predPhogly-Site: Predicting phosphoglycerylation sites by incorporating probabilistic sequence-coupling information into PseAAC and addressing data imbalance

Sabit Ahmed; Afrida Rahman; Md. Al Mehedi Hasan; Md Khaled Ben Islam; Julia Rahman; Shamim Ahmad

doi:10.1371/journal.pone.0249396

Abstract

Post-translational modification (PTM) involves covalent modification after the biosynthesis process and plays an essential role in the study of cell biology. Lysine phosphoglycerylation, a newly discovered reversible type of PTM that affects glycolytic enzyme activities, and is responsible for a wide variety of diseases, such as heart failure, arthritis, and degeneration of the nervous system. Our goal is to computationally characterize potential phosphoglycerylation sites to understand the functionality and causality more accurately. In this study, a novel computational tool, referred to as predPhogly-Site, has been developed to predict phosphoglycerylation sites in the protein. It has effectively utilized the probabilistic sequence-coupling information among the nearby amino acid residues of phosphoglycerylation sites along with a variable cost adjustment for the skewed training dataset to enhance the prediction characteristics. It has achieved around 99% accuracy with more than 0.96 MCC and 0.97 AUC in both 10-fold cross-validation and independent test. Even, the standard deviation in 10-fold cross-validation is almost negligible. This performance indicates that predPhogly-Site remarkably outperformed the existing prediction tools and can be used as a promising predictor, preferably with its web interface at http://103.99.176.239/predPhogly-Site.

Citation: Ahmed S, Rahman A, Hasan MAM, Islam MKB, Rahman J, Ahmad S (2021) predPhogly-Site: Predicting phosphoglycerylation sites by incorporating probabilistic sequence-coupling information into PseAAC and addressing data imbalance. PLoS ONE 16(4): e0249396. https://doi.org/10.1371/journal.pone.0249396

Editor: Ozlem Keskin, Koç University, TURKEY

Received: October 1, 2020; Accepted: March 18, 2021; Published: April 1, 2021

Copyright: © 2021 Ahmed et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All relevant data are within the paper and its Supporting information files.

Funding: The author(s) received no specific funding for this work.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Post-translational modifications (PTM) refer to specific events after the translation stage, where the covalent inclusion of specific functional groups occurs in a protein [1]. These modifications have enormous impacts on biological processes and proteomic analysis, such as cellular signal transduction, subcellular localization, protein folding, protein degradation, and are also responsible for various kinds of diseases [2]. Therefore, accurate identification and effective comprehension of PTM sites are significant for basic research in disease detection, prevention, and various drug developments [3]. Among the 20 standard constituent amino acid residues of cellular proteins, modifications at lysine residue (K) are commonly known as lysine PTM or K-PTM. According to the literature, several K-PTMs such as acetylation, crotonylation, ubiquitination, phosphoglycerylation, glycation, methylation, butyrylation, succinylation, biotinylation can be aided by these covalent modifications [4–8].

Lysine phosphoglycerylation is one of the reversible post-translational modifications, newly discovered in mouse liver and human cells [8, 9]. The formation of 3-phosphoglyceryl-lysine (pgK) takes place when primary glycolytic intermediate (1,3-BPG) interacts with particular lysine residues [8, 10]. A wide variety of diseases, including heart failure, arthritis, and various types of neurodegenerative disorders can be caused by this phosphoglycerylation. Metabolic labeling with substantial glucose indicates that it can be derived from glucose metabolism [9]. It has significant effects on glycolytic enzyme activities and can build up on cells with high glucose exposure [11]. Potential feedback mechanism that contributes to the creation and redirection of glycolytic intermediates to specific biosynthetic pathways is also established [8–11]. Concerning the crucial role of phosphoglycerylation in such biological processes, the effective way to characterize its functional aspects is to identify phosphoglycerylation sites with higher efficacy. Although high throughput experimental procedures to characterize phosphoglycerylation sites are known to achieve higher accuracy, computational methods are getting popularity as an effective alternative because of their laborsaving, time and cost-efficient characteristics.

Recent studies on identifying phosphoglycerylation sites have introduced several computational tools such as, Phogly-PseAAC [9], CKSAAP_PhoglySite [8], iPGK-PseAAC [12] and Bigram-PGK [11]. The first one has applied a KNN-based predictor with the pseudo amino acid feature source [9], where the second one has implemented a fuzzy SVM based predictor with the formation of k-spaced amino acid pairs feature set [8]. iPGK-PseAAC has utilized the pairwise coupling technique with an SVM classifier [11, 12]. The most recently developed predictor, Bigram-PGK has employed SVM with evolutionary information of the sequences for performance improvement [11]. Among these four predictors, only Bigram-PGK can predict phosphoglycerylation sites with an AUC higher than 0.90. However, the overall performance of this predictor needs further improvement in terms of other measurement metrics to be used as a complementary phosphoglycerylation site identification technique.

For constructing an efficient predictor, appropriate informative patterns connected with phosphoglycerylation need to be extracted. In this study, we are introducing a novel computational tool predPhogly-Site for predicting phosphoglycerylation sites by blending vectorized sequence coupling information with PseAAC [3, 13–16]. After generating necessary features from the protein sequences adopted from Bigram-PGK [11], a cost-sensitive SVM [14, 17–19] classifier has been used to predict phosphoglycerylation sites by minimizing class-level imbalance in benchmark dataset. The workflow of our proposed predictor is shown in Fig 1. For validating the statistical significance of the results, 10-fold cross-validation has been repeated ten times, and the average performances of each evaluation metric have been reported in the Results section. It can be observed that our proposed predictor, predPhogly-Site has achieved superior prediction performance than all the existing predictors. The attained performance of predPhogly-Site in terms of specificity, sensitivity, precision, accuracy, MCC, and AUC are 99.97%, 100%, 99.20%, 99.97%, 99.58%, and 99.99%, respectively. The promising results obtained by predPhogly-Site indicates that it can be used as a high-throughput supporting tool for phosphoglycerylation site prediction.

Download:

Fig 1. An overview of predPhogly-Site for phosphoglycerylation site prediction.

https://doi.org/10.1371/journal.pone.0249396.g001

Highlighted in a series of recently published predictors [3, 6, 14, 19–23], to develop an efficient predictor with regards to computational biology, one should go through Chou’s five-step [14, 24, 25] guidelines: i) generating an acceptable benchmark dataset for training and testing the system, ii) formulating the sequences using proper mathematical representations, iii) developing a prediction approach or introducing a robust prediction algorithm, iv) conducting rigorous cross-validation tests to evaluate predictive accuracy, and v) providing an accessible and easy-to-use web-server. Following these steps, details of materials, methods, results, and analysis will be discussed in the following sections.

Materials and methods

Dataset

In this study, verified annotations of phosphoglycerylation sites were obtained from the CPLM version 2.0 [26], one of the reliable repositories of post-translational modification in lysine residue, and corresponding protein sequences were retrieved from UniProt knowledge-base [27] for developing the prediction model. Subsequently, redundant sequences were discarded with 40% similarity cutoff using CD-HIT [28] for avoiding bias in performance evaluation as this level of redundancy removal was widely accepted [11, 24, 29, 30]. As a result, a total of 91 non-redundant proteins were held out for constructing a benchmark dataset. There were 111 experimentally annotated phosphoglycerylated sites and 3249 non-phosphoglycerylated sites, which was identical to the most recent predictor, Bigram-PGK’s [11] dataset (see Table 1). The benchmark dataset containing protein sequences and site positions are given in S1 File. An overview of the dataset preparation as part of the prediction model development is presented in Fig 1. For verifying the statistically significant difference among the positive and negative sites in the obtained dataset, the distribution of amino acid residues in the phosphoglycerylated sites and non-phosphoglycerylated sites are visually analyzed with the help of WebLogo [31] (see Fig 2A and 2B).

Download:

Fig 2. Amino acid frequencies around the K-PTM and non-K-PTM sites.

https://doi.org/10.1371/journal.pone.0249396.g002

Download:

Table 1. Summary of the non-redundant phosphoglycerylation dataset.

https://doi.org/10.1371/journal.pone.0249396.t001

To demonstrate the viability of the proposed predictor predPhogly-Site for new proteins, an independent test set was constructed with recent phosphoglycerylation sites, utterly unknown to the benchmark dataset used for prediction model development. Protein sequences with recent phosphoglycerylation sites were collected from the PLMD database [32] (version 3.0), which is an upgraded version of the CPLM database [26], released nearly 03 years later with many newly discovered PTM sites. For ensuring the non-existence of training proteins in the independent test set, we considered only those proteins which were newly added to the PLMD repository much after the creation of the benchmark dataset with verified phosphoglycerylation sites. Therefore, we obtained 33 proteins with 41 phosphoglycerylated sites and 1334 non-phosphoglycerylated sites for the independent test (available as S2 File). Furthermore, the non-existence of recent test sites was verified manually for avoiding accidental bias in performance benchmarking.

Feature construction

To formulate the phosphoglycerylation site sequences more meticulously and comprehensively, Chou’s scheme [9, 13, 33] was adopted. According to this scheme, a potential phosphoglycerylation site containing sequence fragment could be expressed as: (1) Where Q₁ to Q_ζ denote the leftward and Q_ζ+1 to Q_2ζ+1 denote the rightward amino acid residues, respectively, while ζ being an integer and centered ‘K’ indicating “lysine” [14]. Furthermore, the peptide sequences Θ_ζ(K) can be categorized into two types: and , where the first one denotes phosphoglycerylated peptide and the later one denotes non-phosphoglycerylated peptide with a lysine residue at its center [9, 14]. The sliding window method [9] was adopted to segment the phosphoglycerylation protein sequences with different window size where ζ = 1, 2, 3, …32. Based on the MCC value, window size was selected as (2ζ + 1) = 29 where ζ = 14 (i.e. 14 rightstream and 14 leftstream amino acid residues). It should be mentioned that, only the window sizes less than 65 were taken under consideration due to the compelling protein sequence length [11]. With a sequence fragment of window size 29, Eq (1) could be expressed as: (2)

At the time of segmentation, for making site sequences’ of equal length, the lacking amino acids were filled with ‘X’ residue [9, 34]. As a result, the phosphoglycerylation dataset had taken the following form: (3) where the positive subset could contain only samples, while the negative subset could contain only samples with their center residue K. All the segmented sequences with the expression of Eqs (2) and (3) are provided in S1 File.

For extracting pertinent features hidden in amino acid sequences, different sequence encoding methods such as amino acid composition, pseudo amino acid composition were used initially. However, in the proposed predictor predPhogly-Site, the vectorized sequence-coupled model [3, 14–16, 35] has been incorporated into general PseAAC [3, 14, 33, 35–39] to extract features from the phosphoglycerylation sites conserving the sequence pattern information. According to this conception, the peptide sample in Eq (2) can be expressed as: (4) where, (5) where, Θ⁺(Q₁|Q₂) denotes the conditional probability of amino acid Q₁ at the leftmost position given that its adjacent right member is Q₂ and the same applies for remaining indices of leftward residues [24]. Similarly, Θ⁺(Q₂₈|Q₂₇) denotes the conditional probability of amino acid Q₂₈ at the rightmost position given that its adjacent left member is Q₂₇ and so forth. In contrast, only Θ⁺(Q₁₄) and Θ⁺(Q₁₅) are of non-conditional probability as K is the adjoining member of both amino acids Q₁₄ and Q₁₅ [3, 6, 14, 15, 24]. In order to calculate the probability values of Θ⁺(Q₁₄) and Θ⁺(Q₁₅), firstly, we have to find the frequency of a given amino acid Q₁₄ and Q₁₅ from the set of phosphoglycerylated peptides [15]. Then the obtained values should be divided by the frequency of all amino acids occurring at position 14 and 15 respectively. Accordingly, Θ⁻(K) in Eq (5), with its probabilistic components could also be deduced from the set of non-phosphoglycerylated peptides. A few literature on vectorized sequence-coupling model [3, 13, 15, 16] could provide a better understanding of the procedure of probability calculation out of any dataset. Finally, a 28-dimensional feature vector was obtained by using Eqs 4 and 5 for each potential phosphoglycerylated and non-phosphoglycerylated sample.

For better visualization and insights on the sequence-coupling effects at different positions of any sample, we have stored all possible combinations of conditional probability values extracted from the positive subset i.e. Θ⁺(Q₁|Q₂) to Θ⁺(Q₁₃|Q₁₄) and Θ⁺(Q₁₆|Q₁₅) to Θ⁺(Q₂₈|Q₂₇) in one data frame (available in S3 File) and non-conditional probability values for each amino acid residue extracted from the positive subset i.e. Θ⁺(Q₁₄) and Θ⁺(Q₁₅) in another data frame (available in S4 File) using Pandas library [40], where the columns represent the formulated sample positions and the rows represent the amino acid residues. It should be mentioned that there could be 21 × 21 = 441 (including the dummy amino acid residue ′X′) possible combinations of conditional probability values and 21 non-conditional probability values [15] for each position at any formulated sample. Similarly, the conditional and non-conditional probability values extracted from the negative subset are stored in two separate data frames and provided in S3 and S4 Files, respectively. Fig 3A depicts the conditional probability values of amino acid residue ′A′ which have been calculated from the positive subset, given that its right member is any of the 21 amino acid residues at sample positions 1 to 13 and the conditional probability values of any of the 21 amino acid residue given that the left member is ′A′ at sample positions 16 to 28. Similarly, Fig 3B depicts the conditional probability values of amino acid residue ′A′ which have been calculated from the negative subset, given that its right member is any of the 21 amino acid residues at sample positions 1 to 13 and the conditional probability values of any of the 21 amino acid residue given that the left member is ′A′ at sample positions 16 to 28. The non-conditional probability values of 21 amino acid residues derived from the positive subset at sample positions 14 and 15 are illustrated in Fig 4A and The non-conditional probability values of 21 amino acid residues derived from the negative subset at sample position 14 and 15 are shown in Fig 4B.

Download:

Fig 3. The conditional probability of amino acids at sample positions 1 to 13 and 15 to 28.

https://doi.org/10.1371/journal.pone.0249396.g003

Download:

Fig 4. Probabilistic information of 21 amino acids at sample positions 14 and 15.

https://doi.org/10.1371/journal.pone.0249396.g004

Prediction method and addressing data imbalance

Phosphoglycerylation site prediction problem defined in the previous section is a classification problem. Statistical learning algorithms such as k-nearest neighbor [41], random forest [42] which are widely used in different bioinformatic prediction model development, support vector machine (SVM) [43, 44] is one of the dominant and successful among these algorithms [24, 45]. Apart from that, the structural risk minimization involves a biasing problem where the majority class [24, 46] influences the classification weight. As the set of phosphoglycerylation peptides was highly skewed (i.e. the ratio between positive and negative peptides was approximately 1:29), it could affect the classification model training directly. Inspired by the success of biasing internal decision function during training, as highlighted in recent research [8, 14, 17, 19], different penalty costs C⁺, and C⁻ were assigned for phosphoglycerylated sites and non-phosphoglycerylated sites, respectively for addressing imbalance issue. Therefore, SVM with cost-sensitivity was applied as a core learning algorithm for prediction model development which can be formulated as: (6)

(Subject to: Y_k(w.φ(X_k) + a) ≥ 1 − ξ_k for all, k = 1, 2, ‥, n)

where the training set is denoted by {(X_k, Y_k), k = 1, 2, …, n} and first q samples (i.e. Y_k = 1, k = 1, 2, …, q) are assumed as the positive samples while the rest are assumed as the negative samples (i.e. Y_k = −1, k = q + 1, q + 2, …, n). The non-linear feature mapping and slack variables are denoted by φ(X) and ξ_k(k = 1, 2, …, n), respectively [45, 47]. In our experiments with SVM, as the kernel function, Gaussian RBF was adopted which can be described as: Υ(X_k, X_j) = φ(X_k)^T φ(X_j) = exp(−γ‖x_i − x_j‖²), where γ > 0. However, for effective separation of positive and negative samples, addressing the class imbalance problem, misclassification costs and were assigned for phosphoglycerylated sites and non-phosphoglycerylated sites, respectively.

Formulation of evaluation metrics

To objectively assess the prediction performance of predPhogly-Site, we have utilized five widely used statistical metrics, such as accuracy (ACC),sensitivity (Sn), specificity (Sp), precision (pre) and Matthew’s Correlation Coefficient (MCC) [20, 24, 30, 45, 47–52]. These matrices can be defined in terms of true positive (TP), false positive (FP), true negative (TN) and false negative (FN) prediction made by the predictor as following: (7)

To the best of our knowledge, state-of-the-art phosphoglycerylation site predictors [8, 9, 11, 12] have also estimated their performance based on these metrics. Thus, performance assessment using these metrics was essential to establish a fair comparative benchmarking. Eventually, we have considered the area under the ROC curve (AUC) [24, 53] in addition to MCC for illustrating the stability and robustness of the prediction model.

Validation of the proposed model

To evaluate the statistical significance of a novel predictor’s anticipated performance, three validation schemes, such as k-fold cross-validation, jackknife test, and independent test are widely used [14, 24]. Although the jackknife test can always draw out a unique result for a given dataset and highly desirable, to reduce the computational complexity of model development, researchers prefer k-fold cross-validation over the jackknife test for validating their PTM prediction models [8, 45]. Moreover, existing phosphoglycerylation site predictors validated their anticipated accuracy using k-fold cross validation except Phogly-PseAAC [9]. Even, the most recent predictor, Bigram-PGK [11] validated their model using 10-fold cross-validation and compared with existing predictors. Therefore, to develop and validate our proposed predictor predPhogly-Site, 10-fold cross-validation was adopted. However, as the 10-fold cross-validation involved some arbitrariness, highlighted in [9, 24], to validate the stability, it was repeatedly executed for 10 times. For finding the best performing predictor, a set of prediction models were generated for the hyperparameters C and γ within the grid of C = {2⁰, 2¹, 2², …, 2⁸} and γ = {2⁻¹, 2⁻², 2⁻³, …, 2⁻⁸}. Using 10-fold cross-validation with 10 repeats, the best model with optimal hyperparameters C and γ were selected (see Table 2) depending on the demonstrated AUC.

Download:

Table 2. Selected parameters of 10-fold cross validation (10 iterations).

https://doi.org/10.1371/journal.pone.0249396.t002

The 10-iterations of 10-fold cross-validation were performed according to the following steps:

Step 1: Extract the sequence-coupled features from the segmented sequences provided in S1 File using Eqs (4) and (5).
Step 2: Divide the extracted dataset randomly into 10 disjoint sets.
Step 3: Select 1 set as test set and utilize the remaining 9 sets as training set.
Step 4: Train the RBF kernel based SVM predictor with the training set using the optimal hyperparameters (C, γ) of the respective iteration (see Table 2).
Step 5: Perform prediction on the test set.
Step 6: Repeat steps 2 to 5 until all 10 sets had been used for testing.
Step 7: Merge the prediction outputs and measure the performance with Eq 7.
Step 8: Repeat steps 1 to 7 for 10 times.
Step 9: Measure the average performance of 10 repetitions with corresponding standard deviations.

The predictive decision-making workflow of predPhogly-Site is available at https://github.com/Sabit-Ahmed/predPhogly-Site as a git repository. For additional validation, an independent test was performed on a set of recent phosphoglycerylation sites. It will be discussed thoroughly in the next section.

Results and discussions

Performance of predPhogly-Site

In this work, we employed SVM with variable cost adjustments [14, 19, 24] for suppressing the imbalance between phosphoglycerylated and non-phosphoglycerylated sites. For separating samples by transforming to higher dimensional feature space, radial basis kernel function [14, 22, 24] was utilized. The average results of the considered statistical performance measures with their standard deviations in 10 repeats are presented in Table 3. As shown in Table 3, the proposed prediction model could predict phosphoglycerylation sites with 99.97% accuracy. In addition to that, its sensitivity, specificity, MCC and AUC measure crossed a benchmark of 99%. Moreover, standard deviations were almost negligible in the case of all the measures. However, for constructing the proposed predictor predPhogly-Site to be deployed as a web service, the benchmark dataset and the prediction model’s hyper-parameters with the highest AUC in 10 repetitions (i.e. C = 2⁰ and γ = 2⁻²) were used. An overview of establishing predPhogly-Site is depicted in Fig 1.

Download:

Table 3. Cross-validation performance of predPhogly-Site on the benchmark dataset.

https://doi.org/10.1371/journal.pone.0249396.t003

Comparative analysis of cross-validation performance

To evaluate the effectiveness of the proposed predictor, predPhogly-Site, we compared it with four state-of-the-art phosphoglycerylation site predictors, such as Phogly-PseAAC [9], CKSAAP_PhoglySite [8], iPGK-PseAAC [12] and Bigram-PGK [11]. Among these predictors, the first three i.e. Phogly-PseAAC, CKSAAP_PhoglySite, and iPGK-PseAAC were benchmarked on the same phosphoglycerylation site dataset which was prepared by Xu et al. [9]. Prediction from Phogly-PseAAC and iPGK-PseAAC could be accessed by their web interface. Though CKSAAP_PhoglySite was also accessible by its Matlab interface, there was no such accessibility option in the most recent predictor, Bigram-PGK. However, Bigram-PGK had collected prediction results from these accessible predictors for its benchmark dataset and reported comparative outcomes for all the considered performance metrics. Thus, for conducting a fair comparison with all these predictors, our primary benchmark dataset, which was not resampled as Bigram-PGK’s one, was submitted to the webserver of Phogly-PseAAC and iPGK-PseAAC for getting prediction outcomes. However, CKSAAP_PhoglySite’s predictions were obtained through its Matlab interface. After achieving the prediction outcomes from the Phogly-PseAAC, CKSAAP_PhoglySite, and iPGK-PseAAC on the benchmark dataset constructed for this study, the corresponding performance was measured on the same validation set utilized for evaluating our predictor predPhogly-Site (see Section “Validation of the proposed model”). As we adopted different technique for handling the data imbalance issue and could not obtain the prediction outcomes from the Bigram-PGK predictor on our benchmark dataset, a comparative summary of all the measures was presented in Table 4 in line with Bigram-PGK’s experimental findings [11]. As shown in Table 4 and Fig 5, predPhogly-Site achieved a significant improvement over Phogly-PseAAC, CKSAAP_PhoglySite, and iPGK-PseAAC on the same benchmark dataset used in this study. It remarkably outperformed these predictors in sensitivity, specificity, overall accuracy, and AUC. For instance, predPhogly-Site crossed the milestone of 99% in case of sensitivity, specificity, precision, overall accuracy, MCC and AUC.

Download:

Fig 5. Cross-validation performance of the available predictors.

https://doi.org/10.1371/journal.pone.0249396.g005

Download:

Table 4. Cross-validation performance of the existing prediction systems.

https://doi.org/10.1371/journal.pone.0249396.t004

However, the most recent predictor, Bigram-PGK’s [11] performance was relatively higher in most of the metrics. It obtained a sensitivity of 96.42%, an accuracy of 91.93%, an MCC of 83.30%, and an AUC of 93.06% on the dataset utilized in Bigram-PGK [11]. As demonstrated in Table 4, our proposed predictor predPhogly-Site also outperformed Bigram-PGK [11] by 3.58% in sensitivity, 8.04% in accuracy measure, 16.28% in MCC and 6.93% in AUC. Furthermore, the effectiveness of predPhogly-Site over the recent predictors including Bigram-PGK [11] has been demonstrated in Fig 5.

It can be observed that a comparatively higher specificity and precision of 98.64% and 95.48%, respectively, were obtained by iPGK-PseAAC [12] on the Bigram-PGK’s [11] resampled dataset. Our proposed predictor, predPhogly-Site, has obtained 1.33% and 3.72% increased performance in both specificity and precision, respectively. Both the results represented in Table 4 and Fig 5 indicate that our proposed predictor predPhogly-Site can identify phosphoglycerylation sites more effectively than any other existing predictors.

It is worth mentioning that among these predictors, Phogly-PseAAC [9] has employed the position-specific amino acid propensity which reflects the position-wise occurrence frequency of each amino acid and the K-Nearest Neighbor (KNN) algorithm for prediction, CKSAAP_PhoglySite [8] has utilized the composition of k-spaced amino acid pairs with the fuzzy SVM, iPGK-PseAAC [12] has applied the pairwise coupling technique with the posterior probability-based SVM and Bigram-PGK [11] have considered the SVM engine with the combination of position-specific scoring matrix and profile bigrams for performance improvement.

It might be intuitive to find some insight into why our proposed predictor predPhogly-Site achieved such superior performance. It was possible because of the effective representation of phosphoglycerylation modification in terms of sequence coupling model among the amino acid residues via the conditional probability (see Figs 3 and 4). Suppressing the imbalance ratio of phosphoglycerylated and non-phosphoglycerylated sites using different error costs based SVM also boosted up the performance improvement.

However, the precision calculation measures the believability of a system when it says a peptide sample is phosphoglycerylated. According to Eq 7, the precision measure depends highly on the false positive rate, and a lower false positive rate results in a higher precision rate. In the Bigram-PGK [11] study, the dataset contained only 111 positive samples and 224 negative samples after applying the k-nearest neighbor cleaning treatment [11] and the experimental findings on the resampled dataset might not reflect the false positive rate properly. Moreover, the existing predictors i.e. iPGK-PseAAC, CKSAAPPhoglySite, and Phogly-PseAAC might not handle the real world imbalanced situation of the dataset appropriately. Hence, when we have uploaded the benchmark dataset containing 111 positive instances and 3249 negative instances (see Table 1) to the web or Matlab interfaces of the existing predictors, the false positive rates have come out higher and results in lower precision rates as compared to the experimental findings reported by the Bigram-PGK study (see Table 4). On the other hand, our proposed predictor has obtained a much lower false positive rate and got a higher precision rate as well as higher sensitivity and specificity for having cost-sensitive SVM as an imbalance management technique. By observing all the performance measurements in this study, it can be concluded that our predictor predPhogly-Site could be a high throughput tool for predicting phosphoglycerylation sites more precisely.

Independent test

Existing phosphoglycerylation site, particularly, the most recent predictor assessed their model using 10-fold cross-validation. However, some researchers [54–57] highlighted the necessity of independent test for assessing prediction model in addition to k-fold (e.g. k = 5,10) cross-validation. Thus, in our work, an independent test was conducted for further evaluation of our proposed model predPhogly-Site on an independent set of phosphoglycerylation sites. The same independent test set was uploaded to the web servers of the existing predictors i.e. iPGK-PseAAC, Phogly-PseAAC and predPhogly-Site for obtaining the prediction results. However, the prediction results of CKSAAP_PhoglySite on the independent test set were obtained from the Matlab interface. The predictive performance of predPhogly-Site as well as other predictors were summarized in Table 5. However, as Bigram-PGK [11] had no established web-server, so we could not report the performance of these predictors on the independent test set.

Download:

Table 5. Prediction performance in Independent test.

https://doi.org/10.1371/journal.pone.0249396.t005

As shown in Table 5, predPhogly-Site predicted independent phosphoglycerylation sites with specificity, sensitivity, precision, accuracy, MCC and AUC of 99.93%, 95.12%, 97.50%, 99.78%, 96.19% and 97.52%, respectively, which were almost identical to the cross-validation performance delineated in Table 4. According to the experimental results in Table 5 and the ROC curve illustrated in Fig 6, it was apparent that the proposed predictor predPhogly-Site achieved a significant improvement over their counterparts in terms of all the evaluation metrics.

Download:

Fig 6. Comparative ROC curves between different prediction methods based on the independent test.

https://doi.org/10.1371/journal.pone.0249396.g006

Web-server

For intensifying user accessibility without the concern of experimental implementations, an easy-to-use web-server for predPhogly-Site has been developed. It can be accessed at http://103.99.176.239/predPhogly-Site. Users can submit one or more query protein sequence(s) directly on the web-server as text input in Fasta format or may prefer to upload as a batch to get their predictions. More detailed guidelines on how to use the web-server as well as the working mechanism of this server can also be found there. After submitting a query protein or as a batch, it may take a few moments to get the prediction result, depending on the availability of server resources. Finally, predPhogly-Site will generate a result page based on the user’s submission, i.e., if protein sequences are submitted into the input box, the predictive data will be shown on the result page. Otherwise, it will be sent to the corresponding user through email.

Conclusion

In this study, for identifying phosphoglycerylation sites in protein with higher accuracy, a novel computational tool, predPhogly-Site, has been developed utilizing the coupling effects in a sequence. It exploits probabilistic sequence pattern information with variable cost adjustment in the classifier’s decision function for achieving higher predictive performance compared to the existing phosphoglycerylation site predictors. It has achieved significant performance improvement not only in the 10-fold cross-validation, which has been used as the benchmarking technique in the existing predictors but also in an independent test. Moreover, it has also achieved almost identical performance in both 10-fold cross-validation and independent test, which clearly demonstrates its stability. In the 10-fold cross-validation test, it has achieved more than 0.99 in both AUC and MCC, and in case of the independent test, it has achieved nearly 0.97 in the corresponding measures. These experimental outcomes demonstrate that predPhogly-Site is highly promising compared to the existing state-of-the-art phosphoglycerylation site predictors. It is expected to become a high throughput computational tool for PTM researcher for fast exploration of lysine modifications. Even the experimental scientists would be benefited from this web-based tool without going through its mathematical and implementation details. For further performance improvement and usability of this prediction tool, multiple types of post-translational modification with heterogeneous data would be incorporated simultaneously along with prediction interpretation support.

Supporting information

S1 File. Benchmark dataset.

The phosphoglycerylated proteins as well as the segmented sequences with respective protein ID and positions have been provided.

https://doi.org/10.1371/journal.pone.0249396.s001

(PDF)

S2 File. Independent test dataset.

Proteins which have been recently added to the PLMD database and completely unknown to the proposed system.

https://doi.org/10.1371/journal.pone.0249396.s002

(PDF)

S3 File. All possible combinations of the conditional probability values derived from the positive and negative subset.

https://doi.org/10.1371/journal.pone.0249396.s003

(XLSX)

S4 File. The non-conditional probability values of 21 amino acids derived from the positive and negative subset.

https://doi.org/10.1371/journal.pone.0249396.s004

(XLSX)

References

1. Saraswathy N, Ramalingam P. Concepts and techniques in genomics and proteomics. Elsevier; 2011.
2. McDowell G, Philpott A. New insights into the role of ubiquitylation of proteins. In: International review of cell and molecular biology. vol. 325. Elsevier; 2016. p. 35–88.
3. Qiu WR, Sun BQ, Xiao X, Xu ZC, Chou KC. iPTM-mLys: identifying multiple lysine PTM sites and their different types. Bioinformatics. 2016;32(20):3116–3123. pmid:27334473
- View Article
- PubMed/NCBI
- Google Scholar
4. Freiman RN, Tjian R. Regulating the regulators: lysine modifications make their mark. Cell. 2003;112(1):11–17. pmid:12526789
- View Article
- PubMed/NCBI
- Google Scholar
5. Reddy HM, Sharma A, Dehzangi A, Shigemizu D, Chandra AA, Tsunoda T. GlyStruct: glycation prediction using structural properties of amino acid residues. BMC bioinformatics. 2019;19(13):55–64. pmid:30717650
- View Article
- PubMed/NCBI
- Google Scholar
6. Jia J, Liu Z, Xiao X, Liu B, Chou KC. iSuc-PseOpt: identifying lysine succinylation sites in proteins by incorporating sequence-coupling effects into pseudo components and optimizing imbalanced training dataset. Analytical biochemistry. 2016;497:48–56. pmid:26723495
- View Article
- PubMed/NCBI
- Google Scholar
7. Xu Y, Chou KC. Recent progress in predicting posttranslational modification sites in proteins. Current topics in medicinal chemistry. 2016;16(6):591–603. pmid:26286211
- View Article
- PubMed/NCBI
- Google Scholar
8. Ju Z, Cao JZ, Gu H. Predicting lysine phosphoglycerylation with fuzzy SVM by incorporating k-spaced amino acid pairs into Chou’s general PseAAC. Journal of Theoretical Biology. 2016;397:145–150. pmid:26908349
- View Article
- PubMed/NCBI
- Google Scholar
9. Xu Y, Ding YX, Ding J, Wu LY, Deng NY. Phogly-PseAAC: prediction of lysine phosphoglycerylation in proteins incorporating with position-specific propensity. Journal of Theoretical Biology. 2015;379:10–15. pmid:25913879
- View Article
- PubMed/NCBI
- Google Scholar
10. Moellering RE, Cravatt BF. Functional lysine modification by an intrinsically reactive primary glycolytic metabolite. Science. 2013;341(6145):549–553. pmid:23908237
- View Article
- PubMed/NCBI
- Google Scholar
11. Chandra A, Sharma A, Dehzangi A, Shigemizu D, Tsunoda T. Bigram-PGK: phosphoglycerylation prediction using the technique of bigram probabilities of position specific scoring matrix. BMC molecular and cell biology. 2019;20(2):1–9. pmid:31856704
- View Article
- PubMed/NCBI
- Google Scholar
12. Liu LM, Xu Y, Chou KC. iPGK-PseAAC: identify lysine phosphoglycerylation sites in proteins by incorporating four different tiers of amino acid pairwise coupling information into the general PseAAC. Medicinal Chemistry. 2017;13(6):552–559. pmid:28521678
- View Article
- PubMed/NCBI
- Google Scholar
13. Chou KC. Prediction of signal peptides using scaled window. peptides. 2001;22(12):1973–1979. pmid:11786179
- View Article
- PubMed/NCBI
- Google Scholar
14. Hasan MAM, Ahmad S. mLysPTMpred: Multiple Lysine PTM Site Prediction Using Combination of SVM with Resolving Data Imbalance Issue. Natural Science. 2018;10(9):370–384.
- View Article
- Google Scholar
15. Chou KC. A vectorized sequence-coupling model for predicting HIV protease cleavage sites in proteins. Journal of Biological Chemistry. 1993;268(23):16938–16948. pmid:8349584
- View Article
- PubMed/NCBI
- Google Scholar
16. Chou KC. Prediction of human immunodeficiency virus protease cleavage sites in proteins. Analytical biochemistry. 1996;233(1):1–14. pmid:8789141
- View Article
- PubMed/NCBI
- Google Scholar
17. Veropoulos K, Campbell C, Cristianini N, et al. Controlling the sensitivity of support vector machines. In: Proceedings of the international joint conference on AI. vol. 55; 1999. p. 60.
18. Lin WZ, Fang JA, Xiao X, Chou KC. iDNA-Prot: identification of DNA binding proteins using random forest with grey model. PloS one. 2011;6(9). pmid:21935457
- View Article
- PubMed/NCBI
- Google Scholar
19. Hasan MAM, Ahmad S, Molla MKI. iMulti-HumPhos: a multi-label classifier for identifying human phosphorylated proteins using multiple kernel learning based support vector machines. Molecular BioSystems. 2017;13(8):1608–1618. pmid:28682387
- View Article
- PubMed/NCBI
- Google Scholar
20. Ju Z, Wang SY. Prediction of citrullination sites by incorporating k-spaced amino acid pairs into Chou’s general pseudo amino acid composition. Gene. 2018;664:78–83. pmid:29694908
- View Article
- PubMed/NCBI
- Google Scholar
21. Ju Z, He JJ. Prediction of lysine propionylation sites using biased SVM and incorporating four different sequence features into Chou’s PseAAC. Journal of Molecular Graphics and Modelling. 2017;76:356–363. pmid:28763688
- View Article
- PubMed/NCBI
- Google Scholar
22. Hasan MAM, Li J, Ahmad S, Molla MKI. predCar-site: Carbonylation sites prediction in proteins using support vector machine with resolving data imbalanced issue. Analytical biochemistry. 2017;525:107–113. pmid:28286168
- View Article
- PubMed/NCBI
- Google Scholar
23. Bao W, Yang B, Huang DS, Wang D, Liu Q, Chen YH, et al. IMKPse: Identification of protein malonylation sites by the key features into general PseAAC. IEEE Access. 2019;7:54073–54083.
- View Article
- Google Scholar
24. Hasan MA, Ben Islam MK, Rahman J, Ahmad S. Citrullination Site Prediction by Incorporating Sequence Coupled Effects into PseAAC and Resolving Data Imbalance Issue. Current Bioinformatics. 2020;15(3):235–245.
- View Article
- Google Scholar
25. Qiu WR, Xiao X, Lin WZ, Chou KC. iMethyl-PseAAC: identification of protein methylation sites via a pseudo amino acid composition approach. BioMed research international. 2014;2014.
- View Article
- Google Scholar
26. Liu Z, Wang Y, Gao T, Pan Z, Cheng H, Yang Q, et al. CPLM: a database of protein lysine modifications. Nucleic acids research. 2014;42(D1):D531–D536. pmid:24214993
- View Article
- PubMed/NCBI
- Google Scholar
27. Consortium U. UniProt: a worldwide hub of protein knowledge. Nucleic acids research. 2019;47(D1):D506–D515.
- View Article
- Google Scholar
28. Li W, Godzik A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics. 2006;22(13):1658–1659. pmid:16731699
- View Article
- PubMed/NCBI
- Google Scholar
29. Ju Z, Wang SY. Prediction of lysine formylation sites using the composition of k-spaced amino acid pairs via Chou’s 5-steps rule and general pseudo components. Genomics. 2020;112(1):859–866. pmid:31175975
- View Article
- PubMed/NCBI
- Google Scholar
30. Ning Q, Ma Z, Zhao X. dForml (KNN)-PseAAC: Detecting formylation sites from protein sequences using K-nearest neighbor algorithm via Chou’s 5-step rule and pseudo components. Journal of theoretical biology. 2019;470:43–49. pmid:30880183
- View Article
- PubMed/NCBI
- Google Scholar
31. Crooks GE, Hon G, Chandonia JM, Brenner SE. WebLogo: a sequence logo generator. Genome research. 2004;14(6):1188–1190. pmid:15173120
- View Article
- PubMed/NCBI
- Google Scholar
32. Xu H, Zhou J, Lin S, Deng W, Zhang Y, Xue Y. PLMD: An updated data resource of protein lysine modifications. Journal of Genetics and Genomics. 2017;44(5):243–250. pmid:28529077
- View Article
- PubMed/NCBI
- Google Scholar
33. Du P, Wang X, Xu C, Gao Y. PseAAC-Builder: A cross-platform stand-alone program for generating various special Chou’s pseudo-amino acid compositions. Analytical biochemistry. 2012;425(2):117–119. pmid:22459120
- View Article
- PubMed/NCBI
- Google Scholar
34. Qiu WR, Sun BQ, Xiao X, Xu ZC, Chou KC. iHyd-PseCp: Identify hydroxyproline and hydroxylysine in proteins by incorporating sequence-coupled effects into general PseAAC. Oncotarget. 2016;7(28):44310. pmid:27322424
- View Article
- PubMed/NCBI
- Google Scholar
35. Chou KC. Some remarks on protein attribute prediction and pseudo amino acid composition. Journal of theoretical biology. 2011;273(1):236–247. pmid:21168420
- View Article
- PubMed/NCBI
- Google Scholar
36. Chou KC. Using amphiphilic pseudo amino acid composition to predict enzyme subfamily classes. Bioinformatics. 2005;21(1):10–19. pmid:15308540
- View Article
- PubMed/NCBI
- Google Scholar
37. Ju Z, He JJ. Prediction of lysine crotonylation sites by incorporating the composition of k-spaced amino acid pairs into Chou’s general PseAAC. Journal of Molecular Graphics and Modelling. 2017;77:200–204. pmid:28886434
- View Article
- PubMed/NCBI
- Google Scholar
38. Min JL, Xiao X, Chou KC. iEzy-Drug: A web server for identifying the interaction between enzymes and drugs in cellular networking. BioMed research international. 2013;2013. pmid:24371828
- View Article
- PubMed/NCBI
- Google Scholar
39. Xu Y, Wen X, Wen LS, Wu LY, Deng NY, Chou KC. iNitro-Tyr: Prediction of nitrotyrosine sites in proteins with general pseudo amino acid composition. PloS one. 2014;9(8):e105018. pmid:25121969
- View Article
- PubMed/NCBI
- Google Scholar
40. Reback J, McKinney W, jbrockmendel, den Bossche JV, Augspurger T, Cloud P, et al. pandas-dev/pandas: Pandas 1.2.0rc0; 2020. Available from: https://doi.org/10.5281/zenodo.4311557.
41. Wang D, Liu D, Yuchi J, He F, Jiang Y, Cai S, et al. MusiteDeep: a deep-learning based webserver for protein post-translational modification site prediction and visualization. Nucleic Acids Research. 2020;. pmid:32324217
- View Article
- PubMed/NCBI
- Google Scholar
42. Lv Z, Zhang J, Ding H, Zou Q. RF-PseU: A Random Forest Predictor for RNA Pseudouridine Sites. Frontiers in Bioengineering and Biotechnology. 2020;8. pmid:32175316
- View Article
- PubMed/NCBI
- Google Scholar
43. Cortes C, Vapnik V. Support-vector networks. Machine learning. 1995;20(3):273–297.
- View Article
- Google Scholar
44. Vapnik V. The nature of statistical learning theory. Springer science & business media; 2013.
45. Ju Z, Wang SY. Prediction of lysine formylation sites using the composition of k-spaced amino acid pairs via Chou’s 5-steps rule and general pseudo components. Genomics. 2020;112(1):859–866. pmid:31175975
- View Article
- PubMed/NCBI
- Google Scholar
46. Zhang L, Tan B, Liu T, Sun X. Classification study for the imbalanced data based on Biased-SVM and the modified over-sampling algorithm. In: Journal of Physics: Conference Series. vol. 1237. IOP Publishing; 2019. p. 022052.
47. Ju Z, He JJ. Prediction of lysine glutarylation sites by maximum relevance minimum redundancy feature selection. Analytical biochemistry. 2018;550:1–7. pmid:29641975
- View Article
- PubMed/NCBI
- Google Scholar
48. Al-Barakati HJ, Saigo H, Newman RH, et al. RF-GlutarySite: a random forest based predictor for glutarylation sites. Molecular omics. 2019;15(3):189–204. pmid:31025681
- View Article
- PubMed/NCBI
- Google Scholar
49. Wu M, Yang Y, Wang H, Xu Y. A deep learning method to more accurately recall known lysine acetylation sites. BMC bioinformatics. 2019;20(1):49. pmid:30674277
- View Article
- PubMed/NCBI
- Google Scholar
50. Jia C, Zhang M, Fan C, Li F, Song J. Formator: predicting lysine formylation sites based on the most distant undersampling and safe-level synthetic minority oversampling. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2019;. pmid:31804942
- View Article
- PubMed/NCBI
- Google Scholar
51. Yu J, Shi S, Zhang F, Chen G, Cao M. PredGly: predicting lysine glycation sites for Homo sapiens based on XGboost feature optimization. Bioinformatics. 2019;35(16):2749–2756. pmid:30590442
- View Article
- PubMed/NCBI
- Google Scholar
52. Qu K, Han K, Wu S, Wang G, Wei L. Identification of DNA-binding proteins using mixed feature representation methods. Molecules. 2017;22(10):1602. pmid:28937647
- View Article
- PubMed/NCBI
- Google Scholar
53. Malebary SJ, Rehman MSu, Khan YD. iCrotoK-PseAAC: Identify lysine crotonylation sites by blending position relative statistical features according to the Chou’s 5-step rule. PloS one. 2019;14(11):e0223993. pmid:31751380
- View Article
- PubMed/NCBI
- Google Scholar
54. Li F, Li C, Marquez-Lago TT, Leier A, Akutsu T, Purcell AW, et al. Quokka: a comprehensive tool for rapid and accurate prediction of kinase family-specific phosphorylation sites in the human proteome. Bioinformatics. 2018;34(24):4223–4231. pmid:29947803
- View Article
- PubMed/NCBI
- Google Scholar
55. Adilina S, Farid DM, Shatabda S. Effective DNA binding protein prediction by using key features via Chou’s general PseAAC. Journal of theoretical biology. 2019;460:64–78. pmid:30316822
- View Article
- PubMed/NCBI
- Google Scholar
56. Thapa N, Chaudhari M, McManus S, Roy K, Newman RH, Saigo H, et al. DeepSuccinylSite: a deep learning based approach for protein succinylation site prediction. BMC bioinformatics. 2020;21:1–10. pmid:32321437
- View Article
- PubMed/NCBI
- Google Scholar
57. Liu K, Cao L, Du P, Chen W. im6A-TS-CNN: identifying N6-methyladenine site in multiple tissues by using convolutional neural network. Molecular Therapy-Nucleic Acids. 2020;. pmid:32858457
- View Article
- PubMed/NCBI
- Google Scholar

[ref1] 1. Saraswathy N, Ramalingam P. Concepts and techniques in genomics and proteomics. Elsevier; 2011.

[ref2] 2. McDowell G, Philpott A. New insights into the role of ubiquitylation of proteins. In: International review of cell and molecular biology. vol. 325. Elsevier; 2016. p. 35–88.

[ref3] 3. Qiu WR, Sun BQ, Xiao X, Xu ZC, Chou KC. iPTM-mLys: identifying multiple lysine PTM sites and their different types. Bioinformatics. 2016;32(20):3116–3123. pmid:27334473
View Article
PubMed/NCBI
Google Scholar

[4] View Article

[5] PubMed/NCBI

[6] Google Scholar

[ref4] 4. Freiman RN, Tjian R. Regulating the regulators: lysine modifications make their mark. Cell. 2003;112(1):11–17. pmid:12526789
View Article
PubMed/NCBI
Google Scholar

[8] View Article

[9] PubMed/NCBI

[10] Google Scholar

[ref5] 5. Reddy HM, Sharma A, Dehzangi A, Shigemizu D, Chandra AA, Tsunoda T. GlyStruct: glycation prediction using structural properties of amino acid residues. BMC bioinformatics. 2019;19(13):55–64. pmid:30717650
View Article
PubMed/NCBI
Google Scholar

[12] View Article

[13] PubMed/NCBI

[14] Google Scholar

[ref6] 6. Jia J, Liu Z, Xiao X, Liu B, Chou KC. iSuc-PseOpt: identifying lysine succinylation sites in proteins by incorporating sequence-coupling effects into pseudo components and optimizing imbalanced training dataset. Analytical biochemistry. 2016;497:48–56. pmid:26723495
View Article
PubMed/NCBI
Google Scholar

[16] View Article

[17] PubMed/NCBI

[18] Google Scholar

[ref7] 7. Xu Y, Chou KC. Recent progress in predicting posttranslational modification sites in proteins. Current topics in medicinal chemistry. 2016;16(6):591–603. pmid:26286211
View Article
PubMed/NCBI
Google Scholar

[20] View Article

[21] PubMed/NCBI

[22] Google Scholar

[ref8] 8. Ju Z, Cao JZ, Gu H. Predicting lysine phosphoglycerylation with fuzzy SVM by incorporating k-spaced amino acid pairs into Chou’s general PseAAC. Journal of Theoretical Biology. 2016;397:145–150. pmid:26908349
View Article
PubMed/NCBI
Google Scholar

[24] View Article

[25] PubMed/NCBI

[26] Google Scholar

[ref9] 9. Xu Y, Ding YX, Ding J, Wu LY, Deng NY. Phogly-PseAAC: prediction of lysine phosphoglycerylation in proteins incorporating with position-specific propensity. Journal of Theoretical Biology. 2015;379:10–15. pmid:25913879
View Article
PubMed/NCBI
Google Scholar

[28] View Article

[29] PubMed/NCBI

[30] Google Scholar

[ref10] 10. Moellering RE, Cravatt BF. Functional lysine modification by an intrinsically reactive primary glycolytic metabolite. Science. 2013;341(6145):549–553. pmid:23908237
View Article
PubMed/NCBI
Google Scholar

[32] View Article

[33] PubMed/NCBI

[34] Google Scholar

[ref11] 11. Chandra A, Sharma A, Dehzangi A, Shigemizu D, Tsunoda T. Bigram-PGK: phosphoglycerylation prediction using the technique of bigram probabilities of position specific scoring matrix. BMC molecular and cell biology. 2019;20(2):1–9. pmid:31856704
View Article
PubMed/NCBI
Google Scholar

[36] View Article

[37] PubMed/NCBI

[38] Google Scholar

[ref12] 12. Liu LM, Xu Y, Chou KC. iPGK-PseAAC: identify lysine phosphoglycerylation sites in proteins by incorporating four different tiers of amino acid pairwise coupling information into the general PseAAC. Medicinal Chemistry. 2017;13(6):552–559. pmid:28521678
View Article
PubMed/NCBI
Google Scholar

[40] View Article

[41] PubMed/NCBI

[42] Google Scholar

[ref13] 13. Chou KC. Prediction of signal peptides using scaled window. peptides. 2001;22(12):1973–1979. pmid:11786179
View Article
PubMed/NCBI
Google Scholar

[44] View Article

[45] PubMed/NCBI

[46] Google Scholar

[ref14] 14. Hasan MAM, Ahmad S. mLysPTMpred: Multiple Lysine PTM Site Prediction Using Combination of SVM with Resolving Data Imbalance Issue. Natural Science. 2018;10(9):370–384.
View Article
Google Scholar

[48] View Article

[49] Google Scholar

[ref15] 15. Chou KC. A vectorized sequence-coupling model for predicting HIV protease cleavage sites in proteins. Journal of Biological Chemistry. 1993;268(23):16938–16948. pmid:8349584
View Article
PubMed/NCBI
Google Scholar

[51] View Article

[52] PubMed/NCBI

[53] Google Scholar

[ref16] 16. Chou KC. Prediction of human immunodeficiency virus protease cleavage sites in proteins. Analytical biochemistry. 1996;233(1):1–14. pmid:8789141
View Article
PubMed/NCBI
Google Scholar

[55] View Article

[56] PubMed/NCBI

[57] Google Scholar

[ref17] 17. Veropoulos K, Campbell C, Cristianini N, et al. Controlling the sensitivity of support vector machines. In: Proceedings of the international joint conference on AI. vol. 55; 1999. p. 60.

[ref18] 18. Lin WZ, Fang JA, Xiao X, Chou KC. iDNA-Prot: identification of DNA binding proteins using random forest with grey model. PloS one. 2011;6(9). pmid:21935457
View Article
PubMed/NCBI
Google Scholar

[60] View Article

[61] PubMed/NCBI

[62] Google Scholar

[ref19] 19. Hasan MAM, Ahmad S, Molla MKI. iMulti-HumPhos: a multi-label classifier for identifying human phosphorylated proteins using multiple kernel learning based support vector machines. Molecular BioSystems. 2017;13(8):1608–1618. pmid:28682387
View Article
PubMed/NCBI
Google Scholar

[64] View Article

[65] PubMed/NCBI

[66] Google Scholar

[ref20] 20. Ju Z, Wang SY. Prediction of citrullination sites by incorporating k-spaced amino acid pairs into Chou’s general pseudo amino acid composition. Gene. 2018;664:78–83. pmid:29694908
View Article
PubMed/NCBI
Google Scholar

[68] View Article

[69] PubMed/NCBI

[70] Google Scholar

[ref21] 21. Ju Z, He JJ. Prediction of lysine propionylation sites using biased SVM and incorporating four different sequence features into Chou’s PseAAC. Journal of Molecular Graphics and Modelling. 2017;76:356–363. pmid:28763688
View Article
PubMed/NCBI
Google Scholar

[72] View Article

[73] PubMed/NCBI

[74] Google Scholar

[ref22] 22. Hasan MAM, Li J, Ahmad S, Molla MKI. predCar-site: Carbonylation sites prediction in proteins using support vector machine with resolving data imbalanced issue. Analytical biochemistry. 2017;525:107–113. pmid:28286168
View Article
PubMed/NCBI
Google Scholar

[76] View Article

[77] PubMed/NCBI

[78] Google Scholar

[ref23] 23. Bao W, Yang B, Huang DS, Wang D, Liu Q, Chen YH, et al. IMKPse: Identification of protein malonylation sites by the key features into general PseAAC. IEEE Access. 2019;7:54073–54083.
View Article
Google Scholar

[80] View Article

[81] Google Scholar

[ref24] 24. Hasan MA, Ben Islam MK, Rahman J, Ahmad S. Citrullination Site Prediction by Incorporating Sequence Coupled Effects into PseAAC and Resolving Data Imbalance Issue. Current Bioinformatics. 2020;15(3):235–245.
View Article
Google Scholar

[83] View Article

[84] Google Scholar

[ref25] 25. Qiu WR, Xiao X, Lin WZ, Chou KC. iMethyl-PseAAC: identification of protein methylation sites via a pseudo amino acid composition approach. BioMed research international. 2014;2014.
View Article
Google Scholar

[86] View Article

[87] Google Scholar

[ref26] 26. Liu Z, Wang Y, Gao T, Pan Z, Cheng H, Yang Q, et al. CPLM: a database of protein lysine modifications. Nucleic acids research. 2014;42(D1):D531–D536. pmid:24214993
View Article
PubMed/NCBI
Google Scholar

[89] View Article

[90] PubMed/NCBI

[91] Google Scholar

[ref27] 27. Consortium U. UniProt: a worldwide hub of protein knowledge. Nucleic acids research. 2019;47(D1):D506–D515.
View Article
Google Scholar

[93] View Article

[94] Google Scholar

[ref28] 28. Li W, Godzik A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics. 2006;22(13):1658–1659. pmid:16731699
View Article
PubMed/NCBI
Google Scholar

[96] View Article

[97] PubMed/NCBI

[98] Google Scholar

[ref29] 29. Ju Z, Wang SY. Prediction of lysine formylation sites using the composition of k-spaced amino acid pairs via Chou’s 5-steps rule and general pseudo components. Genomics. 2020;112(1):859–866. pmid:31175975
View Article
PubMed/NCBI
Google Scholar

[100] View Article

[101] PubMed/NCBI

[102] Google Scholar

[ref30] 30. Ning Q, Ma Z, Zhao X. dForml (KNN)-PseAAC: Detecting formylation sites from protein sequences using K-nearest neighbor algorithm via Chou’s 5-step rule and pseudo components. Journal of theoretical biology. 2019;470:43–49. pmid:30880183
View Article
PubMed/NCBI
Google Scholar

[104] View Article

[105] PubMed/NCBI

[106] Google Scholar

[ref31] 31. Crooks GE, Hon G, Chandonia JM, Brenner SE. WebLogo: a sequence logo generator. Genome research. 2004;14(6):1188–1190. pmid:15173120
View Article
PubMed/NCBI
Google Scholar

[108] View Article

[109] PubMed/NCBI

[110] Google Scholar

[ref32] 32. Xu H, Zhou J, Lin S, Deng W, Zhang Y, Xue Y. PLMD: An updated data resource of protein lysine modifications. Journal of Genetics and Genomics. 2017;44(5):243–250. pmid:28529077
View Article
PubMed/NCBI
Google Scholar

[112] View Article

[113] PubMed/NCBI

[114] Google Scholar

[ref33] 33. Du P, Wang X, Xu C, Gao Y. PseAAC-Builder: A cross-platform stand-alone program for generating various special Chou’s pseudo-amino acid compositions. Analytical biochemistry. 2012;425(2):117–119. pmid:22459120
View Article
PubMed/NCBI
Google Scholar

[116] View Article

[117] PubMed/NCBI

[118] Google Scholar

[ref34] 34. Qiu WR, Sun BQ, Xiao X, Xu ZC, Chou KC. iHyd-PseCp: Identify hydroxyproline and hydroxylysine in proteins by incorporating sequence-coupled effects into general PseAAC. Oncotarget. 2016;7(28):44310. pmid:27322424
View Article
PubMed/NCBI
Google Scholar

[120] View Article

[121] PubMed/NCBI

[122] Google Scholar

[ref35] 35. Chou KC. Some remarks on protein attribute prediction and pseudo amino acid composition. Journal of theoretical biology. 2011;273(1):236–247. pmid:21168420
View Article
PubMed/NCBI
Google Scholar

[124] View Article

[125] PubMed/NCBI

[126] Google Scholar

[ref36] 36. Chou KC. Using amphiphilic pseudo amino acid composition to predict enzyme subfamily classes. Bioinformatics. 2005;21(1):10–19. pmid:15308540
View Article
PubMed/NCBI
Google Scholar

[128] View Article

[129] PubMed/NCBI

[130] Google Scholar

[ref37] 37. Ju Z, He JJ. Prediction of lysine crotonylation sites by incorporating the composition of k-spaced amino acid pairs into Chou’s general PseAAC. Journal of Molecular Graphics and Modelling. 2017;77:200–204. pmid:28886434
View Article
PubMed/NCBI
Google Scholar

[132] View Article

[133] PubMed/NCBI

[134] Google Scholar

[ref38] 38. Min JL, Xiao X, Chou KC. iEzy-Drug: A web server for identifying the interaction between enzymes and drugs in cellular networking. BioMed research international. 2013;2013. pmid:24371828
View Article
PubMed/NCBI
Google Scholar

[136] View Article

[137] PubMed/NCBI

[138] Google Scholar

[ref39] 39. Xu Y, Wen X, Wen LS, Wu LY, Deng NY, Chou KC. iNitro-Tyr: Prediction of nitrotyrosine sites in proteins with general pseudo amino acid composition. PloS one. 2014;9(8):e105018. pmid:25121969
View Article
PubMed/NCBI
Google Scholar

[140] View Article

[141] PubMed/NCBI

[142] Google Scholar

[ref40] 40. Reback J, McKinney W, jbrockmendel, den Bossche JV, Augspurger T, Cloud P, et al. pandas-dev/pandas: Pandas 1.2.0rc0; 2020. Available from: https://doi.org/10.5281/zenodo.4311557.

[ref41] 41. Wang D, Liu D, Yuchi J, He F, Jiang Y, Cai S, et al. MusiteDeep: a deep-learning based webserver for protein post-translational modification site prediction and visualization. Nucleic Acids Research. 2020;. pmid:32324217
View Article
PubMed/NCBI
Google Scholar

[145] View Article

[146] PubMed/NCBI

[147] Google Scholar

[ref42] 42. Lv Z, Zhang J, Ding H, Zou Q. RF-PseU: A Random Forest Predictor for RNA Pseudouridine Sites. Frontiers in Bioengineering and Biotechnology. 2020;8. pmid:32175316
View Article
PubMed/NCBI
Google Scholar

[149] View Article

[150] PubMed/NCBI

[151] Google Scholar

[ref43] 43. Cortes C, Vapnik V. Support-vector networks. Machine learning. 1995;20(3):273–297.
View Article
Google Scholar

[153] View Article

[154] Google Scholar

[ref44] 44. Vapnik V. The nature of statistical learning theory. Springer science & business media; 2013.

[ref45] 45. Ju Z, Wang SY. Prediction of lysine formylation sites using the composition of k-spaced amino acid pairs via Chou’s 5-steps rule and general pseudo components. Genomics. 2020;112(1):859–866. pmid:31175975
View Article
PubMed/NCBI
Google Scholar

[157] View Article

[158] PubMed/NCBI

[159] Google Scholar

[ref46] 46. Zhang L, Tan B, Liu T, Sun X. Classification study for the imbalanced data based on Biased-SVM and the modified over-sampling algorithm. In: Journal of Physics: Conference Series. vol. 1237. IOP Publishing; 2019. p. 022052.

[ref47] 47. Ju Z, He JJ. Prediction of lysine glutarylation sites by maximum relevance minimum redundancy feature selection. Analytical biochemistry. 2018;550:1–7. pmid:29641975
View Article
PubMed/NCBI
Google Scholar

[162] View Article

[163] PubMed/NCBI

[164] Google Scholar

[ref48] 48. Al-Barakati HJ, Saigo H, Newman RH, et al. RF-GlutarySite: a random forest based predictor for glutarylation sites. Molecular omics. 2019;15(3):189–204. pmid:31025681
View Article
PubMed/NCBI
Google Scholar

[166] View Article

[167] PubMed/NCBI

[168] Google Scholar

[ref49] 49. Wu M, Yang Y, Wang H, Xu Y. A deep learning method to more accurately recall known lysine acetylation sites. BMC bioinformatics. 2019;20(1):49. pmid:30674277
View Article
PubMed/NCBI
Google Scholar

[170] View Article

[171] PubMed/NCBI

[172] Google Scholar

[ref50] 50. Jia C, Zhang M, Fan C, Li F, Song J. Formator: predicting lysine formylation sites based on the most distant undersampling and safe-level synthetic minority oversampling. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2019;. pmid:31804942
View Article
PubMed/NCBI
Google Scholar

[174] View Article

[175] PubMed/NCBI

[176] Google Scholar

[ref51] 51. Yu J, Shi S, Zhang F, Chen G, Cao M. PredGly: predicting lysine glycation sites for Homo sapiens based on XGboost feature optimization. Bioinformatics. 2019;35(16):2749–2756. pmid:30590442
View Article
PubMed/NCBI
Google Scholar

[178] View Article

[179] PubMed/NCBI

[180] Google Scholar

[ref52] 52. Qu K, Han K, Wu S, Wang G, Wei L. Identification of DNA-binding proteins using mixed feature representation methods. Molecules. 2017;22(10):1602. pmid:28937647
View Article
PubMed/NCBI
Google Scholar

[182] View Article

[183] PubMed/NCBI

[184] Google Scholar

[ref53] 53. Malebary SJ, Rehman MSu, Khan YD. iCrotoK-PseAAC: Identify lysine crotonylation sites by blending position relative statistical features according to the Chou’s 5-step rule. PloS one. 2019;14(11):e0223993. pmid:31751380
View Article
PubMed/NCBI
Google Scholar

[186] View Article

[187] PubMed/NCBI

[188] Google Scholar

[ref54] 54. Li F, Li C, Marquez-Lago TT, Leier A, Akutsu T, Purcell AW, et al. Quokka: a comprehensive tool for rapid and accurate prediction of kinase family-specific phosphorylation sites in the human proteome. Bioinformatics. 2018;34(24):4223–4231. pmid:29947803
View Article
PubMed/NCBI
Google Scholar

[190] View Article

[191] PubMed/NCBI

[192] Google Scholar

[ref55] 55. Adilina S, Farid DM, Shatabda S. Effective DNA binding protein prediction by using key features via Chou’s general PseAAC. Journal of theoretical biology. 2019;460:64–78. pmid:30316822
View Article
PubMed/NCBI
Google Scholar

[194] View Article

[195] PubMed/NCBI

[196] Google Scholar

[ref56] 56. Thapa N, Chaudhari M, McManus S, Roy K, Newman RH, Saigo H, et al. DeepSuccinylSite: a deep learning based approach for protein succinylation site prediction. BMC bioinformatics. 2020;21:1–10. pmid:32321437
View Article
PubMed/NCBI
Google Scholar

[198] View Article

[199] PubMed/NCBI

[200] Google Scholar

[ref57] 57. Liu K, Cao L, Du P, Chen W. im6A-TS-CNN: identifying N6-methyladenine site in multiple tissues by using convolutional neural network. Molecular Therapy-Nucleic Acids. 2020;. pmid:32858457
View Article
PubMed/NCBI
Google Scholar

[202] View Article

[203] PubMed/NCBI

[204] Google Scholar

Figures

Abstract

Introduction

Materials and methods

Dataset

Feature construction

Prediction method and addressing data imbalance

Formulation of evaluation metrics

Validation of the proposed model

Results and discussions

Performance of predPhogly-Site

Comparative analysis of cross-validation performance

Independent test

Web-server

Conclusion

Supporting information

S1 File. Benchmark dataset.

S2 File. Independent test dataset.

S3 File. All possible combinations of the conditional probability values derived from the positive and negative subset.

S4 File. The non-conditional probability values of 21 amino acids derived from the positive and negative subset.

References