The first immunoinformatics tools for vaccine design were developed in the 1980s by DeLisi and Berzofsky and others [17]. Chief among vaccine design informatics tools are epitope-mapping algorithms. Since the T-cell epitopes are bound in a linear form to the human leukocyte antigen (HLA), the interface between ligands and T-cells can now be modeled with accuracy. A large number of T-cell epitope-mapping algorithms have consequently been developed [18, 19]. These tools now make it possible to start with the entire proteome of a pathogen and rapidly identify putative T-cell epitopes. Such information is immensely valuable for the development of new vaccines, diagnostic purposes, and for studying the pathology of infectious diseases [5, 20–27].

Several different routes for vaccine development have been pursued. One method, which has been used by De Groot and Martin [24, 28], is to synthesize the putative T-cell epitopes and screen peripheral blood mononuclear cells (PBMC) isolated from human subjects infected with the target pathogen (or have a target cancer) for immune response to the epitopes. A T-cell in vitro response to a specific peptide epitope (typically measured by ELISA or ELISpot assay) served as an indicator that the protein from which the peptide was derived was expressed, processed, and presented to the immune system in the course of a “natural” immune response. This approach, often considered a means of making epitope-based vaccines (see below), can also be used to identify proteins for use in vaccine development. This approach, described by De Groot and Martin's group as “fishing for antigens using epitopes as bait”, has been used to discover new vaccine antigens for F. tularensis (a bioterror agent) [28], tuberculosis [24], smallpox [29], and H. pylori [30].

The proteome of M. tuberculosis (Mtb), the etiologic agent of TB, contains almost 4,000 proteins. Evaluating each one using the straightforward but expensive and laborious approach of synthesizing and testing overlapping peptides could take decades. Using epitope mapping tools, it is now possible to screen a whole proteome in silico, followed by a finer focus on the resulting sets of peptides [5].

The ability to accurately predict T-cell epitopes from raw genomic data is fundamental to the development of novel vaccines, and serves as the starting point for a number of research projects. Freeing the researchers from the constraints of predetermined sets of “virulence genes” has resulted in some remarkable discoveries. McMurry and De Groot [24] found extraordinary diversity of human immune responses to proteins in the Mtb genome that have yet to be ascribed a function, suggesting that human immune response is omnivorous and is not focused on recognition of a single “immunodominant” protein. In addition, these investigators have found a remarkable similarity between Francisella tularensis (the etiologic agent for Tularemia) and (human) self, at the epitope level [28]. Thus informatics, starting at the genome, may reveal potential antigenic relationships between human proteins and pathogens, or even commensal organisms, which might predetermine individual immune response, that is, prior exposure to a given pathogen may tune immune response to a second pathogen [31].

An alternative approach, “Reverse Vaccinology,” a term coined by Rappuoli, starts with predicting putative vaccine candidates by in silico genomics analysis based on yet different criteria. The predicted vaccine candidates (e.g., bacterial surface proteins) are thought to stimulate protective immunity. Candidate proteins can be evaluated experimentally by demonstrating an immune response that correlates with in vivo protection [10]. The Reserve Vaccinology approach is well discussed later on (see below).

The success rate of vaccine development decreases with the increasing variability of the surface antigens of pathogens and the decreasing ability of antibodies to confer protective immunity [32]. Fortunately, vaccine informatics tools are being developed that increase the accuracy of vaccine target prediction for variable pathogens and help vaccinologists triage antigens.

T-cells are activated by direct interaction with antigen presenting cells (APCs). On the molecular level, the initial interaction occurs between the T-cell receptor and peptides derived from endogenous and exogenous proteins that are bound in the cleft of MHC class I or class II molecules. In general, MHC class I molecules present peptides 8–10 amino acids in length and are predominantly recognized by CD8⁺ cytotoxic T lymphocytes (CTLs). Class I peptides usually contain an MHC I-allele-specific motif composed of two conserved anchor residues [33–35]. Peptides presented by class II molecules are longer, more variable in size, and have more complex anchor motifs than those presented by class I molecules [36–38]. MHC class II molecules bind peptides consisting of 11–25 amino acids and are recognized by CD4⁺ T helper (Th) cells.

MHC class I molecules present peptides obtained from proteolytic digestion of endogenously synthesized proteins. Host- or pathogen-derived intracellular proteins are cleaved by a complex of proteases in the proteasome. Small peptide fragments are then typically transported by ATP-dependent transporters associated with antigen processing (TAPs) and also by TAP-independent means into the endoplasmic reticulum (ER), where they form complexes with nascent MHC class I heavy chains and beta-2-microglobulin. The peptide-MHC class I complexes are transported to the cell surface for presentation to the receptors of CD8⁺ T-cells [39–41].

MHC class II molecules generally bind peptides derived from the cell membrane or from extracellular proteins that have been internalized by APCs. The proteins are initially processed in the MHC class II compartment (MIIC). Inside the MIIC, MHC is initially bound to class II-associated invariant chain peptide (CLIP) which protects the MHC from binding to endogenous peptides. Peptides generated by proteolytic processing within endosomes replace CLIP in a reaction catalyzed by the protein HLA-DM [42, 43]. The class II molecules bound to peptide fragments are transported to the surface of APCs for presentation

To complicate matters further, HLA molecules bind different peptides due to the configuration of their HLA binding pockets. This is the source of genetic diversity of immune responses [34]. Fortunately, there is some conservation between HLA pockets, and both DeLisi and Sette have addressed the issue of HLA coverage for epitope predictions by demonstrating that epitope-based vaccines containing epitopes restricted by selected “supertype” Class I and Class II HLA can provide the broadest possible coverage of the human population [44, 45]. De Groot and Martin have constructed an algorithm, Aggregatrix, which uses the “set cover” method to identify the best set of peptides from a pathogen that would yield the broadest coverage of HLA if included in a vaccine. The Aggregatrix algorithm selects optimized epitope sets which, in terms of immunogenicity and genetic conservation, collectively “cover” a wide variety of known circulating strain variants of a given pathogen and a majority of the common human HLA types [46]. The Conservatrix algorithm is used to identify highly conserved peptide segments contained within multiple isolates of variable pathogens such as retroviruses [47]. The amino acid sequences of protein isolates are parsed into 9 mer frames overlapping by eight amino acids. The resulting peptide set yields a list of unique segments and appearance frequencies. Highly conserved sequences are thought to be important in the evolutionary “fitness” of pathogens and thus are unlikely to change in an attempt to evade the immune system. Conserved sequences can be analyzed using epitope prediction software.

Although textbooks teach that protective immune response is attributed to the development of protective antibodies, the immune response to attenuated intact viruses and subunit vaccines is to a very large degree dependent on T-cell recognition of peptide epitopes bound to MHC. Thus targeting antigens that contain many CD4⁺ T helper epitopes may lead to the selection of good B-cell antigens as well as immunogens for effective CD8 responses—this is because CD4⁺ T helper cells are critically important to the development of memory B-cell (antibody) and memory CTL (cytotoxic T-cell) responses, in addition to being active against pathogens on their own. T helper cells have been called the “conductors of the immune system orchestra” [20]. CTLs generally play a role in the containment of viral and bacterial infection [48], and the prevalence of CTLs usually correlates with the rate of pathogen clearance. Regulatory T-cells are also represented among CD4⁺ T-cells, although some CD8⁺ Tregs have been described.

T-cell epitope algorithms now achieve a high degree of prediction accuracy (in the range of 90 to 95% Positive Predictive Value). For example, epitope mapping tools can now be compared to other available tools, using the Immune Epitope Database “gold standard” as described by Wang et al. [49]. A list of epitope mapping tools, ancillary algorithms, and their comparative features is provided in Table 1. A number of the epitope mapping tools are available to researchers via the web. These include the tool available at the SYFPEITHI website [50] and an HLA binding prediction tool available on at the National Institutes of Health (BIMAS) [51]. A recently developed set of tools has now been made available through the Immunome Epitope Database. Each of these tools has been described and validated [49]. One such proprietary algorithm, EpiMatrix, is in active use in the pharmaceutical industry [52]. While none of these sites yield exactly the same predictions, all predictions are quite accurate, especially when compared to results obtained with early epitope mapping tools (e.g., SYFPEITHI and BIMAS) [49, 52]. In general, the newer and more actively maintained algorithms tend to outperform the older more static predictive methods.

With many machine learning techniques developed since early 1990s for T-cell epitope predictions, it is possible to comparatively evaluate them through prediction performance assessments [49, 53–55]. Lin et al. compared 30 servers developed by 19 groups that can predict HLA-I binding peptides [53]. Their benchmarking study showed that predictions of six out of seven of HLA-I binding peptides achieved excellent classification accuracy. In general, nonlinear predictors outperform matrix-based predictors, and most predictors can be improved by non-linear transformations of their raw prediction scores [53]. While good performance has been achieved for MHC class I predictions, there is still limited success for prediction of epitopes for HLA class II [54, 55]. The low prediction accuracy of HLA-II binding peptides is due to several factors: (a) insufficient or low-quality training data, (b) difficulty in identifying 9-mer binding cores within longer peptides used for training and lack of consideration of the influence of flanking residues, and (c) relative permissiveness of the binding groove of HLA-II molecules for peptide binding, which limits the stringency of binding [54].

Adequate predictors are lacking for predicting epitopes for HLA-C, HLA-DQ, and HLA-DP. However, Wang et al. have made a significant effort in peptide binding predictions for HLA DR, DP, and DQ molecules [56]. Their research with a large-scale datasets of over 17,000 HLA-peptide binding affinities for 11 HLA DP and DQ alleles found that prediction methodologies developed for HLA DR molecules perform equally well for DP and DQ molecules.

The generation of an MHC class-I epitope starts with the degradation of endogenous proteins into oligomeric fragments by cytosolic proteases, mainly the proteasome. These oligomeric fragments may escape from the attack of amino peptidases by entering the endoplasmic reticulum (ER) by the transporter associated with antigen presentation (TAP) [57]. The prediction algorithms for TAP binding and proteasomal cleavage have been developed [58, 59]. For example, Peters et al. used a stabilized matrix method to predict TAP affinity of peptides [58]. This scoring method took advantage of the fact that binding of peptides to TAP is mainly determined by the C terminus and three N-terminal residues of a peptide. Predictions of the MHC class I pathway can be improved by predictions of proteasomal cleavage, TAP transport efficiency, and MHC class I binding affinity [58, 60–62].

While many successes have been made in the area of T-cell epitope prediction, the limitations of all these predictors should be noted. Our goal is to identify good vaccine targets that will induce productive immune responses. However, our ability to measure is usually done indirectly: peptide-binding assays, induction and measurement of immune responses ex vivo, use of animal models, and so forth. Only a small number of HLA-binding peptides are good targets. In clinical vaccine trials, wrong peptides were often selected using indirect methods and tested [63]. For example, the virulence and tumor maintenance capacity of high-risk Human Papillomavirus 16 (HPV-16) is mediated by two viral oncoproteins, E6 and E7. Of 21 E6 and E7 peptides computed to bind HLA-A*0201, 10 were confirmed through TAP-deficient T2 cell HLA stabilization assay. By testing their physical presence among peptides eluted from HPV-16-transformed epithelial tumor HLA-A*0201 immunoprecipitates, only one epitope (E7(11–19)) highly conserved among HPV-16 strains was detected. This 9-mer serves to direct cytolysis by T-cell lines. However, a related 10-mer (E7(11–20)), previously used as a vaccine candidate, was not detected by immune-precipitation or cytolysis assays. These data underscore the importance of precisely defining CTL epitopes on tumor cells and offer a paradigm for T-cell-based vaccine design [63].

It is important to clarify that limited immunoinformatics tools are currently available to identify B-cell antigens (recognized by antibodies). While humoral, or antibody-based, response represents the first line of defense against most viral and bacterial pathogens, the protein target of this arm of defense is usually too complex to model in silico. Antibodies that recognize B-cell epitopes, composed of either linear peptide sequences or conformational determinants, are present only in the three-dimensional form of the antigens. Several B-cell epitope prediction tools, including 3DEX, CEP, and Pepito, are at various stages in development and are in the process of being refined [64–67]. IEDB has collected a list of web prediction tools for B-cell epitope prediction (http://tools.immuneepitope.org/main/html/bcell_tools.html). Unfortunately, the computational resources and modeling complexity required to predict B-cell epitopes are enormous. This complexity is due, in part, to the inherent flexibility in the complementarity-determining regions (CDR) of the antibody and, in part, attributable to posttranslational modifications such as glycosylation, all of which can result in modification of B-cell epitopes.

B-cell epitopes include linear and discontinuous epitopes. Linear epitopes comprise a single continuous stretch of amino acids within a protein sequence. An epitope whose residues are distantly separated in the sequence but have physical proximity through protein folding is named a discontinuous epitope. Although most epitopes are discontinuous [68], experimental epitope detection is primarily for linear epitopes. Tools for prediction of linear B-cell epitopes exist but in general are not predictive [69, 70]. The benchmarking B-cell epitope prediction by Blythe and Flower [69] found that with the best set of scales and parameters, amino acid propensity profiles can predict linear B-cell epitopes only marginally better than random. Such a conclusion has been confirmed by another study where the dismal performance of five predictors was tested against a set of reported linear B-cell epitopes [70].

Although devising accurate B-cell epitope mapping tools remains difficult, the selection of potent B-cell antigens can be accelerated using T-cell epitope mapping tools. When considering B-cell antigens as potential subunit vaccines, it also may be important to also consider their T-cell epitope content since the quality and kinetics of the antibody response is dependent upon the presence of T help. B-cell antigens that contain significant T help may outperform B-cell antigens lacking cognate help. In some cases, an identified T-cell epitope may also contain a B-cell epitope. Different epitopes activate T and B-cells. Despite this observation, it has been widely reported that B-cell epitopes may colocalize near, or overlap, Class II (Th, CD4⁺) epitopes [71, 72].

Different epitope-based vaccine design strategies exist, for example, mosaic vaccines [73], consensus [74, 75], centralized or ancestor immunogen [76, 77], or COT⁺ [78]. Mosaic vaccines are comprised of “mosaic” proteins that are assembled from fragments of natural sequences via a computational optimization method [73]. Many immunogens, such as HIV envelope proteins, have high amino acid sequence divergences. To minimize the genetic differences between vaccine strains and contemporary isolates, immunogenic consensus sequences can be detected and used in vaccine design [74, 75]. Computer programs can also be developed to generate “centralized” vaccine that consists of consensus, ancestor, or center of the tree, modeled from phylogenetic trees. These “centralized” sequences can decrease the genetic distances between the “centralized” and wild-type gene immunogens [76, 77]. In an effort to develop antigens that capture both consensus and mutation sequences among strains, Nickle et al. reconstructed COT⁺ antigens by including the ancestral state sequence at the center of phylogenic tree (COT) and extending the COT immunogen through addition of a composite sequence that includes high-frequency variable sites preserved in their native contexts [78]. These epitope-based vaccine designs have proven effective and provided vaccine researchers with different options in rational vaccine design. It is promising to combine various epitope methods to improve target discovery [56, 60, 79].

Integrated systems and workflows for computational vaccinology are likely to be key for automation of vaccine target discovery [80–82]. For example, Sollner et al. introduced the pBone/pView computational workflow that supports design and execution of immunoinformatics workflow modules, results visualization, and knowledge sharing and reuse [80]. Pappalardo et al. developed ImmunoGrid, an integrative environment for large-scale simulation of the immune system for vaccine discovery, design, and optimization [81]. Feldhahn et al. developed FRED, an extendable, open source software framework for T-cell eptiope detection that integrates many prediction methods and supports implementation of custom-tailored prediction pipelines [82]. The effectiveness of these systems has been demonstrated with different applications.

The EpiVax vaccine design tools (EpiMatrix, ClustiMer, VaccineCAD, EpiAssembler, BlastiMer) are available to researchers through a portal at the Institute for Immunology and Informatics (the iVAX toolkit) [9]. The team of De Groot, Moise, and Martin have implemented the iVAX toolkit to develop four vaccines, a multiepitope TB vaccine [24], a cross-clade HIV vaccine [74], a prototype H. pylori vaccine [83], and a tularemia vaccine [84]. In collaboration with the TRIAD (Translational Immunology Research and Accelerated [vaccine] Development) program at the University of Rhode Island, iVAX is now being used to design additional vaccines including a multipathogen biodefense vaccine against Tularemia and Burkholderia spp, an epitope-based vaccine for HCV, and a vaccine derived from the deer tick saliva to prevent the acquisition of Tick-borne-pathogens. In addition, iVAX has recently been used to scan the entire genome of Salmonella typhi for vaccine candidates. This program is also accessible to researchers working on Neglected Tropical Diseases through the immunome website http://immunome.org/.

Initially, when Reverse Vaccinology (RV) was developed, prediction of putative vaccine candidates was based solely on in silico analysis of the genome of a single strain. Now that selection criteria have been implemented, however, in silico analysis remains the central step in an RV project (see Figure 2).

The first step in the process of genome interpretation, usually referred as gene finding, consists in the prediction and localization of genes onto the chromosome. This is accomplished using prediction programs, which scan the sequence in search of regions that are likely to encode proteins. In prokaryotic systems, the identification of potential coding regions or open reading frames requires implementation of a few basic rules. In the simplest formulation, open reading frames (ORFs) are identified as segments of the same frame comprised between one of the three standard start codons (ATG, TTG, GTG) and one of the three standard stop codons (TAA, TAG, TGA). It is generally accepted that there is approximately one gene for every 1000 DNA base pairs. This suggests that significantly long start-to-stop segments are likely to encode for proteins.

Genome annotation procedures can be automated to different extents. Automated methods for prokaryotic gene finding such as GLIMMER [85], ORPHEUS [86], and GeneMark [87] have been used in genome sequencing projects [88–91]. GLIMMER uses interpolated Markov models, GeneMark uses hidden Markov models, and ORPHEUS is mainly based on codon usage and ribosome binding site statistics derived from annotated genes.

An exhaustive summary of software tools and websites that can be used to obtain bacterial genome annotations was presented by Stothard and Wishart [92].

The annotation procedure allows the translation of the bacterial genome sequence into a list of all the proteins that a bacterium virtually expresses at any time in its life cycle. Each of these amino acid sequences is then compared to the content of public databases of proteins or DNA sequences in an attempt to identify related sequences. When there exist obvious sequence similarities, it is reasonable to transfer this information on the filed sequence to the query. The functional annotation of a protein is sometimes sufficient for the selection of the protein as vaccine candidate, especially when the prediction of protein subcellular localization is uncertain, for example, a protein annotated as fibronectin binding protein may be a good vaccine candidate even when localization algorithms classify it as cytoplasmic. A critical aspect is represented by sequences that lack homologues or contain only remote homologues filed in the databases. ORFs having 20% or less of amino acid identity to any amino acid sequence found in the databases are generally considered to have unreliable homologues. These could represent novel uncharacterized proteins or random open reading frames misidentified as genes. Although homology searches can identify to a limited extent ORFs that are likely to encode functional proteins, experimental authentication by proteomic techniques is usually a more powerful approach for distinguishing genes from random ORFs.

The ensemble of hypothetical proteins can be processed with software programs dedicated to deduce their possible cellular localization. One of the basic assumptions utilized for candidate searches is that a good antigen will be located on the cell surface of a bacterium, where it is readily available for antibody recognition. Several algorithms have been developed that predict the subcellular localization of proteins based solely on the amino acid sequence and composition (see Table 2). The basic assumption made is that the N-terminal sequence of the protein predicts its cellular destination. The presence of a “leader sequence” provides evidence that the proteins will be exported to extra-cytoplasmic compartments. Additional signatures may also be exploited such as the presence of a cleavage site immediately after the leader peptide. Such sites imply that the protein is released into the extra-cellular environment of Gram-positive bacteria or into the periplasmic space of Gram-negatives. Similarly, proteins that contain an LXXC motif, where X is any amino acid, positioned at the end of the leader peptide are often lipoproteins. Anchoring of proteins to the Gram-positive bacterium cell wall often requires a specific carboxy-terminal sorting sequence. This sequence is identified by an LPXTG motif followed by approximately 20 hydrophobic amino acids and a charged tail. In Gram-negative bacteria, additional secretion pathways exist that promote the passage of extracellular proteins across the outer membrane. At least six distinct extracellular protein secretion systems have been reported in Gram-negative and Gram-positive bacterium (type I–VI, T1SS–T6SS) that can deliver proteins through the multilayered bacterial cell membrane and in some instances pass directly into the target host cell [93]. The six secretion systems exist in Gram-negative bacteria and the common Gram positive bacteria. Gram positive bacteria contain an additional specific secretion system (type VII) [94]. This increases the variety and complexity of secretion signals, making the identification of outer membrane and secreted proteins yet more challenging [95, 96].

Several computational methods have been generated to predict extracellular proteins in Gram-negative microorganisms [97].

PSORTb is the most widely used tool for predicting subcellular multiple localizations of organelles in Gram-negative bacteria. This program uses biological knowledge to elaborate “if-then” rules, combining information on amino acid composition, similarity to proteins of known subcellular localization, presence of signal peptides, transmembrane helices, and motifs diagnostics of specific subcellular localization. Recently, two predictive methods CELLO [98] and Proteome Analyst [99] have been proposed for Gram-negative bacteria. These programs are providing comparable performances in terms of accuracy and recall with respect to PSORTb [97].

Despite the recent progress, identification of secretion systems components in silico and their effectors still mainly relies on the detection of amino acid sequence [94] and the structural [100] similarities of selected proteins. Caution is necessary in applying these predictions, as sequence similarities can be very weak and do not necessarily imply any functional analogy.

In conclusion, by knowing the genome sequence it becomes possible to select using bioinformatics tools to generate a list of potential antigens without cultivating the microorganism. This methodology has a huge advantage over conventional vaccinology approaches for two major reasons. First of all, in silico analysis is very fast and cheap, and secondly, proteins not expressed in vitro are also identified. However, this approach only provides a prediction of a protein's subcellular localization and it cannot reveal if a protein is expressed and under what conditions. Therefore, use of a bioinformatics approach may need to be complemented with other techniques, for example, a Mass Spectrometry-based approach to aid vaccine candidate prediction. The first RV project employed a single genome. Indeed, at that time there was only one genome available for N. meningitidis. Nowadays in most cases, there are more than five genomes available for any human pathogen. Therefore, the in silico analysis can take advantage of comparative genomics.

Today, the number of fully sequenced microbial genomes exceeds 1000 (http://www.ncbi.nlm.nih.gov/bioproject/) (Many are not from pathogens). It is clear that microbial diversity has been vastly underestimated, and a single genome does not exhaust the genomic diversity of any bacterial species [101, 102]. In many cases, an extensive genomic plasticity exists. For example, completion of the genome sequence of E. coli O157:H7 revealed that it contains >1,300 strain-specific genes compared to E. coli K12, which encode proteins that are involved in virulence and metabolic capabilities [103, 104]. Additional reports have revealed the occurrence of an extensive amount of genomic diversity among the strains of a single species [105–107].

These early findings were formalized with the definition of the bacterial pangenome, as the sum of the genes present in each individual species. This concept was originally introduced during study of the genome variability in eight isolates of Streptococcus agalactiae (also known as Group B streptococcus or GBS). It was found that each new genome had an average of 30 genes that were not present in any of the previously sequenced genomes. Not every bacterial species has the same level of complexity as GBS. For instance, the pangenome for Bacillus anthracis can be adequately described by four genome sequences. Hence, scientists refer to certain species as having an “open” and others a “closed” pangenome. In species with an open pangenome, there are an unlimited number of new genes found for every genome. In closed pangenomes, there are only a limited number of strain-specific genes. The differences in the nature of the pangenome reflect several factors: differing lifestyles of two organisms, the number of closely related species in the same environment and physiological state, the ability of each species to acquire and stably incorporate foreign DNA (an advantage in niche adaptation from the acquisition of laterally transferred DNA), and the recent evolutionary history of each species. It should be noted that the imperfection of our definition of a bacterial species, for example, B. anthracis and B. cereus, can be considered the same species by some criteria, may render pangenome analysis more complicated. As the definition of a pangenome improves, the coverage of strains included in a bacterial species will change and thus alter the analysis results.

A pangenome can be divided into three elements: (1) a core genome that is shared by all strains (2) a set of dispensable genes that are shared by some but not all isolates, and (3) a set of strain-specific genes that are unique to each isolate. For S. agalactiae, the core genome encodes the basic aspects of S. agalactiae biology and was as such predicted to rapidly converge to 80% of the genome in each isolate. Conversely, dispensable and strain-specific genes, which are largely composed of hypothetical, phage-related and transposon-related genes [108], contribute to its genetic diversity. The concept of the pangenome and comparative genomics has practical applications in vaccine research. In fact, while obviously the ideal vaccine candidate is a conserved protein encoded by a gene present in every isolate of the species, in the case of GBS it was shown that the design of a universal protein-based vaccine against GBS was possible using dispensable genes [109]. Of note, capsular-specificity genes and other pathogenicity traits are often identified in an accessory genome. Moving forward, bacterial taxonomy and epidemiology must take into consideration whole genome sequences and not just a few genetic loci, as has been the case so far with methods such as ribosomal RNA sequences, capsular typing, and multilocus sequence typing (MLST). Comparison of the whole genome sequences of GBS strains has shown that the genomic diversity does not necessarily correlate with serotypes or MLST sequence-types. The application of additional whole genome sequence analysis will require that epidemiology studies have a reliable, systematic correlation between strains and disease and permit a standardization of the classification for clinical isolates. These observations are instrumental for developing a protective vaccine that covers a broad range of pathogenic strains.

Comparative genomics is also important for the identification of pathogenic factors since they potentially represent good vaccine candidates. The level of distinction and the function played by carrier versus virulent strains of streptococci and neisseriae, for example, has been the matter of discussion for a long time and still lacks an answer. There is, as yet, no clear and strict correlation between the presence of apparent virulence factors and the diseases caused by these organisms. The epidemiological evidence is vague and does not provide definitive clues. There may be multiple reasons for this apparent lack of correlation. It is likely that in species that only rarely result in disease, there exist multiple virulence factors and toxins that are uniquely associated with infection. Therefore, comparative genomics can be used to identify the “pathogenicity signature” associated with the most virulent bacterial strains or the strains that are successful in colonization. Comparative genomics can also be used to compare various strains that exhibit different virulence levels, for example, commensal nonpathogenic strains versus virulent ones, to find specific vaccine candidates [110]. The advantages include making a vaccine against commensal strains and narrowing down the pool of vaccine candidates. It is anticipated that most virulence factors will be found in accessory genomes, at least the ones that determine increased pathogenicity. However, presently comparative genomics is not able to identify expression variability that contributes to the different manifestations of pathogenicity of bacterial strains. Hence, functional studies are still critically needed to shed light on the relevance of specific virulence factors.

Another potential application could be studies of certain species of bacterial symbionts such as Mycoplasma, Rickettsiae, and Chlamydiae. These species, instead of acquiring genes during evolution, have actually lost significant levels of their genetic information [111]. Primarily biosynthetic pathway genes have been lost because intracellular bacteria have a relatively constant environment with access to much of what they require for survival. By applying the concept of pangenomics to these species, we would obtain a “microgenome” representative of the set of genes necessary to live in the intracellular niche. Comparing this “microgenome” with the pangenome for free-living species will likely simplify the identification of genes necessary for the microorganism to survive in varying and unfavorable environments. Housekeeping genes specific to the pathogen are considered vaccine candidates.

In comparison to a half decade ago, comparative genomics studies have become incredibly easy to perform. For the most important human pathogens, the average number of genomes for the different available strains is above five. Therefore, for new RV studies, the conservation level of selected antigens can be determined. Antigen conservation level is important since conserved antigens can be used to develop a broad strain protective vaccine [32].

In addition to the basic mechanisms of the RV strategy described above, additional criteria can be added. For example, since outer membrane proteins containing more than one transmembrane helix are difficult to clone and purify [10], the number of transmembrane domains of a candidate protein is often used as an additional filtering criterion. Bacterial adhesins play critical roles in adherence, colonization, and invasion of microbial pathogens to host cells [112]. Therefore, adhesins are essential for bacterial survival and are possible targets for vaccine development. Two RV software programs, NERVE [113] and Vaxign [110], utilize these criteria. Since RV focuses on predicting antigens using protein sequences, immune epitope prediction based on amino acid sequences can also be considered as a criterion for RV vaccine design [110].

Vaxign (http://www.violinet.org/vaxign/) is the first web-based vaccine design program based on genome sequences utilizing the RV strategy. Predicted features in the Vaxign pipeline include protein subcellular location, transmembrane helices, adhesin probability, conservation to human and/or mouse proteins, sequence exclusion from genome(s) of nonpathogenic strain(s), and epitope binding to MHC class I and class II. Vaxign has been demonstrated to successfully predict vaccine targets for Brucella spp. [114, 115] and uropathogenic Escherichia coli [110]. Currently, more than 100 genomes have been precomputed using the Vaxign pipeline and available for query in the Vaxign website. Vaxign also performs dynamic vaccine target prediction based on input sequences.

The availability of three-dimensional structure may facilitate epitope prediction and antigen discovery [116, 117]. It would be ideal to also consider inclusion of analysis of high throughput transcriptomics and proteomics data to aide in complementary identification of vaccine candidates.

Accurate tracking of vaccination history is essential to ensure proper completion of the primary immunization schedule and subsequent booster doses. This seemingly straightforward task is nontrivial system-wide when compounded by an increasingly mobile population, immunization schedules of increasing complexity, multiple vaccine manufacturers of the same vaccine, multiple health care providers and/or health insurance for the same individual (a problem in the U.S.), multiple individual with same name, and so forth. Add in small vaccine vials with hard to read small fonts in a busy pediatric clinic serving many crying babies simultaneously, the opportunities for inaccurate or nonrecording of an administered vaccination is substantial in developed and developing countries.

Computerized immunization information systems (IIS) provide an obvious potential solution to these challenges. In the U.S., the first large IISs were organized in Delaware in the early 1970s [175]. This action was followed by several health maintenance organization (HMOs) with the dual purpose of linking the IIS to medical visits for rigorous studies of vaccine safety [176]. The Robert Wood Johnson Foundation funded the All Kids Count I and II programs in the 1990s in multiple communities. This provided an important impetus to the field [175]. The Centers for Disease Control and Prevention (CDC) now provide some financial and technical assistance for public sector IIS in almost every state (http://www.cdc.gov/vaccines/programs/iis/default.htm). This work is aided by partners such as the American Immunization Registry Association (http://immregistries.org/) and the Public Health Informatics Institute (http://phii.org/). Internationally, Australia [177], Canada [178], and Norway [179] are some of the other countries with active IIS.

IISs also have the potential to provide a foundation of child health registries [175] and electronic health records [180]. While initially focused on routine pediatric immunizations, registries in many locations have been expanded to meet other needs, including adolescent and adult immunizations [181], disasters [182], targeting of at risk populations [182, 183], study vaccine refusal [184], and facilitating accurate and timely reporting of vaccine adverse events [185]. Substantial progress has also been attained in the protection of privacy and confidentiality; in ensuring participation of all immunization providers and recipients, to ensure appropriate functioning of registries, and to ensure sustainable funding for registries [186]. However, challenges remain in exchanging information among different IISs, and across state lines. The National Vaccine Advisory Committee has issued recommendations on how to overcome these challenges (http://www.hhs.gov/nvpo/nvac/IISRecommendationsSep08.html).

Before a vaccine is licensed, it undergoes rigorous testing in preclinical (laboratory and animal) and phased human clinical trials for safety and efficacy [187]. Due mainly to cost (intensive monitoring perprotocol is expensive) and ethical (once a vaccine is determined to be safe and effective, it is no longer ethical to withhold it from others in need) considerations, however, the sample size and duration of followup in prelicensure trials are usually limited. This means surveillance for both vaccine safety and effectiveness [188] in larger immunized population postlicensure and postmarketing is needed. This is challenging because trial conditions (e.g., double-blinding, randomization) that permit straightforward comparison between vaccinated and unvaccinated groups no longer hold. Substantial data collection and adjustments on possible confounders, when possible, are needed to fully analyze and interpret such observational studies.

Post-licensure monitoring for vaccine safety can generally be divided into hypothesis generating and hypothesis testing. Since vaccine coverage for many vaccines can be close to universal, by definition, anyone with a medical adverse event will have previously been vaccinated. Spontaneous reporting or passive surveillance systems like the U.S. Vaccine Adverse Event Reporting System (VAERS, http://vaers.hhs.gov/) in the U.S. [189] and elsewhere [190], where medical problems suspected to be caused by the vaccination can be reported to health authorities, provide the bulk of new vaccine safety hypotheses. Due to the large number of reports (>20,000 annually to VAERS), data mining techniques are beginning to be applied to triage reports worthy of further attention [191].

Once a vaccine safety concern is provisionally identified, based on our understanding of likely pathophysiology and nonrandom clustering of cases in onset time after vaccination, a formal study is usually needed to (1) confirm whether the etiologic link with vaccination is real and not coincidental, and (2) identify the magnitude of the risk (to assist in risk-benefit determination for the immunization). Since these safety concerns are likely to be rare (otherwise they would have been detected pre-licensure), confirmatory pharmacoepidemiologic studies of large vaccinated populations are usually needed to “test the hypothesis”. Large national (e.g., Denmark) or population (e.g., Managed Care Organization (MCO)) health care systems, where members have unique personal identifiers and most of the care for both vaccinations (exposure) and medical visits (outcome) are automated, provide an efficient platform for piggy-backing vaccine safety pharmacoepidemiologic studies [192]. The Vaccine Safety Datalink (VSD) project in the US, a consortium of 8 MCO's representing ∼3% of the population, has been used as a prototype of how such large linked databases can be used for rigorous vaccine safety studies [176, 193]. Examples include rotavirus vaccine and intussusception [194], thimerosal and neurologic adverse events [195], and vaccinations and central demyelination [196]. Similar large-linked vaccine safety databases have been created in England [197] and Vietnam [198].

Safety issues cannot be assessed directly and can only be inferred from the relative absence of multiple adverse events. Therefore, standardizing the case definitions used to assess adverse events is needed to allow for meaningful comparison of vaccine safety data in various settings. Recognizing this need, the Brighton Collaboration (https://brightoncollaboration.org/public) was formed in 1999 as a voluntary global collaboration to facilitate the development, evaluation, and dissemination of high quality information about the safety of human vaccines in both pre- and post-licensure settings. To date, 28 guidelines and case definitions have been developed and are freely available to users. The case definitions are tiered by the level of evidence available and will differ based on whether the data are gathered in prospective clinical trials or passive postmarketing surveillance and on the level of resource availability (e.g., developed versus developing countries). Since its inception, the Collaboration has helped to form a critical mass of experts interested in vaccine safety that can potentially be convened or accessed as new vaccine safety issues arise. The Brighton Collaboration Viral Vector Vaccine Safety Working Group is exploring using the “wiki” model of mass collaboration for completing and maintaining standard templates on characteristics of various viral vectors [199].

Post-licensure monitoring for vaccine effectiveness is usually done by examining the impact on targeted diseases. For reasonable sensitivity and specificity for monitoring trends of the disease, this usually requires the establishment of some type of public health surveillance system. For example, the recent reintroduction of rotavirus vaccine in the US has resulted in delayed onset and diminished magnitude of rotavirus activity [200]. A decline in invasive pneumococcal disease was observed after the introduction of conjugate pneumococcal vaccine [201]. Similar data was obtained in developing countries, as was done with introduction of conjugate Haemophilus influenza type b vaccine in Mali [202]. When disease remains high or an outbreak occurs despite high vaccine coverage, a special epidemiologic study to assess vaccine effectiveness may be needed. This type of action was undertaken after a posthoneymoon period measles outbreak in Burundi [203], the resurgence of diphtheria in the former Soviet Union [204], and the introduction of a monovalent oral type 1 poliovirus vaccine in India [205].

The cyclical nature of epidemics of many infectious diseases such as plague and smallpox in humans (and other animals) was noted by ancient historians prior to the introduction of immunization in modern times [206]. This periodicity was described as a mathematical relationship between susceptible and immune individuals in a population over time that interacted with an external infectious force by Ronald Ross and Anderson Gray McKendrick at the beginning of the 20th Century [207]. It was not until the early 1980's, however, that Anderson and May systematically organized and effectively organized disparate works in population biology, ecology, and epidemiology into mathematical models of infectious diseases that linked the theory with practical translation into public health policy (e.g., vaccinations) [208]. Their 1991 textbook “Infectious Diseases of Humans: Dynamics and Control” [209] has helped to create a cohort of mathematical modelers that have furthered our understanding of transmission of infectious agents within human communities and design programs for their control. As Geographic Information Systems (GISs) that integrate and analyze spatial information become increasingly available for linkage with public health databases [210], this should aid continued refinements in various assumptions used in mathematical models of infectious diseases.

Irrespective of the model or the target disease, a key concept in any mathematical model is the basic reproductive rate or R₀ of a microorganism—the average number of secondary infections produced when one infected individual is introduced into a totally susceptible population. The goal of any control program (e.g., immunizations) is to reduce the R₀ as much as possible. For disease elimination or eradication programs, it must be <1 [211]. Another key concept is “herd immunity”, the indirect effect of some vaccines on reduction of disease transmission beyond the protection in actual vaccine recipients [172]. Most mathematical models attempt to describe as accurately as possible the flow of a human population from susceptibility (usually at birth or with the waning of maternally derived immunity), infected (by wild disease or vaccination), and immune states (adjusting for various variables such as mixing) transmission coefficient, vaccine effectiveness, and duration of protection. Each of these variables in turn can be further modeled (e.g., mixing can differ with age-classes or other subpopulations).

Historically, one of the more successful integrations of mathematical modeling of vaccine-preventable diseases and immunization program policies has been for measles [203, 211, 212] and rubella [213]. Mathematical models have also been critical for understanding how best to (a) introduce newly licensed vaccines like human papillomavirus vaccine [214], (b) control new emerging public health problems, such as pandemic influenza [215], (c) how best to optimize control of a vaccine-preventable disease (such as impact of pneumococcal conjugate vaccine on emergence of penicillin-resistant strains) [216], or (d) how spatio-temporal variation in birth rates may explain the observed patterns of rotavirus disease after the introduction of new rotavirus vaccine [200].

The papers and authors related to vaccine/vaccination have increased exponentially. Only six vaccine-related papers were published and recorded in PubMed before 1900. In the first half of the 20th century, 1,210 vaccine-related papers were published. This number has increased almost 100-fold in the second half of the last century. In addition, the numbers of vaccine publications have increased exponentially (Figure 3). For example, 6,399 vaccine-related papers were published during the period of 1951–1960, and 96,938 in 2001–2010. Therefore, the number of papers published annually in PubMed has increased more than 15-fold during the past 50 years.

It has become increasingly challenging to retrieve useful vaccine data for research purposes from the huge amount of vaccine literature. Literature mining has been used to facilitate the discovery and analysis of potential vaccine targets. For example, cross-matching and analysis of the literature and in silico-derived data allowed the selection of 189 putative vaccine candidates from the entire Mycobacterium tuberculosis genome [217]. In this study, the first step towards the selection of vaccine candidates was to accumulate published experimental data from a literature scan of documented studies with a focus on global analyses. The literature sources were then grouped based on different categories (e.g., macrophage). This literature mining approach detected 189 potential vaccine candidates. These were studied further through in silico functional analysis and immunoinformatics epitope prediction. A qualitative score was designed based on a total of 14 criteria and used to rank and prioritize the gene list [217].

Literature mining can also be used to analyze vaccine-associated host immune response networks. For example, Ozgur et al. recently applied a literature mining and network centrality analysis [218] to analyze the IFN-γ and vaccine-associated gene networks [219]. Among approximately 1,000 genes found to interact with IFN-γ, 102 genes were predicted to be vaccine-associated and 52 of them were verified by manual curation. The production of IFN-γ is crucial for successful immune response induced by vaccines against various viruses and intracellular bacteria. For example, these include HIV [220], M. tuberculosis [221], Leishmania spp. [222], and Brucella spp. [223]. The discovery of the IFN-γ and vaccine-mediated gene network provides a comprehensive view of the vaccine-induced protective immune network and generates new hypotheses for further experimental testing.

Two literature mining programs presented in the Vaccine Investigation and Online Information Network (VIOLIN; see next section) were developed for general vaccine literature searching and analysis [13]. Vaxpresso (http://www.violinet.org/textpresso/cgi-bin/home) is a vaccine literature mining program using natural language processing (NLP) and ontology-based literature searching [224]. For a list of selected pathogens, Vaxpresso contains all possible vaccine-related papers extracted from PubMed (http://www.ncbi.nlm.nih.gov/pubmed). Vaxpresso is able to retrieve and sort article sentences that match specific keywords and ontology-based categories. Vaxmesh (http://www.violinet.org/litesearch/meshtree/meshtree.php) is a vaccine literature browser based on the Medical Subject Headings (MeSH). MeSH is a controlled vocabulary of medical and scientific terms that is used for indexing PubMed articles in a consistent way supporting PubMed literature mining. Vaxmesh enables users to locate articles using MeSH terms in a hierarchical MeSH tree structure.

Many publicly available vaccine databases and online resources exist (Table 3). For example, the USA CDC Vaccine Information Statements (VISs) system (http://www.cdc.gov/vaccines/pubs/vis/) provides information sheets that explain to vaccine recipients, their parents, or their legal representatives both the benefits and risks of a vaccine. Federal law in the US requires that VISs be handed out for all vaccines before their use. The licensed vaccine information is provided by the U.S. FDA (http://www.fda.gov/BiologicsBloodVaccines/Vaccines/default.htm). The Vaccine Resource Library (VRL, http://www.path.org/vaccineresources/) offers various high quality, scientifically accurate documents and links to specific diseases and topics in immunization.

These databases focus primarily on the clinical uses and regulations of existing vaccines for vaccine users. To store and analyze research data concerning commercial vaccines and vaccines under clinical trials, or in early stages of development, the Vaccine Investigation and Online Information Network (VIOLIN, http://www.violinet.org/) was developed. VIOLIN is a web-based vaccine database and analysis system primarily targeted for vaccine researchers [13]. The VIOLIN vaccine database currently contains more than 2,700 vaccines, or vaccine candidates, for more than 160 pathogens through manual curation from >1500 peer-reviewed papers or other reliable sources. The stored vaccine data includes vaccine preparation, pathogen genes used and gene engineering, vaccine adjuvants and vectors, vaccine-induced host immune responses, and vaccine efficacy in host after virulent challenge. VIOLIN curates more than 500 protective antigens (http://www.violinet.org/protegen/) [225]. Vaccine-related pathogen and host genes are also annotated and available for searching through customized BLAST programs. VIOLIN also stores and processes all possible vaccine literature through different text mining programs [13]. Vaxign, a web-based vaccine design program based on reverse vaccinology strategy [110], is also a program in VIOLIN.

Besides the above databases which focused on vaccine awareness and vaccine research, many other databases are available that are useful for vaccine research and development. For example, more than 65,000 antibody and T-cell epitopes have been deposited in the Immune Epitope Database and Analysis Resource (http://www.immuneepitope.org/) since the database was established in 2004 [6]. These immune epitopes cover a broad range of species including humans, nonhuman primates, rodents, and other animal species as related to all infectious diseases [6]. AntigenDB is an immunoinformatics database of pathogen antigens and store sequences, structures, origins, and epitopes [226].

Although public vaccine databases provide help with different aspects of vaccine knowledge and research, it remains a challenge to integrate this disparate body of information on vaccines. Data integration is hampered since the data are often collected using incompatible or poorly described methods for data capture, storage, and dissemination. Integration is also complicated as investigators use independently derived local terminologies and data schemas. These problems can be alleviated through the use of a common ontology, that is, a consensus-based controlled vocabulary of terms and relations, with associated definitions that are logically formulated in such a way as to promote automated reasoning. Ontologies are able to structure complex biomedical domains and relate a myriad of data to allow for a shared understanding of vaccines.

The collaborative, community-based Vaccine Ontology (VO; http://www.violinet.org/vaccineontology/) was recently initiated to promote vaccine data standardization, integration, and computer-assisted reasoning. VO can be used for different applications, including vaccine data integration and literature mining. Currently, VO contains more than 3,000 terms, including more than 700 vaccines and vaccine candidates that are represented in an appropriately structured ontological hierarchy. These vaccines or vaccine candidates are targeted to 70 pathogens and have been studied in more than 20 animal species (e.g., human, mouse, cattle, and fish). VO also stores terms related to different vaccine components (e.g., protective antigens, vaccine adjuvants and vectors), vaccine-induced immune responses, vaccine adverse events, and protection efficacy. The known relations between these terms are also listed. These representations are readable by computer programs and support computer-assisted reasoning. This knowledge is also exchangeable across multiple scientific domains to facilitate hypothesis generation and validation. This approach will undoubtedly lead to new scientific discoveries.

VO has been used for several different applications. For example, VO, in combination with other ontologies, has been used to model and study vaccine protection investigation [15]. Reported vaccine protection data from different reports can be systematically analyzed [227]. VO can also be used to improve vaccine literature mining. For example, a direct PubMed search for “live attenuated Brucella vaccine” returned 69 papers (as of August 2010). VO includes 13 live attenuated Brucella vaccines that are defined as “live” and “attenuated”. When specific “live, attenuated” Brucella vaccine terms are included in a PubMed search, the number of papers found in PubMed increased by more than 10-fold [228, 229]. The application of VO has also enhanced the discovery of IFN-γ and vaccine-associated gene networks [16].