IncRNA-disease association prediction method based on matrix decomposition and heterogeneous graph reasoning

Abstract

The invention provides an lncRNA-disease association prediction system based on matrix decomposition and heterogeneous graph reasoning, and relates to the field of biological information and deep learning. The method comprises the following steps: firstly, on the basis of a known lncRNA-disease incidence matrix, calculating Gaussian interaction profile kernel similarity of lncRNA and a disease by utilizing a Gaussian kernel, and collecting expression similarity of the lncRNA and semantic similarity data of the disease; then, through a nonlinear similarity network fusion strategy, the similarities obtained by different calculation methods are integrated into the comprehensive similarity of the lncRNA and the disease; then, some regularization terms and range constraints are introduced into a target function to construct a constraint matrix decomposition model, and an alternating least square method is used for optimization to obtain an updated lncRNA-disease incidence matrix; then, integrating the updated lncRNA-disease incidence matrix, the updated lncRNA comprehensive similarity matrix and the updated disease comprehensive similarity matrix into an lncRNA-disease heterogeneous network, and defining an adjacent matrix of the network as a target matrix; and finally, based on the network, using a heterogeneous graph reasoning algorithm to obtain a prediction scoring matrix.

Description

Technical Field

[0001] The present invention relates to the technical field of lncRNA-disease association prediction, and in particular to a lncRNA-disease association prediction method based on matrix decomposition and heterogeneous graph reasoning. Background Art

[0002] A growing body of research has demonstrated that microbes play a crucial role in the development, treatment, and prognosis of many diseases, including neurological disorders and cancer. Identifying disease-associated microbes not only helps to reveal the pathological mechanisms of human diseases but also provides potential biomarkers for diagnosis and prognosis. For example, systematically identifying microbes associated with disease pathology can assist physicians and biologists in identifying biomarkers for clinical or experimental diagnosis and treatment, particularly for complex human diseases. Furthermore, computational prediction of pathogenic microbes can help pharmacologists and biologists more effectively narrow the list of candidate compounds, further guiding experimental planning and reducing costs. While biological experiments remain the gold standard for confirming disease-microbe associations, they are time-consuming and costly. Therefore, there is an urgent need to develop more efficient computational methods to identify high-probability pairs of microbe-disease associations, which are crucial for guiding biological experiments. Matrix factorization is an effective method for predicting interactions, but current models fail to fully capture the essential characteristics of microbes and diseases, and the prediction results lack interpretability. Therefore, this study proposes a microbe-disease interaction prediction method based on a matrix factorization algorithm. Summary of the Invention

[0003] The purpose of the present invention is to address the difficulties in the field of lncRNA-disease association prediction mentioned above, and to provide a lncRNA-disease association prediction method based on matrix decomposition and heterogeneous graph reasoning, which helps researchers to make efficient predictions based on the current relevant data information of lncRNAs and diseases. The technical solution of the present invention is as follows:

[0004] 1. A lncRNA-disease association prediction method based on matrix decomposition and heterogeneous graph reasoning, including the following parts:

[0005] A. Based on the known lncRNA-disease association matrix, the Gaussian kernel is used to calculate the Gaussian interaction profile kernel similarity between lncRNAs and diseases. Preprocessing is performed based on the lncRNA expression similarity and the Gaussian interaction profile kernel similarity between lncRNAs and diseases. Preprocessing is performed based on the semantic similarity of diseases and the Gaussian interaction profile kernel similarity between diseases.

[0006] B. Based on the nonlinear similarity network fusion strategy, the similarities of the two lncRNAs and the similarities of the two diseases were fused respectively to obtain a comprehensive similarity network of lncRNAs and diseases.

[0007] C. Based on the constructed constrained collaborative matrix factorization model, the sparse norm and range constraints are introduced into the objective function. D. The model is solved using an iterative algorithm framework based on the alternating least squares method.

[0008] E. Construct the final lncRNA-disease network based on the lncRNA-disease association network, the comprehensive lncRNA similarity network, and the comprehensive disease similarity network. A prediction score matrix is ​​obtained based on the lncRNA-disease network using a heterogeneous graph inference algorithm.

[0009] 2. According to claim 1, the Gaussian interaction profile kernel similarity between lncRNAs and diseases is calculated using a Gaussian kernel based on a known lncRNA-disease association matrix. Preprocessing is performed based on the lncRNA expression similarity and the lncRNA Gaussian interaction profile kernel similarity. Preprocessing is performed based on the disease semantic similarity and the disease Gaussian interaction profile kernel similarity. The present invention operates based on the LncRNADisease2.0 database. Gaussian interaction profile kernel similarities between lncRNAs and diseases are calculated using known lncRNA-disease association data downloaded from this database. In addition, lncRNA expression similarity and disease semantic similarity are calculated based on the lncRNAs and diseases included in this database.

[0010] 3. The nonlinear similarity network fusion strategy described in claim 1 is used to fuse the similarities of the two lncRNAs and the two diseases, respectively, to obtain a comprehensive similarity network for lncRNAs and diseases. Using the nonlinear similarity network fusion strategy, multi-source similarities are effectively integrated and fused, with strong generalization capabilities and rapid convergence.

[0011] 4. The constrained collaborative matrix decomposition model based on the construction of claim 1. This invention uses the constrained collaborative matrix decomposition method, introducing regularization terms and constraints into the objective function, to fully extract the essential characteristics of the two biological feature matrices while ensuring data validity and interpretability. Consequently, this method yields a richer and more robust lncRNA-disease association network.

[0012] 5. Solve the model using the iterative algorithm framework based on alternating least squares (ALS) according to claim 1. The present invention uses an iterative algorithm based on ALS for optimization, alternatingly updating the lncRNA and disease feature matrices until convergence. This algorithm exhibits excellent robustness when processing sparse data, with rapid convergence and low complexity.

[0013] 6. The method of claim 1 constructs a final lncRNA-disease network based on the lncRNA-disease association network, the comprehensive lncRNA similarity network, and the comprehensive disease similarity network. A prediction score matrix is ​​obtained based on the lncRNA-disease network using a heterogeneous graph inference algorithm. The method concatenates the lncRNA-disease association matrix, the comprehensive lncRNA similarity matrix, the transposed lncRNA-disease association matrix, and the comprehensive disease similarity matrix to obtain a final lncRNA-disease network definition target matrix. Based on this network, a heterogeneous graph inference algorithm is used to obtain prediction scores for unknown associations. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a lncRNA-disease association prediction process based on matrix decomposition and heterogeneous graph reasoning. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions and advantages of the present invention more clear, the lncRNA-disease association prediction method of the present invention is further described in detail below with reference to the accompanying drawings.

[0016] During the data preprocessing phase, Gaussian kernels were used to calculate the Gaussian interaction profile kernel similarity between microbes and diseases based on the known microbe-disease association matrix. Multi-source data on microbes and diseases derived from different computational methods were integrated to construct two microbial similarity matrices and two disease similarity matrices.

[0017] A customized similarity fusion strategy is used to fuse the similarities of two microorganisms. If the microorganism-disease association matrix is ​​correlated, the two microorganism similarities are summed and averaged. Otherwise, the Gaussian interaction profile kernel of the microorganism is used to replace it. The same operation is performed for disease similarity fusion.

[0018] The objective function of the collaborative matrix factorization model is constructed and norm and range constraints are introduced. An iterative algorithm based on alternating least squares is used for optimization, alternating between updating the microbial and disease signature matrices until convergence. Finally, the resulting microbial and disease signature matrices are multiplied to obtain the prediction score matrix.

Claims

1. A lncRNA-disease association prediction method based on matrix decomposition and heterogeneous graph reasoning, including the following parts: A. Based on the known lncRNA-disease association matrix, the Gaussian kernel is used to calculate the Gaussian interaction profile kernel similarity between lncRNAs and diseases. Preprocessing is performed based on the lncRNA expression similarity and the Gaussian interaction profile kernel similarity between lncRNAs and diseases. Preprocessing is performed based on the semantic similarity of diseases and the Gaussian interaction profile kernel similarity between diseases. B. Based on the nonlinear similarity network fusion strategy, the similarities of the two lncRNAs and the similarities of the two diseases were fused respectively to obtain a comprehensive similarity network of lncRNAs and diseases. C. Based on the constructed constrained collaborative matrix factorization model, the sparse norm and range constraint terms are specially introduced into the objective function. D. Solve the model using an iterative algorithm framework based on the alternating least squares method. E. Construct the final lncRNA-disease network based on the lncRNA-disease association network, the comprehensive lncRNA similarity network, and the comprehensive disease similarity network. A prediction score matrix is ​​obtained based on the lncRNA-disease network using a heterogeneous graph inference algorithm.

2. The method according to claim 1 calculates the Gaussian interaction profile kernel similarity between lncRNA and disease based on the known lncRNA-disease association matrix using a Gaussian kernel. Preprocessing was performed based on lncRNA expression similarity and lncRNA Gaussian interaction profile kernel similarity. Preprocessing was also performed based on disease semantic similarity and disease Gaussian interaction profile kernel similarity. This method operates based on the LncRNADisease2.0 database. Using known lncRNA-disease association data downloaded from this database, Gaussian interaction profile kernel similarities were calculated for lncRNAs and diseases, respectively. In addition, the expression similarity of lncRNAs and the semantic similarity of diseases were calculated based on the lncRNAs and diseases involved in the database.

3. The nonlinear similarity network fusion strategy of claim 1 is used to fuse the similarities of the two lncRNAs and the two diseases, respectively, to obtain a comprehensive similarity network for lncRNAs and diseases. Using the nonlinear similarity network fusion strategy, multi-source similarities are effectively integrated and fused, with strong generalization capabilities and rapid convergence.

4. The constrained collaborative matrix decomposition model based on the construction of claim 1. The present invention utilizes a constrained collaborative matrix decomposition method, introducing regularization terms and constraints into the objective function to fully extract the essential features of the two biological feature matrices while ensuring data validity and interpretability. Consequently, this method yields a richer and more robust lncRNA-disease association network.

5. The model is solved using the iterative algorithm framework based on alternating least squares according to claim 1. The present invention uses an iterative algorithm based on alternating least squares for optimization, alternatingly updating the feature matrices of lncRNAs and diseases until convergence. This algorithm exhibits excellent robustness when processing sparse data, with rapid convergence and low complexity.

6. The method of claim 1, wherein the final lncRNA-disease network is constructed based on the lncRNA-disease association network, the comprehensive lncRNA similarity network, and the comprehensive disease similarity network. A prediction score matrix is ​​obtained based on the lncRNA-disease network using a heterogeneous graph inference algorithm. The method comprises concatenating the lncRNA-disease association matrix, the comprehensive lncRNA similarity matrix, the transposed lncRNA-disease association matrix, and the comprehensive disease similarity matrix to obtain the final lncRNA-disease network definition target matrix. Based on this network, a heterogeneous graph inference algorithm is used to obtain the prediction scores of unknown associations.

Images