Calculation method for predicting multi-dimensional relationship between miRNA and disease

Abstract

The invention discloses a calculation method for predicting a multi-dimensional relation between miRNA and diseases. The method comprises the following steps: firstly, SMCLMDA constructs a similarity view and a meta-path view based on a similarity matrix and an incidence matrix, and extracts high-order representation based on graph convolution; secondly, the consistency of characterization among different views is optimized based on comparative learning; and finally, predicting the traditional association, up / down regulation and causal / non-causal relationship between the miRNA and the disease by using a multi-layer perceptron. According to the method, an area (AUC) under a working characteristic (ROC) curve of a subject, an area (AUPR) under a precision / recall rate (PR) curve, Accuracy, Precision, Recall, Specific, a Marius correlation coefficient (MCC) and F1-score are used as evaluation indexes. The method comprises the following steps of: taking the AUC under the curve, the AUPR under the curve, the Accuracy, the Precision, the Recall, the Specific, the Marius correlation coefficient (MCC) and the F1-score as evaluation indexes; in addition, case analysis is carried out on lung cancer, breast cancer and colorectal cancer, and the robustness and superiority of the model are further reflected.

Description

Technical Field

[0001] The present invention relates to the technical field of data and text mining in bioinformatics, and specifically to a computational method for predicting the association between miRNAs and diseases. Background Art

[0002] MicroRNA (miRNA) is a subset of non-coding RNAs, usually endogenous short RNA molecules with a length of about 22 nt, which can regulate the cleavage of target mRNAs post-transcriptionally or only inhibit their translation. miRNAs are involved in many important biological processes, and they control cell proliferation, differentiation, metabolism, and apoptosis by inhibiting the expression of mRNAs in organisms. In addition, many biomedical studies have shown that miRNAs are associated with various human diseases, including cancer, cardiovascular diseases, and other complex diseases. For example, compared with normal tissues, the transcription of many microRNAs (miRNAs) is upregulated in papillary thyroid cancer tumors, among which miR-221, miR-222, and miR-146 are the most significantly upregulated, and they can effectively distinguish papillary thyroid cancer from normal thyroid. Gao et al. found that the expression level of let-7 miRNA is decreased in human lung cancer. Recent studies have shown that the expression level of the oncogenic miRNA miR-21-5p is increased in rectal cancer, breast cancer, and lung cancer. miR-21-5p promotes cell proliferation, migration, and invasion, and inhibits apoptosis by downregulating targets such as PDCD4 (programmed cell death protein 4) and PTEN (phosphatase and tensin homolog). Therefore, exploring the multi-dimensional relationship between miRNAs and diseases can help understand the pathogenic mechanism and is beneficial to the diagnosis, treatment, and prognosis of complex human diseases.

[0003] For the prediction of miRNA-disease associations, for a long time in the past, the association information between miRNAs and diseases was mainly obtained through wet biological experiments, which was both time-consuming and costly. In recent years, computational methods based on graph convolutional neural networks have achieved good results. However, most methods only obtain the representations of miRNAs and diseases from a single similarity view. There are also some methods that construct biological meta-paths, such as introducing genes as mediators to form multi-layer networks. Although this method adds a lot of biological prior knowledge, it also introduces more noise and computational complexity. In addition, many studies directly explore whether there is an association between miRNAs and diseases without further in-depth exploration. Summary of the Invention

[0004] The present invention lies in predicting the multi-dimensional relationship between miRNAs and diseases. The constructed statistical meta-path further enhances the representations of miRNAs and diseases in the similarity view through a contrastive learning strategy. In addition, the local information obtained by Node2Vec is used as the initial input for graph convolution, enabling the model to possess graph structure knowledge at the beginning of training and accelerating the convergence process. Finally, a multi-layer perceptron is used to predict the traditional associations, up / down regulations, and causal / non-causal relationships between miRNAs and diseases. The following are the technical solutions for achieving the objectives of the present invention, including the following steps: 1. Dataset arrangement: The HMDD v4.0 and miR2Disease databases are arranged, and the data types include traditional associations, up / down regulations, and causal / non-causal relationships; 2. Construction of the meta-path view: Different from previous biological meta-paths, the meta-path constructed based on statistical methods (matrix multiplication) has low computational complexity and less noise; 3. Acquisition of initial local features: Node2Vec is used to obtain the local information of miRNAs and diseases for the initial features of the graph convolutional neural network; 4. Feature extraction: High-order embedding representations of the similarity view and the meta-path view are obtained based on the graph convolutional neural network.

[0005] 5. Feature enhancement: The embedding information obtained from the meta-path view further enriches the information of the similarity view through contrastive learning; 6. Model evaluation: The HMDD v4.0 dataset is evaluated using ten-fold cross-validation, the miR2Disease dataset is evaluated using ten-fold and five-fold cross-validation, and the HMDD v3.2 dataset is evaluated using ten-fold cross-validation and stratified sampling. Specific evaluation metrics: Area Under the Receiver Operating Characteristic (ROC) Curve (AUC), Area Under the Precision / Recall (PR) Curve (AUPR), Accuracy, Precission, Recall, Specificity, Matthews Correlation Coefficient (MCC), and F1-score.

[0006] Beneficial effects This model uses statistical meta-paths for information enhancement. Compared with previous biological meta-paths, it not only reduces the computational complexity but also reduces the noise in the view. For the initial features of previous graph convolutions, randomization was used. We adopt the Node2Vec algorithm to obtain local information and use this information as the initial feature input for graph convolution, which enables the model to have certain graph structure information at the beginning of training and speeds up the convergence rate. In addition, the features learned in the meta-path further enhance the information in the similar views through the contrastive learning strategy. Compared with existing methods, we have achieved significant improvements. The AUC metrics for all relationship predictions are higher than the state-of-the-art methods, and case studies further demonstrate the effectiveness of the method we proposed. Description of the Drawings

[0007] Figure 1 It is a flowchart of a computational method for predicting the multi-dimensional relationship between miRNA and diseases.

[0008] Figure 2 Detailed information of the dataset Figure 3 Performance comparison of all methods for predicting traditional associations on the unbalanced test set Figure 4 Performance comparison of all methods for predicting traditional associations on the balanced test set Figure 5 Comparison of all methods in predicting up / down regulation Figure 6 Comparison of all methods in predicting causal / non-causal relationships Figure 7 Top ten related miRNAs predicted for three cancers Detailed Implementation Manner

[0009] The following clarifies the detailed implementation manner of the present invention in conjunction with the drawings. The drawings are only for illustrative purposes and should not be construed as a limitation of the present invention. The drawings are for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0010] As Figure 1 shown, Figure 1 It shows the workflow of this work. The statistical meta-path view is constructed through a graph convolutional neural network, and the representation of the similar views is further enhanced based on the contrastive learning strategy. The initial features of the graph convolution adopt the local embedding calculated by Node2Vec, which is convenient for accelerating the convergence of the model. Finally, we predict the multi-dimensional relationship between miRNA and diseases based on a multi-layer perceptron.

[0011] To comprehensively evaluate the relationship between miRNAs and diseases, we collated the HMDD v4.0 and miRDisease databases to obtain the relationships between various miRNAs and diseases. The detailed dataset information is as Figure 1 shown: Note: 0 / 1 indicates traditional presence / absence of association, 1 / -1 indicates up / down regulation, causality / non-causality indicates causal / non-causal, and miR2Disease+

[45] indicates that miR2Disease incorporates the research data of the literature SGNNMD: signed graph neural network for predicting deregulation types of miRNA-disease associations.

[0012] The performance of the model was evaluated using ten-fold cross-validation on the above HMDD v4.0 (0 / 1), miR2Disease+

[43] (0 / 1), and HMDD v3.2 (0 / 1) datasets. Five-fold cross-validation was used to evaluate the model performance on the miR2Disease+

[43] (-1 / 1) dataset. Stratified sampling was used to evaluate the model performance on the yanzheng (causality / non-causality) dataset. We used the area under the receiver operating characteristic (ROC) curve (AUC), the area under the precision / recall (PR) curve (AUPR), Accuracy, Precission, Recall, Specificity, Matthews correlation coefficient (MCC), and F1-score as evaluation metrics. We also conducted tests on balanced and unbalanced test sets to comprehensively evaluate the generalization of the model. Note:

[43] represents the paper: "Sgnnmd: signed graph neural network for predicting deregulation types of mirna-disease associations."

[0013] Finally, our method was compared with the existing state-of-the-art methods. For traditional association prediction, Figure 3 and Figure 4 the tables in Figure 5 and 6 show the results on unbalanced and balanced test sets respectively. The comparison results for up / down regulation and causality / non-causality are as shown in

[0014] In addition, we conducted case studies on three types of cancer: lung cancer, breast cancer, and colorectal cancer. The predicted potential unknown associations were verified through the dbDEMC and miRCancer databases, and the results are as Figure 7 shown. The results indicate that the top ten potential miRNAs for all three types of cancer have been confirmed, further demonstrating the robustness and superiority of our method in predicting potential miRNA-disease associations.

Claims

1. A computational method for predicting the multi-dimensional relationship between miRNA and diseases, the process of which includes the following steps: 1) Dataset collation: The HMDD v4.0 and miR2Disease databases were collated, and the data types include traditional associations, up / down regulations, and causal / non-causal relationships; 2) Construction of meta-path views: Different from previous biological meta-paths, the meta-paths constructed by us based on statistical methods (matrix multiplication) have low computational complexity and less noise; 3) Acquisition of initial local features: Node2Vec was used to obtain the local information of miRNAs and diseases for the initial features of the graph convolutional neural network; 4) Feature extraction: High-order embedding representations of similarity views and meta-path views were obtained based on the graph convolutional neural network; 5) Feature enhancement: The embedding information obtained from the meta-path view further enriches the information of the similarity view through contrastive learning; 6) Model evaluation: The HMDD v4.0 dataset was evaluated using ten-fold cross-validation, the miR2Disease dataset was evaluated using ten-fold and five-fold cross-validations, and the HMDD v3.2 dataset was evaluated using ten-fold cross-validation and stratified sampling. Specific evaluation metrics: Area Under the Receiver Operating Characteristic (ROC) Curve (AUC), Area Under the Precision / Recall (PR) Curve (AUPR), Accuracy, Precision, Recall, Specificity, Matthews Correlation Coefficient (MCC), and F1-score.

2. This method uses Node2Vec to obtain the initial features of miRNAs and diseases as the input of graph convolution, which enables the model to have the basic structure of the graph at the beginning of training and can accelerate the model convergence. To enrich the semantic information of miRNAs and diseases, a statistical meta-path view was constructed under the premise of low computational complexity. Then, a graph convolutional network was used to extract the high-order information of similarity views and meta-path views, and the meta-path view further enriches the information of the similarity view based on the contrastive learning strategy. Finally, a multi-layer perceptron was used for prediction.

3. According to the dataset described in claim 1, this method basically studies all the relationships between miRNAs and diseases, while previous studies only focused on the existence of associations between miRNAs and diseases.

Images