Computational approach for predicting the conserved B-cell epitopes of hemagglutinin H7 subtype influenza virus.

An avian-origin influenza H7N9 virus epidemic occurred in China in 2013-2014, in which >422 infected people suffered from pneumonia, respiratory distress syndrome and septic shock. H7N9 viruses belong to the H7 subtype of avian-origin influenza viruses (AIV-H7). Hemagglutinin (HA) is a vital membrane protein of AIV that has an important role in host recognition and infection. The epitopes of HA are significant determinants of the regularity of epidemic and viral mutation and recombination mechanisms. The present study aimed to predict the conserved B-cell epitopes of AIV-H7 HA using a bioinformatics approach, including the three most effective epitope prediction softwares available online: Artificial Neural Network based B-cell Epitope Prediction (ABCpred), B-cell Epitope Prediction (BepiPred) and Linear B-cell Epitope Prediction (LBtope). A total of 24 strains of Euro-Asiatic AIV-H7 that had been associated with a serious poultry pandemic or had infected humans in the past 30 years were selected to identify the conserved regions of HA. Sequences were obtained from the National Center for Biotechnology Information and Global Initiative on Sharing Avian Influenza Data databases. Using a combination of software prediction and sequence comparisons, the conserved epitopes of AIV-H7 were predicted and clarified. A total of five conserved epitopes [amino acids (aa) 37-52, 131-142, 215-234, 465-484 and 487-505] with a suitable length, high antigenicity and minimal variation were predicted and confirmed. Each obtained a score of >0.80 in ABCpred, 60% in LBtope and a level of 0.35 in Bepipred. In addition, a representative amino acid change (glutamine235-to-leucine235) in the HA protein of the 2013 AIV-H7N9 was discovered. The strategy adopted in the present study may have profound implications on the rapid diagnosis and control of infectious disease caused by H7N9 viruses, as well as by other virulent viruses, such as the Ebola virus.

Introduction

The H7 subtype of avian-origin influenza viruses (AIV-H7) initially emerged in Italy in 1902 as H7N7 (1). During its period of evolution, AIV-H7 has repeatedly caused pandemics among poultry and has been associated with huge losses to livestock. However, AIV-H7 infections in humans have rarely been reported. One of the largest infections of humans occurred in the Netherlands in 2007 (2); the outbreak was caused by the H7N7 subtype and led to keratitis and other minor symptoms in humans, although no mortalities were reported. However, in 2013 an AIV-H7N9 epidemic occurred in China. By May 1st, 2014, China had reported 422 confirmed cases (3), including people suffering from pneumonia, respiratory distress syndrome, septic shock and other life-threatening diseases (4). The exact death toll had not been updated in 2014; however, in 2013, >57 mortalities were reported, with a fatality rate of >33%. Therefore, H7-AIV has aroused global concern.

Similar to other influenza viruses, AIV-H7 can be divided into North American and Euro-Asiatic lineages (5); the 2013 outbreak in China belonged to the Euro-Asiatic lineage (6). The major glycoprotein of AIV is hemagglutinin (HA), which has an important role in binding the virus to sialic acid on the membranes of host cells, including cells in the upper respiratory tract or erythrocytes (7). Phylogenetic tree analyses suggested that the HA of H7N9 was derived from a reassortment of H7N3 in Zhejiang ducks (8). HA, which is a large protein of 560 amino acids, possesses various epitopes that can be divided into linear and conformational types (9). The linear epitopes consist of conserved amino acids, whereas the conformational epitopes contain the adjacent amino acids in space that may lie far away from the linear epitopes in the primary sequence (10). The majority of the HA epitopes are conformational; however, linear epitopes are easier to express and mimic (11). Therefore, the present study aimed to predict the potential linear epitopes of AIV-H7.

The present study was divided into two parts. First, the linear epitopes of HA were predicted using a combination of three epitope prediction softwares, including: Artificial Neural Network based B-cell Epitope Prediction (ABCpred), B-cell Epitope Prediction (BepiPred) and Linear B-cell Epitope Prediction (LBtope). Each software constituted a specific algorithm with the highest predictive accuracy being ≤66% (12). Second, 24 strains of Euro-Asiatic AIV-H7 that had infected humans or had caused avian pandemics in the past 30 years (13) were selected, and the amino acid sequences of HA were compared in order to identify the conserved region of HA. By using epitope prediction softwares and comparing reference strains, the present study effectively screened for invalid and mutant epitopes, which was time-saving and inexpensive. In addition, preliminary investigations on the antigenicity of HA, as well as amino acid mutations and epitopes associated with secondary structure, were performed and may be considered useful for future research. The results of the present study may have a profound impact on the diagnosis and future research of, and design of vaccines for, AIV-H7, as well as for other virulent viruses such as the Ebola virus.

Data and methods

Sequence availability and comparisons

Sequences were obtained from the National Center for Biotechnology Information database (http://ncbi.nlm.nih.gov) and the Global Initiative on Sharing Avian Influenza Data (GISAID) database (http://platform.gisaid.org/epi3/frontend). Since AIV-H7 has frequent variation, 24 strains of Euro-Asiatic lineage H7 subtype AIV that had triggered serious poultry pandemics or had infected humans in the past 30 years were selected (13). Of the 24 strains, two represented the latest H7N9 virus, and were selected from 94 of the latest complete records of H7N9 in the GISAID database.

ClustalW (www.ebi.ac.uk/Tools/msa/clustalw2/) was used to compare the HA amino acid sequence of 24 strains and to identify their conserved region. The accession numbers of the HA proteins of the 24 strains are listed in Fig. 1. The HA gene with 560-amino acids from H7N9 (A/tree sparrow/Shanghai/01/2013 (H7N9), gi|546235348| or EPI439486) was selected as the reference sequence as it was the latest HA gene.

Figure 1.

Conserved regions of the epitopes predicted by three epitope prediction softwares. The accession numbers of HA proteins from 24 strains of Euro-Asiatic avian-origin influenza virus, H7 subtype are listed on the left. The scales on the top show the rank of amino acids within HA. The amino acids in the first line of the graph are the H7N9 reference sequence. Other sequences are presented as dots if they share the same amino acid as the reference sequence or as the one-letter symbols if the amino acids are altered. The black boxes denote the locations of the 11 potential epitopes. Grey shaded areas denote highly conserved sequences. The 235th amino acid is denoted by a short black line. HA, hemagglutinin.

Primary sequence and structure analysis

Various properties of the 560-amino acid HA protein from H7N9, including the theoretical isoelectric point (pI), amino acid composition and molecular weight, were tested using the ProtParam tool from the ExPASy Bioinformatics Resource Portal (http://web.expasy.org/protparam/). Subsequently, a structural search was performed using the Protein Data Bank database (http://rcsb.org). PDB:4N5J was identified as the three dimensional (3D) structure of AIV-H7.

Prediction of linear B-cell epitopes

Linear B-cell epitopes were predicted using the algorithms of ABCped, BepiPred and LBtope. The sequence of HA protein were downloaded and analyzed by each software. ABCpred has developed a systematic method based on a neural network (12). The amino acid length was set at 10, 12, 14, 16, 18 and 20 mer and the scoring threshold at 0.8. BepiPred, which was developed by Larsen et al (14), employs the hidden Markov model and apropensity scale method developed by Parker et al (15). A threshold of 0.350 was selected as it is the point at which sensitivity (0.49)/specificity (0.75) is maximized (14). LBtope is a new method for predicting linear epitopes and was created by Singh et al (16) in 2013. The Lbtope_Confirm dataset was selected since it has previously shown the best performance. The amino acid length was set to 15 mer and the scoring threshold to 60%. The results from the three softwares were assembled and the overlapping regions were considered predicted epitopes.

Analysis of predicted epitopes

Using a combination of software prediction and sequence comparisons, the conserved epitopes were predicted and a domain enhanced lookup time accelerated-Basic Local Alignment Search Tool (BLAST) analysis (http://blast.ncbi.nlm.nih.gov/Blast.cgi) was performed to determine the specificity of the epitopes. Furthermore, the location and structures of the epitopes were denoted in 3D images.

Analysis of glutamine (Gln)235-to-leucine (Leu)235 mutation

The 235th amino acid of HA from 24 AIV-H7 strains, as well as 100 strains of human H7N9 viruses from the GISAID database, were compared. The 235th amino acid of HA from 24 AIV-H7 strains, as well as 100 strains of human H7N9 viruses from the GISAID database, were distinguished based on whether mutations were present.

Results

Similarity searches and primary sequence analyses

Primary sequence analyses revealed that the theoretical pI of HA was pH 6.25 and the molecular weight was 62.062 kDa. The location and variation of HA from 24 AIV-H7 strains are presented in Fig. 1. Following the comparison, the highly conserved sequences included amino acids 18–41, 72–98, 149–164, 253–269, 294–306, 339–356, 358–402, 415–426, 428–454, 463–499 and 507–535. The majority of mutations were located in the N-terminal 339 amino acids, which comprised the HA1 subunit of HA. The HA1 subunit forms the head of HA, which has a vital role in interacting with the environment and host receptor. Thus, the HA1 subunit is readily mutated in order to avoid the attack of antibodies (17).

Epitope prediction by BepiPred

BepiPred software was used to produce a Linear Epitope Prediction map (Fig. 2) and BepiPred Prediction Details (Table I). BepiPred analyzes each amino acid independently and assigns it a score between −3 and 3. A higher score indicates a higher probability for the existing epitope (14). The threshold was set at 0.35 and scores >0.35 were regarded as positive. From the Fig. 2, beyond the axis, the higher of the peak, the higher scores of prediction. Therefore, the 19 peaks of the map formed by consecutive amino acids were regarded as 19 predicted linear epitopes. They were ranked by mean residue score (mean residue score was averaged by every residue in the predicted peptide), as presented in Table I. The peaks 1–4 were the most likely epitopes to be predicted by BepiPred.

Figure 2.

BepiPred linear epitope prediction. The abscissa represents the rank of HA amino acids and the ordinate represents the scores of every residue. Yellow denotes a positive score for linear B-cell epitopes. The horizontal axis denotes the threshold of 0.350. The further amino acids lie above the threshold, the greater possibility they have of becoming linear epitopes. According to their locations and scores, 19 potential epitopes were selected.

Table I.

Epitopes of hemagglutinin predicted by BepiPred.

Rank	Sequence	Location	Mean residue score
1	RENAEEDGTG	466–475	1.130
2	SSNYQQSFVPSPGARPQVNG	215–234	1.126
3	HQNAQGEGTAADYKSTQSAIDQ	365–386	1.086
4	TITGPPQCDQ	77–86	0.857
5	KNVPEIPKGR	330–339	0.850
6	TDNAAFPQMTKSYKNTRKS	165–183	0.732
7	STAEQTKLYG	196–205	0.707
8	NNTYDHSKYREEA	493–505	0.688
9	IRTNGATSACRRSGSS	138–153	0.636
10	VNATETVERT	45–54	0.587
11	MGIQSGVQVDANCEGDCYHS	274–293	0.539
12	VSNGTKVNTLTE	28–39	0.537
13	REGSDVCYPGKFV	99–111	0.523
14	IDSRAVGKCPR	306–316	0.378
15	NDTVTFSFNGAFIA	249–262	0.091
16	KGKRTV	62–67	0.426
17	GGTDK	124–128	0.946
18	FNEVEK	409–414	0.472
19	KLSSG	516–520	0.484

Epitope prediction by LBtope and ABCpred

The output of the LBtope analysis is presented in Fig. 3. LBtope assigns scores between 0 and 100% to each epitope it predicts (16). A higher score indicates a higher probability of the epitope existing (16). According to the mean residue score (Table II), the software selected 14 consecutive amino acids that have a possibility of 61 to 100% of being an epitope. As the threshold was set at 60%, these 14 consecutive amino acids were regarded as 14 predicted epitopes. They were ranked by mean residue score, as shown in Table II. The sequences 1–2 were the most likely linear epitopes as their scores were >80%.

Figure 3.

LBtope linear epitope prediction. The diagram depicts the entire amino acid sequence of hemagglutinin. Box areas have a possibility of 81 to 100% of being the epitope. Underlined areas have a possibility of 61 to 80%. The remaining areas have a possibility of 0 to 20%. The area can be considered as an epitope only when the possibility exceeds 60%.

Table II.

Epitopes of hemagglutinin predicted by LBtope.

Rank	Sequence	Location	Mean residue score
1	RNNTYDHSKYREEAM	492–506	86.80
2	LYGSGNKLVTVGSSNYQQSFVPSPGARP	203–230	82.37
3	IDSRAVGKCPRY	306–317	78.94
4	QITGKLNRLIEK	386–397	77.98
5	IIERREGSDVCYPG	95–108	76.45
6	TNIPRICSKGKRTVDL	54–69	75.26
7	SYKNTRKSPAL	176–186	75.10
8	DGWYGFRHQN	358–367	72.99
9	NTLTERGVEVVNATET	35–50	72.11
10	LRENAEEDGTGCF	465–477	72.04
11	LRGKSMGIQSGVQVDANCEG	269–288	71.31
12	DKEAMGFTYSGIRTNGATSA	127–146	70.84
13	SLLLATGMKNVPEIPK	322–337	69.75
14	SITEVWSYNA	426–435	67.15

ABCpred is able to predict antigens that vary in length from 10 to 20 residues and assigns a score between 0 and 1 to each epitope it predicts. A score that is closer to 1 indicates a higher probability of the epitope existing and a score closer to 0 suggests that the amino acid sequence is unlikely to be an epitope (12). In order to avoid omissions, the amino acid length was set to 10, 12, 14, 16, 18 and 20 mer and the scoring threshold to 0.8. In total, 67 sequences met the requirements.

Prediction of potential epitopes

The overlapping epitopes predicted by the three softwares were considered as potential epitopes. Following a comparison and analysis, 11 epitopes met the requirements. Each of the potential epitopes were above the thresholds set for the 3 software packages, had a suitable length and were highly antigenic. Therefore, they adequately represented the linear epitopes of HA. The 11 epitopes are ranked by ABCpred scores in Table III.

Table III.

Potential epitopes, as ranked by ABCpred scores.

Number	Sequence	Start position	End position	ABCpred score
1	PEIPKGRGLFGAIAGF	333	348	0.91
2	AFPQMTKSYKNTRKSP	169	184	0.91
3	FQNIDSRAVGKCPRYVKQRS	303	322	0.88
4	SMGIQSGVQVDANCEGDCYH	273	292	0.88
5	ATETVERTNIPRICSK	47	62	0.85
6	LRENAEEDGTGCFEIFHKCD	465	484	0.85
7	LRESGGIDKEAMGFTY	120	135	0.84
8	LTERGVEVVNATETVE	37	52	0.83
9	CMASIRNNTYDHSKYREEA	487	505	0.82
10	MGFTYSGIRTNG	131	142	0.82
11	SSNYQQSFVPSPGARPQVNG	215	234	0.80

HA antigenicity and the conservation of predicted epitopes

The locations of the 11 potential epitopes were determined. The entire HA protein can be divided into three parts: Amino acids 1–18 form the signal peptide, amino acids 19–339 form the HA1 subunit and amino acids 340–560 form the HA2 subunit (17). Eight of the 11 epitopes were in the HA1 subunit, two were in the HA2 subunit and one was at the junction of the two subunits. These results suggest that the HA1 subunit is the immunodominant antigen and the HA2 subunit is more conserved.

Five epitopes showing minimal variation were selected from the 11 epitopes by observation and comparison (Fig. 1). Two of the five epitopes were from the HA2 subunit and three were from the HA1 subunit. These five epitopes showed high antigenicity and were highly conserved; thus they were called the conserved predicted epitopes. Three of the five epitopes were 20 amino acids in length, one was 16 amino acids and the other was 12 amino acids. The exact position and composition of amino acids are shown in Table IV.

Table IV.

Five potential conserved epitopes.

Number	Sequence	Start position	End position	ABCpred score
6	LRENAEEDGTGCFEIFHKCD	465	484	0.85
8	LTERGVEVVNATETVE	37	52	0.83
9	CMASIRNNTYDHSKYREEA	487	505	0.82
10	MGFTYSGIRTNG	131	142	0.82
11	SSNYQQSFVPSPGARPQVNG	215	234	0.80

A Venn diagram was used to show the analysis process and overall results (Fig. 4). The 11 epitopes that were detected by all three algorithms were in the overlapping part of the three circles. The middle overlapping part was divided into two parts by a short line. The five dots laying above the line represented the five conserved predicted epitopes among the 11 potential epitopes.

Figure 4.

Venn diagram of epitopes detected by ABCPred, BepiPred and LBtope. Black dots represent epitopes predicted by the three software packages. Epitopes predicted by ABCPred, BepiPred and LBtope are scattered in every circle. The 11 epitopes that were detected by all three algorithms are in the center of the Venn diagram. The center area is divided into two parts by a short line; the five dots that lie above the line represent five conserved sequences among the predicted epitopes.

Secondary structure and specificity of the five conserved predicted epitopes

The secondary structures of the five conserved predicted epitopes were analyzed and are presented in Fig. 5. Epitope 6 contains two β-strands and one turn, epitope 8 contains three β-strands and one β-bridge, epitope 9 consists of one turn and two parts of an α-helix, epitope 10 contains one β-strand and one β-bridge and epitope 11 constitutes one β-strand, one complete β-bridge, one partial β-bridge and one turn. All of the five epitopes are on the surface of HA protein, which exposes them to the environment and makes them more likely to be antigenic (18). Previous studies have demonstrated that β-bridges and turns are more likely to form epitopes (18,19) and, in the present study, at least one β-bridge or turn was identified in every conserved predicted epitope. The exact locations of the epitopes are displayed in Fig. 5.

Figure 5.

The three dimensional structure of the HA protein. HA consists of two subunits: Purple and dark blue regions represent the HA1 and HA2 subunits of the protein, respectively. Yellow bands denote the five conserved epitopes predicted by the three algorithms, including epitopes (A) 10, (B) 11, (C) 6, (D) 8 and (E) 9. The circles show the amplified conserved epitopes and their detailed structures. HA, hemagglutinin.

Epitopes 8, 9 and 11 had no significant similarity with other organisms, as demonstrated by a BLAST analysis. Epitope 6 showed 70% similarity with H5N1 and H2N2. Epitope 9 showed 74% similarity with H3N1 and H3N2, 68% with H5N1, H5N2 and H1N1, and 63% with H12N4. Since changing one amino acid in an epitope can markedly decrease the antigen-antibody interaction involved in antibody recognition, it was concluded that all five epitopes were specific to AIV-H7 (9).

Results of Gln235 to-Leu235 mutation

The present study demonstrated that the 235th amino acid of HA (Gln) was a highly conserved residue for the majority of H7-AIV, including those associated with previous human outbreaks (Fig. 1). However, in the 2013 H7N9 AIV, the 235th residue had mutated to Leu; in the 100 strains of human H7N9 viruses collected from across China, only 16 carried Gln235, with the remaining (84%) all carrying Leu235. These results suggest that the Gln235-to-Leu235 mutation exists in the novel H7N9 viruses.

Discussion

Bioinformatics is a promising and standard approach for the identification of specific and immunogenic epitopes and it has important applications in vaccine design, epitope mapping and antibody research (20,21). Previous studies have demonstrated that the efficiency of discovering novel epitopes is improved 10–20 times by immunoinformatics; the experimental work was reduced by 95% in one study (22). In 2008, Frikha-Gargouri et al (23) predicted a specific and immunogenic antigen of the OmcB protein for the serodiagnosis of Chlamydia trachomatis infections. Their results indicated that the use of sequence alignment tools may be useful for identifying specific regions of an immunodominant antigen (23). Furthermore, Jones and Carter (24) used bioinformatics tools to predict the B-cell epitopes of Listeria monocytogenes and develop immunity to the bacterium in 2013. Their results may be used to investigate the pathogenesis of L. monocytogenes infections, as well as to develop an inexpensive assay. Maksimov et al (25) used in silico-predicted epitopes for the serological diagnosis of Toxoplasma gondii infection in humans, and established a peptide-based microarray assay to assess the diagnostic performance of the selected peptides. The present study used bioinformatics to predict the linear B-cell epitopes of HA of H7-AIV.

ABCpred, which is an algorithm that was created by Saha and Raghava (12) in 2006, is able to predict epitopes with 66% accuracy, 67% sensitivity and 65% specificity. Initially, the recurrent neural network was used, and it was trained with a dataset of 700 experimentally detected B-cell epitopes from the BciPep database (26) and 700 random peptides from the Swiss-Prot database, for which no antibody binding is reported as a negative dataset. The ABCpred algorithm was shown to have a better predictive performance, as compared with various physicochemical properties, including hydrophilicity (15), flexibility (27) and accessibility (28). In addition, Costa et al (29) demonstrated that AAPPred and ABCpred obtained the best results in terms of epitope prediction, as compared with other programs, although AAPPred is no longer available. BepiPred is a traditional algorithm developed by Larsen et al (14) in 2006 using 14 epitope-annotated proteins and a human immunodeficiency virus dataset. BepiPred analyzes each amino acid independently and does not require a minimum or maximum number of amino acids to predict an epitope. While we can distinguish the epitopes by the scores, Reimer (30) demonstrated that the predictions made by BepiPred were better than a random guess for 8/11 proteins. LBtope is a novel tool for the prediction of epitopes that was developed by Singh et al (16) who exploited the availability of several thousands of experimentally verified epitopes and non-epitopes. Singh et al (16) derived five datasets from the Immune Epitope Database called Lbtope_Fixed, Lbtope_Fixed_non_redundant, Lbtope_Variable, Lbtope_Confirm and Lbtope_Variable_non_redundant dataset (13). The greatest advantage of LBtope is the ability to rule out nonepitopes that are neglected by other algorithms, and it has been shown to compensate for the inadequacies of the other two methods (16). However, users are advised to predict linear epitopes using all existing methods and then identify the target predicted by the majority of the methods (12). Therefore, the present study combined the advantages of three algorithms and regarded the overlapped results as the potential epitopes. Due to limitations of epitope prediction methods, the prediction was unable to reach an accuracy of 100% and further studies are required.

In conclusion, the present study identified 11 potential epitopes of the HA protein, and demonstrated that the HA1 subunit was the immunodominant antigen, whereas HA2 was more conserved. Five potential conserved epitopes were selected and were analyzed for secondary structure, software prediction and sequence comparison; they all showed a high antigenicity and low variation. Previous studies demonstrated that the Gln235-to-Leu235 mutation was associated with an improved affinity for human receptors, in particular when sialic acid is 2-6-linked to galactose in novel H7N9 viruses (31,32). This mutation was detected in the present study, thus suggesting that it may have an important role in the increased virulence of novel H7N9 viruses.

In the 2013–2014 period, H7N9 caused an epidemic in humans that was associated with severe morbidity and mortality. Investigation into the epitopes of H7-AIV may accelerate the diagnosis of epidemic disease, permit the prediction of epidemics and allow viral mutation, pathogenesis and recombination mechanisms to be monitored. Future studies should extend to the Euro-Asiatic H5, H1 and other subtypes of AIV as well as to other virulent viruses, such as the Ebola virus.