Enhancer prediction model training method, identification method, and readable storage medium

Abstract

The application discloses an enhancer prediction model training method, comprising the following steps: obtaining a sample set of enhancer sequences; performing sequence feature extraction on the sample set according to a feature extraction algorithm to obtain a sequence feature set; fusing and optimizing the sequence feature set to obtain an optimized feature set; and introducing the optimized feature set into a classification machine for training to obtain an enhancer prediction model. The technical problems that existing enhancers cannot be predicted and the accuracy of the trained model is not high are solved. Through this method, the advantages of feature extraction can be fully utilized, the problems of dimension disaster and model overfitting can be avoided, the robustness of the model is improved after training in the classification machine, and the accuracy of the identification of enhancers and their strengths is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to an enhanced sub-model training method, a recognition method, and a readable storage medium. Background Technology

[0002] Currently, amino acid sequences, i.e., their primary structure, are relatively easy to obtain through the study of corresponding genes. However, due to the complexity of protein folding in its higher-order structure, it is difficult to predict the function of proteins from their amino acid sequences. In recent years, with the development of artificial intelligence technology, researchers have used AI methods to study protein amino acid sequences in order to predict protein function and higher-order structural features. Existing research methods include conventional identification methods such as support vector machines and random forests. However, most current training models and the resulting prediction models cannot specifically predict or identify enhancers and their features. Furthermore, they cannot address the issue of high-dimensional feature sets generated by various feature extraction methods, which leads to the curse of dimensionality in the feature space and overfitting of the model, resulting in low accuracy of the trained model. Summary of the Invention

[0003] In view of the above-mentioned deficiencies of the prior art, the present invention discloses a method to solve the technical problems of existing methods that cannot predict enhancers in a targeted manner and the low accuracy of training models.

[0004] To achieve the above objectives, according to embodiments of the present invention, the present invention provides a method for training an enhancer prediction model, characterized in that it includes:

[0005] Obtain a sample set of enhanced subsequences;

[0006] The sample set is subjected to sequence feature extraction according to the feature extraction algorithm to obtain a sequence feature set;

[0007] The sequence feature set is fused and optimized to obtain an optimized feature set;

[0008] The optimized feature set is imported into a classifier for training to obtain an augmentation prediction model.

[0009] Optionally, a sample set of enhancer sequences is obtained, wherein the sample set includes: a sample set of enhancer sequences, non-enhancer sequences, strong enhancer sequences, and weak enhancer sequences, all of which are fully labeled.

[0010] Optionally, the sample set is subjected to sequence feature extraction according to a feature extraction algorithm to obtain a sequence feature set, wherein the feature extraction algorithm includes: Kmer, Nucleotide Binary Profiles (NBP) and Accumulated Nucleotide Frequency (ANF) algorithms.

[0011] Optionally, the sequence feature set is fused and optimized to obtain an optimized feature set, including the following steps:

[0012] The sequence feature sets are fused to obtain a fused feature set;

[0013] The fused feature set is sorted by weight according to the LGBM algorithm to obtain the selected feature set;

[0014] The selected feature set is split according to the feature extraction algorithm to obtain the Laplacian matrix L corresponding to the feature extraction algorithm. m ;

[0015] According to the Laplace matrix L m This yields an optimized feature set.

[0016] Optionally, the fused feature set is optimized according to the LGBM algorithm to obtain an optimized feature set, wherein the parameters of the LGBM algorithm are:

[0017] The boosting type is GBDT, the learning rate is 0.05, the maximum depth of each tree is 10, the maximum number of leaf trees per tree is 30, 50% of the sample dataset is randomly used each time a new tree is built, 80% of the samples are randomly used each time bagging is performed, and bagging is performed once every 12 trees are built.

[0018] Optionally, the optimized feature set is imported into a classifier for training to obtain an augmenter prediction model, wherein the classifier is a multi-view RBF neural network.

[0019] Optionally, the radial basis functions of the hidden layers of the RBF neural network are:

[0020]

[0021] Where, σ i is a free parameter, which is the width parameter of the function and controls the radial range of the function; xi is the feature point in the optimized feature set, and cj is the cluster center point of the optimized feature set.

[0022] Optionally, the optimized feature set is imported into a classifier for training to obtain an augmented prediction model, wherein the classifier is an RBF neural network, and the hyperparameters of the RBF neural network are determined according to the following steps:

[0023] According to the objective function W * Perform a grid search.

[0024] To minimize the difference between the prediction result and the annotation of the enhancer sequence sample set (e.g., the annotation can be information that can distinguish four different types of enhancer sequences, non-enhancer sequences, strong enhancer sequences, and weak enhancer sequences), the hyperparameters p, λ1, λ2 are obtained, where:

[0025]

[0026]

[0027] 0 < β m <1,

[0028] m = 1, 2, ..., M

[0029] in, L represents the Laplace matrix m The regularization term, The weighted link matrix W = [w1, w2, ..., w q ] p×q L2,1 norm term, β m Let L be the Laplace matrix. m ∈R N ×N The weights are Φ; Φ is the output matrix of the hidden layer, and its expression is as follows:

[0030]

[0031] The present invention also provides a method for identifying enhancers, characterized in that it includes:

[0032] Obtain target sequence samples;

[0033] The target sequence sample is identified using the enhancer prediction model as described in any one of claims 1-8, and the prediction result is obtained.

[0034] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the method steps as described in any one of claims 1-9.

[0035] The beneficial effects of this invention are:

[0036] Enhancer sequence samples are analyzed using a feature extraction algorithm to extract sequence features. These features are then fused and optimized. The optimized feature set is then imported into a classifier for training to obtain an enhancer prediction model. This approach fully leverages the advantages of feature extraction while avoiding the curse of dimensionality and overfitting. Training the model with the classifier improves its robustness and significantly enhances the accuracy of enhancingr and intensity identification. Attached Figure Description

[0037] Figure 1 A flowchart of the enhanced sub-model training method is provided for embodiments of the present invention;

[0038] Figure 2 A flowchart of the enhancer identification method is provided for embodiments of the present invention;

[0039] Figure 3 This invention provides a diagram showing the dimensionality reduction results of various feature subsets based on the LGBM algorithm in an embodiment of the invention.

[0040] Figure 4 The structural diagram of the RBF neural network provided by this invention;

[0041] Figure 5 This is a schematic diagram illustrating the effect of Comparative Embodiment 1 of the present invention;

[0042] Figure 6 This is a schematic diagram illustrating the effect of Comparative Embodiment 2 of the present invention;

[0043] Figure 7 This is a schematic diagram illustrating the effect of Comparative Embodiment 3 of the present invention. Detailed Implementation

[0044] Those skilled in the art will recognize that the embodiments described herein are for the purpose of helping the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention. Those skilled in the art should understand that embodiments of the invention can be provided as methods, systems, or computer program products. Therefore, the invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. The invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce implementations of the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. Obviously, those skilled in the art can make various modifications and variations to the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The embodiments described with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0046] like Figure 1 The present invention provides a method for training an enhancer prediction model, characterized by comprising: S1 obtaining a sample set of enhancer sequences;

[0047] S2 performs sequence feature extraction on the sample set according to the feature extraction algorithm to obtain a sequence feature set;

[0048] S3 fuses and optimizes the sequence feature set to obtain an optimized feature set;

[0049] S4 imports the optimized feature set into the classifier for training to obtain the augmented prediction model.

[0050] The principle of this invention is that a multi-dimensional feature extraction algorithm extracts features from the enhanced subsample set. First, the extracted feature sequences are fused, and then the fused sequences are optimized. Optimization can be achieved by using the LGBM algorithm for weighted filtering, and then further segmenting the sample into corresponding Laplacian matrices L based on different feature extraction algorithms. m The optimized data is then imported into a classifier, particularly an RBF neural network. The RBF neural network then uses the Laplacian matrix L mentioned above. m The regularization term adjusts the parameters and learning process. Clustering algorithms such as k-means are used to complete the initial training of the RBF neural network. Then, methods such as the Lagrange multiplier method are used to solve the subsequent optimization. Finally, training continues until a well-trained enhancer recognition model is obtained. Here, S1, S2, S3, and S4 are only for ease of recognition and do not represent an absolute execution order. Some steps may be performed simultaneously, provided the invention can be implemented.

[0051] Example 1

[0052] Step S1: Obtain a sample set of enhanced subsequences;

[0053] In step S1, in order to ensure that the prediction results include not only whether it is an enhancer, but also the strength of the enhancer, the sample set may include: a sample set of enhancer sequences, non-enhancer sequences, strong enhancer sequences, and weak enhancer sequences that have all been labeled.

[0054] To adjust the sample data and ensure training effectiveness: Before obtaining the original sequence data, the downloaded enhancer, strong enhancer, and weak enhancer sequence data files can be formatted and their content evaluated.

[0055] The specific method for determining the format is as follows: when a line in the read-in enhancer and non-enhancer sequence data file begins with the character ">", then the next line is taken as sequence text data;

[0056] The specific method for content judgment is as follows: when a line in the read-in enhancer and non-enhancer sequence data file consists of the four letters "A", "C", "G", and "T", the sequence data of that line is converted to uppercase. S2 performs sequence feature extraction on the sample set according to the feature extraction algorithm to obtain a sequence feature set;

[0057] In step S2, as Figure 3 As shown, in order to improve the richness and accuracy of the prediction results, the nucleotide composition feature extraction algorithm includes the k-mer extraction algorithm, the NBP extraction algorithm and the ANF extraction algorithm.

[0058] S3 fuses and optimizes the sequence feature set to obtain an optimized feature set;

[0059] The LGBM algorithm was used to optimize feature selection, and the specific process is as follows:

[0060] The k-mer extraction algorithm is set to a parameter k of 6. The feature vectors generated by the 6-mer, NBP, and ANF extraction algorithms are fused. The feature vectors from the three methods are linearly concatenated to form a single feature vector. The concatenation order is: 6-mer, NBP, ANF, i.e., the fused feature vector is (6-mer, NBP, ANF). The output length, i.e., the dimension of the feature set, is 5096. (Specifically, the fusion method assumes that the feature vectors generated by 6-mer, NBP, and ANF are (mer1, mer2, ..., mer4096), (nbp1, nbp2, ..., ...). Given nbp800), (anf1, anf2, ..., anf200), the concatenated vector is: (mer1, mer2, ..., mer4096, nbp1, nbp2, ..., nbp800, anf1, anf2, ..., anf200). Therefore, the concatenation order does not cause significant fluctuations in model performance. Even if the order changes, the numerical information contained in the feature set remains unchanged, and those skilled in the art can adjust it. In this embodiment, the concatenation order is 6-mer, NBP, ANF. The fused feature set will be used as input to the LGBM algorithm; specifically... Figure 3 As shown, the feature vector lengths generated by the feature extraction algorithms 6-mer, NBP, and ANF are 4096, 800, and 200, respectively. After feature filtering by the LGBM algorithm, the feature vector lengths of the three algorithms are 1195, 713, and 198 when identifying enhancers, respectively; and 157, 146, and 197 when identifying the intensity of enhancers, respectively.

[0061] The core parameters of the LGBM algorithm are set as follows: the boosting type is 'gbdt', the learning rate is 0.05, the maximum depth of each tree is 10, the maximum number of leaf trees per tree is 30, 50% of the sample dataset is randomly used each time a new tree is built, 80% of the samples are randomly used each time bagging is performed, and bagging is performed once every 12 trees are built.

[0062] After obtaining the forest generated by the LGBM algorithm, the number of times each feature is used as a partitioning attribute in all trees is used as the basis for judging its importance, and the forest is sorted and output.

[0063] Through experiments, the optimal parameter selection is as follows: when identifying enhancers, select the top 2016 features by feature importance to form a new feature vector; when identifying enhancer strength, select the top 500 features by feature importance to form a new feature vector.

[0064] The new feature set is divided into three feature subsets. When identifying enhancers, the number of features in the three feature subsets is 1195 for 6-mer, 713 for NBP, and 198 for ANF. When identifying enhancer strength, the number of features in the three feature subsets is 157 for 6-mer, 147 for NBP, and 197 for ANF.

[0065] The three feature subsets each form a heat kernel matrix from a different perspective. Each heat kernel matrix can be transformed into a Laplacian matrix through Laplacian transformation, thereby providing a multi-perspective Laplacian regularization effect.

[0066] S4 imports the optimized feature set into the classifier for training to obtain the augmented prediction model.

[0067] Taking the RBF neural network with a multi-view classification machine as an example, its network structure is as follows: Figure 4 As shown, the multi-view RBF neural network consists of an input layer, a hidden layer, and an output layer. The hidden layer is a single layer, and the nonlinear transformation from the input layer to the hidden layer is achieved by relying on the radial basis function of the hidden unit. The weight connection from the hidden layer to the output layer is optimized by regularization and matrix norm, which makes the whole network have better robustness and generalization ability.

[0068] The specific construction method is as follows:

[0069] The number of cluster centers and the iteration precision of the k-means clustering algorithm are set to p and ε, respectively, to obtain a set of cluster centers (c j For clusters $\mathbf{j ∈ {1, 2, ..., p}$, the width of the radial basis function is calculated based on the cluster centers, as shown in the following formula:

[0070]

[0071] Where c max σ represents the maximum distance between the selected cluster centers, p represents the number of cluster centers, and is also the number of hidden layer neurons in the RBF neural network in subsequent steps; i This represents the width of the i-th radial basis function.

[0072] The radial basis functions of the hidden layer of the RBF neural network are obtained based on the cluster centroids and the width of the radial basis functions, as shown in the following formula:

[0073]

[0074] The weight connection matrix W between the hidden and output layers of the RBF neural network is optimized using the least squares method. The objective function of the optimization problem is:

[0075]

[0076] Where Φ is the output matrix of the hidden layer in the RBF neural network, N is the number of samples, p is the number of neurons in the hidden layer, Y is the true label of the dataset, and W is the weight connection matrix between the hidden layer and the output layer. The matrix forms of Φ and W are as follows:

[0077]

[0078] W = [w1, w2, ..., w q ] p×q

[0079] Incorporating the Laplace regularization term and the L2,1 norm term of the matrix into the least squares optimization problem, the objective function of the optimization problem becomes:

[0080]

[0081]

[0082] 0 < β m <1,

[0083] m = 1, 2, ..., M

[0084] in Denotes the Laplace regularization term. Let β represent the L2,1 norm term of the weight link matrix W. m Let L be the Laplace matrix. m ∈R N×N The weights are M; M represents the number of Laplacian matrices. One feature extraction method generates one Laplacian matrix, so in this embodiment, M equals 3.

[0085] To solve the above optimization problem, we first fix and initialize β. m Solving for W, the optimization problem becomes as follows:

[0086]

[0087] in For unconstrained optimization problems, the solution can be obtained by taking the derivative, and the result is as follows:

[0088] W * =(Φ T Φ+λ1Φ T L * Φ+λ2G)-1 Φ T Y

[0089] W obtained * If we treat it as a constant and include it in the optimization problem, then the optimization problem simplifies to:

[0090]

[0091]

[0092] 0 < β m <1,

[0093] m = 1, 2, ..., M

[0094] The optimization problem was solved using the Lagrange multiplier method, and the results are as follows:

[0095]

[0096] Due to the strong convergence capability of the least squares method, the above process can be converged after about 5 overlapping iterations.

[0097] In the process of constructing the multi-view RBF neural network described above, the method for constructing the Laplacian matrix is ​​as follows:

[0098] A Laplacian matrix is ​​generated for each feature subset output by the feature extraction method; therefore, in this invention, M is set to 3, and the Laplacian matrix L... m It can be obtained through the following formula:

[0099]

[0100] L′ m =D m -S m

[0101]

[0102] The above process is the thermonuclear matrix S m The Laplace regularization process, as long as the heat kernel matrix S is known... m The Laplace matrix is ​​obtained, and the heat core matrix S in this invention is obtained. m Constructed using the following formula:

[0103]

[0104] Where γ is a constant term, set to 1 in this invention. Dm can be regarded as the degree matrix of the heat kernel matrix Sm; the dimension of the heat kernel matrix Sm is N*N, where N is the number of samples in the feature set generated by the m-th feature extraction method, xi represents the i-th sample in the feature set, and xj represents the j-th sample in the feature set; there are many ways to construct the heat kernel matrix Sm, and a preferred method is given in this example because it has been found in actual programs that the method of constructing the heat kernel matrix Sm mentioned above is the most suitable, which can avoid the situation of matrix non-invertibility during optimization and will not lead to a decrease in the accuracy of the model.

[0105] Additionally, grid search and 5-fold cross-validation can be used to optimize the model's classification performance, as follows:

[0106] The hyperparameters of the multi-view RBF neural network are p, λ1, and λ2. Different values ​​of p result in different classification performances for the multi-view RBF neural network. The inventors used 5-fold cross-validation to obtain these classification performance parameters, as detailed below:

[0107] The original training samples are divided into 5 subsets, and each subset is used as a validation set. The remaining 4 subsets are used as the training set. This results in 5 models. The average classification accuracy of these 5 models on the validation set is used as the performance metric of the classifier under 5-fold cross-validation to evaluate the classification performance of the model. With the above method for evaluating model performance, the inventors use grid search to select the optimal hyperparameters. It uses exhaustive search, trying every possibility through iterative convenience among all candidate parameter selections, indicating that the best hyperparameter is the final optimal hyperparameter.

[0108] The specific steps for using grid search and 5-fold cross-validation are as follows:

[0109] Ss1. Initialize the hyperparameters of the multi-view RBF neural network, where the number of cluster centers p is initialized to 100, and λ1 and λ2 are initialized to 2 respectively. -5 ,2 -5 The search range of p is [100, 1500], the step size is 50, and the search range of λ1 and λ2 is [2...]. -5 ,2 0 Each time, it doubles in size;

[0110] Ss2. Calculate the classification accuracy obtained by the multi-view RBF neural network in classifying enhancers and their strengths, and use the classification accuracy as the basis for selecting hyperparameters;

[0111] Ss3. Update the hyperparameters p, λ1, λ2 based on the classification accuracy;

[0112] Ss4. Determine whether the grid search is complete. If it is, proceed to S4-5; otherwise, proceed to S4-2.

[0113] Ss5. Substitute the obtained optimal set of hyperparameters p, λ1, λ2 into the multi-view RBF neural network to obtain the multi-view RBF neural network with the best classification performance.

[0114] The classification accuracy is represented by the accuracy ACC, which is calculated as follows:

[0115]

[0116] Where ACC represents the classification accuracy obtained by the multi-view RBF neural network algorithm in classifying enhancers and their strengths, TP represents the number of correctly predicted positive samples, FP represents the number of incorrectly predicted positive samples, TN represents the number of correctly predicted negative samples, and FN represents the number of incorrectly predicted non-negative samples. In the enhancer identification task, enhancers are considered positive samples, and non-enhancers are considered negative samples; when identifying enhancer strength, strong enhancers are considered positive samples, and weak enhancers are considered negative samples.

[0117] Example 2

[0118] like Figure 2 As shown, the steps are: SS1 to obtain target sequence samples; SS2 to use the enhancer prediction model to identify the sequence samples and obtain prediction results.

[0119] The prediction model can be the model trained using the method in Example 1.

[0120] Example 3

[0121] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method steps described above. It may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-readable program code.

[0122] To illustrate the prediction model provided by the present invention and the actual effects of using the prediction model, the inventors also provide the following comparative examples.

[0123] Comparative Example 1

[0124] To verify the impact of feature selection using the LGBM algorithm on the recognition of enhancers and their strength in a multi-view RBF neural network, the inventors input both the feature sets without feature selection and the feature sets with feature selection from independent datasets into the trained multi-view RBF neural network. In this embodiment, the inventors used SN, SP, ACC, and MCC as evaluation metrics, which can be calculated using the following formula:

[0125]

[0126]

[0127]

[0128]

[0129] exist Figure 5 As can be seen from the data, when identifying enhancers, the input feature sets SN, SP, ACC, and MCC without feature selection are 0.695, 0.79, 0.7425, and 0.4872, respectively, while those with feature selection are 0.82, 0.775, 0.7975, and 0.5956, respectively. This shows that SN, ACC, and MCC are improved by 0.125, 0.055, and 0.1084, respectively. When identifying enhancer strength, the input feature sets SN, SP, ACC, and MCC without feature selection are 0.68, 0.52, 0.6267, and 0.2026, respectively, while those with feature selection are 1.00, 0.67, 0.835, and 0.7098, respectively. This shows that SN, SP, ACC, and MCC are improved by 0.32, 0.15, 0.2083, and 0.5072, respectively.

[0130] Comparative Example 2

[0131] Figure 6 The diagram illustrates a comparison of the model's prediction accuracy with other machine learning methods on an independent test set, as provided in this embodiment of the invention. To demonstrate the model's prediction accuracy, the inventors compared its performance with other commonly used machine learning methods on a dataset. The commonly used machine learning methods employed in this paper are: MLR-RBFN, SVM, RF, KNN, and RBFN. MLR-RBFN is the model trained using the method in Embodiment 1, and the prediction result is obtained using the method in Embodiment 2 (the following embodiments are similar). RBFN is a standard RBF neural network. The experimental steps involve performing the following steps to statistically analyze the classification performance of each different method:

[0132] Use appropriate feature extraction algorithms to extract features from the target sequence;

[0133] Use appropriate feature set optimization strategies;

[0134] The optimized feature set is input into the corresponding classification network to obtain the target sequence prediction result, and the classification effect is evaluated.

[0135] Figure 6 The results show the effectiveness of the method used in this project.

[0136] Comparative Example 3

[0137] like Figure 7 As shown, to further illustrate the generalization performance of the inventor's method, the inventor trained a model on a cross-validation set and independently tested it on an independent dataset. The experimental results were then compared with the performance of other existing methods, including iEnhancer-MRBF, iEnhancer-CNN, iEnhancer-EL, iEnhancer-2L, and EnhancerPred. iEnhancer-MRBF is the method of this invention. The experimental steps are as follows: different methods are trained on the training set using selected hyperparameters to obtain a trained classification model. Then, data from the test set is input into this trained model to obtain the final result. Figure 7 As can be seen, when identifying enhancers, the inventor's model achieved optimal performance in SN, ACC, and MCC, with improvements of 0.0375, 0.0225, and 0.0106 respectively compared to the current best performance; when identifying enhancer strength, the inventor's model achieved optimal performance in SN, ACC, and MCC, with improvements of 0.3475, 0.085, and 0.3866 respectively compared to the current best performance.

[0138] The preferred embodiments and comparative embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort, such as changing the order of S1, S2, S3, and S4 or making conventional substitutions, etc. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for training an augmentation sub-prediction model, characterized in that, include: Obtain a sample set of enhanced subsequences, the sample set including: a sample set of enhanced subsequences all labeled, a sample set of non-enhanced subsequences, a sample set of strongly enhanced subsequences, and a sample set of weakly enhanced subsequences; The sample set is subjected to sequence feature extraction according to the feature extraction algorithm to obtain a sequence feature set. The feature extraction algorithm includes: Kmer, Nucleotide Binary Profiles and Accumulated Nucleotide Frequency algorithm. The sequence feature set is fused and optimized to obtain an optimized feature set, including the following steps: The sequence feature sets are fused to obtain a fused feature set; The fused feature set is filtered using the LGBM algorithm to obtain the filtered feature set; The selected feature set is split according to the feature extraction algorithm to obtain the Laplacian matrix corresponding to the feature extraction algorithm. ; According to the Laplace matrix This yields an optimized feature set. The optimized feature set is imported into a classifier for training to obtain an augmenter prediction model, wherein the classifier is a multi-view RBF neural network.

2. The method for training the enhancer prediction model according to claim 1, characterized in that, The fused feature set is optimized using the LGBM algorithm to obtain an optimized feature set. The LGBM algorithm is configured as follows: The boosting type is GBDT, the learning rate is 0.05, the maximum depth of each tree is 10, the maximum number of leaf trees per tree is 30, 50% of the sample dataset is randomly used each time a new tree is built, 80% of the samples are randomly used each time bagging is performed, and bagging is performed once every 12 trees are built.

3. The method for training the enhancer prediction model according to claim 1, characterized in that, The radial basis functions of the hidden layer of the RBF neural network are: ， in, One free parameter is the width parameter of the function, which controls the radial range of the function; xi is the feature point in the optimized feature set, and cj is the cluster center point of the optimized feature set.

4. The method for training the enhancer prediction model according to claim 3, characterized in that, The hyperparameters of a multi-view RBF neural network are determined according to the following steps: According to the objective function W * Perform a grid search. The hyperparameters are obtained by minimizing the difference between the predicted result and the labeled sample set of the enhanced subsequence. ,in: ， ， ， ， in, Represents the Laplace matrix The regularization term, Represents the weighted link matrix L2,1 norm terms, Laplace matrix The weights, These are the hyperparameters of the RBF neural network; The output matrix of the hidden layer is expressed as follows: 。 5. A method for identifying enhancers, characterized in that, include: Obtain target sequence samples; The target sequence sample is identified using the enhancer prediction model as described in any one of claims 1-4, and the identification result is obtained.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method steps as described in any one of claims 1-4.

Images