Allergy is a steadily increasing health problem for all age groups in our society. Food allergies, mostly against milk, eggs, peanuts, soy or wheat affect up to 8% of infants and young children [1–3]. In addition, many allergies to food (including shellfish, nuts, and fruits) and air-borne particulate matter (such as insect residue, tree and grass pollens) can develop later in life. One hypothesis is that this late onset may be the result of the individual being sensitized by long term exposure to environmental factors that contain proteins similar to those in the known triggers of allergenic response [1, 4–6]. Recent studies have identified common molecular features of proteins from quite different sources, which could account for clinically important cross reactivity [7, 8] and sensitivity [9, 10]. For example, major allergenic proteins in peanut have been isolated, and peptides from their sequences that react with IgE from patient sera identified [6, 11, 12]. Proteins similar to these allergens were subsequently found in other foods that are known to elicit clinically significant responses in peanut allergic individuals [13], such as tree nuts [14], soy [15], and legumes [16, 17].

To show how valuable comparing the proteins that are known to cause allergy can be, allergens classified as pathogenesis response proteins have extremely conserved sequences in many different plants [18–21]. Some of the sources of allergens belonging to pathogenesis related proteins of group 5 (PR5) are shown pictorially in Figure 1. The first allergen in this group was isolated from cedar pollen [22, 23], but related proteins that bound to IgE from patients allergic to cherries, bell pepper, apple and tomato were subsequently identified [24]. As the sources of such closely related allergenic proteins are so different, sensitive individuals may indeed feel they are suffering from “multiple allergy syndrome”, when in reality they may have a strong reaction to one protein type that is found in many sources. This means that by identifying the common protein, using bioinformatics tools described in this article, more specific treatments can be defined [25].

Figure 1 Different products can contain similar allergenic proteins. In this case, an allergen, Jun a 3, originally isolated from the pollen of mountain cedar was found to be a member of the PR5 family of proteins, and was modeled based on its similarity to a (more ...)

Allergenic proteins from pollens, particularly from cedar [22, 26–28], birch [29–31], and grass [32–35] are quite similar in overall sequence and structure. Other families of allergenic proteins have been isolated from primary food sources such as milk [36–38] (casein [37, 39, 40] and lactoglobulin [41–43]), egg [44] (ovomucoid [45–47] and lysozyme [48]), shrimp and related species (tropomyosins [49–51]), fish (parvalbumin [52–54]) legumes (albumins [55–58] and glycinins [59–61]).

Many of the allergens in pollen and fruits have similar sequences and structures, as Figure 1 makes clear. Identifying the protein group can be used to guide specific immunotherapy. For example, clinically important cross reactivities between birch pollen and several different types of fruits could be accounted for by demonstrating that they contained similar proteins [62–66]. In a group of patients with oral allergy syndrome to apples, as well as birch pollen sensitivity, the root cause was hypothesized to be the similarity between the allergen Bet v 1 and the apple protein Mal d 1. Specific immunotherapy with a Bet v 1 containing extracts was able to mitigate sensitivity to this fruit [67].

Successes of this sort explain the current interest in developing bioinformatics approaches to interpret the accumulated body of knowledge about the proteins that elicit severe IgE mediated reactions [8, 68–70]. Until recently, much of the information about allergens was distributed in many different literature sources, making oversight and direct sequence comparisons difficult, especially for clinicians dealing directly with patients [71]. However, there are now several databases were that contain sequences and information about allergenic proteins (Table 1). This chapter will discuss the specific features in each of these databases and highlight how our Structural Database of Allergenic Proteins (SDAP, http://fermi.utmb.edu/SDAP/) [72, 73] can be used for comparing the sequence, structure, and epitopes of allergens. SDAP has been designed to be user friendly, and to be of maximum use to clinicians as well as scientists interested in determining the molecular basis of allergen cross reactivity [13, 24, 74].

Table 1 Websites with information about allergens, allergen databases, and allergenicity prediction servers

Other databases are cross-indexed and much more useful to those seeking to identify cross-reacting allergens and their natural sources. Allergome (http://www.allergome.org) is a comprehensive database of clinical, biological and structural information about IUIS and non-IUIS recognized allergens. Allergome records the name, source, biochemical and immunochemical features for each allergen. While no computational tools are integrated in Allergome, there is a list of allergen MEME “sequence motifs” introduced by Stadler and Stadler [75]. Unlike the IgE based sequences we will present later, these motifs are quite long, and could be more properly called conserved domains in the sequences of closely related allergens.

Another database, maintained by the CSL (Central Science Laboratory, UK, http://www.csl.gov.uk/allergen/index.htm), lists official names of allergens with sequence links to Genbank. This is similar to the Biotechnology Information for Food Safety Database (National Center for Food Safety and Technology, http://www.iit.edu/~sgendel/fa.htm), which provides comprehensive lists of allergens with links to sequences from PIR, SwissProt, and Genbank. The Protall project (http://www.ifrn.bbsrc.ac.uk/protall/) developed a database of plant food allergens that contains links to detailed biochemical, structural, and clinical data. Since 2001 the development of the Protall database is part of the InformAll project (http://foodallergens.ifr.ac.uk/).

A few databases allow direct comparison of allergen sequences, using conventional search tools, and permit the use of the WHO guidelines for predicting potential allergenicity [76]. These guidelines specify that a protein might cross-react with an allergen if it is at least 35% identical over a frame of 80 amino acids or contains an exact match of any peptide of 6–8 amino acids with the allergen [77]. The FARRP database (http://allergenonline.com/asp/public/login.asp) contains a searchable list of allergens, sequence links to Genbank, and a FASTA search for related sequences. ADFS (Allergen Database for Food Safety, http://allergen.nihs.go.jp/ADFS/) contains allergen sequences and epitope information, and, implements the WHO allergenicity rules using FASTA. ALLERDB (http://sdmc.i2r.a-star.edu.sg/Templar/DB/Allergen/) lists official names of allergens, has a BLAST search, and a sequence comparison tool that implements the WHO guidelines.

SDAP contains information on sequence, 3D structures and epitopes of known allergens from published literature and compiled databases on the Web [72, 73]. The major uses of SDAP by clinicians are to determine food sources that could induce cross-reactions in sensitive individuals, and help in preparing dietary recommendations for allergic patients with a known sensitivity. For example, patients with a food allergy can be told to avoid other foods contain similar proteins to the ones in that know to trigger an allergic reaction. Alternatively, a recombinant protein can be used to determine the scope of the allergic response, and suggest candidates for specific immunotherapy.

As noted in the Introduction, allergenic proteins belong to discrete groups, or families of structures. SDAP has several different methods for comparing the sequences of allergens within the database. The in-house bioinformatics methods permit almost instantaneous sequence similarity searching, with direct connections to much larger databases. To determine what other sequences are similar to the amino acid sequence of an allergen, one can do a full-length sequences for similarity to the known allergens, by selecting the BLAST and FASTA [78] searches. These searches can be activated directly from the main descriptive page for any of the allergens in SDAP of known sequence.

The Pfam grouping, discussed below in more detail, is available for most of the allergens in SDAP and rapidly indicates which other allergens a given protein resembles. Links to structural models for the various allergens allows one to visualize common areas of the proteins in 3-D. Yet another box on the allergen page indicates what IgE epitopes are known, and each sequence is linked to tools that can be used to automatically identify similar or identical sequences in all the other sequences in SDAP. Other tools allow the user to map the IgE containing peptides onto the experimental modeled structures of allergens. Figure 2 shows two examples of this tool, where the IgE epitopes of two fungal allergens Asp f 1 and Asp f 3 have been mapped on an experimentally determined structure (PDB 1AQZ) and on our MPACK model, respectively.

Figure 2 IgE epitopes mapped on the experimental structure (PDB 1AQZ) of Asp f 1 (a and b) and out MPACK model of Asp f 3 (c and d), two allergens from the fungus causing aspergillosis.

SDAP is also integrated with other bioinformatics servers, allowing the user to investigate structural similarity and neighbors using SCOP (Structural Classification Of Proteins) [79], TOPS (TOpological representation of Protein Structure) [80], CATH (Class, Architecture, Topology and Homologous superfamily) [81], CE (Combinatorial Extension of the optimal path) [82], FSSP (Fold Classification based on Structure-Structure alignment of Proteins) [83], and VAST (Vector Alignment Search Tool) [84].

We have developed special methodology for finding homologues of known epitopes in the sequences of other SDAP entries, the PD search, which will de discussed in some detail in Part III. The PD distance method was developed in our group specifically to detect meaningful similarities when comparing short sequences [72, 73]. PD searches can reveal similarities in IgE epitope sequences, even from different allergenic proteins [13, 24].

Table 2 Output of an automatic FASTA search in SDAP starting from the file for the allergen Jun a 3, an allergen isolated from cedar pollen (see Figure 1). This PR5 protein is related (i.e., the sequence match has an E-score <0.01) to 7 other allergenic (more ...)

Note that the most similar entry in SDAP for Jun a 3 (i.e, that with the lowest E-value) is another pathogenesis related protein (group 5) is from a related cypress tree, Cup a 3. However, similar allergenic proteins have been isolated from various vegetables and fruits, which are shown pictorially in Figure 1. Based on these FASTA alignments, we can anticipate that someone with severe cedar pollen allergy might also develop oral allergy symptoms [88, 89] when eating, for example, apples or cherries. Other clinically relevant cross-reactivities, in addition to those mentioned in the introduction, have been shown, such as that linking dust mite sensitization and development of food allergies to shrimp and other crustaceans. This is postulated based on the similarities of the tropomyosins in these organisms [72, 73, 88].

Routine use of SDAP can prevent certain problems, such as those that may crop up when discoverers of a new allergen give it a name based on their understanding of how many other allergens have been previously isolated from the given biological source. Although this name may be changed by the IUIS, once a protein has been named in the literature, it is often difficult to obtain wide acceptance of a different designation [90]. For example, an allergen identified in Juniper oxycedrus was originally called Jun o 2 by its discoverer. However, when the IUIS examined this allergen, they realized that this was a different protein from the other cypress allergens named type 2, such as Jun a 2 and Cyp a 2 (see Table 3). Thus they assigned the protein an official name of Jun o 4 [91], with relevance to its similarity to the Bet v 4 protein of birch pollen (lower part of Table 3). SDAP names this protein Jun o 4, but alerts the user who types in Jun o 2 of the name change. This is because other databases, including PIR, GenBank and Swissprot, continue to identify Jun o 4 as Jun o 2. If SDAP had been used by initially to name this allergen, this confusing situation (which will get worse if a real Jun o 2 is isolated) would not have arisen.

Table 3 Results of two automatic FASTA searches in SDAP that illustrate how to use this site to correctly name a new allergen. While the cedar pollen allergens Jun a 2 and Cry j 2 are very close in sequence (i.e, have a very low expectation (E) value; top table), (more ...)

The other feature of allergen nomenclature illustrated by Table 3 is that allergens with very closely related sequences, such as those in the bottom section of the table, can have widely differing numbers. Further, the names of closely related proteins can differ. For example, food allergens are generally categorized “seed storage proteins or “albumins”. Depending on their degree of identity to other proteins they may be additionally categorized as vicilins, proglycinins, 7S albumin, etc., which in turn can determine the common name of the allergen. As it is unlikely that a radically different nomenclature scheme will be introduced anytime soon, database searches like this one will provide the best way to truly indicate which allergens are most related to one another.

The user of SDAP can identify allergens that are significantly similar to one another according to their Pfam or enzyme classification. Pfam (http://www.sanger.ac.uk/Software/Pfam/) is a list of multiple sequence alignments of related protein domains, classified in two ways. The Pfam-A database lists protein families that are grouped by their common function as well as sequence, using expert knowledge and experimental data. Pfam-B is computer-generated and contains alignments of proteins sequences selected based on a minimum level of sequence identity, regardless of their protein function. Most SDAP entries have now been classified to one of these groupings. Easy access to this Pfam classification for any allergen can be accessed from the “List SDAP” menu item.

The most common Pfam families for allergens are listed in Table 4, where we show 18 Pfam families that have between 34 and 7 allergens each. The most common Pfam families for allergens are PF00234 (protease inhibitor/seed storage/LTP family, 34 allergens), PF00235 (profilins, 27 allergens), PF00036 (EF hand, 23 allergens), and PF01357 (pollen allergen, 20 allergens). Table 5 lists allergens from these four Pfam families, with representative structures shown in Figure 3 (Pru p 3 from PF00234, Hev b 8 from PF00235, Bet v 4 from PF00036, and Phl p 2 from PF01357).

Table 4 The most abundant Pfam A allergen families from SDAP

Table 5 Allergens from the four most populated with allergens Pfam families

Figure 3 PDB structures for allergens from the four most abundant Pfam families: (a) Pru p 3 (PF00234, protease inhibitor/seed storage/LTP family); (b) Hev b 8, (PF00235, profilin); (c) Bet v 4, (PF00036, EF hand); (d) Phl p 2, (PF01357, pollen allergen).

Cross-reactive allergenic proteins are usually very similar in sequence and structure, at the molecular level, regardless of their source. As noted in the previous section, most allergens can be grouped into discrete sequence families, according to their Pfam classification. However, a typical Pfam will contain allergenic and non-allergenic proteins. Quantitatively discriminating the allergenic members in a group of similar proteins is a difficult task, and, as we will see below, one that eludes the programs currently implemented in popular databases. Experimentally, cross-reactive allergens typically have high sequence identity, which can drop to as low as 35% (hence the WHO guidelines). Still, point mutations are known to eliminate IgE binding [5, 14, 30, 85] and one need only point to the example of isoforms of Bet v 1, which are 98% identical, that are not cross reactive [94–96]. For this reason, the reader should treat the results of current methods for predicting the allergenicity of a given protein with caution. The general consensus of a recent International Bioinformatics Workshop Meeting [97] on this problem concluded that more detailed statistical analysis of the properties of allergenic proteins versus non-allergens are needed and that numerical benchmarks for prediction methods should be developed. Below, we describe the current methodology, and outline other methods, based on identifying motifs of allergenic protein groups, that may have more success in defining cross-reactive allergens and areas for potential IgE recognition. The use of some of the more advanced methods implemented in SDAP, for comparison of IgE epitopes are emphasized, as this is one of the key features of our database that can be used in future to both design allergen vaccines and proteins with reduced allergenic potential.

Table 6 Test of three allergenicity prediction servers. Three database prediction methods are unable to discriminate the allergenic sequences from insects shellfish (bottom 4) from the non-allergenic tropomyosins from animals (top sequences). The former proteins (more ...)

On the other hand, other rapid methods were unable to discriminate the two groups at all. Allermatch, which applies the FAO/WHO allergenicity guidelines, predicts that all eight tropomyosins are allergens. We also tested the MEME classifier based on motifs identified in groups of allergenic proteins [75]. This method, as implemented in the AlgPred server, predicted that all eight tropomyosins are non-allergens. Besides the MEME classification approach, the AlgPred server evaluates the allergenicity by scanning for known IgE epitopes, by a BLAST search, by a support vector machines (SVM) prediction based on amino acid composition or on dipeptide composition, and with a hybrid approach that combines the above five procedures [101]. The SVM dipeptide composition classifier predicts that all eight tropomyosins are allergens. In conclusion, it seems that, consistent with earlier results [102, 103] the WHO guidelines current allergenicity prediction servers cannot discriminate between closely related proteins, such as the non-allergenic and allergenic tropomyosins, as their overall sequences are too similar. However, other studies have shown that these guidelines are useful for detecting proteins that are sufficiently similar to known allergens that they might cross-react [104]. Several reports demonstrate that the succession of bioinformatics and experimental procedures from the FAO/WHO decision tree may be valuable in investigating the protein allergenicity [105–107].

Efforts to improve correlations based on the whole protein sequence are ongoing. Methods based on analyzing the statistics of FASTA alignments with machine learning procedures have been tested [108, 109]. In one case, such an algorithm was able to classify 81% of 91 food allergens and 98% of 367 non-allergens correctly [110]. This level of accuracy could make the method useful for clinically discriminating protein groups to be avoided by patients with a known sensitivity to a related protein.

Another alternative is to detect discrete areas of a protein sequence similar to known IgE epitopes. The SDAP list of IgE epitopes, most which were identified by IgE recognition of linear arrays of synthetic peptides, is unique, as are the tools incorporated at the site for detecting identical and similar sequences (the PD tool) in other known allergens.

Other efforts to predict allergenicity were directed towards identification of linear epitopes [114]. Saha and Raghava used a recurrent neural network for the prediction of continuous B-cell epitopes [115]. While short amino acid sequences matches seem to have little value for allergenicity prediction [116], peptide motifs common to groups of allergens may be a better way to distinguish allergens. An efficient machine learning classification scheme, based on identifying a set of allergen-representative peptides that appear in allergens but have a low or no occurrence in non-allergens, outperformed the FAO/WHO allergenicity rules [117]. Further, as Table 6 shows, the motif-based allergenicity prediction scheme based on wavelet transform did find areas of the allergenic tropomyosins that were not present in the non-allergenic. Further development of this method, taking into account the physicochemical properties of the amino acids, and solvent accessibility, was able to correctly classify 93.0% of 229 allergens tested, and 99.95 of non-allergenic proteins [118].

Table 7 Sequences most closely related to epitope 3 of Jun a 3 from mountain cedar pollen by the PD search reveal other PR-5 proteins from fruits (top); that of epitope 4 reveals another known IgE epitope from latex (bottom; bold sequence). The column lists the (more ...)

Additional data is needed to determine the statistical significance of the identified regions in the other allergens, particularly in a database as large as SDAP. The PD search is designed to compare protein regions with lengths comparable with those of published linear IgE epitopes. Significance levels for the sequence-similarity index PD are set high enough to detect all peptides that are similar to an IgE epitope, but low enough to discriminate them from other regions in the ensemble of allergenic proteins that would match randomly. For the search, each area of all the sequences is individually matched, with a window for the sequence segment that moves progressively by one position. Thus a 200 amino acid protein would have 194 different “sequence windows” of 7 amino acids, and 191 for a 10 mer. All PD searches in SDAP are followed by two histograms. The “lowest scoring window” is determined for the “best matching” peptide in the ~850 protein entries in SDAP. As the data in the heading of Table 7 indicate, for a PD search starting from a known epitope of Jun a 3, the average PD value for the best scoring sequence window in each of the 854 full length entries in SDAP was 12.15 (SD = 1.29) According to this test, values below about 9 (mean value – 2× the standard deviation) would be significantly similar to the test peptide. However, there are many similar sequences and isoforms in SDAP, which tends to skew the statistics for peptides. As a better estimate of what a random match would be, a second histogram summarizes the scores for all ~190,000 windows of a given size in all the SDAP allergen sequences. The average values in this histogram range from 17–26, depending on peptides. For the second example of Table 7, the average PD value for all 190530 possible windows was 17.24 (SD=1.78). According to these statistics for a random match, peptides with PD values below 10 would be clear outliers.

To give the reader a sense of the significance of the match, Z scores (which indicate the quality of the match relative to the database random distribution), are calculated automatically along with the PD value. The lower the PD score, the more closely related the peptide sequences are, but high Z-scores indicate better significance for the match.

The PD searches from two Jun a 3 IgE epitopes (Table 7) illustrate the usefulness of using the PD value to identify potential epitopes and potentially cross reactive allergenic proteins. For both epitopes, areas of thaumatin proteins from other pollens and fruits have the lowest PD value, consistent with their overall similarity according to our FASTA search (Table 2). Note that the order of the sequences in the two tables is a bit different. Although the IgE epitopes have not been identified for the fruit allergens, the Jun a 3 epitope 4 has a low PD value to a known IgE epitope of the latex allergen Hev b 3. The significance of this finding for cross-reactivity has not been determined. Other results with the peanut allergen epitopes [13] have identified epitopes with similar IgE reactivity and predicted structure for PD scores as high as 9.5–10.

We should at this point emphasize that the PD search is a computational way to define the sequence relationship between known IgE epitopes and other sequences in allergenic proteins. The correlation of PD values to meaningful IgE cross-reactivity, and eventually, to clinically relevant ones, is ongoing. However, initial tests indicate that this is a rapid way to quantify local similarities in known allergens. The structure of epitopes and their location on the protein surface (“solvent exposure”) are other possible factors determining whether a given sequence will bind IgE or not [13]. We believe the most promising methodology is to compare not just the sequences, but also the structures of areas before suggesting possible cross reactivity, as described in the next section.

Once similar sequences have been identified by PD values, the structural information in SDAP can be used to understand which parts of an allergen sequence are likely to be surface exposed, and thus likely to form an IgE binding surface [13, 24]. Determining which residues are on the surface of an allergen (and thus most likely to react with an antibody) can also be determined rapidly and automatically using a program developed in this group, GETAREA (http://www.scsb.utmb.edu/cgi-bin/get_a_form.tcl). This program has also been incorporated at the site, where the data can be quickly accessed for SDAP allergens of known or modeled structure. SDAP allows direct access to the experimental structures (out of 586 SDAP allergens, 45 have known PDB structures). We estimate that for more than 90% of allergens where the experimental structure is unknown, we can make reliable models based on results from fold recognition servers such as 3DPSSM, http://www.sbg.bio.ic.ac.uk/~3dpssm/.

To this point, we have only talked about linear epitopes, and occur next to one another in the sequence of a protein. A folded protein in 3D may have epitopes formed from several areas of the sequence. The ConSurf [122] method has been used to detect patches of residues common to many allergens on the surface of allergen structures for Arat 8, Act c 1, Bet v 1, and Ves v 5 [123]. The findings have not yet been tested experimentally. A combined method that uses sequence similarity and comparison of 3D models was used to identify potentially cross-reactive peanut-lupine proteins [124] and to search for potential new latex allergens [125]. Our PCPMer program contains methodology to map conserved residues on the surface of proteins, to detect such common areas. These “stereophysicochemical variability plots” are useful for distinguishing functional areas of viruses [126], and can also be used to identify regions that might constitute IgE epitopes. Alternatively, we are also in the process of developing methods to map peptides that bind IgE to surface areas of allergens of known structure.

Proteins are classified as allergens based on their ability to trigger responses in patients. Allergens may just be more potent forms of other proteins with similar surface areas that may have been the true sensitizing antigens during development of the disease. Thus it is not clear how similar proteins must be to known triggers in order to represent significant risk for cross-reactivity. The problem is made more difficult by the fact that some potent allergens can be rendered non-allergenic by selected point mutations and highly similar proteins, such as the glycinins of soybean and peanut (62% identity), provoke quite different responses [132].

We have outlined here computational methodology to identify cross-reacting proteins at the molecular level, using the databases of allergenic proteins and their structures. Recent identification of the sequence and structure of allergenic proteins from pollen and foods has revealed how similarities which might offer a structural explanation for their allergenicity and cross-reactivity [7, 18, 23, 29, 31, 127, 133]. Some of the recently developed search software tools, such as those implemented in SDAP can help clinicians and patients to find structural and functional relations among known allergens and to identify potentially cross-reacting antigens.

It is clear that available methods cannot, with 100% accuracy, discriminate between closely related proteins according to their allergenicity. Instead, they provide an indication that certain proteins may be cross-reactive. These predictions can certainly be useful in developing dietary guidelines for individual patients, and in designing specific immunotherapy.

Grant support: This work was supported by grants from the Food and Drug Administration (FD-U-002249), the Texas Higher Education Coordinating Board (ATP grant 004952-0036-2003) and the National Institute of Health (R01 AI 064913).