There is a vast, and ever increasing, volume of important information that has accumulated from decades of experimental analysis within immunology. This will only become compounded as high-throughput techniques begin to impinge upon the immunological biosciences. The only efficient way for this information to be properly utilized requires the development of databases that store it and systems that use it. Although the type of data archived may alter from case to case, nonetheless the creation, use, and manipulation of databases containing biologically important information is the most crucial feature of current bioinformatics, both as it supports the genomic and post-genomic revolutions and as a discipline in its own right. There is nothing new in developing databases focusing on immunology: many spotlighting the in-depth sequence analysis of individual immunomacromolecules have existed for some time [1]. Functional or epitope-orientated databases are a more recent development. Examples include the now defunct MHCPEP database [2]http://wehih.wehi.edu.au/mhcpep/, FIMM [3]http://sdmc.krdl.org.sg:8080/fimm, SYFPEITHI [4]http://www.syfpeithi.de, the HIV sequence database [5]http://hiv-web.lanl.gov/, the HLA ligand database [6]http://hlaligand.ouhsc.edu, the EPIMHC database [7]http://bio.dfci.harvard.edu/epimhc/, and the MHCBN database [8]http://www.imtech.res.in/raghava/mhcbn/.

An epitope is any molecular structure that can be recognised by the immune, or other biological, system. Epitopes, or the antigen from which they are derived, can be composed of protein, carbohydrate, lipid, nucleotide, or a combination thereof. It is through recognition of foreign, or non-self, epitopes that the immune system can identify and, hopefully, destroy pathogens. Hitherto, peptide epitopes have been the best studied, and have, traditionally, have been categorized as either T cell or B cell epitopes. T cell epitopes are peptides presented to the cellular arm of the immune system via the MHC-peptide-TCR complex. B cell epitopes represent surface regions of an antigen that are bound by soluble or membrane-bound antibodies. If this region of a protein antigen is comprised of residues distally separated within the primary structure, and brought into local proximity by protein folding, then it is termed a discontinuous or conformational B cell epitope. Linear or continuous B cell epitope residues are sequential in both primary structure and thus as a region on the proteins' surface. Such epitopes are predominantly identified by antigen-specific antibody cross-reactivity with peptides.

There is a need to create a databank for the wider disciplines of immuno-vaccinologists, which can act as a central repository and resource. Our aim is to complement other databanks [2, 3, 4, 5, 6, 7, 8] and thus we have developed AntiJen, a computational information resource for immunology and vaccinology that integrates quantitative kinetic, thermodynamic and biophysical data, with functional and cellular information. AntiJen v2.0, a development of our earlier database system JenPep [9, 10], contains functional data on T cell and B cell epitopes. Moreover, the B cell archive is now sub-divided into linear and conformational epitopes. These epitopes form the basis of the humoral immune response and, unlike T cell epitopes, methods of prediction are often inaccurate [11]. A more in-depth B cell epitope archive should aid the development of prediction strategies. Antigen recognition by the Major Histocompatibility Complex (MHC) is vital to T cell activation hence, the inclusion of thermodynamic data on the binding of peptides to MHC molecules and T Cell Receptor (TCR) binding to peptide-MHC (pMHC) complexes. This data is complemented by kinetic data based on the same molecular interactions. Data on antigen processing and presentation is also included in AntiJen. Binding data derived from peptide interactions with the Transporter Associated with Antigen Processing (TAP transporter) are included in the archive. Additionally, quantitative specificity data from position-specific peptide libraries is included. AntiJen also incorporates thermodynamic data on protein-protein interactions, within an immunological context, such as co-receptor and superantigen binding, plus interactions with the MHC. All of these interactions are, potentially, key factors for the successful computational design of vaccines.

AntiJen also contains biophysical data, including diffusion coefficients and cell surface copy numbers, on a variety of immunological molecules. Such data provides insight into the number of target receptors, which is an important, if under explored, component of binding between cells. Indeed, the number of molecules expressed on the membrane can alter depending on disease. The final addition to the databank focuses upon antigen binding to antibodies. One key innovation is a greatly increased compendium of experimental conditions, which, in conjunction with a greatly enhanced search capacity, consolidates our databases as a unique, value-added data source, fostering developments within both in silico immunology and the wider community of immunovaccinology. The database is available from the URL: http://www.ddg-pharmfac.net/antijen/AntiJen/antijenhomepage.htm.

Relative to the database system used for JenPep [9, 10], the interface to AntiJen is entirely new, having been completely rewritten. AntiJen has been designed and implemented using postgreSQL, a system comprising a relational database and database server, and has thus established increased database robustness, creating an improved infrastructure for foreseeable issues of data storage and data growth. Data within AntiJen is structured into twenty-four normalised tables. Each is category specific and holds either statistical or experimental data. Additional tables accommodate the keyword data – which powers our protein-orientated antigen search and allows integration of the BLAST search – and there is also a structural data table to accommodate links to external structural databases. The user interface consists of a series of HTML forms. The search requests from these forms target PERL scripts integrated with SQL which in turn query the database.

AntiJen contains both generic and dataset-specific data. For each entry, we record the peptide sequence (eg. YTSDYFISY) of the epitope using the standard one-letter code, its length (9 in this case), and, by linking to the sequence database Swiss-Prot http://us.expasy.org/sprot/ or NCBI http://www.ncbi.nih.gov/entrez, the antigen to which the peptide sequence most closely matches (in the case of YTSDYFISY: C-ests-1 (p54), SWISS-PROT code P41156). The description of the antigen is, wherever possible, obtained directly from the literature. AntiJen is also linked to PUBMED. This allows us to record the original citation associated with the data. For example, for YTSDYFISY, we cite: Journal of Immunol 1994 volume 152 pages 3913–3924, PUBMED ID 8144960. For the T cell epitope, MHC ligand, and TCR-pMHC complex categories, we also record, for each peptide, the MHC restriction in terms of the host species, class (class I vs. class II), and, where the data is available, the serotype and allele. For YTSDYFISY, these data would be human, class I, HLA-A1, and A*0101.

Entries within AntiJen are, in turn, linked to external databases, which enables further in-depth cross referencing. As we have said, protein sequence identifiers, which may be the source of an antigenic peptide or immunological co-receptor, link directly to details in the Swiss-Prot database [12] or the NCBI protein database. Journal references can be viewed via a link to the PUBMED database http://www.ncbi.nlm.nih.gov/pubmed, and thus to full literature references, where available. AntiJen also links to structural data, currently derived from the MPID database http://surya.bic.nus.edu.sg/mpid/[13] and the Protein Data Bank https://www.rcsb.org [14]. The database aims to provide access to background data where available.

Allele identifiers serve as a link to the IMGT/HLA database https://www.ebi.ac.uk/ipd/imgt/hla at the European Bioinformatics Institute [15]. Inherent variability in the way MHC alleles are named within the primary literature prevents us from unambiguously standardizing nomenclature within AntiJen. HLA nomenclature follows that of the HLA Informatics Group http://www.anthonynolan.org.uk/HIG/. An allele is named using a defined pattern. For example, for HLA-A*0101: the HLA-A refers to the HLA locus; the first 01 to the serologically recognized A1 antigen and the final 01 to the individual HLA allele protein sequence. AntiJen stores the antigen classification (i.e. HLA-A1) and, when available, the specific allele. We have often encountered problems with the nonstandard allele reporting. A 4-digit HLA name necessarily implies the two digit serological antigen, a two digit classification clearly does not imply a specific allele.

During database compilation, a sequence search allows us to identify the protein from which an epitope sequence originates. However, because epitopes are generally short, their sequences may be present in several potential antigens: in orthologues, paralogues, or in totally unrelated sequences. As epitopes are processed from whole proteins via a complex proteolytic pathway, one can use the sequence context to infer preferred proteasomal or endoplasmic protease cleavage patterns, but not if its context is defined incorrectly. Moreover, assuming that AntiJen is used subsequently to assign the antigenic status of proteins, wrongly identifying particular proteins as antigens can lead to the percolation of annotation errors [16, 17].

AntiJen is, where possible, a quantitative database archiving continuous measures of binding. This is a fundamental feature of several sub databases, such as the MHC ligand and pMHC-TCR databases. The binding of an immunological macromolecule to a peptide or other biomacromolecule is quantified as are other receptor-ligand interactions:

R+L↔RL

Here R is the receptor (an MHC or TCR), L the ligand (peptide or pMHC), and RL, the receptor-ligand complex (pMHC or pMHC-TCR complex). The rate of the forward reaction is proportional to [L] [R]. The rate of the reverse reaction is proportional to [RL] as no other species are involved in dissociation. At equilibrium, the forward and reverse rates are equal, and so using k_on and k_off as the respective constants:

k _on [R][L] = k _off [RL]

Rearranging:

Here K_D is the equilibrium dissociation constant, which represents the concentration of ligand that occupies 50% of the equilibrium receptor population, and K_A is the equivalent association constant.

Experimentally, the measurement of equilibrium dissociation constants is often addressed using radio-ligand binding assays. Saturation radio-ligand binding assays measure equilibrium binding, at a range of peptide concentrations, to establish affinity (K_a) and receptor number (B_max). Competitive binding experiments determine binding at a single labelled ligand concentration in the presence of a range of concentrations of unlabelled ligand. AntiJen records a hierarchy of different binding measures in its different sub-databases. Equilibrium constants are the most dependable and sit atop this hierarchy. Next come IC₅₀ values, which can be obtained from a competitive radio-ligand or fluorescence assay. These are the Fmost commonly reported binding measures.

AntiJen also now contains experimental conditions, such as temperature and pH. A summary of this data is given in Table 4. The accuracy of data depends greatly upon the experimental method used. The grouping of data with respect to specific experimental techniques allows a more thorough assessment of training sets. Figure 1 shows the distribution of data for each type of analysis with respect to each database. The MHC Kinetics and TAP databases highlight the problems outlined above. The kinetics database contains data determined from over 14 methods while the TAP database is derived from 4 methods, with radiolabelled assays accounting for 80% of the data.

DATABASE	TOTAL	pH	Temperature	[standard]	Stand. peptide seq.	[competitor]	Method	[peptide]
MHC Binding	15454	6679	9831	10893	12796	5007	1251
MHC Kinetics	1150	677	1101				1149	606
TAP Binding	1106	22	243	1092	1101	86	981
TCR-pMHC	782	426	632				668
IPPI a	2675	726	1371				2600
Copy Number	414	183	278				414
Peptide Libraries	897		897				897
Diffusion Coefficient	759	321	668				736
Antibody	395	119	115				372

The compilation process has highlighted the considerable inconsistency within the immunological literature regarding the recording of such fundamental data. AntiJen contains, however, a direct, verbatim transcription of data from the primary literature. As such, we do not attempt, as a matter of policy, the comprehensive and retrospective correction of potential errors. To undertake such correction would only compound any errors, introducing the kind of percolating inconsistencies so much a feature of other database systems [16, 17]. Further inaccuracies may stem from our logistic inability to verify data independently, therefore we must trust those values reported in the literature.

Peptide-MHC Kinetics

AntiJen's kinetics sub-database, which contains 1,150 entries, mostly relates to Class I MHC data. It records measurements for forward and reverse rate constants for complexation events. This complements the thermodynamic measurements on peptide-MHC binding described above. The data currently focuses upon both the half-lives of binding interactions, as well as association and dissociation rate constant values (K_on and K_off) for the recorded epitopes. Additionally, concentrations of the peptide, MHC and TAP are archived. The half-life for radioisotope labelled β₂-microglobulin dissociation from an MHC class I complex, as measured at 37°C, is also archived. This is a kinetic measurement rather than a thermodynamic one, although it is often assumed that the greater the half-life the stronger the peptide-MHC complex. The half-life (t_1/2) equals:

Here the t_1/2 corresponds to the dissociation of the MHC-β₂ microglobulin complex rather than the kinetics of the protein-ligand interaction, but is still peptide dependent, as well as kinetic in nature.

Search mechanisms within AntiJen are significantly improved and allow either a detailed or a broad search from a simple user interface. From our experience with JenPep, we recognize that accessibility to the data in a user friendly manner is a vital requirement, and have improved our current search mechanisms and developed new search interfaces. Two different search mechanisms are available. One is based on BLAST [24] and the other is a bespoke system, allowing several alternative searches. Within a typical search, the user-entered search criteria are carried from an HTML form to a category specific PERL/SQL script, which performs the database queries.

The BLAST search allows querying of a peptide or nucleotide sequence against the proteins contained in AntiJen; all entries containing data within AntiJen which are relevant to a protein sequence are linked via the BLAST output. A local database of protein sequences is searched with BLASTP or BLASTX [24] using the BLOSUM64 matrix. All BLAST control variables are fixed. An HTML Front-End (Figure 2), where a sequence can be entered or uploaded, connects to a web server-based PERL/CGI scripts, which interacts with BLAST. An annotated version of the default BLAST output is produced and links to AntiJen entries via SWISS-PROT [12] accession codes, which act as a query within a Keyword search. This allows AntiJen entries to be viewed directly from the BLAST output.

At present, peptide string, keyword, and protein name index searches are available within the bespoke system, which allows querying of individual peptide sequences or at the level of whole protein antigens. An overview of the AntiJen search systems is given in Figure 2 and 3. Epitopes, MHC, TCR, TAP peptide binding and kinetics databases can all be searched using sequence strings. The search protocol first returns an epitope list and a count of epitope matches. Subsequently, experimental criteria can be accessed for each selected epitope. Peptides can be searched using an amino acid orientated query: a sparse peptide string, similar in form to a peptide binding motif [3] or a PROSITE pattern [25], is used to identify all matching sequences. See Figure 4. Alternately, a list of protein antigens within AntiJen can be searched using keywords; the thermodynamic binding data such as MHC-peptide, TCR-pMHC and TAP, plus the B and T cell archives, related to the search criteria can then be selected from the matches based upon the SWISS-PROT accession codes [12], displaying all corresponding entries in the database.The other search method, allows the IPPI, diffusion co-efficients, peptide libraries, antibody-protein and copy number sub-databases to be searched using an index method, within a user-friendly HTML drop-down menu. Each of these methods can also be moderated using subsidiary search filters, data size ranges, and result presentation alternatives, such as peptide length or IC₅₀ values. Minimum and maximum values can be used to restrict results, as can selection of MHC restriction alleles.

The name given to our new database, AntiJen, reflects a shift from a peptide orientated database structure, which was inherent within our earlier JenPep database, to one which can properly balance its focus on both protein antigens and isolated peptides. As such, it represents an important, integrated, immunological data resource. AntiJen now provides broad insight into both T cell and B cell mediated antigen recognition. In addition, through the auspices of the IPPI sub-database, the database also throws light on co-stimulation by co-receptors and gives important insights into the innate immune response. Our approach to protein-protein interactions, focusing on measured affinities, complements other methods, such as the Yeast-2-Hybrid system, which, while giving greater volumes of data, has problems of accuracy [26]. This, while not devoid of experimental artifact, gives a usefully different perspective on cataloguing protein-protein interactions. The further addition of weak binding notation to the MHC-peptide and TAP provides a greater overview of the nature of antigenic epitopes. This is further improved by the addition of the peptide libraries database, whereby key peptide residues can be highlighted. New databases have expanded the breadth of AntiJen to include biophysical data such as diffusion co-efficients and cellular data such as abundance of molecules. The antibody-antigenic protein sub-database will also provide a key resource for in vivo and in vitro studies, aiding in the selection of antibodies and peptide/protein targets.

AntiJen distinguishes itself from the other specific binding databases [2, 3, 4, 5, 6, 7, 8] in several ways. Firstly, more data is recorded; our MHC-peptide database contains over 2,000 more entries than MHCPEP [2] and 10,000 more entries than EPIMHC [7]. Additionally, we have not restricted our archive to only high binders or to a specific category, as seen in EPIMHC and the HIV sequence databases [5]. Furthermore, AntiJen is currently a curated database, which is constantly expanding.

Most obviously, AntiJen is useful in the design of epitope and subunit vaccines. Additionally, AntiJen is helpful in the design of clinical diagnostics and other laboratory reagents, such as the selection of peptides for tetramer design. AntiJen is also useful in the parameterization of mathematical models in theoretical immunology [27]. The redevelopment of the database has focused not only on content, but also on infrastructure. The current system, based on epitope string, keyword and index searches, along with an overall BLAST search, plus the redesigned HTML interface, leads to much greater accessibility and usability. Finally, the database acts as a repository of quantitative, continuous data, for the development of data-driven in silico predictive models, such as prediction of epitopes and MHC binding [18, 19, 28, 29] through QSAR modeling.

Future tasks in the development of AntiJen, fall into two principle categories: eliminating deficiencies, errors, and inconsistencies within the database and simultaneously reinforcing it by expanding its depth, breadth, and scope. We also need to monitor updates within external databases, so that any alterations are mirrored within the archive. Like all other such repositories, AntiJen is prone to both systematic and random errors within the data accumulation process. User feedback and our interactions with immunologists will hopefully address persisting errors. Deficiencies in our database include our current inability to encode chemically or post-translationally modified peptides, non-natural MHC mutants and non-amino acid peptidomimetic MHC ligands. Additionally, it would also be interesting to complement our existing data on TAP binding with information on antigen presentation pathways, such as proteasomal and cathepsin cleavage patterns. Moreover, the compilation of B cell or antibody epitope data is an area ripe for robust development. Linear and conformational B cell epitopes are very much larger in number than our current compilation, leaving us scope to greatly increase recorded epitopes.

The development of a database is always a work in progress. Not simply because the easily accessible literature is typically always increasing, but also because of the desire to capture as much of the existing, but hidden, literature, as possible. In the post-genomic era, the database has formed the bedrock and language of bioinformatics; increasingly databases are coming to underpin our modern understanding of biology as a whole. Traditionally, databases have arisen as a response to need, answering the individual and idiosyncratic questions posed by biologists. However, the history of bioinformatics databases has shown the extraordinarily diverse ways in which archived data can be used.

In creating AntiJen, we were motivated partly by our desire, and the desire of collaborators, to use the data within it to build predictive in silico models [16, 17, 28, 29], and partly by a more altruistic desire to generate a useful, integrated database system with a quantitative focus. AntiJen has many potential uses throughout the immunological discipline, from immunoinformaticans to experimental immunologists and vaccinologists. By increasing the degree to which data is machine readable and web accessible, we open up new, and previously unthought-of, avenues for the bioinformatic exploration of immunological data.

AntiJen is a primary data resource, amongst the most complete of its kind, yet, like SWISSPROT [12] or GenBank [30] decades ago, it is still relatively small and offers much scope for improved annotation. We see the database as a foundation from which to consolidate, through time, thus achieving a comprehensive resource of immunological data.