Abstract

MHC molecule plays a key role in immunology, and the molecule binding reaction with peptide is an important prerequisite for T cell immunity induced. MHC II molecules do not have conserved residues, so they appear as open grooves. As a consequence, this will increase the difficulty in predicting MHC II molecules binding peptides. In this paper, we aim to propose a novel prediction method for MHC II molecules binding peptides. First, we calculate sequence similarity and structural similarity between different MHC II molecules. Then, we reorder pseudosequences according to descending similarity values and use a weight calculation formula to calculate new pocket profiles. Finally, we use three scoring functions to predict binding cores and evaluate the accuracy of prediction to judge performance of each scoring function. In the experiment, we set a parameter in the weight formula. By changing value, we can observe different performances of each scoring function. We compare our method with the best function to some popular prediction methods and ultimately find that our method outperforms them in identifying binding cores of HLA-DR molecules.

1. Introduction

Histocompatibility refers to the degree of antigenic similarity between the tissues of different individuals, which determines the acceptance or rejection of allografts. Transplantation antigen or histocompatibility antigen is the cause of rejection of allografts [1, 2]. MHC (Major Histocompatibility Complex) is present on the chromosome encoding a major histocompatibility antigen, mutual recognition between control cells, and the regulation of immune response.

MHC molecule plays a key role in immunology, and the molecule binding reaction with peptide is an important prerequisite for T cell immunity induced [2, 3]. By detecting a wide variety of microbial pathogens, the immune system protects host against diseases. Because of this, the binding prediction of MHC molecules with peptides has always been a hot topic in bioinformatics. Many researches in this field not only help us to understand the process of immune but also develop the work of vaccine design assisted by computers.

MHC genes produce two different types of molecules, which are MHC I molecules and MHC II molecules [1, 2]. MHC I molecules contain two separate polypeptide chains: the MHC α chain encoded by MHC genes and the MHC β chain encoded by non-MHC genes [4, 5]. MHC I class molecules are expressed in almost all eukaryotic cell surfaces, recognized by CD8+ cells. MHC II class molecules consist of two non-covalently linked polypeptide chains, namely, α chain and β chain. MHC II class molecules are expressed on antigen-presenting cells in general. Foreign MHC II antigens only capture and present on the surface of antigen-presenting cells (APC) TH cell [6]. After that, APC secretes large amounts of cytoplasm, activating cell invasion defensed behavior. Only the binding of antigen peptides and MHC II class molecules can activate CD4+ TH cells (helper T cells) [7]. Then, the activated TH cells would differentiate into effector cells and activate the immune response.

The structures of MHC I molecules and MHC II molecules slightly differ in the binding grooves [5]. Close grooves form on the binding of MHC I molecules and antigenic peptides. On the other hand, MHC II molecules do not have conserved residues, so they appear as open grooves. As a consequence, this will increase the difficulty in predicting MHC II molecules binding peptides [7]. In this paper, we aim to solve more difficult problem of predicting MHC II binding peptides.

The pioneering and most popular pan-specific approach for MHC II binding prediction is the TEPITOPE method [8], and basic idea is the HLA-DR allele having identical pseudosequence. The same pocket will share the same quantitative profile. By using multiple instance learning, the MHCIIMulti method [9] can predict more than 500 HLA-DR molecules. Transforming each DRB allele into a pseudosequence with 21 amino acids and using the SMM-align method to identify binding cores, the NetMHCIIpan method [5] gets an accurate prediction by using an artificial neural network algorithm [10, 11]. Combining NN-align and NetMHCpan with NetMHCIIpan [9, 12], the MULTIPRED2 method [1315] can get a perfect prediction for 1077 HLA-I and HLA-II alleles and 26 HLA supertypes.

In this paper, we propose a novel prediction method for predicting MHC II molecules binding peptides. First, we calculate sequence similarity and structural similarity between different MHC molecules [13, 16]. Then, we reorder pseudosequences according to descending similarity values and use a weight calculation formula to calculate new pocket profiles. Finally, we use three scoring functions to predict binding cores and evaluate the accuracy of prediction to judge performance of each scoring function [17, 18]. In the experiments, we set a parameter in the weight formula. By changing value, we can observe different performances of each of the scoring functions. We compare our method with the best function to some popular prediction methods and ultimately find that our method outperforms them in identifying binding cores of HLA-DR molecule [19]. The work would suggest a novel computational strategy for special protein identification instead of traditional machine learning based methods [20, 21].

2. Materials and Methods

2.1. Data Sets

We find 39 MHC molecules and peptides binding complexes from Protein Data Bank (http://www.rcsb.org/pdb/search/), which constitutes the data set used in this paper. In this data set, lengths are between 11 and 23, and we can find polypeptide-binding sites, namely, binding cores. Table 1 lists the details of these 39 MHC molecules and peptide binding complexes [14, 22, 23].

In Table 1, the first column is PDB ID of 39 complexes from PDB; the second column is the name of corresponding alleles from 39 complexes; the third column is the corresponding polypeptide sequences, in which the enlarged nine positions are the binding cores.

2.2. Methods

There are thousands of allele variants in nature [2, 4]. It is absolutely impossible to measure the binding specificity one by one. Motivated by this perspective, we propose a new computational method to predict the binding specificity of peptides without any biochemical experiment, which combines the sequence and structural information of these known specificity-binding MHC molecules, as showed in Figure 1. We evaluate the method on all general HLA-DRB data sets, and results indicate that our method is close to the state-of-the-art technology and our approach can predict all sequence-known MHC molecules and cost little time, extending the prediction space compared with other time-consuming approaches.

2.3. Crucial Pockets relative to Binding Specificities of HLA-DR Molecules

We mainly use Position Specific Scoring Matrix (PSSM) [13, 24] in our approach, which is a popular technology in the problem of MHC binding. Roughly speaking, there are nine amino acids in MHC binding cores, and each position is a specific pocket as showed in Table 2. We use PSSM to quantify the binding affinity between twenty basic amino acids with these nine pockets.

There are five anchor sites (1, 4, 6, 7, and 9) at the binding core for MHC II molecules, which determine the binding strength of peptides with MHC II molecules. Because site 1 of MHC II is consistent with different MHC II molecules and peptides, it is important to identify the precise quantification of its binding core in site 1, yet we use weights of four anchor sites (4, 6, 7, and 9) to define profiles. For other sites, the same approach, such as TEPITOPE, is to specify their quantitative profiles.

2.4. Computing Similarity between Different MHC Molecules
2.4.1. Sequence-Based Similarity

Sequence-based similarity can be calculated by alignment results. Here, pocket pseudosequences and associated profiles refer to raw pocket pseudosequences and raw pocket profiles, respectively. These raw pseudosequences are composed of several amino acids, whose associated residue indices are shown in Table 3. Eleven representative HLA-DR alleles are adopted to specify different profiles for anchor pockets 4, 6, 7, and 9. These eleven alleles are , , , , , , , , , , and . If two alleles have identical pseudosequences in the same pocket, they will have identical profiles. For a given pocket, we collect all the different raw pocket pseudosequences into one set , , and , where , , is the number of unique pseudosequences, and is the number of amino acids contained in a pseudosequence. Meanwhile, we collect all different raw profiles into one set , , and , where . There is a one-to-one correspondence between and . We use BLOSUM to calculate the sequence similarity between different MHC molecules, defined as . Then, we can get encoded pseudosequence, which is a 20-dimensional real vector . We use Radial Basis Function (RBF) to measure the similarity between encoded predicted pseudosequences and a raw encoded pseudosequence:

2.4.2. Structure-Based Similarity

Using MHC II HLA-peptide complex structure from Protein Data Bank (PDB), we can get the residues 3D-coordinate of the pocket in each MHC molecule, . We define vector , where is the number of amino acids in the pseudocontained sequence; meanwhile, we collect a set , , is the number of different pseudosequences, and there is also one-to-one correspondence between and .

Next, we need to estimate the similarity of three-dimensional structures between a measured MHC molecule and five MHC molecules with known pseudosequence PSSM. Rigid transformation is to compare three-dimensional substructures of two proteins [25, 26].

Intuitively, we fix one of the structures, A, move (translation and rotation) the other structure, B, and find the best movement in three-dimensional space, with two atoms to the nearest structure. We calculate the Euclidean distance between two structures, defined as . We can get encoded pseudosequence and calculate the similarity between 3D structures of encoded predicted pseudosequences and a raw encoded pseudosequence:

2.4.3. Overall Similarity

After that, we have obtained sequence similarity and structural similarity. We calculate final similarity score functions according to the following three formulas:

2.5. Weights Calculation for New Pocket Profiles

We reorder all pseudosequences according to descending similarity values and use a weight calculation formula to calculate new pocket profiles. A new pocket profile is generated as a weighted average over raw pocket profiles in . Next, we use the gamma distribution to generate the weights. The gamma PDF distribution is defined as follows: where and , , and denotes the gamma function.

The weight distribution is generated to discretize the gamma PDF as follows:where is the dimension of the weights and and are the shape and scale parameters, respectively. The gamma distribution generates the weight vector to give a higher weight for more similarity pseudosequences.

After normalizing, the weight vector is defined as follows:

Given a predicted DRB allele , let , where , , and α is a positive number and enhances the weight vector to protect the outstanding contribution of most similarity pseudosequences. Associated raw pocket profiles are . Elements of are sorted in descending order, and the reordered vector of is denoted as . The corresponding weight vector is denoted as . We denote pocket profiles associated with the reordered vector as , . We define the pocket profile for allele as follows:where .

3. Result

First, we design an experiment to choose appropriate scoring function to combine sequence similarity and structural similarity. Then, we compare with other state-of-the-art technologies, which are TEPITOPE, MultiRTA, NetMHCIIpan-2.0, and NetMHCIIpan-1.0. The result indicates that our approach can obtain better prediction and effectively extend current prediction methods. Finally, we test on more data sets.

3.1. Evaluation of Different Scoring Functions

Here, we use 30 of 39 MHC molecules and peptide complexes as test set and get the appropriate scoring functions as showed above. The value of the parameter α is set to 1, 2, 3, 4, 5, 10, 15, and 20, followed by results shown in Figure 2. We find that no significant changes can be found by ; for and , when prediction error number is 10 and 9 and when prediction errors reduced to 8, we set the value of α to 3. Comparing these three functions, the least numbers of errors by three functions are 4, 8, and 8. Details are shown in Tables S1, S2, and S3, in the Supplementary Material available online at http://dx.doi.org/10.1155/2016/3832176.

3.2. Compared with Conventional Well-Known Methods

From the above experimental results, obtains the most accurate prediction, so we will select with α = 3 as our final approach. We compare our current prediction results with conventional well-known methods TEPITOPE [23], MultiRTA [13], NetMHCIIpan-2.0 [12], and NetMHCIIpan-1.0 [12], and these results are shown in Table 4.

TEPITOPE is a relatively early method and is one of the most popular methods for predicting MHC II binding molecules. The basic idea is that if two HLA-DR alleles have the same pseudorandom sequence in the same pocket, they share the same number of profiles. Through multiple instances, MHCIIMulti has predicted over 500 HLA-DR molecules. NetMHCIIpan firstly converts each of the DRB alleles into a pseudorandom sequence of 21 amino acids, then uses the SMM-align method to identify binding residues in the peptide chain and the core side, and finally uses artificial neural network to train the model. MultiRTA makes prediction on HLA-DR and HLA-DP molecules. By thermodynamic method, it calculates a peptide chain and all other residues to predict the average binding affinity of binding strength and the introduction of standardization constraints to avoid overfitting. MULTIPRED2 can predict 1077 HLA-I and HLA-II genes and 26 HLA supertypes. Details are as shown in Figure 3. Our method obtains 4 errors; however, TEPITOPE, MultiRTA, NetMHCIIpan-2.0, and NetMHCIIpan-1.0 get the numbers of errors as 0, 4, 6, and 3, respectively. Because now we only find five MHC II molecules with three-dimensional structural information, we use the scoring matrix with only 5 MHC II molecules. If the three-dimensional structural information of MHC II molecules can be extended to all of the 11 MHC II molecules, our predictions will be more accurate. From the current view, our approach has reached a higher level of prediction.

3.3. Other Prediction Results

When compared with other methods on the above experiments, we only use 30 of 39 MHC molecules and peptide complexes as test set. In this section, we test on the remaining nine MHC molecules. In this experiment, we choose and set the parameter α = 3. As seen in Table 5, eight of nine predictions are accurate. Therefore, our approach produces a considerably great performance.

4. Conclusion

In this paper, we try to solve the problem of predicting MHC II binding peptides with a novel metric and strategy. Sequence similarity and structural similarity between different MHC molecules are calculated to reorder pseudosequences according to descending similarity, and then a weight calculation formula is used to calculate new pocket profiles. Finally, we use three scoring functions to predict binding cores and evaluate the accuracy of prediction to judge performance of each scoring function. In the experiment, we set a parameter in the weight formula. By changing value, we can observe different performances of each scoring function. Then, we compare our method with the best function to some popular prediction methods and ultimately find that our method outperforms them in identifying binding cores of HLA-DR molecules.

Competing Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgments

This work is supported by a grant from the National Science Foundation of China (NSFC 61402326) and Peiyang Scholar Program of Tianjin University (no. 2016XRG-0009).

Supplementary Materials

Using different functions to combine sequence similarity and structural similarity, these are the predicted results with the value of alpha ranging from 1 to 5.

  1. Supplementary Material