Computational Identification of Lysine Glutarylation Sites Using Positive-Unlabeled Learning

Abstract

Background

As a new type of protein acylation modification, lysine glutarylation has been found to play a crucial role in metabolic processes and mitochondrial functions. To further explore the biological mechanisms and functions of glutarylation, it is significant to predict the potential glutarylation sites. In the existing glutarylation site predictors, experimentally verified glutarylation sites are treated as positive samples and non-verified lysine sites as the negative samples to train predictors. However, the non-verified lysine sites may contain some glutarylation sites which have not been experimentally identified yet.

Methods

In this study, experimentally verified glutarylation sites are treated as the positive samples, whereas the remaining non-verified lysine sites are treated as unlabeled samples. A bioinformatics tool named PUL-GLU was developed to identify glutarylation sites using a positive-unlabeled learning algorithm.

Results

Experimental results show that PUL-GLU significantly outperforms the current glutarylation site predictors. Therefore, PUL-GLU can be a powerful tool for accurate identification of protein glutarylation sites.

Conclusion

A user-friendly web-server for PUL-GLU is available at http://bioinform.cn/pul_glu/.

1. Introduction

Protein post-translational modifications (PTMs) are crucial steps in protein synthesis and regulate various biological processes such as protein signaling, localization, and degradation. Among the various types of PTMs, acetylation, succinylation, malonylation, 2-hydroxyisobutyrylation, butyrylation, crotonylation, etc., can all occur at the ε-amino groups of specific lysine residues [1-6] and are known as lysine acylation modification. Recently, Tan et al. [7] discovered a new type of lysine acylation modification, named glutarylation, which is found in both prokaryotic and eukaryotic cells. Lysine glutarylation is a dynamic and evolutionarily conserved modification process, in which a glutaryl group attaches to specific lysine residues of a substrate protein. Similar to succinylation and acetylation, lysine glutarylation has been found to play a crucial role in metabolic processes and mitochondrial functions, such as fatty acid metabolism, amino acid metabolism and cellular respiration [6, 7]. Previous studies have shown that glutarylation of carbamoyl phosphate synthase 1 (CPS1) inhibits its activity but can be reversed by SIRT5 [7]. More importantly, molecular evidence suggested that abnormal glutarylation was closely related to several metabolic disorders, including diabetes, neurodegenerative diseases, glutaric acidemia type I and cancer [7]. Therefore, research on glutarylation would be beneficial for

drug discovery. Although some research work has been done to reveal the biological functions of glutarylation, the regulatory mechanism of glutarylation in cells is still largely unknown.

In order to further investigate the molecular mechanisms of glutarylation, a fundamental and critical task is to identify glutarylation sites with high accuracy. Although several large-scale proteomics methods such as mass spectrometry [7, 8] have been applied to detect glutarylation sites, these experimental approaches are not only time-consuming but also expensive. The majority of lysine glutarylation substrates and glutarylation sites still remain largely unknown. Therefore, it is urgent and necessary to develop computational methods to identify the potential glutarylated proteins and the corresponding glutarylation sites. Up to now, a few computational tools have been proposed to identify glutarylation sites. Ju and He [9] proposed the first glutarylation site predictor named GlutPred based on maximum relevance minimum redundancy (mRMR) feature selection algorithm. Xu et al. [10] developed a predictor, iGlu-Lys, by using the position-specific propensity matrix (PSPM) features around lysine-centered peptides and SVM algorithm. Huang et al. [11] proposed a prediction model by incorporating maximal dependence decomposition (MDD)-identified substrate motifs into an integrated SVM classifier. The cross-validation showed that amino acid composition features were most effective in discriminating between glutarylation and non-glutarylation sites. Recently, Albarakati et al. [12] developed a novel predictor, RF-GlutarySite, by using the physiochemical and sequence-based features and random forest (RF) algorithm.

Note that in the aforementioned four existing prediction methods, the experimentally verified glutarylation sites were treated as the positive samples and the remaining non-verified lysine sites were treated as the negative samples to train classifiers to predict glutarylation sites from unknown proteins. However, due to the limitations of experimental technique and condition, the remaining non-verified lysine sites might contain some glutarylation sites which have not been experimentally identified yet. Thus, the existing predictors were actually built on the noisy dataset. As a result, the accuracy of the existing predictors would not be as good as they were supposed to be.

In contrast to previous methods, experimentally verified glutarylation sites were treated as positive samples and the remaining non-verified lysine sites were treated as unlabeled samples in our study. A novel glutarylation site predictor was developed by using a positive unlabeled (PU) learning technique [13]. Specifically, the algorithm had five stages: stage 1, the composition of k-spaced amino acid pairs (CKSAAP), binary encoding (BE), and amino acid factors (AAF) were combined to encode the glutarylation site; stage 2, the crucial features were refined out using the maximum relevance and minimum redundancy (mRMR) feature selection method [14]; stage 3, a reliable negative set was selected from the unlabeled set by a maximum distance rule; stage 4, the reliable negative set was expanded and a series of SVM classifiers with RBF (Radial Basis Function) kernels were trained iteratively; stage 5, a final SVM model was trained on the positive set and the selected reliable negative set by 10-fold cross-validation. This method was called PUL-GLU (PU Learning for GLUtarylation sites prediction). The experimental results showed that the accuracy of PUL-GLU was 79.77% on the training set evaluated by 10-fold cross-validation and 76.65% on the independent test set.

As demonstrated by a series of recent publications [15, 16] and summarized in three comprehensive review papers [17-19], to develop a really useful predictor for a biological system, one needs to follow “Chou’s 5-steps rule” [17] to go through the following five steps: (1) select or construct a valid benchmark dataset to train and test the predictor; (2) represent the samples with an effective formulation that can truly reflect their intrinsic correlation with the target to be predicted; (3) introduce or develop a powerful algorithm to conduct the prediction; (4) properly perform cross-validation tests to objectively evaluate the anticipated prediction accuracy; (5) establish a user-friendly web-server for the predictor that is accessible to the public. The description of how to deal with these five steps is given below.

2. Materials and Methods

2.1. Dataset

Benchmark dataset was collected from the recent literature (Ju and He, 2018) in this study. The identity of these proteins was reduced to 40% by the CD-HIT program [20]. The training set consisted of 167 proteins with 590 experimentally annotated lysine glutarylation sites and 3498 non-annotated lysine sites; the independent test set consisted of 20 proteins with 56 experimentally annotated lysine glutarylation sites and 428 non-annotated lysine sites. Sliding window method was used to encode every lysine residue K of the dataset because glutarylation only occurred in lysine residues K. Based on our previous work [9], the window size of every training peptide was selected as 35 here. That means every lysine residue in the training dataset and the testing dataset was represented as a peptide segment of length 35 with 17 residues upstream and 17 residues downstream of lysine residue K. The training set and independent testing set are provided in Supplementary Material S1 ^{(670KB, xls)} .

2.2. Feature Construction

2.2.1. Amino Acid Factors

Physicochemical properties of amino acids play a crucial role in the identification of PTMs site. By using multivariate statistical analyses, 544 physicochemical properties of amino acids in the AAIndex have been transformed into five multidimensional patterns of attributes. The five multidimensional patterns of attributes reflect the polarity, secondary structure, molecular volume, codon diversity, and electrostatic charge, respectively [21]. These five generated attributes are named amino acid factors (AAF). For a given peptide, it can be encoded as a 35*5=175-dimensional vector by AAF.

2.2.2. Binary Encoding

The information of amino acid composition and position can be effectively characterized by binary encoding (BE) [22]. Considering 21 amino acids were ordered as ‘ACDEFGHIKLMNPQRSTVWYX’ (‘X’ means virtual amino acid), each amino acid residue in a given peptide was translated into a 21-dimensional binary vector. For example, amino acid ‘A’ is encoded as (100000000000000000000), …, and ‘X’ is encoded as (000000000000000000001). Thus, every training peptide can be expressed as a 35*21=735-dimensional vector by BE encoding.

2.2.3. Composition of k-spaced Amino Acid Pairs

The composition of k-spaced amino acid pairs (CKSAAP) can reflect the short linear motif information by calculating the occurrence frequency of the amino acid pairs in a given sequence fragment [23, 24]. An amino acid pair separated by any k amino acid residues is known as the k-spaced amino acid pair. For example, the CKSAAP of a given peptide for k=1 yields a 441-dimensional numeric vector defined as:

where ‘x’ represents any one of the 21 amino acids, and represents the total number of 1-spaced amino acid pairs. Here, CKSAAP with k= 0, 1, 2, 3 and 4 was utilized to encode the training peptides as 2205-dimensional feature vectors.

2.2.4. The Feature Space

In accordance with our previous work [9], the AAF, BE and CKSAAP were integrated to encode the training samples. Overall, each sample in the benchmark dataset was encoded as a 35*5+35*21+2205=3115-dimensional feature. Since the integrated encoding generated a high-dimensional feature vector, the maximum relevance and minimum redundancy (mRMR) feature selection method [14] and incremental feature selection (IFS) algorithm were used to remove the redundant features. Firstly, each component of 3115 features was ranked by the mRMR method. Then, the IFS algorithm was used to select 50 features with the highest score in each iteration. Here, the top 300 features were selected as optimal input features based on our previous work [9].

2.3. Prediction Method

2.3.1. Support Vector Machine

To facilitate its description, the training set is denoted as. The SVM can be formulated as follows:

where ϕ(x) is the non-linear mapping, and X_i(i=1, 2,...,l) are slack variables. C is the parameter determining the trade-off between model complexity (margin size) and classification errors. The Gaussian kernel function is used in the SVM. The Libsvm toolkit [25] was utilized to carry out the SVM models. Here, C and γ were set to the default values in Libsvm (i.e., C and γ =1 / 300;).

2.3.2. Positive-unlabeled Learning for Bioinformatics

In many fields, the obtainment of negative examples is usually costly and even not possible. Hence, many PU learning algorithms have been developed to deal with the problems which were lacking in negative examples. PU learning originated in text classification [25-28], and has been successfully applied to many biological problems in recent years. For example, Wang et al. [13] developed a PU learning algorithm named PSoL to finding non-coding RNA genes; Zhao et al. [26] designed a PU learning algorithm, AGPS, for gene function prediction; Cerulo et al. [27] used a PU learning algorithm named PosOnly for the derivation of gene regulatory networks; Yang et al. [28] designed a PU learning algorithm named PUDI for disease gene identification; Yang et al. [29] proposed an ensemble-based PU learning method for identifying disease gene by integrating multiple PU learning classifiers; Li et al. [30] proposed a positive unlabelled (PU) learning-based method, PA2DE (V2.0), based on the AlphaMax algorithm for protein glycosylation site prediction [31-33].

2.3.3. Development of PUL-GLU

As mentioned above, in this study, experimentally verified glutarylation sites were treated as positive samples and the remaining non-verified lysine sites were treated as unlabeled samples to build a classifier. In this way, the training dataset is divided into two parts: (1) the positive training dataset P and (2) the unlabeled training dataset U. Thus the prediction of glutarylation sites became learning from positive and unlabeled samples. An effective positive-unlabeled learning algorithm, PSoL [13], was used to construct PUL-GLU. The flowchart of PUL-GLU is shown in Table 1. There are three stages in it:

Table 1.

The flowchart of the PUL-GLU algorithm.

Stage 1. Selection of initial reliable negatives:

PUL-GLU selected the initial reliable negative set RN⁰ from the unlabeled set U based on the formula (3). The formula (3) ensures that the selected initial negative set has the highest reliability because it is farthest from the positive example set.

where d(N,P) is defined as follows:

Stage 2. Expansion of the reliable negative example set:

The initial negative set is gradually extended by iteratively trained SVM classifiers. Let RNⁱ be the current reliable negative training set; and Uⁱ be the current unlabeled set at the ith iteration. An SVM classifier fⁱ was firstly trained on P and RNⁱ; then, fⁱ Uⁱ and calculate its decision value. To ensure the purity of the selected negative set, the selected negative samples with the decision value less than a threshold T (here, T was set to -0.2) were selected as the newly predicted negative set Nⁱ_pred. To avoid the imbalance problem, the size of Nⁱ_pred was controlled less than |P|, and RNⁱ is replaced with the negative support vectors RNⁱ_SV. At the (i+1)th iteration, Uⁱ⁺¹ = Uⁱ \ Nⁱ_pred; RNⁱ⁺¹ = Nⁱ_pred U Nⁱ_SV . An SVM classifier fⁱ⁺¹ was trained on P and RNⁱ⁺¹. As the number of iterations increases, RNⁱ may contain more and more false-positive examples, therefore, iteration should be terminated if the size of r*|P| goes below a threshold RNⁱ⁺¹ (here r was set to 2).

Stage 3. Acquisition of the final classifier:

Let RN be the representative reliable negative training set. A final SVM classifier f was trained on P and RN.

2.4. Cross-validation and Performance Assessment

Jackknife test, K-fold cross-validation, and independent dataset test are three of the most common strategies for the evaluation of the performance of a predictor [17]. Although the jackknife test is the most objective among three evaluation methods, it is the most time-intensive. Therefore, to reduce computational time, we adopted a 10-fold cross-validation test to evaluate the proposed model. The 10-fold cross-validation is repeated 10 times. In addition, an independent dataset test was also adopted to further evaluate our method.

Five widely-accepted measurements, including sensitivity (Sn), specificity (Sp), precision (Pre), accuracy (ACC), and Matthew’s correlation coefficient (MCC), were used to evaluate the prediction performances of PUL-GLU, which are defined as:

where TP, TN, FP and FN stand for the number of true positives, true negatives, false-positives and false-negatives, respectively.

3. Results and discussion

3.1. Performance of PUL-GLU

To select the representative reliable negative samples, a negative set expansion was implemented on the training set. As a result, the representative reliable negative set contains 1199 reliable non-glutarylated lysine sites (Supplementary Material S1 ^{(670KB, xls)} ). Finally, PUL-GLU was trained on the positive set and the representative reliable negative set. The 10-fold cross-validation of PUL-GLU on and is shown in Table 2. As shown in Table 2, the prediction values for Sn, Sp, Pre, ACC, and MCC values reached 66.56%, 86.43%, 70.71%, 79.88% and 0.5384, respectively. The prediction performance of PUL-GLU was much higher than the SVM model trained on positive samples and randomly selected negative samples. This result indicated that the selected representative reliable negative samples could be more effective than those selected randomly. To avoid the overestimation of PUL-GLU, it was performed on the entire training set. The performance of PUL-GLU also achieved a satisfactory performance with an MCC of 0.35.

Table 2.

10-fold cross-validation performance of PUL-GLU and other methods.

Methods	Sn(%)	Sp(%)	Pre(%)	ACC(%)	MCC
SVM1	61.73±0.83	76.45±1.45	56.36±1.53	71.59±0.98	0.3738±0.0173
GlutPred	64.80±0.99	76.60±0.28	31.84±0.49	74.90±0.32	0.3194±0.0087
iGlu-Lys2	50.4±0.88	95.2±0.14	—	88.38±0.15	0.5098±0.0072
MDDGlutar2	67.7	61.9	—	63.8	0.28
RF-GlutarySite	74.9	69.7	71.2	72.3	0.45
PUL-GLU	66.56±0.73	86.43±0.28	70.71±0.45	79.88±0.29	0.5384±0.69
PUL-GLU3	71.69	75.07	32.66	74.58	0.3533

¹ SVM trained on 590 positive samples and 1199 randomly extracted negative samples.

² The values of Pre were not reported for iGlu–Ly and MDDGlutar, therefore, no comparison could be made with respect to this parameter.

³ PUL-GLU was performed on the entire training dataset.

We compared PUL-GLU with existing glutarylation site predictors. As shown in Table 2, in the training dataset, PUL-GLU reaches the highest MCC values of 0.5384 by 10-fold cross-validation. Although iGlu-Lys achieved the highest value of Sp (95.2%), the value of Sn (50.4%) was much lower than that of PUL-GLU (66.6%). It suggests that iGlu-Lys tends to identify a query lysine site as a non-glutaryllysine, and can predict less glutaryllysine sites than PUL-GLU. Moreover, the Sn value of MDDGlutar (66.7%) is slightly higher than that of PUL-GLU (66.6%), but the Sp value of MDDGlutar (61.9%) is much lower than that of PUL-GLU (86.4%). It indicates that PUL-GLU can predict more non-glutaryllysine sites than MDDGlutar at a similar level of Sn. As PUL-GLU and GlutPred were trained on the same training dataset with the same features, the better performance of PUL-GLU suggested that by using the extracted reliable non-glutarylated lysine sites to train model, the prediction performance has been improved effectively. In short, PUL-GLU outperforms the current glutarylation site predictors remarkably on the training dataset.

3.2. Comparison of PUL- GLU with Other Predictors on the Independent Test Set

To further evaluate the effectiveness of PUL-GLU, it was compared with the other current methods on the independent test set. It should be pointed out that RF-GlutarySite [12] did not provide a shared web-server. Hence, RF-GlutarySite was not compared with PUL-GLU. The compared results of existing predictors are shown in Table 3. Although iGlu-Lys achieved the best performance on the independent test dataset, the prediction results were overestimated. In fact, the training set of iGlu-Lys contains all of the samples of our independent test set; whereas PUL-GLU, GlutPred and MDDGlutar were trained and tested on the same dataset. As PUL-GLU was trained by the PU learning algorithm, the performance of PUL-GLU outperforms GlutPred and MDDGlutar. The results of the independent test and cross-validation both demonstrated that PUL-GLU could be an effective predictor for the prediction of glutarylation sites.

Table 3.

Comparison with other predictors on the independent test dataset.

Methods	Sn(%)	Sp(%)	Pre(%)	ACC(%)	MCC
SVM1	51.79	76.87	22.66	73.97	0.2078
GlutPred	51.79	78.50	23.97	75.41	0.2238
iGlu-Lys	89.09	97.67	83.05	96.69	0.8416
MDDGlutar	49.09	84.62	29.03	80.58	0.2715
PUL-GLU	58.93	78.97	26.83	76.65	0.2785

¹ SVM trained on 590 positive samples and 1199 randomly extracted negative samples.

3.3. Prediction Server of PUL-GLU

Building a user-friendly online server can provide convenience for the related experimental researchers to further

investigate the molecular mechanisms of glutarylation. Therefore, PUL-GLU has been implemented as a web-server. The prediction server for PUL-GLU is available at http://bioinform.cn/pul_glu. The style of PUL-GLU is similar to the published webserver iGlu-Lys [10]. As shown in Fig. (1), PUL-GLU accepts single query protein or multiple query proteins in FASTA format. Or users can upload query proteins in FASTA format as a text document for the prediction of glutarylation sites. The predicted results will be written to a CSV-formatted file.

Fig. (1).

The prediction interface of the web-server PUL-GLU. (A higher resolution / colour version of this figure is available in the electronic copy of the article).

3.4. Prediction of the Most Likely Glutaryllysine in Non-annotated Lysine Residues

As mentioned earlier, there are 646 experimentally validated glutarylation sites and 3926 non-validated lysine sites in the training dataset. However, the non-validated lysine residues may contain some glutarylation sites which have not been experimentally identified yet. To find the most likely glutarylation sites from those non-validated lysine residues, all 3926 non-validated lysine sites in the training dataset have been re-predicted by PUL-GLU algorithm. The top 20 most likely glutaryllysine in non-validated lysine residues are listed in Table 4. Here, we just give a possible hypothesis, it remains to be experimentally identified whether those lysine residues can be glutarylated or not. The completed prediction results are given in Supplementary Material S2 and may provide clues for studying glutarylation sites.

Table 4.

The top 20 most likely glutarylation sites in non-validated lysine sites.

Uniprot_AC	Site	SVM Score	Uniprot_AC	Site	SVM Score
P32020	432	3.04	Q8BMS1	411	1.63
Q8BMS1	413	2.82	Q8BWT1	211	1.63
P42125	242	2.51	P26443	386	1.62
Q8BMS1	414	2.48	Q8BMS1	262	1.60
Q8C196	906	2.36	P54869	342	1.59
P32020	442	2.25	Q8C196	856	1.56
Q8C196	908	2.21	Q9D819	238	1.53
Q61425	206	1.93	Q8BMS1	249	1.53
Q61176	39	1.65	Q9D172	155	1.53
Q8BMS1	284	1.64	Q61425	202	1.52

Conclusion

In this study, we developed a bioinformatics tool named PUL-GLU for the prediction of glutarylation sites using the PU learning algorithm and multiple sequence features. To the best of our knowledge, this is the first time PU learning has been applied to predict the glutarylation sites. Experimental results have shown that PUL-GLU outperformed the current glutarylation site predictors. A web-server for PUL-GLU was built, which could provide a great convenience for experimental researchers to investigate glutarylation.

SUPPLEMENTARY MATERIAL

Supplementary material is available on the publisher’s website along with the published article.

Authors' contributions

Zhe Ju wrote the manuscript and was involved in all the experimental steps. Shi-Yun Wang constructed the online web-server of PUL-GLU. Both the authors approved the final version of this manuscript.

Ethics Approval and Consent to Participate

Not applicable.

Human and Animal Rights

No Animals/Humans were used for studies that are the basis of this research.

Consent for Publication

Not applicable.

Availability of Data and Materials

The data supporting the findings of the article is available in the Data repository at 123.206.31.171/GlutPred/, reference number [9].

Funding

This work was supported by the National Natural Science Foundation of China (No. 11701390); the Natural Science Foundation of Liaoning Province (No. 2019-BS-187); and the Scientific Research Fund Project in Liaoning Province Department of Education (No. JYT19027).

CONFLICT OF INTEREST

The authors declare no conflict of interest, financial or otherwise.