CN104484580B - Antibacterial peptide Activity Prediction method based on Multi-label learning - Google Patents

Abstract

Then antibacterial peptide Activity Prediction method based on Multi-label learning obtains corresponding moment characteristics according to physico-chemical properties coding, collectively forms the characteristic vector of peptide sequence by extracting the corresponding aminoacid ingredient of peptide sequence.The characteristic vector of every peptide sequence is made up of two parts, and one is aminoacid ingredient, and two be that the moment characteristics extracted are encoded based on physico-chemical properties.Using the Multi-label learning algorithm computational minimization transformation matrix W of least square, then each mark output of sample to be tested can be drawn by transformation matrix W, obtaining prediction class label vector set according to each mark output closes.The activity for quick and precisely predicting antibacterial peptide sequence is closed according to class label vector set.It is specific therefore, it is possible to the shape that obtains peptide sequence all angles, so as to quick, accurate, automatic marking antibacterial peptide activity.

Description

Antimicrobial Peptide Activity Prediction Method Based on Multi-label Learning

technical field

The invention relates to biomedical engineering, in particular to a method for predicting antimicrobial peptide activity based on multi-label learning that can quickly, accurately and automatically mark antimicrobial peptide activity.

Background technique

Antimicrobial peptides are small molecular polypeptides involved in innate immunity, generally composed of 20 to 60 amino acid residues, and these active peptides have broad-spectrum and high-efficiency bactericidal activity against bacteria. With the deepening of people's research, it is found that these antibacterial peptides have a strong killing effect on some fungi, protozoa, viruses and cancer cells. The wide range of biological activities of antimicrobial peptides shows its good application prospects in medicine.

Determining the activity of antimicrobial peptides by experimental means, whether based on in vivo or in vitro techniques, is not only very time-consuming, but also expensive. At present, researchers have proposed more than ten kinds of antimicrobial peptide predictors. However, these tools are basically used to judge whether the peptide molecule has antibacterial properties, or whether it belongs to the antimicrobial peptide family, and no further research has been made on the specific activity of antimicrobial peptides. predict. Most of them design binary classification models to judge whether peptide molecules belong to antimicrobial peptides; or the proposed method can realize the activity prediction of antimicrobial peptides, but it is limited to 5 kinds of activities, and the prediction accuracy needs to be further improved. Most of the existing methods are binary classification models, which can only be used for antimicrobial peptide identification.

Contents of the invention

Based on this, it is necessary to provide a method for predicting the activity of antimicrobial peptides based on multi-label learning that can quickly, accurately and automatically mark the activities of antimicrobial peptides.

A method for predicting antimicrobial peptide activity based on multi-label learning, comprising the following steps:

Extracting the amino acid composition corresponding to the peptide sequence, and obtaining the corresponding moment feature vector x according to the amino acid composition, wherein the moment feature vector x is used to describe the shape characteristics of each angle of the peptide sequence;

Using the multi-label learning algorithm and calculating the minimized transformation matrix W according to the formula W=(X ^T X) ^-1 X ^T Y, where the class label vector of x is set to be y=[y ₁ ,y ₂ ,...,y _c ] ^T ; the formula for minimizing the transformation matrix W is min||XW-Y||; c is the number of category labels, X represents the training sample matrix, Y represents the class label matrix corresponding to the training sample, and each row vector corresponds to a sample ; Then for the sample x to be tested, its output to each mark is f(x, y)=xW;

According to each label output f(x,y)=xW, obtain the prediction class label vector set h(x)={y|f(x,y)≥0, y∈{1,2,...,c}}.

In one of the embodiments, the step of extracting the amino acid composition and moment feature vector x corresponding to the peptide sequence includes:

The amino acid sequence is digitally coded according to the physical and chemical property indicators of the amino acid;

Convert each amino acid residue of the amino acid sequence into a numerical sequence in one-to-one correspondence;

Calculate the moment feature vector x for the whole, N-terminal and C-terminal of the peptide sequence according to the numerical sequence, wherein the N-terminal refers to the first 5 amino acids of the peptide sequence, and the C-terminal refers to the last 5 amino acids of the peptide sequence.

In one embodiment, the moment feature vector x includes a first-order origin moment, a second-order central moment, a third-order central moment, and a fourth-order central moment.

In one of the embodiments, the class label vector is y=[y ₁ ,y ₂ ,...,y _c ] ^T where y _i =1 means that the sample x has the class label i; y _i =-1 means that the sample x does not have class label i.

In one embodiment, it is judged whether X ^T X is reversible, and if not, the generalized inverse of X ^T X is used instead.

In one of the embodiments, it further includes optimizing the moment eigenvector x by using a genetic algorithm.

In one of the embodiments, the step of optimizing the moment eigenvector x using a genetic algorithm includes:

Select the population size;

encode chromosomes;

Select the fitness function fitness = Hamming loss + sorting loss + 1/10000*number of features;

Adopt elite selection, wherein, the elite selection is the best 2 individuals in the population of the previous generation directly brought into the next generation;

Select a hybridization ratio of 0.8;

When the fitness function value is basically unchanged, the evolution is terminated, and the corresponding moment feature vector set is selected at this time.

In one of the embodiments, when the population evolves to 150 generations, the value of the fitness function remains basically unchanged.

In one embodiment, the method for predicting antimicrobial peptide activity based on multi-label learning is evaluated by using Hamming loss, subset accuracy, sorting loss, coverage, one-bit error and average precision.

In one of the embodiments, the multi-label learning-based antimicrobial peptide activity prediction method is evaluated by using 10-fold cross-validation, and the calculation result is taken as the mean value of 20 times of cross-validation.

The above antimicrobial peptide activity prediction method based on multi-label learning extracts the amino acid components corresponding to the peptide sequence, and then obtains the corresponding moment features according to the physical and chemical attribute encoding, which jointly constitute the feature vector of the peptide sequence. The feature vector of each peptide sequence is composed of two parts, one is the amino acid composition, and the other is the moment feature extracted based on the physical and chemical attribute encoding. Using the least squares multi-label learning algorithm to calculate the minimum transformation matrix W, the label output of the sample to be tested can be obtained through the transformation matrix W, and the predicted class label vector set can be obtained according to the label output. Rapid and accurate prediction of activity of antimicrobial peptide sequences from ensembles of class label vectors. Therefore, the shape specificity of each angle of the peptide sequence can be obtained, so that the antimicrobial peptide activity can be marked quickly, accurately and automatically.

Description of drawings

Fig. 1 is the flow chart of antimicrobial peptide activity prediction method based on multi-label learning;

Fig. 2 is the physicochemical attribute value list of 20 kinds of amino acids;

Figure 3 is an evolutionary diagram of genetic feature selection.

detailed description

As shown in Figure 1, it is a flow chart of the antimicrobial peptide activity prediction method based on multi-marker learning.

A method for predicting antimicrobial peptide activity based on multi-label learning, comprising the following steps:

Step S110, extract the amino acid composition corresponding to the peptide sequence, and obtain the corresponding moment feature vector x according to the amino acid composition, wherein the moment feature vector x is used to describe the shape characteristics of each angle of the peptide sequence.

Step S110 includes:

The amino acid sequence is digitally coded according to the physical and chemical property indicators of the amino acid.

Each amino acid residue of the amino acid sequence is converted into a numerical sequence in one-to-one correspondence.

Calculate the moment feature vector x for the whole, N-terminal and C-terminal of the peptide sequence according to the numerical sequence, wherein the N-terminal refers to the first 5 amino acids of the peptide sequence, and the C-terminal refers to the last 5 amino acids of the peptide sequence.

The moment feature vector x includes a first-order origin moment, a second-order central moment, a third-order central moment, and a fourth-order central moment.

Specifically, the peptide sequence is composed of 20 amino acids, and a sequence from the N-terminal to the C-terminal with a length of L is expressed as follows:

P＝R ₁ R ₂ R ₃ R ₄ …R _L

The amino acid composition (Amino Acid Composition, AAC, that is, the frequency of occurrence of 20 amino acids) and moment feature vectors can be extracted from the sequence. When extracting the moment feature vector, the sequence is first coded according to the physical and chemical property index of the amino acid. Assuming that H _i (i=1,2,...,20) is a certain physical and chemical attribute value of 20 amino acids, each amino acid residue in the protein sequence is converted into a value correspondingly, expressed as [H( R ₁ ), H(R ₂ ),...,H(R _L )]. For this numerical sequence, the moment eigenvalues can be calculated for the whole, N-terminal (first 5 amino acids) and C-terminal (last 5 amino acids), including the first-order origin moment (expectation), the second-order central moment (variance), and the third-order Central moment (skewness) and fourth-order central moment (kurtosis), these moment features can reflect the shape characteristics of the sequence from different angles. In this example, 5 kinds of amino acid physicochemical attributes will be used for amino acid encoding, and these 5 attributes are: hydrophilicity (hydropathy index), molecular weight (Molecular weight), PI, pK1 (alpha-COOH), pK2 ( NH3). The specific values are listed in Figure 2. After the above steps, each amino acid sequence can be expressed as a point in the 80-dimensional feature space, or a vector:

x＝[x ₁ ,x ₂ ,…,x ₈₀ ] ^T

The 80-dimensional feature space can be changed, such as 40-dimensional, 120-dimensional, etc.

Step S120, using the multi-label learning algorithm and calculating the minimized transformation matrix W according to the formula W=(X ^T X) ^-1 X ^T Y, where the class label vector of x is set as y=[y ₁ ,y ₂ ,..., y _c ] ^T ; the formula for minimizing the transformation matrix W is min||XW-Y||; c is the number of category labels, X represents the training sample matrix, Y represents the class label matrix corresponding to the training sample, and each row vector corresponds to a sample; then for the sample x to be tested, its output for each label is f(x,y)=xW.

The class label vector is _{y =} [y ₁ , _{y 2} , ^.

Determine whether X ^T X is reversible, if not, replace it with the generalized inverse of X ^T X.

Specifically, assuming that there are c (here c=10) category labels in total, the class label vector of sample x is y=[y ₁ ,y ₂ ,...,y _c ] ^T , where y _i =1 means that the sample has class Label i, y _i =-1 means that the sample does not have class label i. Then it is necessary to find a transformation matrix W that minimizes the empirical risk on the training sample set, ie min||XW-Y||;

Among them, X represents the training sample matrix, Y represents the class label matrix corresponding to the training sample, and each row vector corresponds to a sample. It can be obtained by the method of least squares:

W＝(X ^T X) ^-1 X ^T Y

If X ^T X is irreversible, replace it with its generalized inverse. Then for the sample x to be tested, its output for each label is:

f(x,y)=xW;

Then we can know the set of predicted class labels. .

Step S130, according to each label output f(x, y) = xW to obtain the predicted class label vector set h(x) = {y|f(x, y)≥0, y∈{1,2,...,c }}.

In this embodiment, the transformation matrix W adopts a value that minimizes the empirical risk on the training sample set, that is, it is used as a classifier parameter of f(x, y).

Because the moment eigenvector can reflect the shape characteristics of the peptide sequence from different angles, therefore, while the peptide sequence is known, the shape characteristics of each angle of the peptide sequence can also be obtained, and then the antimicrobial peptide activity can be marked quickly, accurately and automatically.

Based on all the above-mentioned embodiments, the antimicrobial peptide activity prediction method based on multi-label learning was evaluated by using Hamming loss, subset accuracy, sorting loss, coverage, one-bit error and average precision. In order to verify the validity of the antimicrobial peptide activity prediction method based on label learning.

In the verification process, in order to make the verification results accurate and reliable, the antimicrobial peptide activity prediction method based on multi-label learning was evaluated by using ten-fold cross-validation, and the calculation results were taken as the average of 20 times of cross-validation.

Based on all the above embodiments, in the process of extracting moment feature vectors, in order to eliminate redundant features. To improve the classifier accuracy, therefore, the initial set of moment feature vectors needs to be optimized.

The antimicrobial peptide activity prediction method based on multi-label learning also includes optimizing the moment feature vector x by using a genetic algorithm.

The step of optimizing the moment eigenvector x by using a genetic algorithm comprises:

Choose a population size. Generally, the population size is 50.

Chromosomal coding. Each chromosome is composed of 0-1 strings with a length of 80, 1 indicates that the feature of the corresponding position is selected, and 0 indicates that the feature of the corresponding position is not included.

Select the fitness function fitness = Hamming loss + sorting loss + 1/10000*number of features.

Adopt elite selection, wherein, the elite selection is the best 2 individuals in the previous generation population directly brought into the next generation; others remain unchanged.

Select the hybridization ratio of 0.8; that is, (50-2)*0.8≈38 individuals are produced by hybridization, and the parent individuals for hybridization are selected by the championship method.

In addition to elite selection and hybridization, other offspring are generated through genetic mutation, the mutation method is uniform mutation, and the mutation probability is 0.1.

When the fitness function value is basically unchanged, the evolution is terminated, and the corresponding moment feature vector set is selected at this time.

When the population evolves to 150 generations, the value of the fitness function remains basically unchanged.

Please combine with Figure 3. After the population has evolved for 150 generations, the fitness function value basically does not change. At this time, the feature set corresponding to the best result contains 40 features. In the original feature space and the optimized feature space, ten-fold cross-validation is performed on the data set, and the test results are shown in the following table:

It can be seen from the above table that after optimization, the number of features is reduced by half, and the indicators of each project have been improved instead. Among them, ↓ means the smaller the better, ↑ means the bigger the better. It shows that the genetic algorithm optimization strategy proposed in this example has very good performance. As a computer-aided tool, it can completely become an effective supplement to biological experiments, greatly improve the efficiency of drug development, and reduce costs at the same time.

Based on all the above-mentioned embodiments, the implementation process of the antimicrobial peptide activity prediction method based on multi-label learning is:

Obtain antimicrobial peptide sequences. Antimicrobial peptide sequences correspond to 10 activities. Each peptide sequence has at least one of these 10 activities, and at most seven activities at the same time.

The first step is to extract the moment eigenvectors.

The sequence is digitally coded according to the physicochemical property index of the amino acid. Each amino acid residue of the peptide sequence is then converted into a numerical value in a one-to-one correspondence. This sequence of values allows calculation of moment eigenvectors for the whole, N-terminus and C-terminus, respectively. These moment eigenvectors can reflect the shape characteristics of peptide sequences from different angles. In this example, five kinds of amino acid physicochemical properties are used for amino acid encoding.

Then, the transformation matrix W is sequentially calculated according to the multi-label learning algorithm of least squares.

Assuming that there are c (here c=10) category labels, the class label vector of sample x is y=[y ₁ ,y ₂ ,...,y _c ] ^T , where y _i =1 means that the sample has class label i, y _i =-1 means that the sample does not have the class label i. Then it is necessary to find a transformation matrix W that minimizes the empirical risk on the training sample set, ie min||XW-Y||;

Then, its output for each token can be derived as:

f(x,y)=xW;

It can be seen that the set of predicted class labels is:

h(x)={y|f(x,y)≥0, y∈{1,2,...,c}}.

Finally, according to the class label vector set h(x) corresponding to the moment feature vector x, the corresponding amino acid composition and the peptide sequence corresponding to the amino acid composition can be deduced. And because the moment eigenvector can reflect the shape characteristics of the peptide sequence from different angles, therefore, while the peptide sequence is known, the shape characteristics of each angle of the peptide sequence can also be obtained, and then the antimicrobial peptide activity can be marked quickly, accurately and automatically.

When the antimicrobial peptide activity prediction method based on multi-label learning is used in the development of antimicrobial peptide drugs, it can be used for molecular screening, and it can also inspire new uses of old drugs.

Based on all the above-mentioned embodiments, other multi-label learning algorithms can be used in addition to the least squares method.

The antimicrobial peptide activity prediction method based on multi-label learning extracts the amino acid components corresponding to the peptide sequence, and then obtains the corresponding moment features according to the physical and chemical attribute encoding, which together constitute the feature vector of the peptide sequence. The feature vector of each peptide sequence is composed of two parts, one is the amino acid composition, and the other is the moment feature extracted based on the physical and chemical attribute encoding. Using the least squares multi-label learning algorithm to calculate the minimum transformation matrix W, the label output of the sample to be tested can be obtained through the transformation matrix W, and the predicted class label vector set can be obtained according to the label output. Rapid and accurate prediction of activity of antimicrobial peptide sequences from ensembles of class label vectors. Therefore, the shape specificity of each angle of the peptide sequence can be obtained, so that the antimicrobial peptide activity can be marked quickly, accurately and automatically.

The technical features of the above-mentioned embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, should be considered as within the scope of this specification.

The above-mentioned embodiments only express several implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but should not be construed as limiting the patent scope of the invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (9)

1. A method for predicting antimicrobial peptide activity based on multi-label learning, comprising the following steps: Extract the amino acid composition corresponding to the peptide sequence, and obtain the corresponding moment feature vector x according to the amino acid composition, wherein the moment feature vector x is used to describe the shape characteristics of each angle of the peptide sequence; The sequence is digitally coded; each amino acid residue of the amino acid sequence is converted into a numerical sequence one by one; according to the numerical sequence, the moment feature vector x is calculated for the whole, N-terminal and C-terminal of the peptide sequence, wherein the N-terminal refers to the peptide The first 5 amino acids of the sequence, the C-terminal refers to the last 5 amino acids of the peptide sequence Using a multi-label learning algorithm and calculating the minimized transformation matrix W according to the formula W=(X ^T X) ^-1 X ^T Y, where the class label vector of x is set to be y=[y ₁ ,y ₂ ,...,y _c ] ^T ; the formula for minimizing the transformation matrix W is min||XW-Y||; y ₁ , y ₂ ,..., y _c is the element value of the label vector, c is the number of category labels, X represents the training sample matrix, Y represents the class label matrix corresponding to the training sample, and each row vector corresponds to a sample; then for the sample x to be tested, the output of each label is f(x,y)=xW; According to each label output f(x,y)=xW, obtain the prediction class label vector set h(x)={y|f(x,y)≥0, y∈{1,2,...,c}}. 2. the antimicrobial peptide activity prediction method based on multi-label learning according to claim 1, is characterized in that, described moment feature vector x comprises 1st order origin moment, 2nd order central moment, 3rd order central moment and 4th order central moment . 3. The antimicrobial peptide activity prediction method based on multi-label learning according to claim 1, wherein the class label vector is y in y=[y ₁ , y ₂ ,..., _{y c} ] ^T =1 means sample x has class label i; y _i =-1 means sample x does not have class label i. 4. The antimicrobial peptide activity prediction method based on multi-label learning according to claim 1, characterized in that, it is judged whether X ^T X is reversible, if not, the generalized reverse substitution of X ^T X is used. 5. The antimicrobial peptide activity prediction method based on multi-label learning according to claim 1, further comprising optimizing the moment feature vector x by using a genetic algorithm. 6. the antimicrobial peptide activity prediction method based on multi-label learning according to claim 5, is characterized in that, the described step that adopts genetic algorithm to optimize described moment characteristic vector x comprises: Select the population size; encode chromosomes; Select the fitness function fitness = Hamming loss + sorting loss + 1/10000*number of features; Adopt elite selection, wherein, the elite selection is the best 2 individuals in the population of the previous generation directly brought into the next generation; Select a hybridization ratio of 0.8; When the fitness function value is basically unchanged, the evolution is terminated, and the corresponding moment feature vector set is selected at this time. 7. The antimicrobial peptide activity prediction method based on multi-label learning according to claim 6, wherein the population evolves to 150 generations, and the value of the fitness function is basically unchanged. 8. The antimicrobial peptide activity prediction method based on multi-label learning according to any one of claims 1-7, characterized in that, using Hamming loss, subset accuracy, sorting loss, coverage, one-bit error and average The precision rate was used to evaluate the antimicrobial peptide activity prediction method based on multi-label learning. 9. The antimicrobial peptide activity prediction method based on multi-label learning according to claim 8, wherein the antimicrobial peptide activity prediction method based on multi-label learning is evaluated by using ten-fold cross-validation, and the calculation result is obtained 20 times Cross-validated mean.