Training method of model of prediction of microRNA and disease association, model

Abstract

The application provides a training method of a model for predicting microRNA and disease association. A sample set of the association between microRNA and disease is input into a link propagation network for training to obtain a target model for predicting the association between microRNA and disease. The application solves the problem of low prediction accuracy of the model in the prior art. According to the link propagation principle, if the similarity between a node and another node is high, the association between the microRNA represented by the node and the disease can be propagated to the other node, so that the prediction is more accurate.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of microRNA prediction, and particularly relates to a training method of a model for predicting microRNA and disease association and the model. BACKGROUND

[0002] MicroRNA is a small molecule non-coding RNA (-22 nt) that plays an important role in cells. It is estimated that 1-4% of genes in the human genome are microRNAs, a single microRNA regulates up to 200 mRNAs, and microRNAs usually bind to the 3' untranslated region (UTRs) of target mRNAs through sequence-specific base pairs, inhibit the expression of target mRNAs, and thus participate in a series of important processes in life. In many bioinformatics research projects, identifying potential microRNA and human disease association has always been a key goal, which will help the treatment and prevention of human diseases, the design of molecular tools and personalized diagnosis.

[0003] Traditional biological experiments are expensive and laborious for large-scale association detection. Therefore, effective computational methods for predicting potential microRNA disease associations have attracted increasing attention. Based on the assumption that functionally related microRNAs are more likely to be associated with diseases with similar phenotypes, some computational models based on scoring functions have been proposed, but these models do not perform feature fusion and have low prediction accuracy. SUMMARY

[0004] The present application provides a training method of a model for predicting microRNA and disease association and the model to solve the problem that the existing training method cannot effectively perform feature fusion, resulting in low accuracy of the trained model.

[0005] According to the present application and embodiments, a training method of a model for predicting microRNA and disease association is provided, which inputs a sample set of microRNA and disease association into a link propagation network for training to obtain a target model for predicting microRNA and disease association.

[0006] Further, the link propagation network enables nodes with the same or similar similarity to have the same or similar link strength.

[0007] Further, the method comprises the following steps:

[0008] Obtaining a sample set of microRNA and disease;

[0009] According to the microRNA similarity and disease similarity corresponding to the microRNA in the sample set, a plurality of microRNA spatial similarity kernels and a plurality of disease spatial similarity kernels are obtained.

[0010] Based on the associations between microRNAs and diseases in the sample set, an adjacency matrix was obtained;

[0011] Based on the weights, multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels are fused and normalized to obtain fused microRNA spatial similarity kernels and fused disease spatial similarity kernels.

[0012] The fused microRNA spatial similarity kernel, the fused disease spatial similarity kernel, and the adjacency matrix input link propagation network are trained to achieve the following:

[0013] Nodes in the adjacency matrix with similar or identical RNA similarity and similar or identical disease similarity have similar or identical link strengths, resulting in a model that predicts the association between target microRNAs and diseases.

[0014] Furthermore, the objective function of the link propagation network is:

[0015] ,

[0016] in, This is the link strength matrix; It is an adjacency matrix; It is the balance coefficient parameter.

[0017] Furthermore, the final solution form of the link propagation network is:

[0018] ,

[0019] Among them, the parameters of the b-core fusion method; These are balance coefficient parameters; It is an adjacency matrix; It is a microRNA similarity feature vector. It is a disease similarity feature vector; where, matrix The calculation formula is:

[0020] ,

[0021] Among them, the parameters of the b-core fusion method; These are balance coefficient parameters; These are the eigenvalues ​​after nuclear fusion.

[0022] Optionally, its hyperparameters can be set as follows: when using the Kronecker product, , , When using Kronecker summation, , , .

[0023] Optionally, the hyperparameters are set as follows: when using the Kronecker product, , , , When using Kronecker summation, , , , 。

[0024] The microRNA spatial similarity kernel includes: the microRNA spatial Gaussian correlation summary similarity kernel, the microRNA spatial sequence similarity kernel, and the microRNA spatial functional similarity kernel;

[0025] The disease space similarity kernel includes: the Gaussian association summary similarity kernel of the disease space, the disease semantic similarity kernel of the disease space, and the disease functional similarity kernel.

[0026] Optionally, where,

[0027] Gaussian correlation of microRNA space, summary similarity kernel The calculation formula is as follows:

[0028] ,

[0029] in, Let's consider the adjacency matrix. The row and number The rows are respectively the first The first microRNA and the first Disease overview of microRNAs Adjacency matrix The Line 1 Disease summary of microRNA and No. Line 1 Disease overview of microRNAs; This represents the bandwidth of the Gaussian kernel;

[0030] microRNA functional similarity nucleus The calculation formula is as follows:

[0031] ,

[0032] Among them, diseases In the disease group Chinese; disease In the disease group middle; It is a disease and disease group Functional similarity score; It is a disease and disease group Functional similarity score; Representing the disease group The number of diseases; Representing the disease group The number of diseases;

[0033] Gaussian correlation summary similarity kernel in disease space The calculation formula is as follows:

[0034] ,

[0035] in, and It is an adjacency matrix The Column and number Columns, they represent the first The disease and the first Summary of microRNA associations in a disease This represents the bandwidth of the Gaussian kernel;

[0036] Disease semantic similarity kernel The calculation formula is as follows:

[0037] ,

[0038] in, It is a node in a certain disease. For nodes The set of elements composed of its ancestor nodes, It is a disease semantic score, It is a disease The semantic score; where , It is a disease and disease The relevant semantic values, It is a disease and disease The relevant semantic values;

[0039] Disease functional similarity kernel matrix The calculation formula is as follows:

[0040] ,

[0041] Among them, genes In the genome Chinese; gene In the genome middle; It's a gene. and genome Functional similarity score; It's a gene. and genome Functional similarity score; Representing the genome The number of genes; Representing the genome The number of genes.

[0042] Optionally, multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels are fused and normalized according to their weights to obtain fused microRNA spatial similarity kernels and fused disease spatial similarity kernels, wherein...

[0043] The fusion formula is as follows:

[0044] ,

[0045] ,

[0046] Where n is the number of similarity kernels; It is a spatial similarity kernel; It is a disease similarity kernel;

[0047] The normalization formula is as follows:

[0048] ,

[0049] ,

[0050] Where i and j represent the matrix i and j respectively. Column and number List.

[0051] According to the present invention and embodiments, the present invention also provides a model for predicting the association between microRNA and disease, characterized in that it is trained by the above-described training method for the model for predicting the association between microRNA and disease.

[0052] The beneficial effects of this invention are:

[0053] By inputting a sample set of microRNA and disease associations into a link propagation network for training, a model for predicting the association between target microRNAs and diseases is obtained. Link propagation refers to the phenomenon that if one node has a high similarity to another node, the association between the microRNA represented by that node and the disease can be propagated to the other node. Link propagation has not yet been applied to the research field of microRNA and disease associations. This invention provides a concrete and feasible solution for applying the link propagation concept to this field. In this way, more accurate predictions of the association between unknown microRNAs and diseases can be obtained using a smaller sample size. Attached Figure Description

[0054] Figure 1 The diagram shows a flowchart of a training method for a model that predicts the association between microRNA and disease, provided by an embodiment of the present invention.

[0055] Figure 2 The diagram shown is a schematic representation of a model for predicting the association between microRNA and disease, provided by an embodiment of the present invention.

[0056] Figure 3 The diagram shows a schematic of the link propagation algorithm for a model that predicts the association between microRNA and disease, provided in an embodiment of the present invention.

[0057] Figure 4 The figure shown is a comparison of the performance of the model for predicting the association between microRNA and disease provided in the embodiments of the present invention. Detailed Implementation

[0058] 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 1 The 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 function specified in one or more boxes.

[0059] 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.

[0060] Invention principle:

[0061] To further refine and improve existing microRNA and disease association information, this invention employs a multi-core fusion approach, adding more feature dimensions. Furthermore, it utilizes a link propagation network (or link propagation algorithm) based on matrix factorization and matrix approximation, enabling efficient, accurate, and rapid computation of these fusion kernels.

[0062] This invention can obtain a sample set of microRNA and disease associations from existing databases, and then construct multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels based on these samples. These kernels can respectively characterize the similarity between microRNA and disease in different feature dimensions. Then, according to the weight values ​​of these kernels (the average value is used in Example 1), the multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels are fused and normalized. Finally, the fused microRNA spatial similarity kernels and fused disease spatial similarity kernels are input into a link propagation network based on matrix factorization and matrix approximation for training. The training data are randomly set as training set and validation set (or test set). Finally, based on the objective function of the link propagation network and repeated training, the hyperparameters of the model are confirmed, and a more accurate prediction model is obtained after training.

[0063] Our goal is to predict the probability of a link forming between two nodes. We call this probability link strength and store it in a link strength matrix. In, its size is similar to Same. Elements in the matrix The larger the value, the more the algorithm considers microRNA. and disease The higher the similarity between two nodes, the greater the likelihood of a link. Link propagation refers to the phenomenon where, if one node and another have a high degree of similarity, the association between the microRNA represented by that node and the disease can be propagated to the other node. Link strength matrix A single value in the array represents a node.

[0064] In the masked adjacency matrix, each value represents a node. For example, the value in the i-th row and j-th column of the masked adjacency matrix represents whether there is a correlation between the i-th microRNA and the j-th disease. We obtained the normalized fusion kernels for the microRNA and disease spaces, and the normalized similarity kernel for microRNAs. The microRNA similarity kernel stores the degree of similarity between microRNAs; the higher the value of the kernel, the more similar the two microRNAs are. Similarly, the normalized similarity kernel for diseases stores the degree of similarity between diseases; the higher the value of the kernel, the more similar the two diseases are. The entire link propagation process is achieved by judging the similarity between two nodes, considering nodes with higher similarity to have more similar link strengths.

[0065] In Example 1, the link propagation network was optimized using matrix factorization and matrix approximation methods. The Laplace matrix was calculated using both Kronecker product and Kronecker summation. Furthermore, the principle of link propagation, "two similar nodes may have the same link strength," was applied to the node pairs associated with microRNA and disease, resulting in a minimized objective function. This significantly reduced time and computational costs, making the network run faster and more accurately, and improving the prediction effect of microRNA and disease association.

[0066] Example 1

[0067] Specifically, such as Figures 1-3 As shown, the model training method for predicting microRNA and disease association based on the Hilbert-Schmidt independence criterion and link propagation, and the prediction using the trained model, include the following steps:

[0068] S1. Collect and download the labeled sample set containing associations between microRNAs and diseases;

[0069] S2. Based on the content of the sample set, construct a similarity kernel in the microRNA and disease space;

[0070] S3. The similarity kernels are fused using average weights to obtain similarity kernels that can comprehensively represent microRNA and disease feature information, and the two similarity kernels are normalized respectively.

[0071] S4. Construct a predictor based on the link propagation algorithm using matrix factorization and matrix approximation, and select the optimal parameters for the prediction model using grid search and 5-fold cross-validation;

[0072] S5. Input the test data and similarity kernel into the trained predictor to obtain the link strength prediction results and complete the prediction of the association between microRNA and disease;

[0073] Step S1 can be specifically divided into the following steps:

[0074] In this invention, a sample set containing 495 microRNAs and 383 diseases was collected and downloaded, which contained a total of 5430 microRNA-disease associations. These associations were stored in an adjacency matrix. In the adjacency matrix In the diagram, each row represents a type of microRNA, and each column represents a disease. If the... The first microRNA and the first The diseases are related, that is and If so, then the elements of the adjacency matrix... It is set to 1 in the case of , and set to 0 in the other cases.

[0075] Our goal is to predict the probability of a link forming between two nodes. We call this probability link strength and store it in a link strength matrix. In, its size is similar to Same. Elements in the matrix The larger the value, the more the algorithm considers microRNA. and disease The higher the likelihood of a connection between them.

[0076] To accomplish the semi-supervised learning task and better evaluate performance, we randomly set the adjacency matrix. We use 0 for some elements to mask the adjacency matrix. Then we obtain the masked adjacency matrix. The unmasked portion serves as the training set, while the remainder is the validation or test set. This algorithm predicts... The masked portion is used to calculate the link strength matrix. .

[0077] Step S2 can be specifically divided into the following steps:

[0078] In this step, we will construct three similarity kernels each in the microRNA space and the disease space.

[0079] The Gaussian correlation summary similarity kernel between the microRNA space and the disease space needs to be calculated:

[0080] Gaussian correlation of microRNA space, summary similarity kernel It can be determined by the adjacency matrix. Gaussian association summary similarity kernel of microRNA obtained The calculation formula is as follows:

[0081] ,

[0082] In this embodiment, in the above formula, and It is an adjacency matrix The row and number Lines, they represent the first The first microRNA and the first A summary of disease associations for each microRNA, containing all known associations between a specific microRNA and 383 diseases; This represents the bandwidth of the Gaussian kernel, and its value is set to 1.

[0083] Similarly, the Gaussian correlation summary similarity kernel in the disease space It can also be determined by the adjacency matrix. The formula for obtaining it is shown below:

[0084] ,

[0085] In this embodiment, in the above formula, and It is an adjacency matrix The Column and number Columns, they represent the first The disease and the first A summary of microRNA associations for a specific disease, containing all known association information for the specific disease and 495 microRNAs; The bandwidth of the Gaussian kernel is represented by 1. Ytrain is called the masked adjacency matrix. The adjacency matrix Y stores the association information between microRNAs and diseases. Then, some elements of the adjacency matrix Y are randomly set to 0 to mask the microRNAs, resulting in the masked adjacency matrix Ytrain. Therefore, Ytrain also stores the association information between microRNAs and diseases. A row of Ytrain can be considered as a summary of microRNA associations, allowing the calculation of the Gaussian association summary similarity kernel Km,1 in the microRNA space; a column of Ytrain can be considered as a summary of disease associations, allowing the calculation of the Gaussian association summary similarity kernel Kd,1 in the disease space.

[0086] In addition, the sequence similarity kernel in the microRNA space and functional similarity kernel Disease semantic similarity kernel in disease space Disease functional similarity kernel It can be downloaded from the dataset.

[0087] In the microRNA space, all 495 miRNA sequences were obtained from the miRBase database. Sequence similarity was extracted using the Needleman-Wunsch algorithm. The information was then stored in a kernel matrix. Among them microRNAs and Sequence similarity score between them.

[0088] Constructing miRNA functionally similar nuclei based on the MISIM method This strategy leverages the semantic similarity between miRNAs and diseases, as well as established links, to construct a miRNA functional similarity kernel. In this example, microRNAs... and Functional similarity score .

[0089] First, we define a functional similarity score (DFS) for two diseases, which can be obtained from the MISIM database.

[0090] Then, we defined a disease. With a set of genes The functional similarity score between them is calculated using the formula shown below:

[0091] ,

[0092] In the above formula, It represents disease With disease group Functional similarity score between them; The meaning expressed is: disease With disease group The maximum functional similarity score for all diseases in the study.

[0093] Then, the association between microRNA and disease can be downloaded from the MISIM database, allowing us to obtain the association between microRNA and disease. Related disease groups , with microRNA Related disease groups Then we can obtain the microRNA functional similarity matrix. The formula is as follows:

[0094] ,

[0095] In the above formula, disease In the disease group Chinese; disease In the disease group middle; It is a disease and disease group Functional similarity score; It is a disease and disease group Functional similarity score; Representing the disease group The number of diseases; Representing the disease group The number of diseases.

[0096] In the disease space, the disease semantic similarity kernel is constructed using the MeSH database. In this database, a disease can be represented as a directed acyclic graph (DAG). A DAG representing a disease consists of three parts: first, the disease node; second, the set containing the node and its ancestor nodes; and third, the set of corresponding links. We can calculate the semantic score for each disease using the DAG, and then calculate the disease semantic similarity kernel from the disease semantic score, which is stored in a matrix. middle.

[0097] In the disease space, the first step is to construct a disease semantic similarity kernel. In the MeSH database, for a given disease... It can be represented as a directed acyclic graph, and its formula is as follows:

[0098] ,

[0099] In the above formula, It is a node in a certain disease. For nodes The set of elements composed of its ancestor nodes, For the set of corresponding links, It is a disease The directed acyclic graph. The correlation semantic value measures the correlation between two diseases; the higher the correlation semantic value, the stronger the correlation. Let... ,disease and disease The formula for calculating the relevant semantic value is as follows:

[0100] ,

[0101] In the above formula, It is a semantic contribution factor, and in this invention, its value is set to 0.5.

[0102] disease The formula for calculating the semantic score is as follows:

[0103] ,

[0104] In the above formula, It is a disease The semantic score. Then the disease semantic similarity kernel. The calculation formula is as follows:

[0105] ,

[0106] In the above formula, It is a disease semantic score, It is a disease The semantic score. Among them , It is a disease and disease The relevant semantic values, It is a disease and disease The relevant semantic value measures the correlation between two diseases; the higher the value of the relevant semantic value, the stronger the correlation between the two diseases.

[0107] Calculate the functional similarity kernel of diseases This involves the correlation between genes and diseases. First, the LLS score can be obtained from the HumanNet database. This score measures the probability that two genes in the database have a functional relationship. The LLS score is then normalized using the formula shown below:

[0108] ,

[0109] In the above formula, Representative gene and genes The LLS score is obtained from the HumanNet database; This represents the minimum value of all LLS scores in the HumanNet database; This represents the maximum value of all LLS scores in the HumanNet database; Represents the normalized genes and genes The LLS score.

[0110] Then, the functional similarity score between the two genes is calculated using the formula shown below:

[0111] ,

[0112] In the above formula, Representative gene and genes The functional similarity score between two genes.

[0113] Then, we defined a gene. With a set of genes The functional similarity score between them is calculated using the formula shown below:

[0114] ,

[0115] In the above formula, Represents genes With genome Functional similarity score between them; The meaning expressed is that genes With genome The maximum functional similarity score of all genes in the dataset.

[0116] Then, the association between diseases and genes can be downloaded from the SIDD database, allowing us to obtain information about the relationship between diseases and genes. Related genomes , with disease Related genomes Then we can obtain the disease functional similarity kernel matrix. The formula is as follows:

[0117] ,

[0118] In the above formula, genes In the genome Chinese; gene In the genome middle; It's a gene. and genome Functional similarity score; It's a gene. and genome Functional similarity score; Representing the genome The number of genes; Representing the genome The number of genes.

[0119] LLS scores can be obtained from the HumanNet database. After normalizing the LLS scores, they can be used to calculate the disease functional similarity kernel matrix by utilizing the relationships between genes. .

[0120] like Figure 2 As shown, the adjacency matrix after masking Gaussian correlation summative similarity kernels for microRNA and disease spaces are calculated. Sequence similarity and functional similarity kernels for microRNAs, and semantic similarity and functional similarity kernels for diseases are then constructed. Kernel fusion is performed using average weights. The fused microRNA and disease kernels are then compared with the masked adjacency matrix. The results are fed into the link propagation algorithm and stored in the link strength matrix. middle.

[0121] Step S3 can be specifically divided into the following steps:

[0122] In step S2, we obtained three similarity kernels each for the microRNA space and the disease space. We then fused these kernels using average weights, as shown in the following formula:

[0123] ,

[0124] ,

[0125] Then we analyzed the microRNA similarity kernel. Disease similarity kernel The normalization process is performed, and the formula is shown below:

[0126] ,

[0127] ,

[0128] Among them, such as Figure 3 As shown, step S4 can be specifically divided into the following steps:

[0129] In step S3, normalized fusion nuclei of microRNA and disease space were obtained, respectively. and First, the diagonal matrix. and It can be calculated using the following formula:

[0130] ,

[0131] ,

[0132] For similarity kernel matrix Perform singular value decomposition, using and Let its left singular vector and right singular vector be represented by respectively. Let represent singular values. According to the Eckart-Young-Mirsky theorem, if we take its singular values... The smallest Set the singular values ​​to zero and take the corresponding left singular vector. and right singular vectors Among the parameters This refers to the approximation degree of the matrix. This parameter needs to be specified by the user. The optimal parameter value for model performance is obtained through grid search experiments, i.e., when using the Kronecker product. The value should be set to 150 when using Kronecker summation. The value should be set to 175. Then, through the diagonal matrix... By standardizing it, a standardized microRNA similarity kernel matrix can be obtained. The low-rank approximation of is given by the following formula:

[0133] ,

[0134] Similarly, we can also obtain standardized disease similarity and matrices. The low-rank approximation of is given by the following formula:

[0135] ,

[0136] In the above formula, the parameters This refers to the approximation degree of the matrix. This parameter needs to be specified by the user. The optimal parameter value for model performance is obtained through grid search experiments, i.e., when using the Kronecker product. The value should be set to 350 when using Kronecker summation. The value should be set to 300.

[0137] In this invention, the Laplacian matrix is ​​calculated using two methods: Kronecker product and Kronecker summation. These two operations are represented by the symbols... and symbols The formula for calculating the normalized Laplace matrix is ​​shown below:

[0138] ,

[0139] When using the Kronecker product, The value is 1. When using Kronecker summation, The value is 3. .

[0140] right After performing eigenvalue decomposition, we can obtain the similarity kernel matrix. eigenvalues The eigenvectors corresponding to its eigenvalues ​​are Similarly, for By performing eigenvalue decomposition, we can also obtain the similarity kernel matrix. eigenvalues The eigenvectors corresponding to its eigenvalues ​​are .

[0141] When using Kronecker product and when using Kronecker summation The eigenvectors are the same, and their formulas are as follows:

[0142] ,

[0143] When using the Kronecker product The formulas for the eigenvalues ​​are shown below:

[0144] ,

[0145] When using Kronecker summation, The formulas for the eigenvalues ​​are shown below:

[0146] ,

[0147] The principle of link propagation we use is that "two nodes that are similar to each other may have the same link strength".

[0148] Adjacency matrix after masking In this context, a single value in the matrix represents a node, such as in the masked adjacency matrix. The i-th row and j-th column represents whether there is a correlation between the i-th microRNA and the j-th disease.

[0149] In the preceding text, we obtained the normalized fusion cores of microRNA and disease space, respectively. and microRNA normalized similarity kernel The matrix stores the degree of similarity between microRNAs; the higher the value of the matrix, the more similar the two microRNAs are. For example, the similarity kernel... The value in the i-th row and j-th column represents the similarity between the i-th microRNA and the j-th disease. Similarly, the similarity kernel after disease normalization... It preserves the degree of similarity between diseases; the higher the value of the matrix, the more similar the two diseases are.

[0150] By fusing the Kronecker product and Kronecker summation, the similarity kernels of the microRNA space and the disease space can be integrated, thus measuring the similarity between two nodes. If two nodes have high similarity, they are considered to have the same link strength. Link strength refers to the probability of an association forming between two nodes, and it is stored in the link strength matrix that preserves the model results. middle.

[0151] Applying the principle of link propagation, we obtain the objective function to be minimized as follows:

[0152] ,

[0153] The obtained minimized objective function pair Differentiate and convert the Laplace matrix Substitute the values, and use the Woodbury formula and the vec-trick method to define... Using the Hadamard product, the final solution can be obtained as follows:

[0154] ,

[0155] In the above formula, 'b' is a parameter that controls the use of different kernel fusion methods. Specifically, when using the Kronecker product, the value of 'b' is 1; when using the Kronecker summation, the value of 'b' is 1. The value is 3; σ is a parameter used to balance the coefficients. Its value is specified by the user, and the optimal parameter values ​​for model performance are obtained through grid search experiments. Specifically, when using the Kronecker product, the value of σ should be set to 0.2, and when using the Kronecker summation, ... The value should be set to 0.25; It is called the masked adjacency matrix, which is achieved by randomly setting the adjacency matrix. This was achieved by setting certain elements to 0 to mask the result. It is the eigenvector obtained by eigenvalue decomposition of the low-rank approximation matrix of the microRNA similarity kernel matrix. These are the eigenvectors obtained by performing eigenvalue decomposition on the approximate matrix of the disease similarity kernel matrix. The two eigenvectors mentioned above are obtained by performing eigenvalue decomposition on the approximate matrix of the corresponding similarity kernel matrix; the calculation method is detailed above. The matrix... The calculation formula is:

[0156] ,

[0157] In the above formula, 'b' is a parameter that controls the use of different kernel fusion methods. Specifically, when using the Kronecker product, the value of 'b' is 1; when using the Kronecker summation, the value of 'b' is 1. The value is 3; σ is a parameter used to balance the coefficients. Its value is specified by the user, and the optimal parameter values ​​for model performance are obtained through grid search experiments. Specifically, when using the Kronecker product, the value of σ should be set to 0.2, and when using the Kronecker summation, ... The value should be set to 0.25; These are the eigenvalues ​​after nuclear fusion; the calculation method is detailed above.

[0158] The principle behind using grid search and 5-fold cross-validation to optimize the model's classification performance is as follows:

[0159] The hyperparameters of the link propagation algorithm based on matrix decomposition and matrix approximation are: , and When the values ​​are different, the link propagation algorithm based on matrix factorization and matrix approximation will achieve different classification performances; we use 5-fold cross-validation to obtain the classification performance here, and the specific method is as follows:

[0160] The original training samples are divided into 5 subsets, each of which 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 for the classifier under 5-fold cross-validation to evaluate the model's classification performance. Using this method to evaluate model performance, we employ grid search for optimal hyperparameter selection. Grid search is one of the most common, intuitive, and effective parameter tuning methods in machine learning. The basic idea of ​​grid search is to set a sequence of parameter values ​​for each freely adjustable parameter with a certain step size, arranging and combining all parameters, iterating through each parameter combination, and using 5-fold cross-validation to find the parameter combination that best performs the model. In this invention, the parameter combination that best performs the model, as determined experimentally, is generally: when using the Kronecker product... , , When using Kronecker summation, , , .

[0161] In this invention, we use two evaluation criteria to assess model performance: AUPR and AUC, both of which are widely used in related studies. The area under the receiver operating characteristic (ROC) curve is calculated by plotting the true positive rate and false positive rate at different thresholds, thus generating the AUC value. The AUC ranges from 0 to 1; generally, an AUC value greater than 0.5 is considered better than random guessing. A higher AUC value indicates better model performance. The area under the precision-recall curve at different threshold levels is called AUPR. Similarly, AUPR should be greater than 0 and less than 1; a higher value indicates a better model.

[0162] The specific steps of step S5 are as follows:

[0163] The sequence similarity kernel, functional similarity kernel, and Gaussian association summary similarity kernel of the constructed microRNA space, and the semantic similarity kernel, functional similarity kernel, and Gaussian summary similarity kernel of the disease space are fused together. The fused kernels are then normalized, and the matrix storing microRNA and disease association information is then processed. Inputting a predictor based on a link propagation algorithm using matrix factorization and matrix approximation yields a matrix containing microRNA and disease association predictions. And evaluate the classification results. For example... Figure 3 As shown, five existing methods with the best performance were selected and compared with the method of this invention on the same dataset. The evaluation metric used was the AUPR value. The existing methods included: CKA-HGRTMF, CKA-MKL+HGBLM, FKL-Spa-LapRLS, MDA-SKF, and LRSSLMDA. It can be seen that the AUPR value of this invention is the highest, reaching 0.9800, surpassing the existing methods.

[0164] It is worth noting that the steps labeled S1, S2, S3, S4, etc., do not indicate the execution order of the present invention. They can be performed simultaneously or in different orders as long as the effect of the invention is not affected. In addition, the specific implementation methods mentioned above, such as the various optional methods mentioned in S1, S2, S3, S4, all have possible implementation methods that can be replaced by those skilled in the art without creative effort. Furthermore, these implementation methods can be combined or combined with each other to form steps that can achieve the purpose of the present invention, and these steps should also be within the protection scope of the present invention.

Claims

1. A method for training a model to predict the association between microRNA and disease, characterized in that, The sample set of microRNA and disease associations is input into the link propagation network for training to obtain a model that predicts the association between microRNA and disease. The link propagation network enables nodes with the same or similar similarity to have the same or similar link strength. Includes the following steps: Obtain microRNA and disease sample sets; Based on the microRNA similarity and disease similarity in the sample set, multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels are obtained; Based on the associations between microRNAs and diseases in the sample set, an adjacency matrix was obtained; Based on the weights, multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels are fused and normalized to obtain fused microRNA spatial similarity kernels and fused disease spatial similarity kernels. The fused microRNA spatial similarity kernel, the fused disease spatial similarity kernel, and the adjacency matrix input link propagation network are trained to achieve the following: Nodes in the adjacency matrix that have the same or similar RNA similarity and the same or similar disease similarity have the same or similar link strength, thus obtaining a model for predicting the association between target microRNA and disease. The objective function of the link propagation network is: ， in, This is the link strength matrix; It is an adjacency matrix; It is the balance coefficient parameter.

2. The training method for the model predicting the association between microRNA and disease as described in claim 1, characterized in that, The final solution form of the link propagation network is: ， Among them, the parameters of the b-core fusion method; These are balance coefficient parameters; It is an adjacency matrix; It is a microRNA similarity feature vector. It is a disease similarity feature vector; where, matrix The calculation formula is: ， Among them, the parameters of the b-core fusion method; These are balance coefficient parameters; These are the eigenvalues ​​after nuclear fusion.

3. The training method for the model predicting the association between microRNA and disease as described in claim 2, characterized in that, Its hyperparameter settings are as follows: when using the Kronecker product, , , , When using Kronecker summation, , , , .

4. The training method for the model predicting the association between microRNA and disease as described in claim 1, characterized in that, in, The microRNA spatial similarity kernel includes: the microRNA spatial Gaussian correlation summary similarity kernel, the microRNA spatial sequence similarity kernel, and the microRNA spatial functional similarity kernel; The disease space similarity kernel includes: the Gaussian association summary similarity kernel of the disease space, the disease semantic similarity kernel of the disease space, and the disease functional similarity kernel.

5. The training method for the model predicting the association between microRNA and disease as described in claim 4, characterized in that, in, Gaussian correlation of microRNA space, summary similarity kernel The calculation formula is as follows: ， in, Let's consider the adjacency matrix. The row and number The rows are respectively the first The first microRNA and the first Disease overview of microRNAs Adjacency matrix The Line 1 Disease summary of microRNA and No. Line 1 Disease overview of microRNAs; This represents the bandwidth of the Gaussian kernel; microRNA functional similarity nucleus The calculation formula is as follows: ， Among them, diseases In the disease group Chinese; disease In the disease group middle; It is a disease and disease group Functional similarity score; It is a disease and disease group Functional similarity score; Representing the disease group The number of diseases; Representing the disease group The number of diseases; Gaussian correlation summary similarity kernel in disease space The calculation formula is as follows: ， in, and It is an adjacency matrix No. Column and number Columns, they represent the first The disease and the first Summary of microRNA associations in a disease This represents the bandwidth of the Gaussian kernel; Disease semantic similarity kernel The calculation formula is as follows: ， in, It is a node in a certain disease. For nodes The set of elements composed of its ancestor nodes, It is a disease semantic score, It is a disease The semantic score; where , It is a disease and disease The relevant semantic values, It is a disease and disease The relevant semantic values; Disease functional similarity kernel matrix The calculation formula is as follows: ， Among them, genes In the genome Chinese; gene In the genome middle; It's a gene. and genome Functional similarity score; It's a gene. and genome Functional similarity score; Representing the genome The number of genes; Representing the genome The number of genes.

6. The training method for the model predicting the association between microRNA and disease as described in claim 1, characterized in that, Multiple microRNA spatial similarity kernels and multiple disease spatial similarity kernels were fused and normalized according to their weights to obtain fused microRNA spatial similarity kernels and fused disease spatial similarity kernels. The fusion formula is as follows: ， ， Where n is the number of similarity kernels; It is a spatial similarity kernel; It is a disease similarity kernel; The normalization formula is as follows: ， ， Where i and j represent the matrix i and j respectively. Column and number List.

7. A model for predicting the association between microRNA and disease, characterized in that, The model was trained using the training method described in any one of claims 1-6 for predicting microRNA and disease association.

Images