Biocell DOI:10.32604/biocell.2021.013770
Article

iPhosD-PseAAC: Identification of phosphoaspartate sites in proteins using statistical moments and PseAAC

Alaa Omran Almagrabi1, Yaser Daanial Khan2 and Sher Afzal Khan3,*

1Faculty of Computing and Information Technology, Department of Information Systems, King Abdulaziz University, Jeddah, 80200, Saudi Arabia
2Department of Computer Science, School of Systems and Technology, University of Management and Technology, Lahore, 54770, Pakistan
3Department of Computer Sciences, Abdul Wali Khan University Mardan, Mardan, 23200, Pakistan
*Address correspondence to: Sher Afzal Khan, sher.afzal@awkum.edu.pk
Received: 20 August 2020; Accepted: 18 February 2021

Abstract: Phosphoaspartate is one of the major components of eukaryotes and prokaryotic two-component signaling pathways, and it communicates the signal from the sensor of histidine kinase, through the response regulator, to the DNA alongside transcription features and initiates the transcription of correct response genes. Thus, the prediction of phosphoaspartate sites is critical, and its experimental identification can be expensive, time-consuming, and tedious. For this purpose, we propose iPhosD-PseAAC, a new computational model for predicting phosphoaspartate sites in a particular protein sequence using Chou’s 5-steps rues: (1) Benchmark dataset. (2) The feature extraction techniques such as pseudo amino acid composition (PseAAC), statistical moments, and position relative features. (3) For the classification, artificial neural network AAN will be used. (4) In this step, 10-fold cross-validation and self-consistency testing will be used for validation. For self-consistency testing, 100% Acc is achieved, whereas, for 10-fold cross-validation 95.14% Acc, 95.58% Sn, 94.70% Sp and 0.95 MCC are observed. (5). The final step is the development of a user-friendly web server for the ease of users. Thus, the iPhosD-PseAAC is the first and novel predictor for accurate and efficient identification of phosphoaspartate sites.

Keywords: Phosphorylation; Phosphoaspartate; Prediction; 5-step rule; Statistical moments; PseAAC

Introduction

Proteins are the basic and key part of the human body and perform many kinds of major functions in and outside of a cell. The proteins are translated or synthesized from messenger RNA, which is first codified into ribosomes and makes a chain of amino acids called a polypeptide chain. Later, the polypeptide passes through the process of folding and makes it an active protein. In the translation process, some of the amino acids can experience chemical changes at the C- or N- terminal of amino acid side chains, this process of alteration is called post-translation modifications (PTLM or PTM). The post-translation modification can modify or may introduce a new functional group to the protein, such as phosphate, so it plays a key role in the making of protein products (Xu et al., 2017).

Among all PTMs, phosphorylation is of great importance. It deactivates or activates the protein target and affects the speed at which a protein can be degraded. Also, it enables translocation of the protein from one subcellular compartment to another and helps protein binding (Mok and Snyder, 2009). Phosphorylation has exhibited pathological implications in diseases like Parkinson’s and Alzheimer’s alongside other neurodegenerative disorders.

Various eukaryotic proteins experience phosphorylation which causes modifications in localization, conformation, stability, function, and so forth (Hubbard and Cohen, 1993). It occurs on threonine (T), serine (S), histidine (H), tyrosine (Y), and aspartate (D) residuals in eukaryotes usually. However, as mentioned above, histidine (H) and aspartate (D) are unusual and least studied (Khan et al., 2018; Mann et al., 2002; Thomason and Kay, 2000).

The phosphorylated form of aspartate is known as phosphoaspartate (PhosD) and has a key role in multiple biological processes (Fig. 1). Phosphorylation of active-site aspartate residues also supports many enzyme-catalyzed reactions. It is observed that sometimes both dephosphorylation and phosphorylation of aspartate residues occur in proteins (Attwood et al., 2011; Knowles, 1980).

Phosphoaspartate is one of the major components of eukaryotic and prokaryotic two-component signaling pathways. The two-component signaling pathways, which have two parts, the histidine kinase and response regulator protein (RR), communicate from the sensor of histidine kinase to the DNA alongside transcription features and begin transcription of correct response genes (Capra and Laub, 2012; Falke et al., 1997).

These two signaling pathways are common in microorganisms and also in plants in order to regulate ripening and the circadian rhythm. Cytokinin and ethylene are the key components of plant-specific hormones, which normally regulate the entire life cycle of a plant from the germination of seeds to the development of flowers to set the new seeds. During evolution, plants follow the bacterial type of basic signal transduction strategy to regulate such important and mature biological mechanisms (Lohrmann and Harter, 2002; Mizuno, 2005).

Figure 1: Structural transformation of aspartate to phosphoaspartate.

Phosphoaspartate is involved in many important biological processes and plays an important role in both prokaryotic and eukaryotic organisms, especially in plants. Finding these positions is, therefore, a fundamental task. There are several experimental approaches to determine these positions, and of these approaches, high throughput mass spectrometry (MS) is one of the most common techniques (Akmal et al., 2017). However, the test results are time-consuming, tedious, and expensive.

To overcome these problems, various machine learning methods, algorithms, and techniques for predicting phosphorylation sites such as neural NN, SVM, and ANN have been proposed (Jiang et al., 2016; Li et al., 2016). However, no procedure has been proposed for PhosD sites.

Several studies have been proposed for other types of phosphorylation. Ingrell et al. (2007) developed NetPhosYeast with ANN. Huang et al. (2005) developed the KinasePhos web server using Hidden Markov Models (HMM) to predict specific phosphorylation sites of kinases. Lin et al. (2015) developed a server called Rice_Phospho to predict the phosphorylation sites of rice using SVM. The cluster-based Phosphorylation Scoring and Prediction (GPS) webserver was developed for the prediction of kinase-specific phosphorylation sites. He can find almost 70 types of phosphorylation, kinase-specific (Senawongse et al., 2005; Xue et al., 2008).

Senawongse et al. (2005) used an HMM-based model that focused on extracting features based on the proteomic primary structure for forming function-based clusters of proteins. The feature vector was formulated from both positive and negative samples (Cheng et al., 2018b). In 2018, Khan et al. (2018) proposed a strategy named iPhosT-PseAAC for an expectation of phosphothreonine destinations utilizing PSeAAC, measurable minutes, and different position relative highlights. ANN was utilized for classification while testing was performed by 10-overlay Cross-Validation and Jackknife testing.

Until this point in time, no predictor so far has been proposed for the identification of phosphoaspartate sites in proteins. This study proposes a novel prediction methodology iPhosD-PseAAC for the identification of phosphoaspartate sites in a given protein arrangement. The essential objectives for the proposed model are to outline a predictor that features the significance of phosphoaspartate and formulation of a methodology for in-silico examinations to their pertinent sequences. To address these objectives, we pursue Chou’s 5-step rule (Chou, 2011) as established in a progression of various studies (Cai et al., 2018; Chen et al., 2018; Cheng et al., 2018a; Cheng et al., 2018c, Cheng et al., 2018d; Cheng et al., 2018e; Chou et al., 2018; Liu et al., 2015; Liu et al., 2016a; Liu et al., 2016b; Xiao et al., 2017; Xuao et al., 2018). Formulation of the proposed predictor based on the 5-step rule brings a large dividend. It renders the model clarity of rationale, sets a benchmark for improvements, and makes it easily accessible to the wide-spread scientific community. The rules of Chou’s 5-step model are given as: (1) benchmark dataset development; (2) Transformation to equivalent mathematical form; (3) prediction algorithm; (4) Model Validation; and (5) development of a webserver. From here onwards, let us address these strategies one by one.

Materials and Methods

Benchmark dataset

In this study, we used Chou’s peptide formulation (Chou, 2001c) to facilitate the description of samples in the dataset. In computational biology, Chou’s peptide formulation has been widely used for the prediction of phosphothreonine sites (Khan et al., 2018), methylation sites (Qiu et al., 2014b), lysine ubiquitination sites (Qiu et al., 2015), signal peptide cleavage sites (Shen and Chou, 2007), hydroxyproline and hydroxylysine sites (Qiu et al., 2016a; Xu et al., 2014a), lysine succinylation sites (Jia et al., 2016b), phosphorylation sites (Qiu et al., 2016c), lysine PTM sites (Qiu et al., 2016b) and protein-protein binding sites (Jia et al., 2016a).

Uniprot is a huge database of proteins that contain annotated descriptions based on several experimental studies. The advanced query option of Uniprot was used to query PTM annotated protein sequences to form a benchmark dataset. The term 4-aspartyl phosphate was specified as modified residue to filter the protein having phosphoaspartate sites. Subsequently, to increase the reliability of data, only the proteins which have been reviewed and their post-translation modification identification is based on experimental assertion, were selected. As a result of this query, 1043 proteins were listed having 1052 phosphoaspartate sites. After removing duplicate sequences and reducing homology using CD-HIT, only 985 proteins were left. A negative dataset was generated using exactly the converse query, which yielded a multitude of samples. Negative samples were only collected from 1043 proteins which were identified to have positive sites as well. Even limiting the number of proteins yielded 7193 unique sequences having low homology. Out of these 7193 sites, only 1000 were randomly filtered out. For each occurrence of aspartic acid residue in a sequence, that particular residue and its associated upstream and downstream amino acids were extracted. Hence, for every selected instance, the length was 41 generated, comprising D residue, 20 residues upstream, and 20 residues downstream. For better results, both negative and positive datasets were pre-processed to remove duplicate occurrences. The dataset is constructed using only sequences that have been experimentally proven. Computationally annotated data and uncategorized data is left out as it may yield even more redundancies. In pre-processing all special characters, spaces, characters that are not amino acids (Z, X, U, O, J, and B) were removed. After pre-processing, the benchmark dataset contained 1994 samples (994 positive and 1000 negative samples).

By following Chou’s method (Chou, 2001b), the peptide having a potential PhosD can be expressed generally by

Pn(D)=R−nR−(n−1)R−(n−2)…R−2R−1DR+1R+2…R+(n−2)R+(n−1)R+n (1)

where D in Eq. (1), having double struck, is used to highlight the significance of amino acid D in the current study, and n is an integer value. R+n means the nth downstream residue of amino acid, and R−n means the nth upstream amino acid residue, and so forth. The further classification of (2n+1)−tuple size peptide is

Pn(D)∈{P+n(D),if the center of peptide isaPhosD siteP−n(D), otherwise (2)

In Eq. (2), P+n(D) represent a true PhosD sample in dataset, having D at its center, P−n(D) represent a false PhosD sample, and P represents peptide, and the ∈ symbol denotes a member of as described in the set theory.

In prediction models, the selected benchmark dataset normally contains both training data and testing data: the purpose of training and testing datasets are training the prediction model and testing of the prediction model, respectively. As extensively reviewed (Chou and Shen, 2008), it is noted that there is no need for these two separate datasets if one is validating the model using extensive validation methods, i.e., k-fold cross-validation or jackknife as its outcome is obtained by using a number of the different exclusive dataset for multiple tests. Moreover, unbiased results are proved by curating datasets that have the least homology. A dataset encompassing significant homologous samples will only yield biased results, and the predictor hence derived may not be as assiduous. Furthermore, during the initial exploration, there was found that the best value for n is 20; as a result, a sample consists of (2n+1)=41 residues (Eq. (1)). Correspondingly, the benchmark dataset minimized to TS total samples

TS=PS∪NS (3)

where PS denotes the 994 positive samples and NS denotes 1000 negative samples, while ∪ defines union (from set theory). Thus, total samples in the dataset TS are 1994, as 994∪1000 = 1994 (Supplementary Information S1 are available at https://www.biopred.org/iphosd/supl).

Sample formulation

Today, biological data and sequences are growing enormously. The most difficult task for us is to express the biological data and sequences into a vector or discrete form without dropping sequence pattern information and its characteristics for final analysis. As because the machine learning algorithms such as Covariance Discrimination or CD algorithm (Chou and Elrod, 2002; Lin et al., 2012), KNN (Cai and Chou, 2004; Chou and Cai, 2006), SVM (Feng et al., 2013a; Feng et al., 2013b) and RF- Algorithms (Jia et al., 2016b; Lin et al., 2011) can only process vectors (Chou, 2015). However, a vector characterized by a discrete model can lose all design data completely. To avoid completely losing protein grouping design data, the composition of pseudo amino acids (Chou, 2001a) or PseAAC (Chou, 2005) has been proposed. At this point, Chou’s PseAAC was already used in almost all regions of computer-aided proteomics (see, e.g., Akbar and Hayat, 2018; Arif et al., 2018; Contreras-Torres, 2018; Chou, 2017; Javed and Hayat, 2019; Ju and Wang, 2018; Krishnan, 2018; Liang and Zhang, 2018; Mei et al., 2018; Mei and Zhao, 2018a, 2018b; Qiu et al., 2018; Rahman et al., 2018; Sabooh et al., 2018; Sankari and Manimegalai, 2018; Srivastava et al., 2018; Zhang and Kong, 2018; Zhang and Duan, 2018; Zhang and Liang, 2018; Zhao et al., 2018). According to the general PseAAC (Chou, 2011), a sample protein sequence can be formulated as follows:

Pn=7(C)=[ℵ1ℵ2⋯ℵu⋯ℵΩ]T (4)

where ℵu=(u=1,2,3,⋯,Ω) and T for the transpose of features vector. The components in the above Eq. (4) will be defined by extracting the useful information from the corresponding peptide sequence. According to Eq. (1), the defined length of the peptide sequence in the benchmark dataset is 41; it can be modified as

P=R1,R2,R3⋯R19,R20,R21⋯R39,R40,R41 (5)

In Eq. (5) R21=D the targeted aspartate and Rj(j=1,2,3,4…,41;j≠21) can be any other amino acid or dummy code X as explained above. From now onwards, we use amino acid numerical codes as per their alphabetical order according to their first letter, 1, 2, 3 …… 20 for all 20 as amino acids, and dummy X will use 21 as code.

Statistical moments’ calculation

Here we use the statistical moments’ approach for the sequence to define the dimensions and its components of Eq. (4). Using different orders of moments, multiple kinds of data features are described; from those moments, few can be used to indicate the eccentricity and orientation of data, and others can be used for data size evaluation. Numerous moments were defined and described by statisticians and mathematicians primarily based on distribution functions and polynomials, actually well know (Khan et al., 2012; Khan et al., 2014a).

For the iPhosD-PseAAC prediction model, Hahn, raw, and central moments are computed. The Hahn moments are calculated, which are location and scale variant, using the Hahn polynomial (Khan et al., 2014c). Raw moments calculated, which can be a place and scale variant, are used for the mean, asymmetry, and variance calculation, using the probability distribution of the benchmark dataset. Besides, the central moment calculates the asymmetry, mean, and variance, but those are vicinity invariant because the calculation is done concerning the centroid (Butt et al., 2016; Butt et al., 2017), and these calculations are scale variant.

There is a particular reason for working with statistical moments; it maintains the sequence order sensitive information, which is an important point as described above. Furthermore, moments based on scale variants were not used. In the form of quantified value, data is defined on its own by every method (Khan et al., 2014b). In this current method, moments were used in 2-dimensional (2D) matrix P′ having k*k dimensions to accommodate all the residues from a protein P.

P′=⌊P11P12⋮Pk1P12…P22…⋮Pk2…P1kP2k⋮Pkk⌋ (6)

The function ω used for matrix transformation into P′ as described by Akmal et al. (2017). Using the elements of P′ all statistical moments were calculated up to 3rd degree. First of all, the (a+b)th order raw moments are calculated as

Mxy=∑a=1k∑b=1kaxby∂aba,b=1,2.…,k (7)

where the degree of moments is x+y and M00, M01, M02, M10, M11, M12, M20, M21, M30, and M03 are the calculated raw moments and ∂ab={1ifa=b0ifa≠b . The calculation of central moments is done as

Ixy=∑a=1k∑b=1k(a−r¯)x(a−s¯)y∂ab (8)

The P matrix transformed into 2-dimensional square matrix P′ which helps to calculate Hahn moments because Hahn moments can be easily calculated from square dimensional data. The distinct Hahn moments are the orthogonal moments of 2-dimension, which requires an even matrix in the input. The orthogonal Hahn moments have the feature that they can be reversed by using the inverse method of Hahn moments; it is feasible to reconstruct the material, and consequently, the data regarding relative positions, and the sequence composition, can also be conserved in those moments.

N order Hahn polynomial is calculated by using this equation

Hnu,z(r,M)=(M+Z−1)n(M−1)n×∑k=0n(−1)k(−n)k(−r)k(2M+u+z−n−1)k(M+z−1)k(M−1)k×1k!¯ (9)

The Eq. (9), Gamma operator and pochhammer symbol are explained in Akmal et al. (2017). The following equation used to calculate the orthogonal normalized Hahn moments

hxy=∑a=1M−1∑b=1M−1∂xyHxu,z∼(b,M)Hyu,z∼(a,M),n=0,1,2,3…M−1 (10)

Constructing PRIM and RPRIM

The model based on residue information of protein related to relative positions and the primary protein sequence is the central paradigm. It is necessary to quantize the relative positions of amino acids, thus the 20 × 20 matrix of Position Relative Incidence Matrix (PRIM) is constructed to extract the information, from all the instances of the benchmark dataset, about the relative position of amino acids residue of proteins as

DPRIM=[A1→1A1→2⋯A1→j⋯A1→1A2→1A2→2⋯A2→j⋯A2→20Ai→1⋮Ai→2⋮⋯Ai→j⋮⋯Ai→20⋮AN→1⋮AN→2⋮⋯AN→j⋮⋯AN→20⋮] (11)

In the DPRIM matrix, each Ai→j holds the sum of the relative position of the ith element in accordance with the first appearance of the jth element is represented. Similarly, the Reverse Position Relative Incidence Matrix (DRPRIM) is calculated with the reverse protein sequence sample. The DRPRIM is calculated as

DRPRIM=[D1→1D1→2⋯D1→j⋯D1→1D2→1D2→2⋯D2→j⋯D2→20Di→1⋮Di→2⋮⋯Di→j⋮⋯Di→20⋮DN→1⋮DN→2⋮⋯DN→j⋮⋯DN→20⋮] (12)

Both DPRIM and DRPRIM yield the 400 coefficients. To reduce the number of coefficients, the statistical moments are calculated for both, which yield the 30 coefficients.

Determination of Frequency Matrix (FM)

To find out the frequency of each amino acid residue and mine compositional information from the sequence in benchmark dataset instances, the frequency matrix is calculated. The frequency matrix is calculated as

FM={f1,f2,⋯,f20 (13)

Herein, f1 is the frequency of each amino acid residue occurring in the sequence arranged by their alphabetical order.

Constructing AAPIV and RAAPIV

To accumulate the positional information, the Accumulative Absolute Position Incidence Vector is calculated for the length of 20 native amino acids as

K={μ1,μ2,μ3,…,μ20} (14)

Here, μi is an arbitrary element of AAPIV that can be calculated as

μi=∑k=1npk (15)

Similarly, the Reverse Accumulative Absolute Position Incidence Vector (RAAPIV) is computed by using the reverse sequence of the protein as

K′={μ1,μ2,μ3,…,μ20} (16)

Prediction model

Billions of neurons in the human brain process and transmit information about a certain aspect when they are activated. Whenever you learn things through patterns without having a certain inclination, you take action based on the situation. The artificial neural network (ANN) inspired by the human processing system is made up of several highly interconnected neurons that work together to find a specific solution to a problem. It takes the information from the neuron and processes it using different patterns in the examples given. The ANN has two modes of working: training the ANN on the given input labeled data. In the second mode, when the input pattern matches the learned pattern, it becomes the current output, when the input pattern does not match any of the learned patterns it finds the closest and outputs according to that pattern (Khan et al., 2018), as shown in Fig. 2.

Figure 2: ANN for the proposed prediction model.

In this study, we also used AAN with backpropagation to reduce the output error. Using the benchmarks dataset, which contains 994 positive data samples and 1000 negative samples, the features vector was calculated for all data samples. Every feature vector contains Hahn, central, and raw moments of two-dimensional protein sample representation, PRIM, and RPRIM. For positional and compositional information, the calculated FM, AAPIV, RAAPIV, and Site Vicinity Vector also in the feature vector. The final feature vector contains 194 features based on positional and compositional information for each instance of the benchmark dataset. Thus, both an input matrix comprising all feature vectors and a label matrix were used to train an ANN (Jiang et al., 2016; Khan et al., 2018). Furthermore, extensive probing and experimentation showed that the ANN exhibited optimal results with 50 neurons in the hidden layer while adaptive gradient descent was used for learning.

Results and Discussion

Estimated accuracy

The important process for the new predicting model is, how to justifiably measure the success rate (Chou, 2011). We have to consider two issues to address the justifiable evaluation of the model: (1) to reflect the model quality, what kind of metrics should be used? (2) to score the metrics, what kind of test methods should be used?

Formulation of metrics

Generally, the following metrics are used from four different viewpoints to evaluate the prediction model accuracy: (1) MCC for model stability (2) Sp for model specificity (3) Sn for model sensitivity (4) Acc to measure the total accuracy of a prediction model. Thus, a set of four intuitive metrics were derived (Feng et al., 2013a; Xu et al., 2013) as given in Eq. (17).

For PhosD sites prediction, N+ is the true positive value, N−+ is the false negative value. Moreover, N− is the true negative value and N+− is the false positive value. Using Eq. (17), it can be seen that when N−+=0 , not a single PhosD site is predicted as the non-PhosD site so we have sensitivity Sn=1 . If N−+=N+ , it means all PhosD sites are incorrectly predicted as non-PhosD sites, so we have sensitivity Sn=0 . Moving forward, if N+−=0 , means not a single the non-PhosD sites are predicted as PhosD sites, so we have specificity Sp=1 ; whereas N+−=N− , means all the non-PhosD sites are incorrectly predicted as PhosD sites, so we have specificity Sp=0 . If N−+=N+−=0 , means not a single PhosD site in the positive dataset and non-PhosD site in the negative dataset incorrectly predicted, so we have MCC=1 and Acc =1 ; if N−+=N+ and N+−=N− means all PhosD sites in the positive dataset and non-PhosD sites in the negative dataset are incorrectly predicted, so we have MCC=−1 and Acc =0 . Whereas, if N+−=N−/2 and N−+=N+/2 then we have MCC=0 and Acc =0.5 , nothing but a guess. So, Eq. (17) explains overall-accuracy, sensitivity, specificity, and stability more easily to understand and intuitive, particularly about MCC (Chen et al., 2016b; Qiu et al., 2017b; Xiao et al., 2016).

{Sn=1−N−+N+0≤Sn≤1Sp=1−N+−N−0≤Sp≤1Acc=1−N−++N+−N++N−0≤Acc≤1MCC=1−(N−+N++N+−N−)(1+N+−−N−+N+)(1+N−+−N+−N−)−1≤MCC≤1 (17)

The set of intuitive metrics have been concurred and applauded by a series of recent publications see, e.g., Chen et al. (2016a); Chen et al. (2017b); Ehsan et al. (2018); Feng et al. (2013a); Feng et al. (2017); Feng et al. (2018); Jia et al. (2016c); Lin et al. (2014); Liu et al. (2017a); Liu et al. (2017b); Liu et al. (2018); Xu et al. (2014b); Zhang et al. (2016).

In Eq. (17), defined equations set are only effective for single-label data. For multi-label, which becomes more popular in biological (Chou et al., 2012; Lin et al., 2013; Xiao et al., 2011) and biomedicine (Xiao et al., 2013), is a completely different problem and need different metrics (Chou, 2013).

Self-consistency testing

A self-consistency test on the same dataset was performed to validate the iPhosD-PseAAC predictive model. The validation method was carried on the already known actual positive and negative sample dataset, and the result is shown in Tab. 1.

Validation of Model

Commonly, the experimentally proven datasets are used for model prediction; sometimes for testing, we do not have an experimentally proven dataset to test the model against the actual available data. By chance, if the data is available, it might be possible that data are not sufficient to test the accuracy of the predicting model. Only those samples could be incorporated into the model which exists in nature as for such a biological problem, it is not possible to build hypothetical datasets. To score the four metrics of Eq. (17), what kind of testing should be done to meet sufficient accuracy reliability? Usually, the dataset is split into 3 partitions. One partition is used for training while another is used for testing, and the leftover partition is used for validation. However, if the dataset is limited, then it is recommended that a predictor should be tested against k-folds (Subsampling), jackknife, and independent test (Chou and Zhang, 1995). Prediction model testing against jackknife is very exhausted and can always give a different outcome for a given benchmark dataset. It has been widely used to validate the prediction model by investigators (Chen et al., 2017a; Chen et al., 2017b; Dehzangi et al., 2015; Dou et al., 2014; Feng et al., 2005; Kumar et al., 2015; Mondal and Pai, 2014; Nanni et al., 2014; Qiu et al., 2014a; Shen et al., 2007; Wu et al., 2011; Zhou and Doctor, 2003). If an obvious dataset is not available to validate the model prediction, cross-validation is the best option to choose and to give the validation that the developed model is working fine.

Figure 3: 10-fold cross validation for benchmark dataset.

Herein, we performed 10-fold cross-validation and calculated accumulated accuracy by adding the accuracy of each fold. The average accuracy was 95.14%, as shown in Tab. 2 and Fig. 3. We also validate the prediction model using jackknife to verify the quality of iPhosD-PseAAC. For jackknife validation training, every instance of both the datasets is used for training and testing for unique output and received 94.46% of the prediction validation accuracy.

Comparative analysis

In a comparative analysis, the results of iPhosD-PseAAC for the metrics are compared with already existing PTM site prediction models, i.e., iPhosT-PseAAC (Khan et al., 2018) and PhosphoSVM (Dou et al., 2014). Both the models iPhosT-PseAAC and PhosphoSVM are merely used benchmarks for comparison of accuracy metrics. Since no earlier model for identification phosphorylation sites of aspartic acid has been found in texts. Considering these benchmark values, the metrics yielded by iPhosD-PseAAC has higher values than iPhosT-PseAAC and PhosphoSVM for all Acc, Sp, Sn, and MCC. This indicates better prediction as compared to others.

From Tab. 3, there can be seen that iPhosD-PseAAC outperforms other benchmark models. iPhosD-PseAAC uses different kinds of compositional and positional features to perform the prediction of PhosD sites. Firstly, it uses PseAAC and trims the modified residue by 20 downstream and upstream, then construct AAPIV, RAAPIV, PRIM, RPIRM, and moments are calculated, using the compositional and positional features. The feature extraction technique derived to furnish the ANN bears great significance in the vibrant performance of the predictor. Results showing high accuracy are a testament to its potential. High accuracy rate yielded in all validation tests compared with other state-of-the-art methods connote that the feature extraction technique is proficient in extracting obscure and pivotal traits of data peculiar to each class. Subsequently, the multilayer ANN is also well equipped to partition classes based on these intricate features.

Webserver

The 5th step is the user-friendly public web server, as documented in many recent publications (Cheng et al., 2017a, 2017b; Cheng et al., 2017c; Cheng et al., 2016; Feng et al., 2017; Liu et al., 2017b; Qiu et al., 2017a; Qiu et al., 2017b). The webservers are of great importance; thus, the webserver for iPhosD-PseAAC is available at https://www.biopred.org/iphosd, which is developed in Django framework with Python 3.6 and scikit-Learn (Fig. 4).

Figure 4: The graphical user interface of the iPhosD-PseAAC available at biopred.org/iphosd.

Conclusion

In this study, we have proposed a prediction model named iPhosD-PseAAC for phosphoaspartate site prediction using Chou’s 5-steps rule. Phosphoaspartate plays many fundamental roles in a number of biological processes, including signal transduction pathways, energy metabolism, various cellular processes, and ripening and circadian rhythms in plants. The aim of the current study was to propose a new and more accurate phosphoaspartate sites predictor and make it easy-to-use, user-friendly, and publicly available to experimental biologists to get their desired results. In PseAAC, it uses the various compositional and positional features of the protein sequence. The proposed model was validated against different exhaustive validation techniques, i.e., self-consistency, jackknife, and cross-validation. Using self-consistency, the accuracy is 100%, for cross-validation 95.14%, and jackknife gives 94.46% accuracy. The overall accuracy of the proposed model is 95.14%, sensitivity value 95.58%, and specificity 94.70%. It is concluded that the proposed model for the prediction of the PhosD sites has the ability of accurate and efficient predictions for phosphoaspartate sites in proteins, but it still can be improved in computational ways as the protein sequences may rapidly increase, day by day.

Acknowledgement: This project was funded by the Deanship of Scientific Research (DSR) at King Abdulaziz University, Jeddah, under Grant No. G:136-611-1441. The authors, therefore, acknowledge with thanks DSR for technical and financial support.

Availability of Data and Materials: The data that support the findings of this study are available in https://www.biopred.org/iphosd/supl.

Author Contribution: AO ALMAGRABI and YD Khan designed the model and the computational framework and analyzed, preprocessed, and extracted the data by SA Khan and YD Khan. The classification was carried out by AO ALMAGRABI and YD Khan. The evaluation of the model was carried out by AO ALMAGRABI and YD Khan under the overall supervision of SA Khan.

Ethics Approval: The work/experiments do not involve human subjects, animals, or plants.

Funding Statement: This project was funded by the Deanship of Scientific Research (DSR) at King Abdulaziz University, Jeddah, https://dsr.kau.edu.sa/Default-305-EN under Grant No. G:136-611-1441.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.

This work is licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.