Long-chain non-coding RNA-disease association prediction system based on multi-view graph convolutional network

Abstract

The long-chain non-coding RNA-disease association prediction system based on a multi-view graph convolutional network belongs to the technical field of gene-disease association prediction, and aims to solve the problem that the existing long-chain non-coding RNA-disease association prediction cannot sufficiently mine node features and thus has limited prediction capability. The system first processes a long-chain non-coding RNA original feature matrix and a disease original feature matrix to obtain long-chain non-coding RNA-miRNA features, long-chain non-coding RNA-GO features and long-chain non-coding RNA-Gene features, and disease-miRNA features and disease-Gene features; then obtains respective corresponding deep features based on graph convolution and SENet, and then obtains respective corresponding comprehensive features of long-chain non-coding RNA and diseases by using convolution and SENet; and finally obtains a long-chain non-coding RNA-disease association relationship matrix based on a multi-layer perception network. The system is used for long-chain non-coding RNA-disease association prediction.

Description

Technical Field

[0001] This invention belongs to the field of gene-disease association prediction technology, specifically relating to long non-coding RNA-disease association prediction methods, systems, storage media, and devices. Background Technology

[0002] Long non-coding RNAs (lncRNAs) are non-coding RNAs longer than 200 nucleotides. Increasing evidence suggests that lncRNAs are closely related to many vital functions in the human body. LncRNAs participate in dose compensation effects, epigenetic regulation, transcriptional regulation, post-transcriptional regulation, cell cycle regulation, and cell differentiation regulation. Alterations in many lncRNAs are also associated with complex diseases, including cancer and neurological disorders. LncRNAs can be classified into antisense lncRNAs, intronic lncRNAs, large intergenic lncRNAs, promoter-associated lncRNAs, and untranslated region (UTRassociated lncRNAs). Therefore, understanding the relationship between lncRNAs and diseases is crucial. Current methods for studying the association between lncRNAs and diseases mainly fall into two categories: traditional methods and predictive methods. Traditional methods mainly refer to in situ hybridization and overexpression techniques. Although these methods yield accurate results, they are very inefficient. Predictive methods, on the other hand, use high-throughput screening techniques combined with bioinformatics prediction tools to predict the relationship between long non-coding RNAs and diseases, which greatly improves research efficiency.

[0003] Current prediction methods can be categorized into three types: biological network-based methods, machine learning-based methods, and others. Biological network-based methods are based on the assumption that long non-coding RNAs with similar functions may be associated with diseases with similar phenotypes. Machine learning-based methods train classifiers using features of known disease-associated long non-coding RNAs and unknown long non-coding RNAs, ranking candidate long non-coding RNAs based on differences in biological characteristics between datasets. The first two types of methods require predictions based on known long non-coding RNA-gene associations. Since experimentally validated long non-coding RNA-disease associations are currently limited, researchers have begun to focus on predicting long non-coding RNA-disease associations based on known disease-associated genes, miRNAs, and the relationships between long non-coding RNAs and genes or miRNAs.

[0004] The three methods mentioned above predict long non-coding RNA-disease associations from different perspectives. Although most methods have achieved some success in predicting long non-coding RNA-disease associations, they only superficially obtain the characteristics of long non-coding RNA and diseases, and cannot fully explore their deep features, namely node features. Summary of the Invention

[0005] This invention addresses the problem that existing long non-coding RNA-disease association prediction methods have limited predictive power due to their inability to fully exploit node features.

[0006] A long non-coding RNA-disease association prediction system based on multi-view graph convolutional networks includes a data processing module, a feature extraction module, a feature synthesis module, and a multilayer perceptron module.

[0007] Data processing module: used to process the original feature matrix of long non-coding RNA and the original feature matrix of disease to obtain long non-coding RNA-miRNA features, long non-coding RNA-GO features and long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features;

[0008] Feature extraction module: Based on long non-coding RNA-miRNA features, long non-coding RNA-GO features, long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features, the module obtains their respective deep features; the specific processing includes the following steps:

[0009] For the five perspectives of long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene, the compressed features corresponding to each perspective are processed as follows:

[0010] First, a graph convolution process is performed to obtain a graph convolution feature;

[0011] Then, the first-order graph convolution features are passed through a first-order graph convolution network to obtain second-order graph convolution features;

[0012] The first-order graph convolutional features and the second-order graph convolutional features are respectively processed by SENet to obtain first-order graph convolutional weighted features and second-order graph convolutional weighted features. Then, the first-order graph convolutional weighted features and the second-order graph convolutional weighted features are combined to form the deep features of the current viewpoint.

[0013] For each of the five perspectives, we obtained the deep features corresponding to each perspective: long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene.

[0014] Feature synthesis module: Based on the obtained deep features, it obtains the comprehensive features corresponding to long non-coding RNAs and diseases respectively; the specific processing includes the following steps:

[0015] For three perspectives of long non-coding RNA, the deep features corresponding to long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene are input into the convolution module for synthesis, and then SENet is used to obtain the long non-coding RNA comprehensive feature matrix.

[0016] From the two perspectives of disease, the deep features corresponding to disease-miRNA and disease-Gene are input into the convolutional module for integration, and then processed by SENet to obtain the comprehensive disease feature matrix;

[0017] Multilayer perceptron module: Based on the comprehensive feature matrix corresponding to long non-coding RNA and the comprehensive feature matrix corresponding to disease, the long non-coding RNA-disease association matrix is ​​obtained.

[0018] Furthermore, the data processing module's processing procedure includes the following steps:

[0019] The original feature matrix of long non-coding RNA includes the original feature matrices of long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene; the original feature matrix L of long non-coding RNA is of size ln×lu; L is divided to obtain the original feature matrix of long non-coding RNA-miRNA of size ln×lm, the original feature matrix of long non-coding RNA-GO of size ln×lgo, and the original feature matrix of long non-coding RNA-Gene of size ln×lge;

[0020] The original disease feature matrix includes the original disease-miRNA feature matrix and the original disease-Gene feature matrix; the original disease feature matrix D has a size of dn×dv; D is divided to obtain the original disease-miRNA feature matrix of size dn×dm and the original disease-Gene feature matrix of size dn×dge.

[0021] Then each original feature is compressed. The size of the long non-coding RNA-miRNA feature, the long non-coding RNA-GO feature, and the long non-coding RNA-Gene feature is ln×fc, and the size of the disease-miRNA feature and the disease-Gene feature is dn×fc.

[0022] Furthermore, before compressing each original feature, principal component analysis is used to reduce the dimensionality and noise of each segmented original feature matrix.

[0023] Furthermore, the processing procedure of the multilayer perceptron module includes the following steps:

[0024] A multilayer perceptron consists of an input layer, at least two hidden layers, and an output layer.

[0025] The input layer concatenates the comprehensive feature matrix corresponding to long non-coding RNA and the comprehensive feature matrix corresponding to the disease;

[0026] At least two hidden layers are used to learn the relationship between long non-coding RNA features and disease features;

[0027] The output layer has 1 neuron. It calculates the correlation between long non-coding RNAs and diseases to obtain a score that characterizes the degree of relationship between a long non-coding RNA and a disease.

[0028] A computer storage medium storing at least one instruction, the at least one instruction being loaded by a processor and used to run a long non-coding RNA-disease association prediction system based on a multi-view graph convolutional network.

[0029] A long non-coding RNA-disease association prediction device based on a multi-view graph convolutional network is disclosed. The device includes a processor and a memory, wherein the memory stores at least one instruction, which is loaded by the processor to run the long non-coding RNA-disease association prediction system based on the multi-view graph convolutional network.

[0030] A method for predicting long non-coding RNA-disease associations based on multi-view graph convolutional networks includes the following steps:

[0031] S1. Obtain the original feature matrix of long non-coding RNA and the original feature matrix of disease, and process them to obtain long non-coding RNA-miRNA features, long non-coding RNA-GO features and long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features;

[0032] S2. Based on long non-coding RNA-miRNA features, long non-coding RNA-GO features, long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features, the corresponding deep features are obtained; the specific processing includes the following steps:

[0033] For the five perspectives of long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene, the compressed features corresponding to each perspective are processed as follows:

[0034] First, a graph convolution process is performed to obtain a graph convolution feature;

[0035] Then, the first-order graph convolution features are passed through a first-order graph convolution network to obtain second-order graph convolution features;

[0036] The first-order graph convolutional features and the second-order graph convolutional features are respectively processed by SENet to obtain first-order graph convolutional weighted features and second-order graph convolutional weighted features. Then, the first-order graph convolutional weighted features and the second-order graph convolutional weighted features are combined to form the deep features of the current viewpoint.

[0037] For each of the five perspectives, we obtained the deep features corresponding to each perspective: long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene.

[0038] S3. Based on the obtained deep features, the comprehensive features corresponding to long non-coding RNAs and diseases are obtained; the specific processing includes the following steps:

[0039] For three perspectives of long non-coding RNA, the deep features corresponding to long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene are input into the convolution module for synthesis, and then SENet is used to obtain the long non-coding RNA comprehensive feature matrix.

[0040] From the two perspectives of disease, the deep features corresponding to disease-miRNA and disease-Gene are input into the convolutional module for integration, and then processed by SENet to obtain the comprehensive disease feature matrix;

[0041] S4. Input the comprehensive feature matrix corresponding to long non-coding RNA and the comprehensive feature matrix corresponding to disease into the multilayer perceptron to obtain the long non-coding RNA-disease association matrix.

[0042] Furthermore, the processing procedure of S1 includes the following steps:

[0043] The original feature matrix of long non-coding RNA includes the original feature matrices of long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene; the original feature matrix L of long non-coding RNA is of size ln×lu; L is divided to obtain the original feature matrix of long non-coding RNA-miRNA of size ln×lm, the original feature matrix of long non-coding RNA-GO of size ln×lgo, and the original feature matrix of long non-coding RNA-Gene of size ln×lge;

[0044] The original disease feature matrix includes the original disease-miRNA feature matrix and the original disease-Gene feature matrix; the original disease feature matrix D has a size of dn×dv; D is divided to obtain the original disease-miRNA feature matrix of size dn×dm and the original disease-Gene feature matrix of size dn×dge.

[0045] Then each original feature is compressed. The size of the long non-coding RNA-miRNA feature, the long non-coding RNA-GO feature, and the long non-coding RNA-Gene feature is ln×fc, and the size of the disease-miRNA feature and the disease-Gene feature is dn×fc.

[0046] Furthermore, before compressing each original feature, principal component analysis is used to reduce the dimensionality and noise of each segmented original feature matrix.

[0047] Beneficial effects:

[0048] The MVGCNLDA proposed in this invention utilizes graph convolutional networks to obtain node information from the input information and uses SENet to assign attention weights to the acquired features, which can fully extract deep features of long non-coding RNAs and diseases, ultimately achieving good prediction results. In 10-fold cross-validation, the obtained AUC value is 0.8925, AUPR value is 0.0300, F1 score is 0.0171, sensitivity is 0.8963, and the ROC curve is as follows. Figure 2 As shown, the PR curve is as follows Figure 3 As shown. Attached Figure Description

[0049] Figure 1 This is a flowchart of the MVGCNLDA algorithm for predicting long non-coding RNA-disease associations based on multi-view graph convolutional networks.

[0050] Figure 2 This is the ROC curve for the 10-fold cross-validation method.

[0051] Figure 3 This is the PR curve for the 10-fold cross-validation method. Detailed Implementation

[0052] To address the problems of existing methods, this invention proposes a method for predicting long non-coding RNA-disease associations based on multi-view graph convolutional networks. This method can utilize graph convolutional networks to extract deep features of long non-coding RNAs and diseases from multiple data relationships, thereby improving the accuracy of predicting long non-coding RNA-disease associations. The invention will be described below with reference to specific embodiments.

[0053] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0054] This implementation is a long non-coding RNA-disease association prediction system based on multi-view graph convolutional networks, namely MVGCNLDA. MVGCNLDA includes a data processing module, a feature extraction module, a feature synthesis module, and a multilayer perceptron module.

[0055] Data processing module: The input consists of two feature matrices: a raw feature matrix of long non-coding RNAs and a raw feature matrix of diseases. These two features represent the relationships between long non-coding RNAs, diseases, and various biomolecules.

[0056] The original feature matrix of long non-coding RNAs includes the original feature matrices of long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene. It describes the associations of long non-coding RNAs with miRNAs, GOs (Gene Ontology), and Genes. If a long non-coding RNA is associated with a miRNA, the intersection of these two in the matrix is ​​marked as 1; conversely, if the long non-coding RNA is not associated with a miRNA, the intersection is marked as 0. This is the original feature matrix of long non-coding RNA-miRNA. Correspondingly, the original feature matrices of long non-coding RNA-GO and long non-coding RNA-Gene have the same form.

[0057] The original disease feature matrix includes the disease-miRNA original feature matrix and the disease-Gene original feature matrix.

[0058] The input long non-coding RNA raw feature matrix L has a size of ln×lu, and the disease raw feature matrix D has a size of dn×dv. The input matrices L and D are partitioned as follows:

[0059] Knowing the feature matrix sizes of long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene beforehand, L and D are then divided according to the known sizes of each part. The original feature matrix sizes of the resulting long non-coding RNA-miRNA are ln×lm, long non-coding RNA-GO are ln×lgo, long non-coding RNA-Gene are ln×lge, disease-miRNA are dn×dm, and disease-Gene are dn×dge.

[0060] Principal component analysis (PCA) was used to reduce the dimensionality and noise of the original feature matrix after the above segmentation. To facilitate subsequent processing of the algorithm, the original features were uniformly compressed to fc dimension. After compression, the feature size of long non-coding RNA-miRNA, long non-coding RNA-GO and long non-coding RNA-Gene is ln×fc, and the feature size of disease-miRNA and disease-Gene is dn×fc.

[0061] Feature extraction module: The input is the compressed feature matrix, and the deep features corresponding to the long non-coding RNA and the disease are obtained respectively;

[0062] The feature extraction module includes a graph convolutional network and SENet, which aims to obtain deep features from each perspective: long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene.

[0063] For the five perspectives of long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene, the compressed features corresponding to each perspective are processed as follows:

[0064] For a node in the compressed features corresponding to the current viewpoint (a point in the feature matrix), the graph convolutional network multiplies its neighboring nodes by the given weights and sums them, then adds the node itself. After dimensionality reduction, normalization, and aggregation, a brand new node is obtained. The node information at this time includes not only the previous node information, but also the information of the neighboring nodes of the previous node. The node information obtained at this time is called the first-order graph convolutional feature.

[0065] Then, the first-order graph convolution features are passed through a second-order graph convolution network to obtain second-order graph convolution features. The node information obtained at this time not only includes the previous node information, but also the information of two layers of neighboring nodes (including the neighboring node information of the previous node and the neighboring node's neighboring node information).

[0066] The first-order graph convolutional features and the second-order graph convolutional features are respectively weighted by SENet (the weighting operation is the weight allocation within a single matrix) to obtain the first-order graph convolutional weighted features and the second-order graph convolutional weighted features. Then, the first-order graph convolutional weighted features and the second-order graph convolutional weighted features are combined to form the deep features of the current view.

[0067] For each of the five perspectives, we obtained the deep features corresponding to each perspective: long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene.

[0068] Feature synthesis module: Based on the obtained deep features, new long non-coding RNAs and corresponding comprehensive features of diseases are obtained;

[0069] The feature synthesis module includes a convolutional module and SENet. In deep learning, convolution is an ordered process that interweaves multiple information sources while preserving the original features. Based on this, the convolutional module is used to process long non-coding RNAs and disease-related perspectives, respectively:

[0070] For the three perspectives of long non-coding RNA, the deep features corresponding to long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene are input into the convolution module for synthesis. Since long non-coding RNA has features from three perspectives, the convolution module is used to perform convolution synthesis on these three deep features to obtain a long non-coding RNA feature matrix of size ln×fc.

[0071] The long non-coding RNA feature matrix is ​​weighted by SENet to obtain the long non-coding RNA comprehensive feature matrix;

[0072] Similarly, for the two perspectives of the disease, the deep features corresponding to the disease-miRNA and the disease-Gene are input into the convolution module for synthesis. Since the disease has features from two perspectives, the two deep features are combined by convolution using the convolution module to obtain a disease feature matrix of size dn×fc.

[0073] The disease feature matrix is ​​weighted by SENet to obtain the comprehensive disease feature matrix;

[0074] Multilayer perceptron module: Input the comprehensive feature matrix corresponding to long non-coding RNA and the comprehensive feature matrix corresponding to the disease into the multilayer perceptron module to obtain the long non-coding RNA-disease association matrix.

[0075] The multilayer perceptron consists of an input layer, two hidden layers, and an output layer. The first layer is the input layer, which concatenates the long non-coding RNA features and disease features obtained from the previous module. The input dimension is 2×fc. The two hidden layers are used to learn the relationship between the long non-coding RNA features and the disease features. The output of the first hidden layer serves as the input of the second hidden layer. The last layer is the output layer, with only one neuron. The output layer calculates the correlation between the long non-coding RNA and the disease, and finally gives a score that represents the degree of relationship between a certain long non-coding RNA and a certain disease. The closer the score is to 1, the greater the probability of a relationship between the long non-coding RNA and the disease. Specific Implementation Method Two:

[0077] This embodiment is a computer storage medium that stores at least one instruction, which is loaded and executed by a processor to create a long non-coding RNA-disease association prediction system based on a multi-view graph convolutional network.

[0078] It should be understood that any method described in this invention can be provided as a computer program product, software, or computerized method, which may include a non-transitory machine-readable medium on which instructions are stored, which can be used to program a computer system or other electronic device. The storage medium may include, but is not limited to, magnetic storage media, optical storage media; magneto-optical storage media include: read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers; or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0080] This embodiment is a long non-coding RNA-disease association prediction device based on a multi-view graph convolutional network. The device includes a processor and a memory. It should be understood that this includes any device including a processor and a memory described in this invention. The device may also include other units or modules that perform display, interaction, processing, control, and other functions through signals or instructions.

[0081] The memory stores at least one instruction, which is loaded by the processor and used to run a long non-coding RNA-disease association prediction system based on a multi-view graph convolutional network. Specific implementation method four:

[0083] This embodiment describes a long non-coding RNA-disease association prediction method based on multi-view graph convolutional networks. The method involves a processor loading at least one instruction and implementing the following process:

[0084] S1. Obtain the original feature matrix of long non-coding RNA and the original feature matrix of disease, and process them to obtain long non-coding RNA-miRNA features, long non-coding RNA-GO features, long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features; the specific processing includes the following steps:

[0085] The original feature matrix of long non-coding RNA includes the original feature matrices of long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene; the original feature matrix L of long non-coding RNA is of size ln×lu; L is divided to obtain the original feature matrix of long non-coding RNA-miRNA of size ln×lm, the original feature matrix of long non-coding RNA-GO of size ln×lgo, and the original feature matrix of long non-coding RNA-Gene of size ln×lge;

[0086] The original disease feature matrix includes the original disease-miRNA feature matrix and the original disease-Gene feature matrix; the original disease feature matrix D has a size of dn×dv; D is divided to obtain the original disease-miRNA feature matrix of size dn×dm and the original disease-Gene feature matrix of size dn×dge.

[0087] Then, Principal Component Analysis (PCA) was used to reduce the dimensionality and noise of the original feature matrix after the above segmentation. To facilitate subsequent processing of the algorithm, we uniformly compressed the original features to fc dimension. After compression, the feature size of long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene is ln×fc, and the feature size of disease-miRNA and disease-Gene is dn×fc.

[0088] S2. Based on long non-coding RNA-miRNA features, long non-coding RNA-GO features, long non-coding RNA-Gene features, as well as disease-miRNA features and disease-Gene features, the corresponding deep features are obtained; the specific processing includes the following steps:

[0089] For the five perspectives of long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene, the compressed features corresponding to each perspective are processed as follows:

[0090] First, a graph convolution process is performed to obtain a graph convolution feature;

[0091] Then, the first-order graph convolution features are passed through a first-order graph convolution network to obtain second-order graph convolution features;

[0092] The first-order graph convolutional features and the second-order graph convolutional features are respectively processed by SENet to obtain first-order graph convolutional weighted features and second-order graph convolutional weighted features. Then, the first-order graph convolutional weighted features and the second-order graph convolutional weighted features are combined to form the deep features of the current viewpoint.

[0093] For each of the five perspectives, we obtained the deep features corresponding to each perspective: long non-coding RNA-miRNA, long non-coding RNA-GO, long non-coding RNA-Gene, disease-miRNA, and disease-Gene.

[0094] S3. Based on the obtained deep features, the comprehensive features corresponding to long non-coding RNAs and diseases are obtained; the specific processing includes the following steps:

[0095] For three perspectives of long non-coding RNA, the deep features corresponding to long non-coding RNA-miRNA, long non-coding RNA-GO, and long non-coding RNA-Gene are input into the convolution module for synthesis, and then SENet is used to obtain the long non-coding RNA comprehensive feature matrix.

[0096] From the two perspectives of disease, the deep features corresponding to disease-miRNA and disease-Gene are input into the convolutional module for synthesis, and then processed by SENet to obtain the comprehensive disease feature matrix.

[0097] S4. Input the comprehensive feature matrix corresponding to long non-coding RNA and the comprehensive feature matrix corresponding to disease into the multilayer perceptron to obtain the long non-coding RNA-disease association matrix.

[0098] Example

[0099] In this invention, two datasets—long non-coding RNA (LNI) raw features and disease raw features—are used to illustrate the effectiveness of the proposed graph convolutional network-based LNI-disease association prediction algorithm. The LNI raw feature data comes from the databases LncRNA2Target, GeneRIF, and StarBase v2.0. Specifically, LNI-Gene features are obtained from LncRNA2Target, GO features from GeneRIF, and miRNA features from StarBase v2.0. After preprocessing the collected data, 5347 genes (Gene), 487 gene functional annotations (GO), and 232 miRNAs associated with 240 LNIs were obtained. Finally, a 240×6066 LNI raw feature matrix L was obtained, where Lij is 1 if a LNI i is confirmed to be related to a biologically related data j, and 0 if no relationship is confirmed. Disease feature data were obtained from the DisGeNet and HMDD databases. Disease-gene features were obtained from DisGeNet, and disease-miRNA features were obtained from HMDD. After preprocessing, 10146 genes and 495 miRNAs associated with 412 diseases were obtained, resulting in a final disease feature matrix D of size 412 × 10⁶²¹. In this matrix, Dij is set to 1 if disease i is confirmed to be related to a gene or miRNA j; otherwise, Dij is set to 0.

[0100] The attributes of the two datasets, long non-coding RNA raw features and disease raw features, are shown in Table 1. This invention employs 10-fold cross-validation to verify the accuracy of the algorithm, and uses the area under the ROC curve (AUC value), the area under the PR curve (AUPR value), the F1 score, and sensitivity as metrics. Higher values ​​for these four metrics indicate better algorithm accuracy.

[0101] Table 1. Detailed attributes of the long non-coding RNA and disease datasets.

[0102]

[0103] To verify the effectiveness of the proposed neural network algorithm, it was run on the PyTorch framework with a learning rate (lr) of 0.0001. The Adam optimizer was used to optimize the network parameters, the random number seed (n) was set to 1, the number of training epochs was set to 1000, the compression dimension (fc) of the principal component analysis (PCA) method was 128, the dimension of the first hidden layer of the multilayer perceptron was set to 64, the dimension of the second hidden layer of the multilayer perceptron was set to 64, the mean squared error loss function was used, and an L2 regularization term was used to prevent overfitting. The L2 regularization penalty coefficient (weight_decay) was 0.00005. Experimental hardware platform: CPU: i5-8300H, GPU: GTX-1050Ti, Memory: 4GB.

[0104] In the 10-fold cross-validation method, the obtained ROC curve is as follows: Figure 2 As shown, the PR curve is as follows Figure 3 As shown, the AUC value was 0.8925, the AUPR value was 0.0300, the F1 score was 0.0171, and the sensitivity was 0.8963. The AUC value is an important indicator for evaluating the quality of a method; the closer the value is to 1, the higher the accuracy of the method. In the 10-fold cross-validation method, the AUC value was 0.8925, indicating that the algorithm proposed in this invention has high prediction accuracy.

[0105] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A long-chain non-coding RNA-disease association prediction system based on multi-view graph convolutional network, characterized in that, The data processing module, the feature extraction module, the feature synthesis module and the multi-layer perception module are included. The data processing module is used for processing the long-chain non-coding RNA original feature matrix and the disease original feature matrix to obtain long-chain non-coding RNA-miRNA features, long-chain non-coding RNA-GO features and long-chain non-coding RNA-Gene features, and disease-miRNA features and disease-Gene features. The feature extraction module is based on the long-chain non-coding RNA-miRNA features, the long-chain non-coding RNA-GO features and the long-chain non-coding RNA-Gene features, and the disease-miRNA features and the disease-Gene features to obtain respective corresponding deep features; the specific processing process includes the following steps: For the five perspectives of long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO, long-chain non-coding RNA-Gene, disease-miRNA and disease-Gene, the compressed features corresponding to each perspective are processed as follows: First, a graph convolution feature is obtained through a graph convolution process; Then, a second graph convolution feature is obtained through a graph convolution network; The first graph convolution feature and the second graph convolution feature are respectively subjected to SE-Net to obtain a first graph convolution weighted feature and a second graph convolution weighted feature, and then the first graph convolution weighted feature and the second graph convolution weighted feature are combined to form a deep feature of the current perspective; For the five perspectives, the deep features corresponding to each perspective of long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO, long-chain non-coding RNA-Gene, disease-miRNA and disease-Gene are obtained respectively; The feature synthesis module is based on the obtained deep features to obtain respective corresponding synthesis features of long-chain non-coding RNA and diseases; the specific processing process includes the following steps: For the three perspectives of long-chain non-coding RNA, the deep features corresponding to long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO and long-chain non-coding RNA-Gene are input into a convolution module for synthesis, and then a long-chain non-coding RNA synthesis feature matrix is obtained through SE-Net; For the two perspectives of diseases, the deep features corresponding to disease-miRNA and disease-Gene are input into a convolution module for synthesis, and then a disease synthesis feature matrix is obtained through SE-Net; The multi-layer perception module is based on the long-chain non-coding RNA corresponding synthesis feature matrix and the disease corresponding synthesis feature matrix to obtain a long-chain non-coding RNA-disease correlation relationship matrix.

2. The long non-coding RNA-disease association prediction system based on multi-view graph convolutional network according to claim 1, wherein, The processing process of the data processing module includes the following steps: The long-chain non-coding RNA original feature matrix includes long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO and long-chain non-coding RNA-Gene original feature matrix; the long-chain non-coding RNA original feature matrix L has a size of ln×lu; L is segmented to obtain a long-chain non-coding RNA-miRNA original feature matrix with a size of ln×lm, a long-chain non-coding RNA-GO original feature matrix with a size of ln×lgo and a long-chain non-coding RNA-Gene original feature matrix with a size of ln×lge; The disease original feature matrix includes disease-miRNA original feature matrix and disease-Gene original feature matrix; the disease original feature matrix D has a size of dn×dv; D is segmented to obtain a disease-miRNA original feature matrix with a size of dn×dm and a disease-Gene original feature matrix with a size of dn×dge; Then each original feature is compressed, and the long-chain non-coding RNA-miRNA feature, the long-chain non-coding RNA-GO feature and the long-chain non-coding RNA-Gene feature all have a size of ln×fc, and the disease-miRNA feature and the disease-Gene feature all have a size of dn×fc.

3. The long non-coding RNA-disease association prediction system based on multi-view graph convolutional network according to claim 2, characterized in that, Before each original feature is compressed, principal component analysis is used to reduce dimension and reduce noise of each original feature matrix after segmentation.

4. The long non-coding RNA-disease association prediction system based on multi-view graph convolutional network according to claim 1, 2 or 3, characterized in that, The multi-layer perception module processing process includes the following steps: The multi-layer perception includes an input layer, at least two hidden layers and an output layer; The input layer splices the comprehensive feature matrix corresponding to the long-chain non-coding RNA and the comprehensive feature matrix corresponding to the disease; The at least two hidden layers are used to learn the relationship between the long-chain non-coding RNA feature and the disease feature; The output layer has a neuron number of 1, and the output layer obtains a score for representing the relationship degree between a long-chain non-coding RNA and a disease by calculating the correlation degree between the long-chain non-coding RNA and the disease.

5. A computer storage medium, characterized in that, The storage medium stores at least one instruction, and the at least one instruction is loaded and run by the processor to execute the long-chain non-coding RNA-disease correlation prediction system based on the multi-view graph convolutional network in any one of claims 1 to 4.

6. A long-chain non-coding RNA-disease association prediction device based on a multi-view graph convolutional network, characterized by, The device includes a processor and a memory, and the memory stores at least one instruction, and the at least one instruction is loaded and run by the processor to execute the long-chain non-coding RNA-disease correlation prediction system based on the multi-view graph convolutional network in any one of claims 1 to 4.

7. A long-chain non-coding RNA-disease association prediction method based on a multi-view graph convolutional network, characterized in that, The method includes the following steps: S1, obtaining long-chain non-coding RNA original feature matrix and disease original feature matrix and processing to obtain long-chain non-coding RNA-miRNA feature, long-chain non-coding RNA-GO feature and long-chain non-coding RNA-Gene feature, and disease-miRNA feature and disease-Gene feature; S2, based on long-chain non-coding RNA-miRNA features, long-chain non-coding RNA-GO features and long-chain non-coding RNA-Gene features, and disease-miRNA features and disease-Gene features, to get the corresponding deep features respectively; The specific processing process includes the following steps: For long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO, long-chain non-coding RNA-Gene, disease-miRNA and disease-Gene five perspectives, the compressed features corresponding to each perspective are processed as follows: First, a graph convolution feature is obtained by one graph convolution processing; Then the first graph convolution feature is obtained by a graph convolution network to get the second graph convolution feature; The first graph convolution feature and the second graph convolution feature are respectively processed by SE Net to obtain the first graph convolution weighted feature and the second graph convolution weighted feature, and then the first graph convolution weighted feature and the second graph convolution weighted feature are combined to form the deep feature of the current perspective; For the five perspectives, the deep features corresponding to each perspective of long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO, long-chain non-coding RNA-Gene, disease-miRNA and disease-Gene are obtained respectively; S3, based on the obtained deep features, the comprehensive features corresponding to each long-chain non-coding RNA and disease are obtained; The specific processing process includes the following steps: For the three perspectives of long-chain non-coding RNA, the deep features corresponding to long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO and long-chain non-coding RNA-Gene are input into the convolution module for comprehensive processing, and then the SE Net is used to obtain the long-chain non-coding RNA comprehensive feature matrix; For the two perspectives of disease, the deep features corresponding to disease-miRNA and disease-Gene are input into the convolution module for comprehensive processing, and then the SE Net is used to obtain the disease comprehensive feature matrix; S4, input the comprehensive feature matrix corresponding to long-chain non-coding RNA and the comprehensive feature matrix corresponding to disease into the multi-layer perception machine to obtain the long-chain non-coding RNA-disease correlation matrix. 8.The method of claim 7, wherein the method comprises: The processing process of S1 includes the following steps: The long-chain non-coding RNA original feature matrix includes the original feature matrix of long-chain non-coding RNA-miRNA, long-chain non-coding RNA-GO and long-chain non-coding RNA-Gene; The size of the long-chain non-coding RNA original feature matrix L is ln×lu; L is segmented to obtain the original feature matrix of long-chain non-coding RNA-miRNA with a size of ln×lm, the original feature matrix of long-chain non-coding RNA-GO with a size of ln×lgo, and the original feature matrix of long-chain non-coding RNA-Gene with a size of ln×lge; The disease original feature matrix includes a disease-miRNA original feature matrix and a disease-Gene original feature matrix; the disease original feature matrix D has a size of dn×dv; the D is segmented to obtain a disease-miRNA original feature matrix with a size of dn×dm and a disease-Gene original feature matrix with a size of dn×dge; Then, each original feature is compressed, and the long-chain non-coding RNA-miRNA feature, the long-chain non-coding RNA-GO feature and the long-chain non-coding RNA-Gene feature each have a size of ln×fc, and the disease-miRNA feature and the disease-Gene feature each have a size of dn×fc. 9.The long non-coding RNA-disease association prediction method based on multi-view graph convolutional network according to claim 8, characterized in that, Before each original feature is compressed, principal component analysis is used to reduce the dimension and reduce noise of each original feature matrix after segmentation.

Images