ETMB-RBF: Discrimination of Metal-Binding Sites in Electron Transporters Based on RBF Networks with PSSM Profiles and Significant Amino Acid Pairs

Background Cellular respiration is the process by which cells obtain energy from glucose and is a very important biological process in living cell. As cells do cellular respiration, they need a pathway to store and transport electrons, the electron transport chain. The function of the electron transport chain is to produce a trans-membrane proton electrochemical gradient as a result of oxidation–reduction reactions. In these oxidation–reduction reactions in electron transport chains, metal ions play very important role as electron donor and acceptor. For example, Fe ions are in complex I and complex II, and Cu ions are in complex IV. Therefore, to identify metal-binding sites in electron transporters is an important issue in helping biologists better understand the workings of the electron transport chain. Methods We propose a method based on Position Specific Scoring Matrix (PSSM) profiles and significant amino acid pairs to identify metal-binding residues in electron transport proteins. Results We have selected a non-redundant set of 55 metal-binding electron transport proteins as our dataset. The proposed method can predict metal-binding sites in electron transport proteins with an average 10-fold cross-validation accuracy of 93.2% and 93.1% for metal-binding cysteine and histidine, respectively. Compared with the general metal-binding predictor from A. Passerini et al., the proposed method can improve over 9% of sensitivity, and 14% specificity on the independent dataset in identifying metal-binding cysteines. The proposed method can also improve almost 76% sensitivity with same specificity in metal-binding histidine, and MCC is also improved from 0.28 to 0.88. Conclusions We have developed a novel approach based on PSSM profiles and significant amino acid pairs for identifying metal-binding sites from electron transport proteins. The proposed approach achieved a significant improvement with independent test set of metal-binding electron transport proteins.


Introduction
Cellular respiration is the process by which cells obtain energy from glucose. During respiration, cells break down simple food molecules, such as sugar, and release the energy they contain [1]. The point of cellular respiration is to harvest electrons from organic compounds such as glucose and use that energy to make a molecule called ATP (adenosine triphosphate). ATP in turn is used to provide energy for most cellular reactions.
As cells do cellular respiration, they need a pathway to store and transport electrons, the electron transport chain. The function of the electron transport chain is to produce a trans-membrane proton electrochemical gradient as a result of oxidation-reduction reactions. If protons flow back through the membrane, ATP synthase converts this mechanical into chemical energy by producing ATP, which is provided energy in many cellular processes. The architecture of the electron transport chain with complex I-IV is given in Figure 1.
As Figure 1 shows, at the mitochondrial inner membrane, electrons from NADH and succinate pass through the electron transport chain to oxygen (Complex I(NADH dehydrogenase) and Complex II(succinate dehydrogenase)). Electrons pass from complex I to a carrier (coenzyme Q) embedded by itself in the membrane. From coenzyme Q electrons are passed to a Complex III (cytochrome b, c1 complex). Note that the path of electrons is from Complex I to Coenzyme Q to Complex III. Complex II, the succinate dehydrogenase complex, is a separate starting point, and is not a part of the NADH pathway. From Complex III the pathway is to cytochrome c then to a Complex IV (cytochrome oxidase complex). In the end, the proton electrochemical gradient allows ATP synthase to use the flow of H + to generate ATP.
There are many oxidation-reduction reactions in the electron transport chain, such as NADH dehydrogenase, coenzyme Qcytochrome c reductase, and succinate -coenzyme Q reductase. In these oxidation-reduction reactions in electron transport chains, metal ions play very important role as electron donor and acceptor. For example, Fe ions are in complex I and complex II, and Cu ions are in complex IV. Therefore, to identify metalbinding sites in electron transporters is an important issue in helping biologists better understand the workings of the electron transport chain. In this work, we try to develop a method based on Position Specific Scoring Matrix (PSSM) profiles and significant amino acid pairs to identify metal-binding residues in electron transport proteins.
In recent years, several methods have been proposed for predicting metal-binding sites (MBS) in proteins based on neural networks and support vector machines [2][3][4][5]. These work are major from A. Passerini and his co-workers except the work from Lin [2]. Prof. Passerini has proposed a two-stage machine-learning approach on their work [4]. The first stage consists of a support vector machine classifier, and the second stage consists of a bidirectional recurrent neural network. The authors of the work [4] have also published their web server as MetalDetector [5], which is the most popular web server for prediction metal-binding sites in proteins.
According to a recent comprehensive review 6], to establish a really useful statistical predictor for a protein system, we need to consider the following procedures: (i) construct or select a valid benchmark dataset to train and test the predictor; (ii)formulate the protein samples with an effective mathematical expression that can truly reflect their intrinsic correlation with the attribute to be predicted; (iii) introduce or develop a powerful engine to operate the prediction; (iv) properly perform cross-validation tests to objectively evaluate the anticipated accuracy of the predictor; (v) establish a user-friendly web-server for the predictor that is accessible to the public.
In this work, we propose a method based on PSSM profiles and significant amino acid pairs to identify metal-binding residues in electron transport proteins. We have selected a non-redundant set of 55 metal-binding electron transport proteins as our dataset. The proposed method can predict metal-binding sites in electron transport proteins with an average 10-fold cross-validation accuracy of 93.2% and 93.1% for metal-binding cysteine and histidine, respectively. Comparing with the general metal-binding predictor from A. Passerini et al., the proposed method can improve over 9% of sensitivity, and 14% specificity on the independent dataset in identifying metal-binding cysteines. The proposed method can also improve almost 76% sensitivity with same specificity in metal-binding histidine, and MCC is also improved from 0.28 to 0.88. The proposed approach achieved a significant improvement with independent test set of metalbinding electron transport proteins.

Materials and Methods
This work focuses on identifying metal-binding sites efficiently in electron transport proteins. As Figure 2 shows, the analyzing flowchart includes three sub-processes: data collection, feature set generation, and model evaluation. Following this model, we have developed a novel approach based on PSSM profiles and significant amino acid pairs for identifying metal-binding sites from electron transport proteins. The details of the proposed approach are described as follows.

Data collection
First of all, we selected electron transport proteins with metal binding sites from UniProt database [7]. Then, we removed the sequences without the evidence at protein level and experimental metal-binding sites. Next, by using BLAST [8], the sequences with sequence identity more than 20% were excluded from the dataset. Since sequences falling below a 20% sequence identity can have very different structure [9], it is difficult to get a high success rate when tested by dataset in excluding homologous sequences with 20% sequence identity. Finally, 55 electron transport proteins are surveyed in this work.
The collected electron transport protein sequences were divided into two datasets: the training dataset and the independent test dataset. The training dataset is used for identifying metal binding sites and evaluating significant amino acid pairs in electron transport proteins. The training dataset includes 44 electron transport protein sequences which contain 79 metal-binding cysteine, 77 metal-binding histidine and 368 non-metal-binding cysteine and histidine. The independent test dataset, which includes 11 electron transport proteins which contain 22 metalbinding cysteine, 21 metal-binding histidine and 103 non-metalbinding cysteine and histidine, is used to evaluate the performance of the proposed method. The details of two datasets are listed in Table 1 and Table 2. Table 3 summarizes the statistics of structural topology and molecular function on 55 electron transporters in this work.  Feature set generation Position Specific Scoring Matrix Profiles. In the structural point of view, several amino acid residues can be mutated without altering the structure of a protein, and it is possible that two proteins have similar structures with different amino acid compositions. Hence, the Position Specific Scoring Matrix (PSSM) profile is adopted, which have been widely used in protein secondary structure prediction, subcellular localization, classification of transporters, prediction of transport targets and other bioinformatics problems with significant improvement [10][11][12][13][14][15][16][17][18][19][20]. The PSSM profiles are obtained by using PSI-BLAST and nonredundant (NR) protein database.
PSSM profiles can be a useful feature set to represent evolutionary information in protein sequences [11,21]. Life on Earth originated and then evolved from a common ancestor approximately 3.7 billion years ago, sequences are more similar among species that share a more recent common ancestor, and can be used to reconstruct evolutionary histories. In this work, we searched a very large sequences database (NR database) by using PSI-BLAST to find similar sequences of the target sequence. Then, we adopted the evolutionary information contained in PSSM profiles as input to radial basis function networks.
In the identification of metal binding sites on electron transport proteins, the generated PSSM profiles contained the probability of occurrence of each type of amino acid residues at each position.
Each element in PSSM profile is scaled by 1 1ze {x for normalizing the values between 0 and 1. The window size of 13 residues which the central residue is metal-binding site and 6 residues along both sides of the central residue is used for encapsulating an amino acid residue. Finally, 13 X 20 elements are used as PSSM feature set for identifying metal-binding sites. Features of non-metal-binding sites are generated by using the same approach as features of metal-binding sites.
In addition, we also generated different feature sets for comparison. There feature sets are generated by amino acid types(AA), BLOSUM62 matrix (BLOcks of Amino Acid SUbstitution Matrix) [22], and PAM250 matrix [23]. A matrix of 13 X 20 elements is used to represent each residue in a training dataset, where 13 denotes the window size and 20 elements from each row of the type of amino acids, BLOSUM62 matrix and PAM250 matrix.
Significant amino acid pairs. The significant amino acid pairs (SAAPs) around the metal-binding sites are identified based on the training dataset. These SAAPs are adopted to construct learning model for improving performance [24]. In order to make further investigations of substrate sites specificity, these SAAPs are identified based on statistical measurement of hypergeometric distribution. Each amino acid pairs surrounding metal-binding site is calculated p-value of hypergeometric distribution. The hypergeometric distribution is defined as: where N denotes the number of sequences in the whole dataset, M denotes the number of sequences in the positive dataset, and (N-M) denotes the number of sequences in the negative dataset; n, x and n-x denotes the number of sequences which include the k-th SAAP in the whole dataset, in the positive dataset,and in the negative dataset respectively. The amino acid pair surrounding metal-binding sites is significant when p-value is less than the significance level. It indicates that central residue is the metal-binding site with higher probability if significant amino acid pairs appear. As shown in Table 4, the most significant amino acid pair on cysteine is (24C, 1P). (24C, 1P), which suggests that the cysteine(C) on position 24 and the proline(P) on position +1 surrounding metal-binding sites is significant with p-value 2:7|10 {4 . The illustration of calculating p-value for identifying significant amino acid pairs was shown in Figure 3.
After calculating p-value for each amino acid pair surrounding metal-binding sites, the ranked SAAPs added into the feature set by using forward feature selection based on 10-fold crossvalidation for improving predictive performance. Finally, 25 and 90 SAAPs are added into feature set of identifying metal binding cysteine and histidine, respectively. The final model was evaluated by using the independent dataset of 11 electron transporters.  The topmost 25 of SAAPs with p-value surrounding metalbinding cysteine and histidine are listed respectively in Table 4.

Model evaluation
Design of the Radial Basis Function Networks. We have employed the QuickRBF package [25] to construct RBFN classifiers in this work. Also, the fixed bandwidth of 5 for each kernel function is employed in the network. In addition, we used all training data as centers. Then, the RBFN classifier identifies metal-binding sites based on the output function value. The details about network structure and design have been explained in our earlier article [26].
Classification based on radial basis function (RBF) networks has several applications in bioinformatics. It has been widely used to predict the cleavage sites in proteins [27], inter-residue contacts [28], protein disorder [29], the discrimination of b-barrel proteins [13], the classification of transporters [14,16],the identification of O-linked glycosylation sites [24] , and the identification of ubiquitin conjugation sites [30].
The general mathematical form of the output nodes in an RBFN is as follows: g j (x) is the function corresponding to the j-th output node and is a linear combination of k radial basis functions w(m i ) with center m i and bandwidth s i ; Also, w ji denotes the weight associated with the correlation between the j-th output node. Assessment of predictive ability. The prediction performance was examined by 10-fold cross validation test, in which the training data were randomly divided into ten subsets of approximately equal size. The data were trained with nine subsets and the remaining set was used to test the performance of the method. This process was repeated 10 times so that every subset had been used as the test data once.
Sensitivity, specificity, accuracy, and MCC (Matthew's correlation coefficient) were used to measure the prediction performance. TP, FP, TN, FN are true positives, false positives, true negatives, and false negatives, respectively.

Results and Discussion
Predictive performance of metal-binding sites identification in electron transport proteins with different feature sets We developed a variety of methods for metal-binding sites identification in electron transport proteins. The results obtained from the AA, BLOSUM62, PAM250, PSSM, and the combination of PSSM and SAAPs are presented in Table 5. The results showed that PSSM with SAAPs properties was successful in identifying metal-binding sites with an average 10-fold crossvalidation accuracy of 93.2% and 93.1% for metal-binding cysteine and histidine, respectively. Our analysis showed that PSSM profiles and SAAPs properties had marginally improved the accuracy of identification, compared with the other feature sets.
Combining the significant amino acid pairs with the sequence of amino acids increases the predictive accuracy specificity for metalbinding sites identification from 90.1% to 93.2% with metal-  binding cysteine, and from 91.0% to 93.1% with metal-binding histidine. In addition, the sensitivity, precision specificity, and MCC are also improved. Consequently, according to the evaluation of 10-fold cross validation, the identified significant amino acid pairs can increase the predictive performance.
In statistical prediction, the following three cross-validation methods are often used to examine a prediction: independent dataset test, subsampling test, and jackknife test [6]. However, of the three test methods, the jackknife test is deemed the least arbitrary that can always yield a unique result for a given benchmark dataset. However, to reduce the computational time, we adopted the 10-fold cross validation and independent testing dataset test in this study.

Comparison the performance with other method with independent test set
The independent test dataset, which includes 11 electron transport proteins which contain 22 metal-binding cysteine, 21 metal-binding histidine and 103 non-metal-binding cysteine and histidine, is used to evaluate the performance of the proposed method. As Table 6 shows, comparing with the general metalbinding predictor from A. Passerini et al., the proposed method can improve over 9% of sensitivity, and 14% specificity on the independent dataset in identifying metal-binding cysteines. The proposed method can also improve almost 76% sensitivity with same specificity in metal-binding histidine, and MCC is also improved from 0.28 to 0.88. This results shows that our method  could be effectively used for indentifying metal-binding sites in electron transport proteins.

The statistical analysis of amino acid compositions in electron transporters and general proteins
We have analyzed metal-binding cysteine and cysteine residues on electron transporters and general proteins. Using the sequences of electron transporters in Table 3, we generated the sequence logos of metal-binding cysteine and cysteine residues in electron transporters with flanking amino acids (26 , +6) by WebLogo [31,32]. Also, we generated the sequence logos of metal-binding cysteine and cysteine residues in general proteins with the dataset in A. Passerini's work [4]. These four sequence logos are listed in Figure 4.
We also statistically analyzed the amino acid compositions with standard T-test of metal-binding cysteine and cysteine residues in electron transporters and general proteins. As Figure 5 shows, seven residues, Q, Y, K, M, I, D and G, surrounding metalbinding cysteine have significant difference between electron transporters and general proteins. Also, 8 residues, H, K, D, E, V, Q, F, and R, surrounding metal-binding histidine have significant difference between electron transporters and general proteins.
According the statistical analysis, the distribution of amino acids surrounding metal-binding residues are different between electron transport proteins and general proteins. This may be the reason why our proposed method performs better than the general metalbinding predictor.

Conclusions
Cellular respiration is the process by which cells obtain energy from glucose, and is a very important biological process in living cell. As cells do cellular respiration, they need a pathway to store and transport electrons, the electron transport chain. The function of the electron transport chain is to produce a trans-membrane proton electrochemical gradient as a result of oxidation-reduction reactions. In these oxidation-reduction reactions in electron transport chains, metal ions play very important role as electron donor and acceptor. Therefore, to identify metal-binding sites in electron transporters is an important issue in helping biologists better understand the workings of the electron transport chain.
In this work, we proposed a method based on PSSM profiles and significant amino acid pairs to identify metal-binding residues in electron transport proteins. We have selected a non-redundant set of 55 metal-binding electron transport proteins as our dataset. The proposed method can predict metal-binding sites in electron transport proteins with an average 10-fold cross-validation accuracy of 93.2% and 93.1% for metal-binding cysteine and histidine, respectively. Comparing with the general metal-binding predictor from A. Passerini et al., the proposed method can improve over 9% of sensitivity, and 14% specificity on the independent dataset in identifying metal-binding cysteines. The proposed method can also improve almost 76% sensitivity with same specificity in metal-binding histidine, and MCC is also improved from 0.28 to 0.88. Our proposed approach achieved a significant improvement with independent test set of metalbinding electron transport proteins. The result shows that our method could be effectively used for indentifying metal-binding sites in electron transport proteins to help biologists better understand the workings of the electron transport chain.
Since user-friendly and publicly accessible web-servers represent the future direction for developing practically more useful models, simulated methods, or predictors, we will make efforts in our future work to provide a web-server for the method presented in this paper.