iSNO-PseAAC: Predict Cysteine S-Nitrosylation Sites in Proteins by Incorporating Position Specific Amino Acid Propensity into Pseudo Amino Acid Composition

1The post-translational modifications (PTMs) play a key role in providing proteins with structural and functional diversity, as well as in regulating cellular plasticity and dynamics. As illustrated in Fig. 1, the PTMs are covalent processing events that change the properties of a protein by proteolytic cleavage for adding a modifying group to one or more amino acids. One of the most important and universal PTMs is S-nitrosylation (SNO). Recent reports have indicated that SNO can modulate protein stability and activities,, as well as play an important role in a variety of biological processes, including cell signaling, transcriptional regulation, apoptosis, and chromatin remodeling.

2Meanwhile, increasing evidences have indicated that SNO also plays an important role in various major diseases, such as cancer, Parkinson's,, Alzheimer's, and Amyotrophic Lateral Sclerosis (ALS).

3Therefore, identifying the SNO sites in proteins is very important to both basic science and drug development.

4Many experimental methods have been developed for identifying SNO sites, such as BST (biotin switch assay), SNOSID,, and SNO-RAC. These methods have indeed provided very useful information in this area. Unfortunately, as pointed out by Seth and Stamler, experimental identification of SNO sites with a site-directed mutagenesis strategy is laborious and low-throughput due to the labile nature and the low-abundance of SNO. Particularly, with the avalanche of protein sequences generated in the postgenomic age, it is highly desired to develop computational method for timely and reliably identifying the SNO sites in proteins.

5Actually, some computational methods have been proposed in this regard. For instance, based on a benchmark dataset consisting of 65 positive and 65 negative samples, Gross and co-workers developed a computational method called SNOSID for identifying the SNO sites in proteins. A few years later, based on 549 experimentally verified SNO sites in 363 proteins, Xue et al proposed a different method called GRS-SNO for the same purpose. Shortly afterwards, Li et al. tried to improve the prediction performance by introducing the SVM (support vector machine) algorithm. Recently, Li et al. proposed a predictor by means of the nearest neighbor algorithm (NNA) with the maximum relevance minimum redundancy (mRMR) approach. Each of the aforementioned methods has its own merit and did play a role in stimulating the development of this area although bearing various limits. For example, no web-server has been provided for the most recent method, and hence its usage is quite limited, particularly for the majority of experimental scientists.

6The present study was initiated in an attempt to develop a new and more powerful method to identify the SNO sites in proteins in hopes that it may become a useful tool for both basic research and drug development in the relevant areas.

7As summarized in and demonstrated in a series of recent publications (see, e.g.,,,, ), to establish a really useful statistical predictor for a protein or DNA system based on the sequence information, we usually need to consider the following procedures: (i) construct or select a valid benchmark dataset to train and test the predictor; (ii) formulate the protein or DNA sequence samples with a feature vector that can truly reflect the intrinsic correlation with the target to be predicted; (iii) introduce or develop a powerful algorithm (or engine) to operate the prediction; (iv) properly perform cross-validation tests to objectively evaluate the anticipated prediction accuracy; (v) establish a user-friendly web-server for the predictor that is accessible to the public. Below, let us describe how to deal with these procedures one by one.

8The benchmark dataset used in this study was derived from the dbSNO (http://dbsno.mbc.nctu.edu.tw/), a database that integrates the experimentally verified cysteine SNO sites in 1,757 proteins from different species. To reduce the redundancy and avoid homology bias, we randomly picked 438 proteins in which none had pairwise sequence identity to any other. Based on these proteins and the annotations in the dbSNO database, a total of 731 experimentally verified SNO sites were collected. Meanwhile, to construct a corresponding negative dataset, a total of 810 experimentally verified non-SNO sites were randomly collected from the 438 proteins as well. The corresponding peptide fragments for the 731 SNO sites and 810 non-SNO sites were derived from UniProt database (release 2012_08), as can be generally formulated bywhere the subscript is an integer, is the downstream amino acid residue from cysteine (C), the upstream amino acid residue, and so forth. Hereafter let us call a peptide as SNO or non-SNO peptide if its center is a SNO or non-SNO site, respectively. In the current study, we choose. If the upstream or downstream in a protein was less than 10, the lacking residues were filled with the dummy code X. Thus, the benchmark dataset can be formulated aswhere the positive dataset contains SNO peptide fragments, while the negative dataset contains non-SNO peptide fragments (cf. Eq.1 ), respectively. For reader's convenience, their sequences as well as the corresponding sites and protein codes are given in Supporting Information S1.

9To develop a sequence-based predictor for identifying the attribute of a protein or peptide, one of the keys is to formulate its sequence with an effective mathematical expression that can truly reflect the intrinsic correlation with the attribute to be predicted. The most straightforward method to formulate the sample of a protein or peptide is to use its entire amino acid sequence. To identify its attribute, the tools for computing amino acid sequence similarity, such as BLAST,, were utilized to search the database for those targets that have high sequence similarity to the query protein or peptide. Subsequently, the attribute annotations of the target proteins or peptides thus found were used to infer the attribute for the query protein or peptide. Unfortunately, this kind of straightforward sequential model, although containing the entire sequence information, failed to work when the query protein or peptide did not have any significant sequence similarity to the attribute-known proteins or peptides.

10To avoid the above difficulty, which is inherent to the sequential model, various non-sequential or discrete models to formulate protein or peptide samples were proposed in hopes to enhance the prediction power.

11Among the discrete models, the simplest one is the amino acid (AA) composition or AAC. However, if using AAC to represent a peptide sample, its sequence-order or position-specific information would be totally lost, and hence might considerably limit the prediction quality.

12To avoid completely losing the sequence-order information, the pseudo amino acid composition (PseAAC) was proposed to represent the sample of a protein or peptide,. The idea of PseAAC has been widely used in bioinformatics, proteomics, and system biology, such as predicting protein structural class, predicting metalloproteinase family, predicting protein subcellular localization, predicting DNA-binding proteins, identifying allergenic proteins, identify recombination spots, identifying bacterial virulent proteins, predicting protein folding rate, predicting GABA(A) receptor proteins, predicting protein supersecondary structure, predicting cyclin proteins, classifying amino acids, predicting enzyme family class, identifying risk type of human papillomaviruses, identifying protein quaternary structural attributes, identifying GPCRs and their types, and discriminating outer membrane proteins, among many others (see a long list of references cited in ). Because of its wide and increasing usage, in 2012 a powerful software called “PseAAC-Builder” (http://www.pseb.sf.net) was established for generating various special modes of PseAAC for protein or peptide sequences.

13According to a recent review, the general form of PseAAC for a protein or peptide is formulated bywhere the subscript is an integer, and its value as well as the components,, … will depend on how to extract the desired information from the amino acid sequence of (cf. Eq.1 ). Below, let us describe how to extract useful information from the benchmark dataset to define the peptide samples concerned via Eq.3.

14It is obvious from Eq.1 that when, the corresponding peptide contains amino acid residues. Since the residue at the center of the sequence is always C, we can omit it. Thus, for the convenience of formulation, Eq.1 can be reduced toAlso, as mentioned above, besides the 20 native amino acids, the sequence may also contain a dummy amino acid X. Here, let us use the numerical codes 1, 2, 3, …, 20 to represent the 20 native amino acids according to the alphabetic order of their single letter codes, and use 21 to represent the dummy amino acid X.

15Thus, we can introduce the following matrix, the so-called “Position Specific Amino Acid Propensity” (PSAAP) matrix, to define the components of Eq.3 where the elementwhere is the occurrence frequency of the amino acid ( = 1,2,21) in the column in the positive benchmark dataset that can be easily derived using the method described in from the sequences in the Supporting Information S1, while is the corresponding occurrence frequency but derived from the negative benchmark dataset.

16Thus, the components in Eq.3 can be uniquely defined bySince the components of the feature vector in Eq.3 are now derived from the benchmark dataset, its correlation with SNO sites and non-SNO sites are self-evident.

17In this study, the “Conditional Random Field” (CRF) algorithm was adopted to operate the prediction. It is a discriminative probabilistic model that inherits the advantages of “Maximum Entropy Markov Models” (MEMMs), often used for labeling and segmenting sequence data. The CRF operation engine has been quite successfully utilized in various areas of bioinformatics and computational proteomics, such as gene prediction, SNP array analysis, and protein structure.

18In this study, the CRF software was downloaded from the web-site at http://www.di.ens.fr/~mschmidt/Software/crfChain.html. When used in the current study, the input of CRF is the query peptide fragment as formulated by the feature vector of Eq.3 as well as Eqs. 5–7, and the output is, thus the query peptide is identified aswhere is a threshold obtained by optimizing the overall success rate for the peptides in the benchmark dataset as done in.

19The predictor thus established via the above procedures is called iSNO-PseAAC, which can be used to identify the nitrosylated proteins and their SNO sites. To provide an intuitive picture, a flowchart is provided in Fig. 2 to illustrate the prediction process of iSNO-PseAAC.

20One of the important procedures in developing a useful statistical predictor is to objectively evaluate its performance or anticipated success rate. To provide a more intuitive and easier-to-understand method to measure the prediction quality, here the criteria proposed in was adopted. According to those criteria, the rates of correct predictions for the SNO peptides in dataset and the non-SNO peptides in dataset are respectively defined bywhere is the total number of the SNO peptides investigated while the number of the SNO peptides incorrectly predicted as the non-SNO peptides; the total number of the non-SNO peptides investigated while the number of the non-SNO peptides incorrectly predicted as the SNO peptides. The overall success prediction rate is given by It is obvious from Eqs. 9 – 10 that, if and only if none of the SNO peptides and the non-SNO peptides are mispredicted, i.e., and, we have the overall success rate. Otherwise, the overall success rate would be smaller than 1.

21On the other hand, it is instructive to point out that the following equation is often used in literatures for examining the performance quality of a predictorwhere TP represents the true positive; TN, the true negative; FP, the false positive; FN, the false negative; Sn, the sensitivity; Sp, the specificity; Acc, the accuracy; MCC, the Mathew's correlation coefficient.

22The relations between the symbols in Eq.10 and those in Eq.11 are given bySubstituting Eq.12 into Eq.11 and also considering Eq.10, we obtainFrom the above equation, we can see: when meaning none of the SNO peptides was mispredicted to be a non-SNO peptide, we have the sensitivity; while meaning that all the SNO peptides were mispredicted to be the non-SNO peptides, we have the sensitivity. Likewise, when meaning none of the non-SNO peptides was mispredicted, we have the specificity; while meaning all the non-SNO peptides were incorrectly predicted as the SNO peptides, we have the specificity. When meaning that none of SNO peptides in the dataset and none of the non-SNO peptides in was incorrectly predicted, we have the overall accuracy; while and meaning that all the SNO peptides in the dataset and all the non-SNO peptides in were mispredicted, we have the overall accuracy. The MCC correlation coefficient is usually used for measuring the quality of binary (two-class) classifications. When meaning that none of the SNO peptides in the dataset and none of non-SNO peptides in was mispredicted, we have; when and we have meaning no better than random prediction; when and we have meaning total disagreement between prediction and observation. As we can see from the above discussion, it is much more intuitive and easier-to-understand when using Eq.13 to examine a predictor for its sensitivity, specificity, overall accuracy, and Mathew's correlation coefficient.

23In statistical prediction, the following three cross-validation methods are often used to examine a predictor for its effectiveness in practical application: independent dataset test, subsampling (K-fold cross-validation) test, and jackknife test. However, as elaborated in and demonstrated by Eqs.28–32 of, among the three cross-validation methods, the jackknife test is deemed the least arbitrary and most objective because it can always yield a unique result for a given benchmark dataset, and hence has been increasingly used and widely recognized by investigators to examine the accuracy of various predictor (see, e.g.,,,,,,, ). However, to reduce computational time, here let us adopt the 10-fold cross-validation to examine the prediction quality as done by many investigators for PTM sites prediction,,,. The cross-validations were performed 50 times for different subsampling combinations, followed by averaging their outcomes.

24The results thus obtained on the benchmark dataset for the four metrics as defined in Eq.13 are given in Table 1, where for facilitating comparison the corresponding results obtained by GPS-SNO are also given. As can be seen from the table, the overall success, sensitivity and MCC rates achieved by iSNO-PseAAC are all significantly higher than those by the GPS-SNO predictor regardless its threshold was set at “high”, “medium”, or “low”. As for the method proposed in and the method recently proposed in, the former web-server was not working, while the latter had no web-server at all, and hence no corresponding data can be given in Table 1 for comparison.

25Listed in Supporting Information S2 are the predicted results by iSNO-PseAAC for a set of 461 independent nitrosylated proteins, none of which occurs in the 438 proteins used to train the current predictor. They were taken from Xue et al. and known belonging to nitrosylated proteins as verified by experiments. As we can see from Supporting Information S3, of the 461 proteins, 416 were predicted containing at least one SNO sites meaning belonging nitrosylated proteins. The overall success rate was

26For the convenience of the vast majority of experimental scientists, a web-server for iSNO-PseAAC was established. Below, let us give a step-by-step guide on how to use the web-server to get the desired results without the need to follow the mathematic equations that were presented just for the integrity in developing the predictor.

27Open the web server at at http://app.aporc.org/iSNO-PseAAC/ and you will see the top page of the predictor on your computer screen, as show in Fig. 3. Click on the Read Me button to see a brief introduction about iSNO-PseAAC predictor and the caveat when using it.

28Either type or copy/paste the query protein sequences into the input box shown at the center of Fig. 3. The input sequence should be in the FASTA format. A sequence in FASTA format consists of a single initial line beginning with a greater-than symbol (“>”) in the first column, followed by lines of sequence data. The words right after the “>” symbol in the single initial line are optional and only used for the purpose of identification and description. All lines should be no longer than 120 characters and usually do not exceed 80 characters. The sequence ends if another line starting with a “>” appears; this indicates the start of another sequence. Example sequences in FASTA format can be seen by clicking on the Example button right above the input box.

29Click on the Submit button to see the predicted result. For example, if you use the query protein sequences in the Example window as the input, after clicking the Submit button, you will see on your screen the predicted SNO site positions and the corresponding sequences segments as formulated by Eq.1. All these results are fully consistent with the experimentally verified results. It takes about a few seconds for the above computation before the predicted results appear on the computer screen; the more number of query proteins and longer of each sequence, the more time it is usually needed.

30Click on the Citation button to find the relevant papers that document the detailed development and algorithm of iSNO-PseAAC.

31Click on the Data button to download the benchmark datasets used to train and test the iSNO-PseAAC predictor.

32To obtain the predicted result with the expected success rate, the entire sequence of the query protein rather than its fragment should be used as an input. A sequence with less than 50 amino acid residues is generally deemed as a fragment. Also, the size of your input for each submission should be less than 100K; if greater than 100K, please contact Yan Xu at xuyan@ustb.edu.cn.