Using epitope predictions to evaluate efficacy and population coverage of the Mtb72f vaccine for tuberculosis

Although the Bacille Calmette Guérin (BCG) vaccine for tuberculosis (TB) is the most widely used vaccine worldwide, TB continues to be a tremendous public health problem [1]. A third of the world's population is estimated to be infected with Mycobacterium tuberculosis and 2-3 million people die of the disease each year [1, 2]. Key among the reasons for the unabated spread of TB is the inability of the BCG vaccine to provide adequate protection against pulmonary TB in adults, the most contagious form of TB [1]. Developing an improved vaccine for TB, whether a replacement for BCG, a booster to the existing vaccine, or a vaccine specifically directed against latent TB, is of crucial importance in the battle to defeat the disease [3, 4].

While several TB vaccine candidates have demonstrated protective efficacy in animal models and have proceeded to clinical trials in humans [1, 3], even a successful clinical trial cannot guarantee that a vaccine can protect all members of the diverse worldwide human population against all variants of M. tuberculosis. One promising vaccine candidate is the Mtb72f subunit vaccine, a polyprotein composed of the M. tuberculosis proteins PepA and PPE18 [5]. The PPE18 antigen has been demonstrated to contain at least 10 epitopes [6] and the vaccine has been shown to provide protection against TB in cynomolgus monkeys [7]; it is currently under clinical trials in humans. The peptide sequence of Mtb72f, like that of many vaccines, is based on a laboratory strain of the pathogen whose antigens may differ from those found in variable clinical strains in immunologically critical ways [8]. Furthermore, the diversity of the Human Leukocyte Antigen (HLA) and other immune genes results in variable vaccine efficacy in different individuals even in the absence of pathogen variation [9]. While clinical trials include individuals of many HLA genotypes, genotype frequencies vary dramatically in different regions; genotypes that are common in TB-endemic regions may be underrepresented in clinical trial populations. It is therefore necessary to incorporate information on host and pathogen genetic diversity in vaccine design, development, and testing.

The PPE protein family of M. tuberculosis is a large, 69-member protein family with a currently unknown function [10]. Previous work by Hebert and colleagues [8] demonstrated that the PPE18 protein of Mtb72f, but not the PepA protein, had significant variability in a sample of clinical M. tuberculosis isolates. Such variation, however, would be important only if the variation were in regions of the protein that were vital for the human immune response to M. tuberculosis. To determine whether variation in these proteins might impact the efficacy of Mtb72f, we must consider interactions between the M. tuberculosis proteins and the human immune system.

Of the many genes involved in the immune response, the Major Histocompatibility Complex (MHC) genes - HLA in humans - are among the most crucial and yet most variable. MHC proteins present foreign peptide epitopes from intracellular (MHC Class I) and extracellular (MHC Class II) pathogens to CD8+ cytotoxic (Class I) or CD4+ helper (Class II) T cells to initiate the immune response. The Mtb72f vaccine was found to stimulate both CD4+ and CD8+ T cell-mediated responses in a mouse model [5], and the CD4+ T cell response in particular is thought to be essential for preventing M. tuberculosis infection [1, 2]. There are thousands of HLA alleles [11], however, and variation in these alleles can significantly impact individual responses to vaccination [9].

Recently, several algorithms to predict the affinity of peptide sequences for various Class I and Class II MHC alleles have been developed [12, 13] and several have been extensively validated [11‐15]. A consensus approach incorporating three or more programs has been shown to increase the accuracy of MHC Class II epitope predictions [15, 16]. To investigate the impact of host and pathogen variation on TB vaccine efficacy, we have used previously reported protein sequences from clinical isolates of M. tuberculosis[8] and in silico HLA epitope prediction programs to assess the protection offered by the Mtb72f subunit vaccine against diverse strains of M. tuberculosis in human populations suffering from a high burden of the disease. This investigation revealed that due to variation in both M. tuberculosis protein sequences and epitope-binding capabilities of different HLA alleles, certain human populations with a high burden of tuberculosis may not be optimally protected by the Mtb72f vaccine.

Despite the striking variation in the protective efficacy of the BCG vaccine for TB observed among different world populations[18], the joint impact of host and pathogen variation on novel TB vaccine candidates has never been characterized. Using epitope prediction programs, we investigated the impact of clinical variations in the Mtb72f TB vaccine components, the PPE18 and PepA proteins, on epitope binding to alleles of the Class II HLA DRB1 gene. We identified conserved and unconserved promiscuous epitopes in the PPE18 and PepA proteins using the clinical variants described previously [8] and determined that while 60% of potential CD4+ T-cell epitopes in the PPE18 protein were unconserved, 65% of the actually predicted T-cell epitopes in this protein were unconserved. We furthermore found several DRB1 alleles that bound few vaccine epitopes overall and others that bound predominantly unconserved or non-TB epitopes, allowing us to determine individual genotypes as well as broader populations where the Mtb72f vaccine may not offer maximum protection.

Our finding that 60% of potential CD4+ T-cell epitopes in PPE18 are unconserved is consistent with previous research that has found the PPE gene family to be particularly variable among tuberculosis isolates [19]. The significant increase in the proportion of unconserved T-cell epitopes among the actually predicted epitopes compared to the potential epitopes (all nonamer sequences in the vaccine) likely reflects the selective pressure that the immune system places upon the bacterium to alter antigenic protein regions. Based on the clinical protein sequences previously described [8], no potential epitopes in the PepA protein were classified as unconserved. Although to our knowledge no other studies have compared PepA protein conservation to that of other M. tuberculosis proteins, previous findings that PepA is highly conserved among Mycobacterium species [20] suggest that PepA may be a relatively well-conserved protein. The high conservation of the PepA protein even under selective pressure from the immune response may indicate that little variation in this protein is compatible with protein function, and thus PepA may be a particularly good vaccine target [21].

We also examined the conservation of promiscuous epitopes. Epitope promiscuity is commonly evaluated in the generation of epitope-based vaccines [22‐24], but since they are antigenic, promiscuously binding epitopes should be selected against by a large range of host immune systems. We would therefore expect to find many unconserved promiscuous epitopes. This was found to be the case: seven of the ten promiscuous epitopes predicted for PPE18 were unconserved. However, all five of the promiscuous PepA epitopes were conserved, and thus these epitopes could be particularly good candidates for epitope-based vaccines.

Our finding that three of the twelve DRB1 alleles selected, DRB1*0301, DRB1*0802, and DRB1*1301, were predicted to bind few total and conserved epitopes in the Mtb72f vaccine suggests that the vaccine might not be maximally protective in people who have one or more of these alleles. By contrast, DRB1*0101 was consistently predicted to bind far more total (median 147) and conserved (median 93) vaccine epitopes than any other DRB1 allele (median 16 total or 9 conserved), suggesting that individuals with this allele might be particularly well protected by the vaccine. DRB1*0101 is a common allele in the US and Belgium, where Phase I clinical trials for the Mtb72f vaccine were conducted, but unfortunately is rarer in Brazil, China, Indonesia, and many other countries where TB is endemic according to the Allele*Frequencies in Worldwide Populations database [25]. Therefore, while the vaccine may be effective in the US and Belgium, it may not be equally effective in other regions.

Our analysis of all the DRB1 alleles in the Allele*Frequencies in WorldWide Populations database [25] that were found to be one of the top three most common alleles in any population in any of the twenty-two countries designated as TB high-burden countries by the WHO [26] generated a list of the alleles of concern in high-burden TB countries. These alleles include DRB1*0301, 0302, 0403, 0411, 0802, 0803, 0807, 1202, 1401, 1403, 1404, 1405, and 1504. Although the epitope predictions for these alleles should be treated with more caution, as predictions for most could be generated by only one or two prediction programs, the importance of these alleles merits their consideration.

While the number of epitopes capable of binding to a particular MHC allele does not necessarily indicate how strong the immune response will be in that MHC background [27], the fact that many of these alleles bind very few conserved epitopes is of concern for two reasons. First, the analysis conducted here includes only predictions of epitope binding to the MHC molecule and not the cellular processing that must take place before MHC presentation occurs. Because of this processing, many potential antigenic peptides will not be generated in vivo. For MHC alleles to which few epitopes are predicted to bind, it is possible that none of the potentially binding epitopes will be generated in vivo and thus that no epitopes in the vaccine will be presented on DRB1 proteins. Furthermore, for some DRB1 alleles the number of conserved epitopes predicted to bind is much smaller than the number of unconserved epitopes predicted to bind. In the immune response to a vaccine or pathogen, a single or small number of epitopes is usually immunodominant [27] even though many epitopes could potentially bind to the MHC alleles in question. For MHC alleles that are predicted to bind many fewer conserved than unconserved epitopes, it is likely that the immune response to the vaccine would be dominated by responses to unconserved epitopes and therefore would not be optimally protective against all M. tuberculosis strains. Most (8/13, ~62%) of the alleles of concern were predicted to bind at least as many unconserved as conserved epitopes, and one was predicted to bind no conserved or unconserved epitopes at all. Only five of the 21 alleles predicted to bind more than four conserved epitopes were predicted to bind at least as many unconserved as conserved epitopes. Individuals with these alleles - DRB1*1101, 0901, 1402, 1502, and 1602 - might also be less efficiently protected by the Mtb72f vaccine.

While this analysis should be useful for assessing population coverage of the Mtb72f vaccine, it is important to recognize the limitations inherent in the use of epitope prediction programs. No epitope prediction program is perfectly accurate, and there is substantial disagreement among the epitopes predicted by each program [14, 15]. We increased the accuracy of our analysis by using multiple prediction programs for each DRB1 allele when possible, but there are nevertheless likely to be differences between the predicted and actual epitopes for many alleles. However, as we obtained good agreement among programs as far as which alleles were predicted to bind relatively many or few vaccine epitopes, the lists of alleles and populations of concern would not likely be substantially changed by improvements in the prediction programs or experimental confirmation.

Although we compared epitope predictions across eight different programs, these programs are not perfectly comparable. Several of the programs, such as Vaxign, provided information on peptide affinity predictions only for epitopes predicted to bind each allele. As we could not obtain affinity information for peptides predicted not to bind, we were unable to evaluate the binding threshold to which the prediction program was automatically set. Furthermore, the binding cutoffs used by several of the programs likely differ from the IC₅₀ ≤ 500 nM cutoff that we imposed on programs generating IC₅₀ predictions. However, as the number of epitopes predicted by programs with non-IC₅₀-based cutoffs generally though not always fell within the range of programs that did use the IC₅₀ cutoff, it is unlikely that the differing thresholds among programs severely skewed the results.

Finally, it should be noted that the HLA allele frequencies in different populations [25] that were utilized to generate the list of alleles of concern were in some cases generated through only a few small studies. It is likely that further study of HLA genotypes would alter our predictions of which populations might be more or less effectively protected by the Mtb72f vaccine.

We conclude that the Mtb72f vaccine may be less protective in certain populations that suffer from a high burden of TB and thus that additional protective measures may be needed in these populations in addition to this promising vaccine candidate. While the findings from this in silico analysis should be verified with immunological studies, this analysis complements in vitro and in vivo experimentation by vastly expanding the range of host and pathogen factors that can be examined and incorporating an analysis of many more genotypes than a laboratory or even clinical study could include. Furthermore, this type of analysis can be used to aid rational selection of populations in which to conduct clinical trials for vaccines. This method of vaccine evaluation could also be usefully extended to other subunit vaccine candidates for TB, as well as vaccine candidates for HIV and other diseases, and might provide a means of comparing different vaccines to select the best vaccine to use in a given population. However, continued improvements to epitope prediction software, refinement of programs to predict pre-presentation epitope processing, and further study of the HLA genotypes of world populations could help to increase the accuracy and utility of such vaccine evaluations.