A method for predicting plant long noncoding RNA based on sequence features

Abstract

A plant long non-coding RNA prediction method based on sequence features, including positive and negative data set construction de-redundancy, sequence k-mer frequency feature extraction, feature fusion and selection, and classifier comparison. The construction de-redundancy of the positive and negative data sets is intended to eliminate the influence of sequence similarity on the prediction results. Sequence feature extraction realizes the extraction of gene sequence k-mer features. Feature fusion adds sequence length features and sequence GC content features on the basis of sequence k-mer feature extraction to fuse these three types of features, then uses chi-square test for optimal feature selection, and finally uses gradient boosting decision tree algorithm to solve the identification and classification problem of plant lncRNA. The present invention deeply excavates gene sequence features, effectively fuses these features and selects them, and provides a feasible method for the problem of how to accurately identify plant lncRNA from a large number of transcripts.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics and natural language processing, and specifically relates to a method for predicting plant long-chain non-coding RNA based on sequence features. Background Art

[0002] In recent years, a large amount of biomedical data has shown that lncRNA plays an important role in various life activities of human beings. Among them, the prediction problem of plant lncRNA, as an intersection of bioinformatics and natural language processing, has also attracted much attention. However, its accuracy is far from satisfactory. The development of computer programs that can accurately identify plant lncRNA is still considered a valuable and necessary task. This work combines various subtasks in natural language processing, such as positive and negative sample data collection and processing, sequence feature extraction and fusion, optimal feature subset selection and comparison of the results of each classifier.

[0003] In the prediction of plant lncRNA, the computer solves the prediction problem of plant lncRNA based on the sequence features obtained from sufficient data or big data. Although researchers have proposed many methods, plant lncRNA prediction has always been an open problem, and the accuracy and robustness of the proposed models need to be further improved. Plant lncRNA prediction algorithms can be divided into the following categories: 1) models based on gradient boosting decision trees; 2) models based on Bayesian; 3) models based on support vector machines; 4) models based on random forests; 5) models based on adaptive enhancement, etc. At present, the adaptive enhancement model is a research hotspot. However, a large number of studies have shown that focusing only on the attention model does not seem to be enough. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a solution to the problem of predicting plant long non-coding RNA based on sequence features. The present invention is based on a gradient boosting decision tree model, and solves the technical difficulties in plant lncRNA prediction by fusing plant lncRNA sequence features and selecting optimal sub-features. It is necessary to first perform redundancy processing on the positive and negative sample sequences, select gene sequences with a similarity of less than 60% to reduce the impact of sequence similarity on the prediction results, then perform feature extraction and feature fusion on the retained sequences, and then select the optimal feature subset based on the fused features.

[0005] A plant long non-coding RNA prediction method based on sequence features includes redundancy removal of positive and negative data sets, sequence k-mer frequency feature extraction, feature fusion and selection, and classifier comparison. The redundancy removal of positive and negative data sets is aimed at eliminating the influence of sequence similarity on the prediction results. Sequence feature extraction realizes the extraction of gene sequence k-mer frequency features. Feature fusion adds sequence length features and sequence GC content features on the basis of sequence k-mer frequency feature extraction and then fuses these three types of features. Then, the chi-square test is used to select the optimal feature. Finally, the gradient boosting decision tree algorithm is used to solve the identification and classification problems of plant lncRNA.

[0006] The specific steps include:

[0007] Step 1: Given the positive and negative sample data sets, perform redundancy removal on them and retain gene sequences with similarity less than 60%;

[0008] Step 2, given the de-redundant positive and negative sample data sets, extract the sequence GC content features and sequence length features;

[0009] Step 3: Given the positive and negative sample data set after redundancy removal, use the sliding window to extract the k-mer subsequence. When k is 1, 2, 3, or 4, the sliding window diagram is as follows: Figure 3 shown.

[0010] k-mer refers to a subsequence of k bases in a DNA or RNA sequencing sequence. The base at each position in the subsequence can be any of the four bases A, U, C, and G. In this case, the number of k-mer subsequences is: 4 1 +4 2 +4 3 +...+4 k The value range of k is a positive integer. If k = 1, 2, 3, 4, a total of 340 k-mer subsequences can be extracted. Specifically, there are 4 1-mer series; 16 2-mer sequences; 64 3-mer sequences; 256 4-mer sequences.

[0011] Step 4, the present invention calculates the frequency characteristics of each k-mer subsequence of the positive and negative sample data according to the TF-IDF algorithm, and the calculation formula is:

[0012]

[0013]

[0014] F i =tf ij *idf i

[0015] Among them, Fi represents the frequency characteristic value of the i-th k-mer subsequence; tf ij represents the frequency of occurrence of the i-th k-mer subsequence in the j-th sequence of a data set; n i,j Indicates the number of occurrences of the i-th k-mer in the j-th sequence of the data set; ∑ k n k,j Indicates the number of occurrences of all k-mers in the jth sequence of the data set. The value range of k is k = 1, 2, 3, 4…, k is a positive integer; idf i represents the logarithm of the frequency of occurrence of the i-th k-mer subsequence in all sequences in the data set; |D| represents the number of all sequences in a data set; |j:t i ∈d j | represents the number of sequences containing the i-th k-mer subsequence; if the i-th k-mer is not in the data set, the denominator will be zero, so 1+|j:t is generally used. i ∈d j |.

[0016] In step 5, the k-mer frequency features corresponding to the positive and negative sample data sets, the sequence GC content features extracted in step 2, and the sequence length features are combined to obtain the initial features.

[0017] Step 6: Use the chi-square test to select the feature subset that can best distinguish positive and negative samples from the feature set containing all the initial features of positive and negative samples. These feature subsets are called optimal feature subsets. The chi-square test is a commonly used χ2 test. 2 The hypothesis testing method based on distribution can be used to analyze and compare the correlation between two categorical variables. Its fundamental idea is to compare the degree of agreement between the theoretical frequency and the actual frequency. The calculation formula is:

[0018]

[0019] Among them, A is the actual value, T is the theoretical value, and the chi-square value χ 2 It is used to measure the difference between the actual value and the theoretical value. The larger the chi-square value, the greater the deviation between A and T, and vice versa.

[0020] Step 7: Apply the GBDT classifier to the optimal feature subset as input to give the answer to the question. GBDT is an iterative decision tree algorithm. T = (x1, y1), (x2, y2), …, (x n ,y n ), x=(x1,...,x n ) is the input training data set, and the algorithm model of GBDT is:

[0021]

[0022] Among them, x represents the input, α t is the weight of the tth tree, h t (x) represents the tth Classification And Regression Trees (CART), and T is the number of trees built in the gradient boosting decision tree. The gradient boosting decision tree algorithm uses the forward distribution algorithm to first determine the initial value of the model F0(x), which is generally a constant. From this, we can know that the model formula for the mth step is:

[0023] F m (x) = F m-1 (x)+α m h m (x)

[0024] where F m-1 (x) is the current model. The newly added classification and regression tree h m (x) is obtained by minimizing the loss function:

[0025]

[0026] Where N represents the number of samples, y i is the input data set x i The corresponding label value, when x i Belongs to the positive class y i =1, otherwise y i = 0. L represents the loss function. In the two-classification model of GBDT, L is generally expressed as a logarithmic function, and the expression formula is:

[0027] L(y,F)=log(1+exp(-2yF))

[0028] The GBDT algorithm uses the gradient descent method to obtain the optimal model. The gradient descent direction is the loss function L in the current model F. m-1 The negative gradient value of (x):

[0029]

[0030] in, Represents the direction of gradient descent for F(x) in the loss function L, that is, the partial derivative of the loss function L; α m Obtained through linear search, the calculation formula is:

[0031]

[0032] The learning rate v can regulate the regularization of the gradient boosting decision tree, and the calculation formula is:

[0033] F m (x) = F m-1 (x)+v α m h m (x)

[0034] Among them, v represents the learning rate of this model, v∈(0,1]. The learning rate controls the speed and effect of parameter optimization of the model. The smaller the learning rate, the slower the model optimization speed.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention uses sequence-based plant lncRNA features as input to the GBDT classifier to construct a method for predicting plant lncRNA. Among them, the sequence-based plant lncRNA features include sequence GC content features, sequence length features, and sequence k-mer frequency features. Different from the traditional lncRNA prediction, this invention integrates the biologically relevant features of lncRNA and uses the chi-square test method to select the optimal feature subset therefrom, solving the problem of identifying and predicting plant lncRNA from massive gene sequences. The selected optimal feature subset provides input features for the GBDT classifier, allowing the classifier to accurately and efficiently predict plant lncRNA from a small number of sequence features.

[0037] (2) The present invention proposes a plant long non-coding RNA prediction model based on sequence features. On the basis of fusion sequence features, the chi-square test is used to achieve feature optimization. The optimal feature subset is selected from a large number of fused sequence features and input into the classifier to generate an answer. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The figure is a flowchart of the steps of the method for predicting plant long non-coding RNA based on sequence features in an embodiment of the present invention.

[0039] Figure 2 The figure is a flow chart of obtaining the optimal feature subset in an embodiment of the present invention.

[0040] Figure 3 Schematic diagram of a sliding window in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0042] A method for predicting plant long non-coding RNA based on sequence features, including construction and de-redundancy of positive and negative data sets, extraction of sequence GC content features, length features and k-mer frequency features, feature fusion and selection, and classifier comparison. Construction and de-redundancy of positive and negative data sets aims to eliminate the influence of sequence similarity on prediction results. Sequence feature extraction realizes the extraction of gene sequence k-mer features. Feature fusion adds sequence length features and sequence GC content features on the basis of sequence k-mer feature extraction, and then fuses these three types of features. Then, the chi-square test is used to select the optimal feature. Finally, the gradient boosting decision tree algorithm is used to solve the identification and classification problems of plant lncRNA. Figure 1 As shown in FIG. 1 , the entire model architecture includes the construction of positive and negative data sets to remove redundancy, sequence GC content feature extraction, sequence length feature extraction, sequence k-mer frequency feature extraction, feature fusion, optimal feature subset construction, and final prediction result generation.

[0043] The present invention constructs a plant lncRNA feature set based on sequence features, then applies the chi-square test as the optimal feature selection method to construct an optimal feature subset, and finally uses the optimal feature subset as the input of the prediction model to complete the plant lncRNA prediction task.

[0044] Based on the above plant lncRNA prediction model, the implementation method of the plant lncRNA model proposed in the present invention comprises the following steps:

[0045] Step 1: Given the positive and negative sample data sets, perform redundancy removal on them and retain gene sequences with similarity less than 60%.

[0046] Step 2: Given the de-redundant RNA sequence, extract the sequence GC content features and sequence length features respectively.

[0047] Step 3: Extract k-mer subsequences from the RNA sequence after redundancy removal, where k is 1, 2, 3, or 4. The schematic diagram of the sliding window is shown in Figure 2 shown.

[0048] In bioinformatics, k-mer refers to a subsequence of k bases in a DNA or RNA sequence. The base at each position in the subsequence can be any of the four bases A, U, C, and G. In this case, the number of k-mer subsequences is: 4 1 +4 2 +4 3 +...+4 k. The value range of k is a positive integer. If the value of k is k=1,2,3,4, a total of 340 k-mer subsequences can be extracted. Specifically, there are 4 1-mer series: A%,U%,C%,G%; 16 2-mer sequences: AA%,AG%,AC%,AU%,GA%,GG%,GC%,...; 64 3-mer sequences: AAA%,AAG%,AAC%,AAU%,AGA%,AGG%…; 256 4-mer sequences: AAAA%,AAAG%,AAAC%,AAAU%,AAGA%,AAGG%,AAGC%,AAGU%,….

[0049] Step 4: Calculate the importance of each k-mer subsequence according to the TF-IDF algorithm, and the calculation result is the frequency feature of the k-mer. The calculation formula is:

[0050]

[0051]

[0052] F i =tf ij *idf i

[0053] Among them, F i Indicates the frequency feature value of the i-th k-mer subsequence. tf ij represents the frequency of occurrence of the i-th k-mer subsequence in the j-th sequence of a data set; n i,j Indicates the number of occurrences of the i-th k-mer in the j-th sequence of the data set; ∑ k n k,j It indicates the number of occurrences of all k-mers in the jth sequence of the data set. The value range of k is k=1,2,3,4…, k is a positive integer, and idf i represents the logarithm of the frequency of occurrence of the i-th k-mer subsequence in all sequences in the data set; |D| represents the number of sequences in a data set; |j:t i ∈d j | represents the number of sequences containing the i-th k-mer subsequence; if the i-th k-mer is not in the corpus, the denominator will be zero, so 1+|j:t is generally used. i ∈d j |.

[0054] Step 5: Calculate and save the k-mer frequency features corresponding to the positive and negative sample data sets, and then combine them with the GC content features and length features of the corresponding data to obtain the initial features.

[0055] Step 6: Use the chi-square test to select the feature subset that can best distinguish positive and negative samples from the feature set containing all the initial features of positive and negative samples. These feature subsets are called optimal feature subsets. 2 The hypothesis testing method based on distribution can be used to analyze and compare the correlation between two categorical variables. Its fundamental idea is to compare the degree of agreement between the theoretical frequency and the actual frequency. The calculation formula is:

[0056]

[0057] Among them, A is the actual value, T is the theoretical value, and the chi-square value χ 2 It is used to measure the difference between the actual value and the theoretical value. The larger the chi-square value, the greater the deviation between A and T, and vice versa.

[0058] Step 7: Apply the classifier to give the answer to the question.

[0059] like Figure 2 Shown is a flowchart of constructing the optimal feature subset of the present invention.

[0060] The present invention uses sequence features to construct a plant lncRNA prediction method based on sequence features. Starting from the perspective of sequence features, the GC content features, sequence length features and k-mer frequency features based on the sequence are fused, and the optimal feature subset is selected from the fused features using the chi-square test. The method applies statistical hypothesis testing to the selection of gene sequence features. The GBDT algorithm is used as a classifier, the optimal feature subset is input, and multiple rounds are performed. Each round of iteration generates a weak classifier, which is trained on the basis of the residual result of the previous round of classifiers. The data is classified by continuously reducing the residual generated by the training process. The problem model is solved by machine learning. This model maximizes the use of sequence information. The extracted features are completely derived from the gene sequence, and no other external feature information needs to be cited. On this basis, accurate prediction results are obtained. The present invention deeply explores the sequence structure of the gene, effectively extracts the GC content and length features of the fused sequence, and the k-mer frequency features, and provides a feasible way to solve the prediction problem of plant lncRNA.

[0061] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for predicting plant long non-coding RNA based on sequence features, characterized in that: The method includes constructing positive and negative data sets to remove redundancy, extracting sequence k-mer frequency features, merging and selecting features, and comparing classifiers; Among them, the construction of positive and negative data sets removes redundancy and eliminates the influence of sequence similarity on the prediction results. Sequence feature extraction realizes the extraction of gene sequence k-mer features. Feature fusion is to add sequence length features and sequence GC content features on the basis of sequence k-mer frequency feature extraction and then fuse these three types of features. Then, the chi-square test is used to select the optimal feature. According to the p-value of different features, the optimal features are selected to form feature subsets, and then the optimal subset features are input into the gradient boosting decision tree classifier to solve the identification and classification problems of plant lncRNA. The method specifically comprises the following steps: Step 1: Given the positive and negative sample data sets, perform redundancy removal on them and retain gene sequences with similarity less than 60%; Step 2, given the de-redundant positive and negative sample data sets, extract the sequence GC content features and sequence length features; Step 3, extracting k-mer subsequences given the de-redundant positive and negative sample data sets; Step 4: Calculate the importance of each k-mer subsequence in the positive and negative sample data sets according to the TF-IDF algorithm, and the calculation result is the frequency feature of the k-mer; Step 5, the k-mer frequency features corresponding to the positive and negative sample data sets, the sequence GC content features extracted in step 2, and the sequence length features are combined to obtain the initial features; Step 6: Use the chi-square test to select the feature subset that can best distinguish positive and negative samples from the feature set containing all the initial features of positive and negative samples. These feature subsets are called optimal feature subsets. Step 7: The selected optimal feature subset is input into the classifier as input to obtain the final classification result; The classifier in step 7 is a gradient boosted decision tree GBDT classifier; In the gradient iteration algorithm of GBDT, let T = {(x1,y1),(x2,y2),…,(x n ,y n )},x=(x1,...,x n ) is the input training data set, and the algorithm model of GBDT is: Among them, x represents the input, α t is the weight of the tth tree, h t (x) represents the t-th classification and regression tree, and T is the number of trees constructed in the gradient boosting decision tree. The gradient boosting decision tree algorithm uses the forward distribution algorithm. First, the initial value F0(x) of the model is determined as a constant. The model formula of the m-th step is: F m (x)=F m-1 (x)+α m h m (x) Among them, F m-1 (x) is the current model; the newly added classification regression tree h m (x) is obtained by minimizing the loss function: Where N represents the number of samples; y i is the input data set x i The corresponding label value, when x i Belongs to the positive class y i =1, otherwise y i =0; L represents the loss function. In the binary classification model of GBDT, L is generally expressed as a logarithmic function, and the expression formula is: L(y,F)=log(1+exp(-2yF)) The GBDT algorithm uses the gradient descent method to obtain the optimal solution. The gradient descent direction is the loss function L in the current model F m-1 The negative gradient value of (x): in, Represents the direction of gradient descent for F(x) in the loss function L, that is, the partial derivative of the loss function L; α m Obtained through linear search, the calculation formula is: The regularization of the decision tree is improved by regulating the gradient through the learning rate v, and the calculation formula is: F m (x)=F m-1 (x)+v α m h m (x) Among them, v represents the learning rate of this model, v∈(0,1]; the learning rate controls the speed and effect of parameter optimization of the model. The smaller the learning rate, the slower the model optimization speed.

2. A method for predicting plant long non-coding RNA based on sequence features according to claim 1, characterized in that: In step 3, k-mer refers to a subsequence of length k bases in a sequencing sequence; the base at each position in the subsequence is any one of the four bases A, U, C, and G, and the total number of k-mer subsequences is 4 1 +4 2 +4 3 +...+4 k , k is a positive integer.

3. A method for predicting plant long non-coding RNA based on sequence features according to claim 1, characterized in that: In step 4, the calculation formula is: F i =tf ij *idf i Among them, tf ij represents the frequency of occurrence of the i-th k-mer subsequence in the j-th sequence of a data set; n i,j Indicates the number of occurrences of the i-th k-mer in the j-th sequence of the data set; ∑ k n k,j Indicates the number of occurrences of all k-mers in the jth sequence of the data set. The value range of k is k = 1, 2, 3, 4…, k is a positive integer; idf i represents the logarithm of the frequency of occurrence of the i-th k-mer subsequence in all sequences in the data set; |D| represents the number of all sequences in a data set; |j:t i ∈d j | represents the number of sequences containing the i-th k-mer subsequence; F i Represents the frequency feature value of the i-th k-mer subsequence.

Images