Assessment of 3D models for allergen research

Abstract

Allergenic proteins must cross-link specific IgE molecules, bound to the surface of mast cells and basophils, to stimulate an immune response. A structural understanding of the allergen-IgE interface is needed to predict cross-reactivities between allergens and to design hypoallergenic proteins. However, there are less than 90 experimentally determined structures available for the approximately 1500 sequences of allergens and isoallergens catalogued in the Structural Database of Allergenic Proteins (SDAP). To provide reliable structural data for the remaining proteins, we previously produced over 500 3D-models using an automated procedure, with strict controls at template choice and model quality evaluation. Here we assessed how well the fold and residue surface exposure of 10 of these models correlated with recently published experimental 3D structures determined by X-ray crystallography or NMR. We also discuss the impact of intrinsically disordered regions on the structural comparison and epitope prediction. Overall, for seven allergens with sequence identities to the original templates higher than 27%, the backbone root-mean square deviations were less than 2Å between the models and the subsequently determined experimental structures for ordered regions. Further, the surface exposure of known IgE epitopes on the models of three major allergens, from peanut (Ara h 1), latex (Hev b 2) and soy (Gly m 4) was very similar to the experimentally determined structures. For three remaining allergens with lower sequence identities to the modeling templates, the 3D folds were correctly identified. However the accuracy of those models is not sufficient for a reliable epitope mapping.

Introduction

Structural comparisons of allergenic proteins are needed to supply a molecular explanation for clinically observed cross-reactivities between proteins from different organisms^1–14, to design hypoallergenic proteins ^15–19 and to predict whether new proteins or other biotechnology products are potential allergens^20–22. However, there are only 86 experimental 3D structures in the Protein Data Bank (PDB) for allergens²³, a small fraction of the 1499 allergen and isoallergens sequences collected in the Structural Database of Allergenic Proteins (SDAP)²⁴. Thus reliable, template-based models of allergens are needed to compare allergen features, and to determine potential cross-reactive IgE binding surfaces²⁵.

In 2008, we used an automated procedure to determine template-based models for 500 of the (at that time) 850 allergenic proteins in SDAP²⁶. Although template-based modeling is a well established and reliable method for predicting the global fold of proteins as documented in recent CASP competitions ^27–30, less is known about how reliably one can use 3D models in the context of allergy research. Recently, experimentally determined structures, based on X-ray and NMR data, for several of the proteins we modeled have been published ^3,31–39. Here we systematically compared these 10 models, which differed in their length, % ID to the template, and % coverage, to recently determined 3D structures for the same allergen. We evaluated the similarities of the 3D folds, the surface exposure of experimentally established epitopes and the location of glycosylation sites. Our results from this evaluation indicate that the published models in SDAP are useful, not just for determining the overall structure of ordered regions, but also for determining conformational epitopes and other characteristics of allergenic proteins, if the identity of the allergen sequence to the template sequence was higher than 27%.

Materials and Methods

The automated modeling of allergenic proteins previously performed²⁶ is briefly summarized here. We collected fold recognition results from the three fold recognition servers (FUGUE⁴⁰, mGenThreader⁴¹ and 3DPSSM⁴²). The top hits from each server were classified by the SCOP hierarchy and rated as 0 for “reliable”, 1 for “medium” and 2 for “difficult” according to the E-value or the Z-score of the servers. The confidence scores of all templates with the same SCOP fold were added, and a 3D model generated for templates with a confidence score less than 3 and no gap in the alignment greater than 20 residues. In a second round we analyzed the targets which did not pass the SCOP classification filter and used a structural comparison of aligned regions by the program CE⁴³ to find structural similar templates. If the aligned template regions had an RMSD of less than 3Å, they were considered as the same fold and the template with the longest alignment was chosen to generate a 3D model by MPACK⁴⁴.

To provide objective benchmarks for our automatic modeling procedure ²⁶, we analyzed 3D models for which experimental 3D structures were released to the PDB (http://www.pdb.org/pdb) after November 2008. These 10 models are listed in Table 1 in order of decreasing sequence identity to the template chosen for model generation. Other data in the table include the allergen name (nomenclature of the International Union of Immunological Societies, IUIS), the protein designation (vicilin, lipocalin, etc.), the template used for modeling with the PDB entry and the protein name, sequence identity between target and the template used, the PDB entry of the experimentally determined 3D structure, the backbone root mean square deviation (RMSD) in Å between the model and template, and the coverage of the structural comparison as determined with the tools of DaliLite⁴⁵ or FATCAT⁴⁶. The structural superpositions of 3D models (red and yellow in Fig 1) with the experimental structures (green in Fig 1) were prepared with MolMol ⁴⁷ using the “Fit to first” command. Solvent accessible surface areas (SASA) of individual residues were determined with the GETAREA program⁴⁸ (http://curie.utmb.edu/getarea.html). All 3D models and links to the NMR or X-ray crystal structures in the PDB are available from the SDAP website (http://fermi.utmb.edu/SDAP/). Our assessment of the quality of the 3D models using the available native structures was also compared with results from three assessment software tools, QMEAN⁴⁹ (http://swissmodel.expasy.org/qmean/cgi/index.cgi), ProSA⁵⁰ (https://prosa.services.came.sbg.ac.at/prosa.php) and Verify_3D⁵¹ (http://nihserver.mbi.ucla.edu/SAVES/) that provide quality scores based only on the 3D structures of the protein models. They use different scoring functions based on analysis of the statistical properties of the 3D models, such as packing, amino acid specific C^a-C^a distances, environmental propensities of residues, relative to the same general properties of experimental structures.

Table I.

Comparison of 3D models deposited in SDAP to experimentally determined structures

		Homology Modeling				Experimental Structure(s)	DaliLite/FATCAT: Model vs. Experimental
Allergen Name	Protein	Template		Sequence Identity (%)	SDAP Model	PDB ID	Structural RMSD (Å)	Coverage (%)
Hev b 2	rubber (latex), 1,3-glucanase	1AQ0.pdb	barley, 1,3–1,4-beta-glucanase	45.9	802.pdb	3EM5.pdb	1.3 1.2*	95.5 (300/314) 98.4 (309/314)*
Gly m 4	soybean, stress-induced protein SAM22	1FM4.pdb	birch tree, pollen allergen Bet v 1L	44.9	371.pdb	2K7H.pdb	2.3 2.7*	98.1 (153/156) 100.0 (156/156)*
Sol i 3	fire ant, venom allergen III	1QNX.pdb	yellowjacket, venom allergen 5; Ves v 5	42.8	256.pdb	2VZN.pdb	1.5 1.7*	92.6 (199/215) 96.3 (207/315)*
Ara h 1	peanut, vicilin	2CAV.pdb	jack-bean, canavalin	38.7	34.pdb	3S7E.pdb	1.5 1.6*	79.8 (332/416) 83.9 (349/416)*
Dau c 1	carrot, pathogenesis-related protein, PR-10	1BV1.pdb	birch tree, pollen; allergen Bet v 1	37.9	496.pdb	2WQL.pdb	1.3 1.3*	95.4 (146/153) 99.3 (152/153)*
Der f 1.0106	American house dust mite, cysteine protease	1CS8.pdb	human, procathepsin L	31.6	784.pdb	3D6S.pdb	1.9 1.9*	90.1 (191/212) 99.1 (210/212)*
Can f 2	salivary canine allergen, lipocalin	1MUP.pdb	house mouse, major urinary protein	26.8	82.pdb	3L4R.pdb	1.5 1.6*	89.8 (141/157) 94.3 (148/157)*
Ara h 2	peanut, conglutin	1PSY.pdb	castor bean, 2S albumin storage protein; allergen Ric c 1	22.4	283.pdb	3OB4.pdb	3.1 2.2*	41.0 (55/134) 63.4 (85/134)*
Bla g 4	German cockroach, calycin	1I4U.pdb	lobster crustacyanin, subunit C1	15.4	63.pdb	3EBK.pdb	3.0 3.0*	35.2 (64/182) 77.0 (141/182)*
Aed a 2	mosquito, salivary, odorant binding protein, protein D7	1OOH.pdb	Drosophila melanogaster, odorant binding protein LUSH	9.8	4.pdb	3DXL.pdb	3.2 2.5*	53.7 (66/123) 94.0 (116/123)*
				11.3	356.pdb		3.2 2.9*	65.3 (81/124) 94.0 (116/124)*

^*Flexible structural superposition with FATCAT.

Figure 1.

Structural overlays of aligned regions of the SDAP allergen model structures (red, yellow) with the corresponding experimental structures (green) from Table 1. A) 1,3-glucanase from latex rubber (Hev b 2) B) Stress-induced protein SAM22 from soybean (Gly m 4) C) Venom allergen III from fire ant (Sol i 3) D) Pathogenesis-related protein PR-10 from carrot (Dau c 1) E) Cysteine protease from American dust mite (Der f 1.0106) F) Canine salivary lipocalin (Can f 2) G) Conglutin from peanut (Ara h 2) H) Calycin from German cockroach (Bla g 4) I) Salivary odorant binding protein D7 from mosquito (Aed a 2) J) Vicilin from peanut (Ara h 1). The regions depicted in grey could not be aligned due to missing residues in either the model or the experimental structure. For Ara h 2, a disordered loop between helix 1 and 2, which was differently oriented in the Ric c 1 NMR structure used as template, is also depicted in grey since significant residues (shown with a thin line) are missing from the experimental structure.

Results and Discussion

Comparison of the backbone folds

The ten models used in this comparison are a good sampling, as they include proteins with many different sizes and a wide range of sequence identities from 9.8% to 45.9% between target and template used for modeling (Table 1). Backbone folds were correctly predicted with high precision for all 7 allergens where the sequence identities between templates and targets was >27%. The RMSD values between the models and the experimental structures were in most cases below 2Å, with a coverage for the structural alignment between models and experimental structures of 80–98% determined by DaliLite⁴⁵ or FATCAT⁴⁶. For the three allergens, Ara h 2, Bla g 4 and Aed a 2, the global folds were correctly identified by our modeling procedure despite low sequence identities. All 10 chosen templates were structurally similar to the experimental structures with significant P-values as determined by the flexible structure comparison method FATCAT⁴⁶. Two of the best models were for proteins larger than 300 residues, for the major latex allergen Hev b 2⁵², a 1,3-glucanase ⁵³, and the peanut vicilin, Ara h 1. The Hev b 2 model, based on the X-ray crystal structure of 1,3–1,4-beta-glucanase from barley⁵⁴, captured the (α/β)₈ TIM barrel structure in the X-ray crystal structure (PDB code 3EM5) ³¹ with remarkable accuracy (Fig. 1A). The RMSD was only 1.3Å for a structural alignment covering 95% of the protein, including the irregular loop structures. Similar results (RMSD value of 1.5AǺ with 89% coverage) were obtained for the canine lipocalin allergen Can f 2 model, where the sequence was only 27% identical to the template (Fig 1 F).

Despite low sequence identities to their templates, the global folds of the 3D models of the cockroach allergen Bla g 4 and the mosquito salivary protein antigen Aed a 2 had a sizable fraction of structural overlap. About a third of the backbone fold of the Bla g 4 model was structurally equivalent to the crystal structure 3EBK³⁸ with an RMSD value of 3AǺ (Fig. 1H) with rigid superposition by DaliLite and more than 2/3 (77%) of the global fold was structurally similar in a flexible superposition using FATCAT⁴⁶. The structurally similar regions in that flexible superposition include the 8-strand β-barrel characteristic of the lipocalin fold and part of an α-helix between K154-K162. Thus, even though we used templates with low sequence identities, the only ones available at that time, the model did correctly predict the 3D fold of Bla g 4. The rigid superposition of the model and the X-ray crystal structure by MolMol is shown in Fig. 1 H, where three 3 β-strands of Bla g 4 (V98-T101, Y106-G111 and I121-R126 in the 3EBK labeling) were found as structurally equivalent.

The fold prediction was also surprisingly good for models determined for the two domains of the mosquito Aed a 2. As is also the case for the two cupin domains of Ara h 1⁵⁵, Aed a 2 has two similar domains. Although only 15% of the residues in the two domains are identical, alignment of the two sequences suggested they had a common fold. The same template, the Drosophila protein LUSH (1OOH)⁵⁶, was thus used for modeling both domains, although its sequence was only 9% and 11% identical to either of the Aed a 2 domains. Structural alignments of the models of the N and C-terminal domains to the corresponding domains of Aed a 2 in the X-ray crystal structure (3DXL³⁹) gave an RMSD of 3.2AǺ with a coverage of 53.7% and 65.3%, respectively (Figure 1, I). Flexible superposition with FATCAT recognized more than 90% of the folds as structurally similar, with RMSD values of less than 3Ǻ.

The successful template choices depended on the internal quality control of our automated modeling procedure, which required that the 3 fold recognition servers consistently agree on a template. The alignments for modeling were also selected based on a scoring function that included a gapping penalty, and overall length of the alignment of the template with the target sequence, and low conformational energy²⁶.

Mapping IgE epitopes on the models

In many applications of 3D modeling for allergenic proteins, the goal is to determine the solvent accessibility of side chains on the allergens, and possibly predict the structure of (or at least the residues involved in) conformational epitopes²². Most data on IgE epitopes comes from dot spot assays, which measure the binding of IgE in allergic patient sera to linear peptides synthesized on membranes^7,22 or microarrays ⁵⁷. IgE epitope information is available for a few of the allergens in this study, including those from peanut Ara h 1^58,59 and Ara h 2⁶⁰, latex, Hev b 2⁶¹, and soybean, Gly m 4⁶². The Ara h 1 model was based on the jack-bean canavalin crystal structure 2CAV⁶³; the overall sequence identity to this template was below 40%. The model correlated very well with the subsequently determined X-ray crystal structure of Ara h 1³³ (3S7E); the RMSD was 1.5Ǻ with 80% coverage (Fig. 1J). Further, the orientation and solvent accessible surface area (SASA) of the linear IgE epitopes (i.e., the peptides that bound to IgE in dot spot assays) were nearly identical in the model and the crystal structure (Fig. 2), and the conformations of the eight individual epitopes were also quite similar, with RMSD values ranging from 0.3Å to 0.9Å. Indeed, as Fig. 3 shows graphically, the solvent exposures of more than 90% of the residues in the model were within 10 Ǻ² of those in the crystal structure. There was a significant difference (>50 Ǻ²) in solvent exposure of only 27 residues (Fig. 3B) most of them for surface exposed residues (see Supplementary Material S1). The single glycosylation site (residues 521–524, NASS; arrow in Fig. 2) was also located at similar surface exposed areas in both the model and crystal structure, as was its distance to the side chains of the IgE epitopes that surround it.

Figure 2.

Comparison of the structural regions of Ara h 1 that correspond to the linear peptides that bind IgE antibodies. Epitopes 10–21, listed in the Ara h 1 entry page of SDAP (http://fermi.utmb.edu/SDAP/), were mapped on the 3D model (A) and X-ray crystal structure (B) of Ara h 1.

Figure 3.

Comparison of the solvent accessible surface areas (SASA) per residue in the Ara h 1 model from SDAP and the crystal structure (3S7E,³³). (A) Correlation plot (B) Histogram for the absolute differences of the SASA of residues in the 3D model and the crystal structure.

Similarly high structural equivalence was found for the locations of the individual epitopes of the latex allergen Hev b 2 (Fig. S2) and the soybean allergen Gly m 4 (Fig. S3). The structural deviations of the individual epitope backbone segments were near 1Å with an average over all RMSD values of the segments of 0.7Å for Hev b 2 and slightly higher, 1.0Å, for Gly m 4. The highest deviating segments were found for 3 epitopes of Gly m 4 (1.4Å, 1.7Å and 2.7Å) in loop regions. For the more challenging model of Ara h 2 we found larger variations (Fig. S4). Whereas the surface positions of epitopes 3, 4, 8 and 10 were similar in the model and crystal structure, epitopes 6 and 7 showed large deviations. Note, however, that only 4/8 residues of Epitope 7 could be compared, as epitope 7 is part of a flexible loop (R59 to H84) that is missing in the X-ray crystal structure (thin line in Fig. 1G).

Weaknesses and Strengths of the 3D models by quality assessment tools

The above results show that, before mapping epitopes on a SDAP model, the user should take into account three factors included in all the structure files: the % identity of the allergen sequence to the template used and the overall coverage of the sequences. In addition, the user should determine the degree of disordered structure in the template. The user can also use several different, publicly assessable tools to further determine the quality of the models, based only on the coordinates. In supplementary material(Table S1), we show the results for assessing the quality of our 10 models using three current software tools, including QMEAN ⁴⁹, ProSA ⁵⁰ and Verify_3D ⁵¹. The quality scores of all three assessment tools for the seven high quality models of Hev b 2, Gly m 4, Sol i 3, Ara h 1, Dau c 1, Der f 1 and Can f 2 are in a range that one would expect for native structures, and are thus consistent with the analysis using the experimentally known 3D structures. Also all servers rank the qualities of the three models for Ara h 2, Bla g 4 and Aed a 2 consistently lower. However, there are some differences in the degree of quality assessment of those models. Whereas Verify_3D flags all three models of lower quality with “Fail”, the OMEAN Z-scores and the ProSA Z-scores for the models of Ara h 2 and Aed a 2 are in a borderline region when compared to the scores of native structures. Thus in practice those 3D models might be useful for the structural classification of allergens in super families, but not for reliable epitope mapping or the design of hypoallergenic proteins. As those tools provide quality scores for the 3D allergen models, they can be applied by end-users to the remaining predicted models in SDAP to give them further confidence for use of the models or reasons for further analysis by local quality plots.

Impact of structural flexibility on 3D models and experimental structures

Flexible regions were the major reason for differences between the models and the experimental structures. Intrinsically disordered regions are found prominently in the protein universe⁶⁴. Although the overwhelming majority (80%) of allergens have a structured 3D fold, a sizable fraction contain locally disordered regions⁶⁵. Peanut Ara h 2 is a good illustration of how disordered regions present problems for both experimentalists and modelers. Despite its importance as an allergen^60,66, and its small size, the protein eluded crystallization efforts by several groups. An X-ray crystal structure was only recently determined, using a fusion protein with an N-terminal deleted form of Ara h 2 with the maltose binding protein (MBP) ³⁷ (PDB structure 3OB4). The MBP fusion partner might impact the structure of the N-terminal region of the protein. Our model of Ara h 2 was based on an NMR structure for 2S albumin Ric c 1 from castor bean (PDB code 1PSY; ⁶⁷). Although the Ric c 1 template was only 22% identical, the Ara h 2 model’s four helices overlay well with the corresponding ones in the crystal structure (Fig. 1G). The relatively high backbone RMSD (3.1 AǺ) reflects mainly differences in the conformation of the kinked helix 1. The NMR ensemble showed considerable variation especially in the loop region connecting helices 1 and 2 (gray line in Fig. 1G), which was also predicted by DisProt⁶⁸ to be disordered. The crystal structure lacks also structural information for 15 of the residues (R59 to H84) in that loop (thin line in Fig. 1G). Two IgE epitopes (epitope 6 and 7 in the SDAP epitope table for Ara h 2) have been located in that area of the protein⁶⁰, illustrating that IgE epitopes can overlap with flexible loop regions as documented by a recent statistical study ⁶⁵. Thus determining the actual structure of these two epitopes will probably require a co-crystal structure of Ara h 2 with relevant monoclonal antibodies.

We also saw some structural differences due to flexibility even in allergen structures where the identity to the template was much higher. Soybean Gly m 4^69,70, a member of the pathogen related protein family PR-10, was modeled on an X-ray crystal structure of the birch pollen allergen Bet v 1 ¹⁵. Although the overall fold is correct (RMSD 2.3Å), the model deviates from the NMR solution structures of Gly m 4³ in flexible areas, where the structures in the bundle also differ from each other.

Despite the problems noted above for flexible regions, which may not be completely resolved by any technique, experimental structures are still the most direct and reliable procedure to map conformational epitopes or to graft conformational epitopes between different allergens^9,71,72. Thus it is very encouraging that many structures have been published recently for newly identified allergens, or those for which we were unable to find a suitable modeling template in our 2008 study. These include NMR structures of the major mugwort pollen allergen Art v 1 ⁷³ and X-ray crystal structures of the mold allergen Alt a 1⁷⁴, pollen allergen Che a 3¹³, dust mite allergen Der p 7 ⁷⁵, and antibody complexes of the house dust mite allergens Der p 1 and Der f 1⁷⁶ and the cockroach allergen Bla g 2¹⁸. At least one of these, the fungal allergen Alt a 1, showed a novel fold, a unique β-barrel dimer structure⁷⁴. However, based on the number of allergens for which no structure exists, the analysis of IgE epitopes will continue to rely on 3D models for some time to come ^{4,5,10,11,77–83}. As the accuracy of prediction methods for template based modeling ^{27,30,84–86} and intrinsically disordered regions^87,88 is continually improving, as is the number of potential templates in the PDB for the protein families of most allergens^89,90, we can anticipate that future models will provide even better descriptions of epitope surfaces.

Conclusion

Our comprehensive 3D modeling effort for the allergens in SDAP was a major effort towards determining the 3D structures of all allergens. This analysis shows that the automatic, template based modeling procedure, probably thanks to the quality controls implemented, yielded reliable models if the identity of the allergen sequence to the template sequence was higher than 27%. In those cases the 3D models could be considered as high-quality models with backbone deviations of typically between 1Å and 2 AǺ in structured regions. However, not all 3D models have the same reliability. Current quality assessment tools can help end-users from over-interpreting the structural information. The user should be especially wary of regions that do not align to the template, or those based on templates with a large content of disordered or flexible loops. Proteins with large disordered regions will require improved tools to determine accurate models. The examples of three major allergens show that it is possible to use 3D models to determine the IgE epitopes and glycosylation sites with great accuracy if a reliable template can be found. This structural information will allow a better comparison of allergens from different organisms for potential cross-reactivity and to design hypoallergenic proteins than just using sequence information alone.