Biological sequence prediction method and system based on correlation entropy kernel sparse representation model

By stacking bidirectional long and short memory networks and neural tangent nuclear models, combining the relevant entropy loss function and L2,1 norm optimization method, the problems of model complexity and noise sequence impact in biological sequence prediction are solved, and more efficient feature extraction and Prediction accuracy.

CN118609644BActive Publication Date: 2024-11-05YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411080541.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-05
Estimated Expiration
2044-08-08

AI Technical Summary

Technical Problem

The existing biological sequence prediction method has complex model structure, high computing resource requirements, and it is difficult to integrate sequence evolution information. The noise sequence has a great impact on the classifier, which reduces the generalization ability of the classifier.

Method used

The stacked bidirectional long and short memory network and neurotangent nuclear model are used, combined with the correlation entropy loss function and the L2,1 norm optimization method, and the characteristics of the therapeutic peptide sequence are extracted, and the combination of semiquadratic optimization and L2,1 norm optimization is trained. Nuclear sparse represents a classifier, reducing the impact of noise sequences and improving the generalization ability of the classifier.

Benefits of technology

A simpler model structure is realized, reducing the demand for computing resources, improving the prediction accuracy of therapeutic peptide sequences and the generalization ability of classifiers, and reducing the impact of noise sequences.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118609644B_ABST
    Figure CN118609644B_ABST
Patent Text Reader

Abstract

The present invention provides a biological sequence prediction method and system based on a correlation entropy kernel sparse representation model, including: preparing a biological sequence data set; inputting the biological sequence into a stacked bidirectional long short-term memory network to extract features of the biological sequence and obtain a second feature vector; inputting the second feature vector into a neural tangent kernel model to generate a kernel matrix; using the kernel matrix as input and using a correlation entropy loss function to train a kernel sparse representation classifier to generate a final classification model. In order to solve the problem of difficult feature extraction of sequences such as therapeutic peptides and the influence of noise sequences on classifiers, this solution uses a stacked bidirectional long short-term memory network and a neural tangent kernel to extract therapeutic peptide sequences and uses a correlation entropy loss function to train the classifier, which can effectively reduce the influence of noise sequences on the classifier and improve the generalization ability of the classifier.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer bioinformatics, and in particular relates to a biological sequence prediction method and system based on a correlation entropy kernel sparse representation model. Background Art

[0002] At present, the prediction of specific biological sequences such as therapeutic peptides mainly includes traditional methods (mass spectrometry-based technology, bioinformatics technology, cell experiments, etc.) and machine learning-based methods. Traditional biochemical experimental detection methods are known for their time-consuming characteristics. In order to supplement these methods, computational model methods based on machine learning have become an effective tool for improving the detection of therapeutic peptides. Traditional machine learning methods require the code to be input into a deep learning model, and cannot integrate sequence evolution information. In this regard, the applicant has proposed a specific biological sequence prediction method based on sequence homology and its system (publication number CN117953973B). The scheme proposes a fuzzy membership function based on sequence homology scoring to convert sequence data into feature vectors, which can integrate sequence evolution information. However, the model structure of the scheme is relatively complex, and the demand for computing resources is relatively high. It is more suitable for understanding the evolutionary history of the sequence, functional conservation, or evolutionary relationships between species. In this regard, the applicant continues to research and develop, and proposes a new biological sequence prediction method that can achieve the same effect, but the model structure is simpler and the computing resource requirements are smaller. Summary of the invention

[0003] The purpose of the present invention is to provide a new method for predicting biological sequences, solve the problem of difficulty in extracting sequence features such as therapeutic peptides and the influence of noise sequences on classifiers, use stacked bidirectional long short-term memory networks and neural tangent kernels to extract therapeutic peptide sequences, and use the relevant entropy loss function to train the classifier, and use the joint optimization method of HQ optimization and L2,1 norm optimization to optimize the objective function, which can effectively reduce the influence of noise sequences on the classifier and improve the generalization ability of the classifier.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A biological sequence prediction method based on a correlation entropy kernel sparse representation model, the method comprising:

[0006] Prepare biological sequence datasets;

[0007] Inputting the biological sequence into a stacked bidirectional long short-term memory network to extract features of the biological sequence and obtain a second feature vector;

[0008] Inputting the second eigenvector into a neural tangent kernel model to generate a kernel matrix;

[0009] The kernel matrix is ​​used as input to train the kernel sparse representation classifier using the relevant entropy loss function to generate the final classification model.

[0010] In the above-mentioned biological sequence prediction method based on the correlation entropy kernel sparse representation model, the stacked bidirectional long short-term memory network includes a language model pre-trained with protein data and a stacked BiLSTM encoder;

[0011] The biological sequence is first subjected to feature extraction by the language model to obtain a first feature vector;

[0012] The first feature vector and its corresponding biological sequence are input into the stacked BiLSTM encoder to obtain a second feature vector related to the corresponding biological sequence.

[0013] In the above-mentioned biological sequence prediction method based on the correlation entropy kernel sparse representation model, the pre-training data of the language model is the protein sequence in the Pfam database;

[0014] The language model includes a forward LSTM layer and a reverse LSTM layer to extract the biological sequence forward and backward and output a first feature vector.

[0015] In the above-mentioned biological sequence prediction method based on the correlation entropy kernel sparse representation model, the stacked BiLSTM encoder includes three layers of Bilstm and one linear layer, the first eigenvector passes through the three layers of Bilstm in sequence, and finally outputs the second eigenvector through the linear layer.

[0016] In the above biological sequence prediction method based on the correlation entropy kernel sparse representation model, the correlation entropy criterion and the L2,1-matrix norm are used to construct the objective function for training the kernel sparse representation classifier.

[0017] In the above biological sequence prediction method based on the correlation entropy kernel sparse representation model, the objective function of the kernel sparse representation classifier is as follows:

[0018] (6)

[0019] Among them, the X matrix represents the training set feature matrix, Yi is the test sample (Y is the test set feature matrix, there are m samples, i represents the i-th one), and g is a function g(x) = exp(-x 2 / 2σ 2 ), σ is the standard deviation. λ is the regularization term ||A || 2,1 The coefficient (a constant between 0 and 1, set based on experience), A i represents the i-th row of matrix A.

[0020] n is the number of biological sequence training samples, and m is the number of biological sequence test samples;

[0021] ϕ(Yi)∈R 1×dp , ϕ(X)∈R n×dp , A ∈ R m×n , dp is the dimension after projection.

[0022] In the above biological sequence prediction method based on the correlation entropy kernel sparse representation model, the matrix A is made sparse in the following way:

[0023] (7)

[0024] n is the number of biological sequence training samples, and m is the number of biological sequence test samples;

[0025] A 2 i,j Represents the element A in the matrix i,j The square of (the element in the i-th row and j-th column of matrix A).

[0026] In the above biological sequence prediction method based on the correlation entropy kernel sparse representation model, during the training process of the kernel sparse representation classifier, a joint optimization method of semi-quadratic optimization and L2,1 norm optimization is used to solve the objective function.

[0027] In the above biological sequence prediction method based on the correlation entropy kernel sparse representation model, the method of solving the objective function by the joint optimization method includes:

[0028] Introduce an auxiliary variable τ and transform the optimization problem through a convex conjugate function φ;

[0029] By fixing the value of z, the auxiliary objective function JbCKSRC is optimized to obtain the auxiliary parameter τi;

[0030] After τi is determined, these values ​​are fixed and the objective function is redefined;

[0031] (11)

[0032] Re-express equation (11) as:

[0033] (12)

[0034] and is a diagonal matrix,

[0035] Solve equation (12) for the partial derivative with respect to the matrix A:

[0036] (15)

[0037] Update the weight matrix A by solving (15).

[0038] A biological sequence prediction system based on a correlation entropy kernel sparse representation model deploys a trained prediction model and performs biological sequence prediction in the following manner:

[0039] receiving a biological sequence to be predicted;

[0040] Input the biological sequence into a language model pre-trained with protein data to obtain a first feature vector;

[0041] Inputting the first feature vector and its corresponding biological sequence into a stacked BiLSTM encoder to obtain the second feature vector;

[0042] Input the second eigenvector into the neural tangent kernel model to generate a kernel matrix;

[0043] The kernel sparse representation classifier outputs predictions based on the kernel matrix.

[0044] The advantages of the present invention are:

[0045] 1) This scheme uses stacked bidirectional long short-term memory networks and neural tangent kernels to extract therapeutic peptide sequences, and uses pre-trained stacked bidirectional long short-term memory networks to extract features from amino acid sequences, providing a new technical means to achieve biological sequence prediction without feature encoding.

[0046] 2) This scheme introduces the neural tangential kernel NTK, and inputs the output values ​​of the above-mentioned stacked bidirectional long short-term memory network into the NTK model to extract deep features. It can identify specific patterns, motifs or domains in the sequence, and analyze the complex interactions and regulatory relationships in the sequence, such as transcription factor binding sites, signal peptide sequences, etc. By extracting the multi-level features of the sequence at multiple levels, understanding its features and patterns at different levels, it can be used to reveal the biological functions and regulatory mechanisms of the sequence or design biological molecules with specific functions.

[0047] 3) The stacked bidirectional long short-term memory network used in this scheme takes the original sequence and the output of the pre-trained language model as the combined input, so that the stacked BiLSTM encoder can combine the evolutionary information from the language model and the information of the therapeutic peptide sequence itself to generate a feature vector that contains richer information and can more comprehensively describe the characteristics of the therapeutic peptide sequence, so that the neural tangential kernel can construct an information-rich kernel matrix to facilitate the kernel sparse representation classifier to learn and predict;

[0048] 4) This scheme uses the L2,1-matrix norm to make the matrix A sparse, which helps to improve the generalization ability and interpretability of the model;

[0049] 5) This scheme uses the correlation entropy loss function to optimize the model parameters of the kernel sparse representation classifier. The correlation entropy loss function can estimate the distribution of prediction errors and approximate the actual prediction error probability density function. In the optimization process, it suppresses the problem of gradient jitter caused by noise data to the gradient descent algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the training process of the biological sequence prediction method based on the correlation entropy kernel sparse representation model of the present invention;

[0051] Figure 2 It is a schematic diagram of the biological sequence prediction method based on the correlation entropy kernel sparse representation model of the present invention;

[0052] Figure 3 It is a schematic diagram of stacking bidirectional long short-term memory networks in the biological sequence prediction method based on the correlation entropy kernel sparse representation model of the present invention;

[0053] Figure 4 It is a flow chart of the prediction process of the biological sequence prediction method based on the correlation entropy kernel sparse representation model of the present invention;

[0054] Figure 5 The comparison results of this method and existing methods on the CPP dataset are shown;

[0055] Figure 6 The comparison results of this method and existing methods on the QSP dataset are shown;

[0056] Figure 7 The comparison results of this method and its variants with existing methods on the CPP dataset are shown in Figure 2.

[0057] Figure 8 The figure shows the comparison results of our method and its variants with existing methods on the QSP dataset. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] This solution discloses a biological sequence prediction method and system based on the correlation entropy kernel sparse representation model, such as Figure 1 and Figure 2 As shown, the method comprises the following steps:

[0060] S1. Use a trained stacked bidirectional long short-term memory network to represent the characteristics of a biological sequence in the form of a string, and use a stacked bidirectional long short-term memory network composed of a pre-trained language model and a stacked BiLSTM encoder. The biological sequence referred to in this embodiment is mainly a therapeutic peptide sequence.

[0061] like Figure 3 As shown, the language model includes a forward LSTM layer and a reverse LSTM layer to extract the therapeutic peptide sequence forward and backward and output the first feature vector. The pre-training data of the language model is the protein sequence in the Pfam database, which contains more than 20 million protein sequences.

[0062] The stacked BiLSTM encoder includes three layers of Bilstm and one linear layer. The first feature vector passes through the three layers of Bilstm in sequence, and finally outputs the second feature vector through the linear layer.

[0063] Bilstm is the abbreviation of Bidirectional Long Short-Term Memory Network.

[0064] The pre-trained language model can adopt the SBE-PLM pre-trained language model that has been trained in the prior art, and this solution can directly use it. The stacked BiLSTM encoder is constructed for this solution and trained using the Pfam database and the pre-trained language model.

[0065] The therapeutic peptide sequence is first feature extracted by a language model to obtain a first feature vector; then the first feature vector and its corresponding therapeutic peptide sequence are input into a stacked BiLSTM encoder to obtain a second feature vector.

[0066] The output of the language model contains the evolutionary information between the therapeutic peptide sequence and the protein family. The output obtained by the language model and the corresponding therapeutic peptide sequence are input into the stacked BiLSTM encoder constructed in this scheme, so that the stacked BiLSTM encoder can combine the evolutionary information from the language model and the information of the therapeutic peptide sequence itself to generate a feature vector that contains richer and more comprehensive descriptions of the characteristics of the therapeutic peptide sequence. Moreover, doing so can reduce the model's dependence on specific training data, thereby reducing the risk of overfitting.

[0067] Stacked bidirectional long short-term memory networks can predict the sequence of therapeutic peptides i amino acids at positions, by minimizing the cross entropy loss, it estimates all amino acids, which are amino acids a i in the forward and reverse directions, and the log predicted probability is given by:

[0068] (1)

[0069] in, and denote the probabilities of the forward and reverse LSTMs respectively. The peptide sequence is represented in 1-hot format. Assume that the hidden state of the language model (LM) at position i is , the 1-hot format of this position is , using the ReLU nonlinear function to calculate the input vector of the stacked BiLSTM encoder:

[0070] (2)

[0071] in, and are the parameters of the stacked BiLSTM encoders, respectively.

[0072] Through the pre-trained language model and stacked BiLSTM encoder, the vector embedding of the therapeutic peptide sequence, i.e., the second feature vector, is obtained, and the vector embedding of the therapeutic peptide sequence is input into the neural tangent kernel for processing.

[0073] S2. Input the extracted second eigenvector into the neural tangent kernel NTK to generate the kernel matrix

[0074] During training, the NTK remains constant and converges to a fixed kernel in the infinite width limit. Imagine a fully connected neural network with multiple hidden layers and an output layer, with layer 0 representing the input layer and layer L representing the output layer. For an input vector x, the output of each layer in the network can be recursively calculated as follows:

[0075] (3)

[0076] Among them, h represents the hth layer of the neural tangent kernel network, f (h) (x) represents the output of the X variable at the hth layer without the activation function, o (h) (x) is f (h) (x) Output after the activation function.

[0077] W(h)∈R dh×dh−1 is the weight matrix of the hth layer, and dh represents the number of neurons in the hth layer. In this network, the activation function γ(·) used is the rectified linear unit (ReLU), and there is a scaling factor cγ to match it.

[0078] The output layer of the neural network can be expressed as:

[0079] (4)

[0080] The connection weights of the neural network ω = [W(1), ..., W(H)] are randomly initialized before training. To optimize the network parameters, the gradient descent method is used with a small learning rate to minimize the squared loss. When the number of neurons in the hidden layer approaches infinity (i.e., d1, ..., dH → ∞), the hidden state remains unchanged and equal to its initial value. In this case, NTK converges the kernel to a specific limiting form. The NTK kernel in this limiting form can be explicitly expressed as K NTK (x, x'), which represents the multi-layer structure of the neural network:

[0081] (5)

[0082] Among them, K NTK (x, x') represents the NTK kernel, which represents a neural network with a multi-layer structure. The NTK kernel provides a more powerful representation capability for complex data. In this scheme, a three-layer neural network, i.e., L = 3, is considered, and the corresponding NTK kernel of the therapeutic peptide sequence is generated based on this.

[0083] S3. Take the kernel matrix generated by NTK as input;

[0084] The final classification model is generated by the Correntropy-based Kernel Sparse Representation Classiffer (CKSRC) built using the correlation entropy function. Correlation entropy is a method to measure the similarity between the predicted output and the actual label, which helps to reduce the impact of noise.

[0085] Define an objective function that includes the relevant entropy loss and the L2,1-matrix norm. The L2,1-matrix norm is used to promote the sparsity of the weight matrix A to improve the generalization ability of the model. In this process, ϕ ( Yi ) and ϕ ( X ) represent the representation of the test sample and the training sample in the high-dimensional space, and A is the weight matrix of the model. The objective function is expressed as follows:

[0086] (6)

[0087] Among them, ϕ(Yi) ∈R 1×dp , ϕ(X)∈R n×dp , A ∈ R m×n . dp is the dimension after projection. m is the number of test samples, and n is the number of training samples.

[0088] L2,1-matrix norm Make the matrix A sparse by:

[0089] (7).

[0090] S4.CKSRC optimization

[0091] The objective function is non-convex for the weight matrix A, and direct optimization is difficult. This solution uses a semi-quadratic optimization method to optimize CKSRC. An auxiliary variable τ is introduced, and the optimization problem is transformed through a convex conjugate function φ, which converts the non-convex problem into a more tractable form. The specific implementation method is as follows:

[0092] Assume that there exists a convex conjugate function φ, related to the function g(z):

[0093] (8)

[0094] In this case, τ represents a scalar variable. By fixing the value of z, the auxiliary objective function JbCKSRC is optimized to obtain the auxiliary parameters τi. The purpose of this step is to find a value of τ that maximizes the objective function in the extended parameter space. The objective function (6) is expressed in an extended parameter space as follows:

[0095] (9)

[0096] Here, τ = [τ1, τ2, . . . , τn] represents the auxiliary vector, and by keeping A unchanged and maximizing JbCKSRC with respect to τ, the auxiliary parameters τi can be obtained:

[0097] (10)

[0098] Once τi is determined, these values ​​are fixed and the objective function is redefined to further optimize the weight matrix A, resulting in the following objective function:

[0099] (11)

[0100] The above equation can be further reformulated as:

[0101] (12)

[0102] in, and is a diagonal matrix:

[0103] (13)

[0104] and

[0105] (14)

[0106] Solve equation (12) for the partial derivative with respect to A

[0107] (15)

[0108] By solving 15 to update the weight matrix A, the matrix equation can be effectively processed to optimize the model parameters.

[0109] Finally, the reconstruction residual of the test sample is calculated to evaluate the prediction performance of the model. The smaller the reconstruction residual, the more accurate the prediction of the model:

[0110] (16)

[0111] c represents the category, Y i Represents a test sample, such as Figure 2 As shown, the hyperplane of category 1 is: ;

[0112] The class 2 hyperplane is: , the closest hyperplane criterion is used to determine which category it belongs to. The closer the distance, the smaller the reconstruction residual, and the corresponding sample is identified as this category.

[0113] In the above method, first, a pre-trained language model is prepared, then a stacked BiLSTM encoder is constructed, and the stacked BiLSTM encoder is trained using the aforementioned language model and biological sequence dataset. The language model and the stacked BiLSTM encoder form a stacked bidirectional long short-term memory network that can directly extract feature representations of biological sequences in string form. Then a neural tangent kernel NTK is prepared. This network does not need to be trained and is only responsible for generating a kernel matrix for the input. Then a kernel sparse representation classifier based on correlation entropy is constructed, and the input of the classifier is the kernel matrix output by NTK. In the process of training the kernel sparse representation classifier, the stacked bidirectional long short-term memory network and NTK are unchanged. The objective function used in the training process combines the correlation entropy criterion and the L2,1-matrix norm, and the joint optimization method of semi-quadratic optimization and L2,1 norm optimization is used to solve the objective function, and finally a trained kernel sparse representation classifier is obtained.

[0114] So far, a biological sequence prediction system based on the correlation entropy kernel sparse representation model has been obtained. The system deploys a prediction model consisting of a stacked bidirectional long short-term memory network, NTK, and a kernel sparse representation classifier. Figure 4 As shown, the biological sequence to be predicted is predicted in the following way:

[0115] receiving a biological sequence to be predicted;

[0116] Input the biological sequence into the language model to obtain a first feature vector;

[0117] Inputting the first feature vector and its corresponding biological sequence into a stacked BiLSTM encoder to obtain the second feature vector;

[0118] Input the second eigenvector into the neural tangent kernel model to generate a kernel matrix;

[0119] The kernel sparse representation classifier outputs predictions based on the kernel matrix.

[0120] like Figure 5 and Figure 6 As shown, in order to verify the effect of this scheme, this scheme method is compared with the existing method. The vertical axis represents the values ​​of evaluation indicators ACC-accuracy, SN-sensitivity, Spec-specificity, MCC-Matthews correlation coefficient, and AUC-area under the receiver operating characteristic curve.

[0121] CKSRC-NTK, indicating the method of this scheme;

[0122] PPTPP, a Physicochemical Property-based Therapeutic Peptide Predictor

[0123] SVM,is a prediction model built using support vector machines;

[0124] LSTM, a prediction model built using long short-term networks;

[0125] ITP-Pred, Interpretable Therapeutic Peptide Prediction

[0126] from Figure 5 and Figure 6 It can be seen that this method has the highest prediction effect and shows the best prediction effect.

[0127] Furthermore, this scheme also compares this scheme method and its variants with existing methods:

[0128] CKSRC-RBF, similar to this method, but uses RBF (radial basis kernel function) instead of neural tangential kernel NTK;

[0129] CKSRC-Lap, similar to this method, but uses Lap (Laplacian kernel function) instead of the neural tangential kernel NTK;

[0130] NTKCKSRC-Poly, similar to this method, but uses Poly (polynomial kernel function and linear kernel function) instead of the neural tangential kernel NTK;

[0131] CKSRC-Linear, similar to this method, but uses Linear instead of the neural tangential kernel NTK;

[0132] KSRC-NTK, similar to this method, but uses the mean square error function instead of the correlation entropy loss function;

[0133] KSRC-RBF, similar to this method, but uses the mean square error function instead of the correlation entropy loss function, and uses RBF instead of the neural tangential kernel NTK;

[0134] KSRC-Lap, similar to this method, but uses the mean square error function instead of the correlation entropy loss function, and uses Lap instead of the neural tangential kernel NTK;

[0135] KSRC-Poly, similar to this method, but uses the mean square error function instead of the correlation entropy loss function, and uses Poly instead of the neural tangential kernel NTK;

[0136] KSRC-Linear is similar to this method, but uses the mean square error function instead of the correlation entropy loss function, and uses Linear instead of the neural tangent kernel NTK.

[0137] from Figure 7 and Figure 8 It can be seen that the method of this scheme is the best among all methods, including the existing technology and variants of the method of this scheme, and the variants of the method of this scheme also have obvious advantages over most of the existing technology methods.

[0138] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0139] Although the terms biological sequence, therapeutic peptide sequence, stacked bidirectional long short-term memory network, first eigenvector, second eigenvector, kernel sparse representation classifier, language model, relevant entropy loss function, neural tangent kernel, etc. are used more frequently in this article, the possibility of using other terms is not excluded. The use of these terms is only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitation is contrary to the spirit of the present invention.

Claims

1. A biological sequence prediction method based on a correlation entropy kernel sparse representation model, characterized in that: The method includes: Prepare biological sequence datasets; Inputting the biological sequence into a stacked bidirectional long short-term memory network to extract features of the biological sequence and obtain a second feature vector; Inputting the second eigenvector into a neural tangent kernel model to generate a kernel matrix; Taking the kernel matrix as input and using the relevant entropy loss function to train the kernel sparse representation classifier to generate a final classification model; The stacked bidirectional long short-term memory network includes a language model pre-trained with protein data and a stacked BiLSTM encoder; The biological sequence is first subjected to feature extraction by the language model to obtain a first feature vector; Input the first feature vector and its corresponding biological sequence into a stacked BiLSTM encoder to obtain a second feature vector for the corresponding biological sequence; The stacked BiLSTM encoder includes three layers of Bilstm and one linear layer. The first feature vector passes through the three layers of Bilstm in sequence, and finally outputs the second feature vector through the linear layer.

2. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 1 is characterized in that: The pre-training data of the language model is the protein sequence in the Pfam database; The language model includes a forward LSTM layer and a reverse LSTM layer to extract the biological sequence forward and backward and output a first feature vector.

3. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 1 is characterized in that: The objective function for training the kernel sparse representation classifier is constructed using the relevant entropy criterion and the L2,1-matrix norm.

4. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 3 is characterized in that: The objective function of the kernel sparse representation classifier is as follows: Among them, the X matrix represents the training set feature matrix, Yi is the test sample, and g is a function g(x) = exp(-x 2 / 2σ 2 ),σ is the standard deviation; λ is the regularization term ||A|| 2,1 The coefficient of i represents the i-th row of matrix A; n is the number of biological sequence training samples, and m is the number of biological sequence test samples; ϕ(Yi)∈R 1×dp , ϕ(X)∈R n×dp , A ∈ R m×n , dp is the dimension after projection.

5. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 4 is characterized in that: The matrix A is made sparse as follows: n is the number of biological sequence training samples, and m is the number of biological sequence test samples; A 2 i,j Represents the square of the element Ai,j in the matrix.

6. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 5 is characterized in that: During the training process of the kernel sparse representation classifier, a joint optimization method of semi-quadratic optimization and L2,1 norm optimization is used to solve the objective function.

7. The biological sequence prediction method based on the correlation entropy kernel sparse representation model according to claim 6 is characterized in that: The joint optimization method for solving the objective function includes: Introduce an auxiliary variable τ and transform the optimization problem through a convex conjugate function φ; By fixing the value of z, the auxiliary objective function JbCKSRC is optimized to obtain the auxiliary parameter τi; After τi is determined, these values ​​are fixed and the objective function is redefined; Put the equation Rephrased as: and is a diagonal matrix For the equation Solve for the partial derivatives with respect to the matrix A: Update the weight matrix A by solving the above equation.

8. A biological sequence prediction system based on a correlation entropy kernel sparse representation model, characterized in that: Deploy the trained prediction model and perform biological sequence prediction in the following ways: receiving a biological sequence to be predicted; Input the biological sequence into a language model pre-trained with protein data to obtain a first feature vector; The first feature vector and its corresponding biological sequence are input into the stacked BiLSTM encoder to obtain the second feature vector; Input the second eigenvector into the neural tangent kernel model to generate a kernel matrix; The kernel sparse representation classifier outputs prediction results based on the kernel matrix; The stacked BiLSTM encoder includes three layers of Bilstm and one linear layer. The first feature vector passes through the three layers of Bilstm in sequence, and finally outputs the second feature vector through the linear layer.