets for the Mtb genome, which is a reasonable strategy because one of the main differences between the Dminus and the Dplus groups are the number of predicted cytoplasmic versus exported proteins: 111 and 10 for Dminus versus 35 and 72 for Dplus, respectively. It is more likely that exported proteins interact with the host cells than membrane and cytoplasmic proteins [6, 9, 10, 30]. However, it is important not to neglect proteins that could be exported via non-classical mechanisms. This conclusion can also be drawn out by analyzing the middle panel of Figure 2, where the majority of Mtbplus proteins are classified as cytoplasmic. Medpipe allows the prediction of cytoplasmic targets, but this is the major part of any bacterial genome; medpipe still does not allow differentiating between cytoplasmic proteins without classical exportation motifs and those exported via non-classical pathways.

In addition, it is quite difficult to compare MED scores with previous trained software for antigenic features as such programs tend to be binary classifiers [31–33]. For instance, two control datasets used here were split into training sets (75 proteins) and test sets (25 proteins). Such division does not make sense for MED score because it does not depend on training steps; instead, the MED technique searches for immunological features based on a probable immunological memory concerning epitopes from known pathogens. In this regard, the results obtained with the evidence dataset is more informative because they represent experimental evidence of predictive strengthens or weaknesses of the method.

Evidence dataset

An extensive literature search for proteins from the well-studied Mtb organism gave experimental indication to validate our hypothesis that promising proteins for reverse vaccinology can be revealed based on the overall set of predicted epitopes. When searching for literature evidence, regarding the proteins within the evidence dataset, experimental results of other proteins were also found but not included in this work. This approach was chosen because it is not possible to determine a mean value for MED scores in order to use it as a binary classifier because the number of epitopes predicted per protein can vary significantly. This limitation was less difficult to work with when considering only 60 proteins: the 30 higher and the 30 lowest MED scored proteins out of 553 Mtb's predicted exported proteins (Figure 3).

NetMHC version

The newest NetMHC software (version 3.2) offers the ability to predict epitopes for 57 MHC alleles (http://www.cbs.dtu.dk/services/NetMHC/), but there is not yet a stand-alone version available to download. The NetMHC version (3.0) used here is the previous version and offers the possibility to predict epitopes for 55 MHC alleles [34]. However, the changes in version 3.2, compared to version 3.0, include a small increment in the number of MHC alleles and the possibility to predict epitopes of lengths ranging from 8 to 14mers. The authors of version 3.2 advise that predictions of peptides longer than 11mers have not been extensively validated. They also advise caution regarding predictions involving 8mers, as some alleles might not bind 8mers to any significant extent (http://www.cbs.dtu.dk/services/NetMHC/). Moreover, most MHCs prefer peptides of 9mers and the alleles' set from the version 3.0 are still present in version 3.2 [18]. Therefore, epitope predictions based on version 3.0 are still valid to answer relevant biological queries.

Are these pathogenic proteins?

The method presented here was initially conceived to predict antigenic proteins. Our approach is based on the fact that both antigenic and pathogenic proteins can be useful for vaccines and diagnosis and such targets can be revealed by the overall set of predicted epitopes and their concentrations in mature proteins. As related in the methods section, the in silico predicted exoproteins were ordered by decreasing MED score values. Following this sorting, the literature was searched for evidence proving or denying the contribution for the bacterial pathogenicity of each protein. The majority of the true positives presented here (Table 1) showed pathogenic instead of antigenic evidence (16 out 21 true positives), as detailed in the additional file 1. One protein (Rv3018c) has evidence for both antigenicity and pathogenicity simultaneously. In the same way, this criterion was also applied to the true negatives (Table 2), where seven out of 14 contrary cases fit into the pathogenic class instead of the antigenic one. Could these apparently unexpected results have a rationale? Could pathogenomics explain these findings? Pathogenomics is defined as the analysis, at genomic level, of the processes involved in bacterial pathogenesis caused by the interaction of pathogenic microbes and their hosts [35]. The identification of mutants showing altered pathology may be a useful framework to understand tuberculosis, but it is not clear how these phenotypes relate to the human disease [36]. Here, we presented evidence that Mtb pathogenic proteins have some of the highest MED scores within the Mtb genome.

Genome data

The complete genome of Mtb H37Rv was obtained from the GenBank database under the NCBI identifier NC_000962. All coding sequences were selected and exported as amino acids in FASTA format using the annotation software ARTEMIS from the Sanger Institute.

Prediction schema

Our software environment for MED predictions integrates SurfG+ [37], TMHMM [38] and NetMHC [18]. As seen in Figure 6, an amino acid MULTIFASTA file is first processed by SurfG+ to filter sequences predicted to be secreted (SEC) or potentially surface exposed (PSE). The SEC sequences have then their signal peptide intervals removed from the original sequence, maintaining only the predicted mature protein sequences for further processing. This step is also performed for PSE predicted sequences; however, another TMHMM prediction step is used on these sequences as SurfG+ does not store the TMHMM results concerning the mature portion of the sequences. These steps result in the creation of an artificial amino acid sequence from each original amino acid sequence predicted as SEC and PSE, containing only the concatenated original amino acid sequence portions that were predicted as the mature portions. The artificial amino acid sequences are then submitted to NetMHC, configured to predict all 55 possible MHC alleles within the software (version 3.0), and only the predicted strongly binding peptides are filtered for further processing. Finally, the MED score is calculated for each amino acid sequence according to Equation 1.

Equation 1 divides the number of linear predicted epitopes from each amino acid sequence by the number, for instance, of possible 9mers peptides overlapping windows. To ensure qualitative differentiation for this ratio calculation, the epitopes' MHC binding affinity average is also multiplied, after being normalized according to the maximum MHC strong binding affinity (50 nM). This calculation returns the Mature Epitope Density (MED), a number measured in nanomolar per mer (nM/mer) units. All amino acid sequences are ordered by descending MED score and presented as the final result. The prediction schema was implemented using a Linux shell script. The web server is hosted on Ubuntu OS, release 11.10 and the whole processing takes approximately 90 minutes for Mtb H37Rv amino acid sequences using a standard personal desktop computer.

Control datasets

100 antigen and 100 non-antigen swissprot identifiers were obtained from a previous work [31]. These protein identifiers were retrieved from the Uniprot database [39], culminating in 107 and 121 amino acid sequences used as positive (Dplus) and negative (Dminus) control groups, respectively. To enrich our tests, a set of 38 Mtb's proteins (Mtbplus) were similarly retrieved from the AntigenDB [40] and from Uniprot. The Mtbplus control group was obtained selecting the antigenic proteins from M. tuberculosis and filtering for those known as eliciting immune cellular responses.

Evidence dataset

Sixty proteins out of the 553 in silico predicted as exported were chosen for detailed investigation of experimental proof concerning their capacity to induce cellular responses. In this regard, based on MED, 30 out of 60 proteins were designated as the lowest scored, and the other 30 were designated as the highest scored. An extensive literature search was carried out to look for evidence concerning whether these proteins were related to antigenicity or contribute to the bacterial pathogenicity. Supporting evidence for 39 out of 60 proteins was found, depending on whether a protein induces a cellular response, has evidence of frame shifts, has evidence of differential expression, is part of a known pathogenic protein family or has a cloning experiment that has failed. The complete evidence dataset and corresponding published evidence can be found in the additional file 1.