Use Chou’s 5-Step Rule to Classify ProteinModification Sites with Neural Network

Lysine malonylation is a novel-type protein post-translational modification and plays essential roles in many biological activities. Having a good knowledge of malonylation sites can provide guidance in many issues, including disease prevention and drug discovery and other related fields. )ere are several experimental approaches to identify modification sites in the field of biology. However, these methods seem to be expensive. In this study, we proposed malNet, which employed neural network and utilized several novel and effective feature description methods. It was pointed that ANN’s performance is better than other models. Furthermore, we trained the classifiers according to an original crossvalidation method named Split to Equal validation (SEV). )e results achieved AUC value of 0.6684, accuracy of 54.93%, andMCC of 0.1045, which showed great improvement than before.


Introduction
Protein post-translational modification (PTM) is a key mechanism to regulate protein functions by the covalent and generally enzymatic modification. Hundreds of types of PTMs have been discovered and reported in this field [1][2][3][4][5][6]. ey played vital roles in influencing almost all aspects of cell biology and pathogenesis, e.g., gene expression, cell division, and cell signaling [7][8][9][10]. As one of a newly identified PTM type in both eukaryotic and prokaryotic, Lysine malonylation (Kmal) has wide connections with various biological processes, where some Kmal sites are potentially associated with cancer. erefore, it is very critical to identify and understand Kmal sites in the studies of biology and diseases [11][12][13][14][15][16].
Different from traditional experimental methods, computational approach of PTMs provides a fast and low-cost strategy for experimental designing, as the PTM site prediction can be abstracted as a typical classification problem. Meanwhile, there are a list of machine learning approaches which can be successful utilized in this field. For instance, Logistic Regression (LR) was used in ModPred for 23 different modifications using sequence-based features, physicochemical properties, and evolutionary features as features [17].
Musite, which is a general and kinase-specific protein phosphorylation site prediction, applied Support Vector Machine (SVM) models utilizing three types of features: K-nearest neighbor score, disorder scores, and amino acid frequencies. In the previous work, Wei et al. presented PhosPred-RF for predicting phosphorylation sites, which utilized the evolutionary information features from position specific scoring matrices [18,19]. Deep learning method was also applied in this area, such as the recently published tool MusiteDeep, which was utilized for general and kinase-specific phosphorylation site prediction [20].
In this article, we employed artificial neural network (ANN) classifier based on Stochastic Gradient Descent (SGD) algorithm for protein Kmal site prediction. It was pointed that we investigated a wide range of types of feature extraction schemes and finally choose EBAG + Profile and EAAC methods to train our predictors. Furthermore, we employed another two classifiers, including SVM and kNN for comparative experiments, with the same feature extraction schemes. Besides, in view of the fact that the Kmal prediction problem can be regarded as a binary classification problem, we adopted the original SEV method to solve the inherent imbalance problem of positive and negative samples in the training set. e result of our experiment shows ANN performed better than SVM and kNN predictors. Overall, ANN can be a useful tool for identifying Kmal sites.

Methods and Materials
ere are 4 steps in our research, which is depicted in Figure 1. e first step is dataset construction and procession, where the training set and testing set were generated. And then we encode the dataset according to two feature extraction methods. e next step is to construct three models, which were trained by the training set. Finally, all the classifiers would be tested by crossvalidation and independent testing set. Five assessment metrics would be used to evaluate the performance of our classifiers.

Dataset Construction.
In this work, we derived lots of Kmal peptides from mice and human species according to a proteomic assay. Referring to the procedure established by Chen et al. [21], we built a benchmark dataset. ere are 67322 Kmal sites in the training set, where the sites with high confidence were regarded as positive sites and other lysine residues were collected as negative sites. For each sample, we extracted 31-residue peptides (−15 to +15) with the lysine site in the center from the representatives. As a result, 5023 positive peptides and 62299 negative peptides were retained for further analyses. We can easily find that the ratio of positive and negative samples in training set approaches 1 to 12. erefore, the trained models would be tested by Split to Equal validation and independent test, where 35955 peptides (including 2798 positive peptides and 33157 negative peptides) were employed as the independent testing dataset.

EBAG + Profile
Encoding. EBAG + Profile encoding is an integration scheme consisted of two different feature encoding method utilized by Han et al. [22]. One is Encoding Based on Attribute Grouping (EBAG) [23], which divides 20 types of amino acids into 5 groups depending on various physical and chemical properties. Table 1 shows the grouping result based on EBAG. e other encoding method is Profile, which counts the frequency of each amino acid residue occurred in the protein peptides. And then the frequency was used as the representation of this residue in the sequence so that each peptide with 31 residues can be transformed into a 31-dimension vector. e way to combine EBAG and Profile is replacing source amino acid peptides into EGBA sequence and then encoding the sequence according to the Profile method. As a result, a peptide with 31 residues was converted to a vector of 31 dimensions as the EBAG + Profile encodings.

EAAC Encoding.
A typical encoding scheme of prediction for PTMs was AAC encoding [24], which reflects the frequency of 20 amino acid residues surrounding the modification site. In this work, we coded each amino acid by the EAAC method proposed by Zhen et al. [25], which is based on the AAC encoding. As 8-size window continuously slides from the N-terminus to C-terminus of each peptide in the dataset, the EAAC method counted the frequency of the 20 amino acid residues. Accordingly, the dimension of features can be calculated as follows: where L p refers to the length of each peptide, L s is the length of sliding windows, and D eaac is the dimension of feature vector. As we set L s to 8, a peptide with 31 residues would be corresponded to 24 (31 -8 + 1) sliding windows and converted to a matrix of 24 × 20 dimensions.

Construction of Classifiers
ANN is a traditional machine learning algorithm that was widely utilized in lysine PTM prediction applications. In this article, we construct an ANN model with four layers, i.e., input layer, output layer, and two hidden layers. e input layer received the feature sequence generated from different encoding method. e two hidden layers owe both 100 neurons and adopt "reLu" as their activation function. e output layer owes a single unit, outputting the probability score of each site.    [20]. SVM can transform the samples into a high-dimensional feature space and then construct an Optimal Separating Hyperplane (OSH) to maximize its distance from the closest training samples. Here, based on Tensorflow [26] and Scikit-learn [27], we employed SCV as our SVM model, where the applied kernel function was linear kernel.

K-Nearest Neighbor
Algorithm. kNN algorithm is another widely employed algorithm that calculates the distances of samples to cluster them [28]. If we obtain the training dataset D � {v 1 , v 2 , ..., v n }and a testing sample x, we can utilize KNN to calculate the distances between x and all the instances in D. erefore, as the nearest neighbor (shortest distance) in the training dataset, the query sample will be assigned to the same class. In this work, we also construct a kNN model implemented by Tensorflow and Scikit-learn. e parameters of our kNN were set to their default values.

Crossvalidation Methods
In general, when the classification model is built, researchers will divide the dataset into two parts as training set and testing set. Process of dataset usage partition is depicted in Figure 2. To make full use of the training set samples, we usually train the model through 10-fold crossvalidation. e samples in training set are divided into training set and validating set by the crossvalidation method. And then, the training set is used to train the model, while the validating set is used to verify the effect of the model and obtain the validation scores. After the crossvalidation is completed, the trained model will pass the testing set to evaluate its performance and get the testing scores.
In view of the fact that the classifier is always more sensitive to the category containing more samples and less sensitive to the category containing fewer samples in binary classification problems, it is necessary to preprocess the training set before inputting the data with unbalanced positive and negative samples into the classifier. In the previous work, we proposed a new feature extraction method named SEV (Split to Equal Validation), which can well solve the problem of imbalanced training samples in PTM sites prediction research studies. In this experiment, we also adopt the SEV method and at the same time used 10-fold crossvalidation to do comparative experiments. e working flow of SEV is as follows.
(Note: the pos means the positive and the neg means the negative in Figure 3.) e experiment shown in Figure 3 is a part of the whole experiment, which corresponds to the process of using training set to obtain classifier and its validation scores after the dataset is divided into training set and testing set in Figure 3. In details, SE validation consists of five steps. Assuming that the ratio of negative samples to positive samples in the training set is close to n:1, (1) the first step is to divide the negative samples into n groups; (2) in the second step, each positive sample is combined with the positive sample to generate n balanced subsets; (3) subsequently, model 1 will be trained by subset 1 and verified by subset 2; model 2 will be trained by subset 2 and verified by subset 3, and so on; (4) according to the n balanced subsets, n models were trained and validated; (5) finally, each model will be tested by independent testing sets, and the average of their scores will be utilized to evaluate their performance.

Performance Assessment of Predictors.
ere are a set of four metrics [29] directly that are often utilizing to quantitatively evaluate the performance of predictors: Sn (sensitivity), also known as TPR (True Positive Rate), reflects the proportion of true positive samples (TP) determined by the model to all the positive samples in the dataset; Sp (specificity), also known as TNR, reflects the proportion of true negative samples judged by the model in all negative samples; Acc (accuracy) is the proportion of correct samples determined by the model to the total samples; and MCC (Mathew's Correlation Coefficient) reflects the correlation coefficient between the actual predicted samples and the expected predicted samples: where TP, FP, TN, and FN represent the true positives, false positives, false negatives, and true negatives, respectively. Scientific Programming e pre means precision and the rec means recall in the classification model. On the contrary, ROC curves and AUC value were also adopted to evaluate the performance of the predictors.

Performance of the ree Classification Models Based on Different Encoding Schemes.
In this study, we firstly constructed three machine learning models, i.e., ANN, SVM, and kNN algorithm, and then trained them according to Amino Acid Composition (AAC) encoding scheme that considered the hydrophobicity and charged character of the amino acid. Split to Equal Validation (SEV) and independent training sets were utilized to assess the performance of models above, where AUC, Acc, MCC, Sn, and Sp were adopted as assessment metrics. e results of the independent testing results were depicted in Table 2.
Based on the results of this experiment, we speculate that feature extraction schemes are very important factors affecting the final classification accuracy. erefore, we adopt the EBAG + Profile encoding method, which utilized the physical and chemical properties of amino acids. EAAC encoding method, which is based on AAC encoding and accords to the probability of occurrence of specific amino acids in the peptide sequence, was also adopted in this experiment. e testing scores were depicted in Tables 3 and  4, respectively.
As we know, a larger AUC value means that the current classification algorithm is more likely to rank positive samples in front of negative samples, so as to get better classification results. erefore, it is obvious that classifiers under EAAC encoding scheme plays better performance than the other two schemes for getting higher AUC values. And other experimental results such as MCC and Acc value and EAAC encoding also obtained higher scores and showed similar advantages than others.
On the contrary, there is one thing certain that ANN's classification effect is better than SVM and kNN under EAAC encoding scheme. As for independent test, when taking EAAC, the AUC value of ANN is 0.7471, while SVM and kNN algorithms obtain the AUC value of 0.6322 and 0.6317. All of these results in Tables 2-4 show that different types of classifiers have great impact on the prediction performance. In this work, ANN is the best classifier.

Comparing the Results of SEV with 10-Fold Crossvalidation and Scaling the Number of Training Samples by SEV.
In this research, we utilize the SEV verification method to preprocess the training samples, and thus train the classifier model. In addition, on the premise that other conditions remain unchanged, we used 10-fold crossvalidation instead of SEV to do the same experiment. e experimental results are shown in Table 5: Among them, the experiment is based on the neural network model, using 10-fold crossvalidation and SEV methods, respectively. It can be seen that although the Acc equivalent of the 10-fold cross is too high, its AUC value does not perform well. is is because, in the case of extremely unbalanced positive and negative samples, the classifier will guess the kind of samples with higher probability in the training set, but its classification ability is not outstanding. e SEV verification method can overcome this problem well. Although Acc equivalence is not as good as 10-fold cross verification, SEV has more advantages in the AUC value, which can best represent the classification ability of the model in the real sense.
More importantly, in order to further explore the influence of imbalance between positive and negative samples in the training set, we used the SEV method to scale the training samples. We verified the fact that unbalanced training data would finally lead to very low Sn and very high Sp of the classifier, further causing these evaluation metrics to lose its significance. In order to further explore how far  reaching the positive and negative sample ratios affect the classifier's performance, we calculate the five metrics, i.e., AUC, Acc, MCC, Sn, and Sp by adjusting different positive and negative sample ratios from 1 : 1 to 1 : 12 under SEV. As can be seen from Table 6, as the proportion of positive and negative samples in the training set increases, the AUC value of the model gradually decreases from 0.7471 to about 0.7000, the MCC value gradually decreases to about 0.1300, while the ACC value continuously increases to 89.43%. is also verifies our previous conclusion: when the positive and negative samples of the training set are extremely unbalanced, the classifier will tend to guess the kind with more samples, but its classification effect is not good.
In addition, Table 6 also shows another information, that is, when the ratio of positive and negative samples in the training set reaches 1 : 9, the AUC value of the classifier will also tend to be stable, only fluctuating around 0.7000 without further decline. is means that, in this experiment, although the ratio of positive and negative samples has been changing towards a more unbalanced direction, the performance of the classifier will not decrease indefinitely, but will tend to be stable after reaching a certain threshold.
In a word, through this experiment, we can further verify that the positive and negative sample ratios have farreaching influence on the results in the binary classification problem, and the SEV method can solve this problem well.

Conclusions
e currently available PTM prediction approaches are mainly based on ML that requires preprocessing amino acid data into digital features. Here, we adopted two feature extraction schemes according to different ideas of physical and chemical characteristics and occurrence frequency and then constructed three ML classifiers, while the application of the SEV method solves the problem of sample imbalance in binary classification.
e results not only showed that     feature extraction methods and classifier types play important roles on prediction results but also indicated the direction of our next work. In addition to proposing new feature encoding schemes, more classifiers can be utilized in this field, including depth learning (DL) classifiers such as CNN or RNN. In addition, the SEV method will be improved and applied to the new machine learning model. In total, the outstanding performance of ML in prediction of Kmal sites suggests that computational methods can be applied widely to this field.

Data Availability
e data utilized to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare no competing interests.