Method for predicting disease through circular RNA based on graph neural network

Abstract

The invention relates to the technical field of bioinformatics, and provides a CDA prediction model composed of an original channel and an interaction channel. The method comprises the following steps: firstly, acquiring original data from two open source databases of circ2Disease and circR2Disease; the method comprises the following steps: processing original data, screening out human circRNA and disease association, constructing a sparse matrix, carrying out multiple similarity calculations based on the sparse matrix to obtain initial features, and then entering two channels: extracting high-order features and retaining important information by using a double-attention mechanism in a channel I; in the second channel, degree segmentation and Wave modules are used, and potential information association of different sample features is fused. And then, the information of the two channels is integrated through a channel fusion module to generate final output, and finally, a model with optimal performance is found through five-fold cross validation and evaluation indexes. Experimental results based on two data sets show that the performance of the model in a CDA prediction task is superior to that of an existing method, and further case research and generalization analysis verify the effectiveness and wide applicability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for predicting diseases using circular RNA based on a graph neural network. Background Art

[0002] Circular RNA (circRNA) was discovered decades ago, but it has not received much attention. With the continuous advancement of technology, high-throughput sequencing and bioinformatics analysis have found that circRNA is closely related to cancer and is considered to be one of the regulators of various cancers. circRNA can directly or indirectly affect cell cycle regulators or related signaling pathways, thereby inhibiting or promoting cancer progression. At the same time, it also plays an important role in regulating cancer cell proliferation and metastasis. circRNA is an attractive therapeutic target in cancer treatment with prospective diagnostic and prognostic value.

[0003] For example, hsa_circ_0001320, which has been confirmed by research, is not expressed in lung cancer cells, but it will be expressed after treatment and its expression level will increase in combined drug therapy. Hsa_circ_0002762 has been identified as an important regulator of cervical cancer, and dynamic monitoring after surgery has revealed its practical utility in predicting clinical prognosis. CircCSPP1 was found to be closely related to gastric cancer and has high sensitivity and specificity in distinguishing gastric cancer patients from healthy individuals. CircLIFR can delay the progression of thyroid carcinoma through the TIMP2 axis and inhibit tumor growth in vivo.

[0004] Typically, biotechniques such as CLIP-seq, high-throughput sequencing, and immunoprecipitation are used for identification. However, they are usually time-consuming and labor-intensive, making it difficult to screen CDAs on a large scale. In contrast, computational methods are not bound by these limitations and have attracted great attention. As an effective solution for large-scale screening of CDAs, computational methods help prioritize CDA candidates and their subsequent experimental confirmation.

[0005] In recent years, many computational methods for predicting circRNA-disease associations have been proposed. For example, Niu et al. proposed the GMNN2CD model, which uses graph Markov neural network (GMNN) and variational inference, calculates Gaussian and disease similarity networks as raw feature inputs, integrates the correlation between labels into the learning process through Markov neural network, and then optimizes the parameters in the model propagation process through variational inference. Deng et al. proposed the KATZCPDA model, which integrates different biological data to construct a heterogeneous network, aggregates the various similarity matrices calculated, and then uses KATZ to calculate the affinity score of circRNA-Disease association pairs. Ge et al. proposed the LLCC model, which uses local constrained linear coding (LLC) on the circRNA disease matrix, circRNA similarity matrix and disease similarity matrix to obtain the reconstructed circRNA and disease similarity matrix. Then, the LP method is used for the similarity matrix of circRNA and disease and the reconstructed similarity matrix, respectively, to obtain four affinity score matrices. Then, the four affinity matrices are integrated to form the final prediction matrix. Li et al. proposed the Bi-SGTAR model, which splits the association matrix into a circRNA view and a disease view, uses a sparse quality control (SQC) module to suppress unreliable associations, and then uses a true association regression (TAR) module to establish the probability of true associations, quantify the authenticity of the associations, and finally integrate the reliable and true associations of the two views. For example, using deep learning methods, Peng et al. proposed the GATCL2CD model, which first calculated four similarity networks to construct a heterogeneous graph based on circRNA, disease, and circRNA-disease, and then used a multi-head attention mechanism and convolutional networks (CNNs) of different sizes to extract high-level feature representations of high-order nodes from the heterogeneous graph, and finally connected to a multi-layer perceptron neural network (MLP) for prediction and classification. Liu et al. proposed the RDGAN model, which first constructed four similarity matrices, used resistance distance to extract features from the similarity matrices, and then cited three modules, GAT, MFConv, and ResGatedGraphConv, to learn the feature representation of association pairs, and finally predicted associations through linear regression.

[0006] Although existing methods have achieved some results in the field of CDA, they still have two limitations: they always ignore the important information of the original input sample features and the extraction of high-order features, or lack the potential information association between different sample features. Even if there is an interactive feature extraction process, it may only be limited to homogeneous networks, ignoring the information aggregation process between heterogeneous networks. Therefore, there is still a lot of room for improvement. Summary of the invention

[0007] The present invention aims to overcome the shortcomings of the above-mentioned prior art and proposes a method for predicting diseases based on circular RNA based on graph neural network to solve the problems in the prior art. The original data is collected, the similarity matrix is ​​calculated, a dual-channel network architecture is constructed, and the best model is selected according to the evaluation index.

[0008] The detailed technical scheme of the present invention is as follows:

[0009] A method for predicting diseases by circular RNA based on graph neural network, comprising the following steps:

[0010] Step 1: Obtain the original data from the open source circular RNA (circRNA) and disease association database, filter the original data set, retain only human disease information, and then construct the initial sparse matrix.

[0011] Step 2: Construct multiple similarity matrices based on the initial sparse matrix, namely disease semantic similarity matrix, functional similarity matrix, Gaussian similarity matrix, Jaccard similarity matrix, and cosine similarity matrix, and stack and combine the above similarity matrices as the initial features of the model input.

[0012] Step 3: Build a dual-channel model for training. Channel 1 is composed of a dual attention mechanism to extract high-order features. Channel 2 first divides the samples into piles through degree segmentation, and then uses the wave module to aggregate the potential correlation information between different samples. Then it enters the channel fusion module to integrate the information from the two channels for the final output.

[0013] Step 4: To make the experiment more complete, five evaluation indicators are used to quantify the results, namely accuracy, F1 score (F1), Matthews correlation coefficient (Mcc), area under the precision-recall curve (AUPR), and area under the receiver operating characteristic curve (AUC). The best model is found using a five-fold cross-validation method, and the results of the five rounds are averaged.

[0014] Preferably, in step one, the association between circular RNA and disease in the open source database is taken out and repeated associations are removed, the information of animals such as mice is deleted, and only the human part is retained, and then an initial matrix is ​​constructed, the horizontal axis of the matrix is ​​the name of the disease, and the vertical axis is the name of the circular RNA. If the two are associated, the matrix is ​​1, otherwise it is 0.

[0015] Preferably, in step 2, multiple similarity matrices are calculated by using information such as the initial matrix, and a medical theme (Mesh) description is downloaded from a medical library. The semantic information of different diseases is included to calculate the disease semantic similarity matrix, and then the Gaussian similarity matrix of circular RNA and disease is obtained by using Gaussian kernel similarity calculation, and the circular RNA functional similarity matrix is ​​obtained based on the disease semantic similarity matrix and the initial sparse matrix calculation, and the residual similarity matrix is ​​obtained according to Jaccard similarity and cosine similarity.

[0016] Preferably, in step two, the similarity matrix stacking refers to performing a column splicing operation on the similarity matrices of circular RNA and diseases, merging 8 similarity matrices into 4 matrices through column splicing, and finally using the column-spliced ​​RNA functional similarity matrix and the disease semantic similarity matrix as the input features of channel two, and the remaining spliced ​​matrices as the input features of channel one.

[0017] Preferably, in step three, the dual attention mechanism in channel one refers to channel attention and spatial attention. Channel attention can better read the stacked input matrix and autonomously allocate the importance of each channel. Spatial attention can remove redundant information in features, strengthen the expression of main features, and weaken edge features.

[0018] Preferably, in step three, the degree segmentation means that the degree of a node in a graph represents the activity of this point, and different samples are divided into blocks with different activities, that is, samples with similar information are grouped together.

[0019] Preferably, in step three, the wave module proposes a new patch fusion method based on the traditional MLP architecture, changes the patch fusion process into dynamic fusion, combines the semantic information of various patches, and obtains related information.

[0020] Preferably, in step three, the fusion module refers to stacking the matrices output by the two channels and then feeding them into the dilated convolution to expand the receptive field and reduce the risk of overfitting.

[0021] Preferably, in step 4, the five-fold crossover means randomly dividing the two data sets into five equal parts, four of which will be used as training sets for model training, and the remaining part will be used for testing, and the five parts will be used for training and testing in turn. This can prevent risks such as overfitting.

[0022] Preferably, in step 4, the five evaluation indicators can more accurately display the performance of each model and select the best model.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1) The present invention constructs an adjacency matrix for the original data of circular RNA and disease association, and simultaneously calculates multiple similarity matrices as features, alleviating the problem of insufficient information in the original data of RNA and disease association.

[0025] 2) The dual-channel network model constructed by the present invention can not only extract high-order features and retain important information from the original similarity network through the dual attention mechanism in channel one, but also fuse the potential information associations between different sample features that are not composed of similarity networks through the wave module in channel two, which has advantages in both reasoning speed and accuracy.

[0026] 3) Based on the reasoning of the model, this study solved the time-consuming and labor-intensive problems of biotechnologies such as CLIP-seq, high-throughput sequencing and immunoprecipitation, and made large-scale screening of circular RNA and disease associations faster and more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The overall flow chart provided by the present invention;

[0028] Figure 2 A schematic diagram of dual attention provided by the present invention;

[0029] Figure 3 A schematic diagram of a channel attention module provided by the present invention;

[0030] Figure 4 A schematic diagram of a spatial attention module provided by the present invention;

[0031] Figure 5 Wave module architecture diagram provided by the present invention;

[0032] Figure 6 Patm module flow chart provided by the present invention;

[0033] Figure 7 A schematic diagram of the Patm module provided by the present invention;

[0034] Figure 8 A structural diagram of the dilated convolution module provided by the present invention; DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be described clearly and comprehensively below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The present invention proposes a method for predicting diseases by circular RNA based on graph neural network. The overall flow chart of the method is as follows: Figure 1 As shown, the following steps are included:

[0037] Step 1: Obtain the original data from the open source circular RNA and disease association database, filter the original data set, retain only human disease information, and then construct the initial sparse matrix.

[0038] Step 2: Based on the initial sparse matrix, multiple similarity matrices are constructed, namely disease semantic similarity matrix, functional similarity matrix, Gaussian similarity matrix, Jaccard similarity matrix, and cosine similarity matrix, and the above similarity matrices are stacked and combined as the initial features of the model input.

[0039] Step 3: Build a dual-channel model for training. Channel 1 is composed of a dual attention mechanism to extract high-order features. Channel 2 first divides the samples into piles through degree segmentation, and then uses the wave module to aggregate the potential correlation information between different samples. Then it enters the channel fusion module to integrate the information from the two channels for the final output.

[0040] Step 4: To make the experiment more complete, five evaluation indicators are used to quantify the results, namely accuracy, F1 score (F1), Matthews correlation coefficient (Mcc), area under the precision-recall curve (AUPR), and area under the receiver operating characteristic curve (AUC). The best model is found using a five-fold cross-validation method, and the results of the five rounds are averaged.

[0041] Specifically, as described in step 1, Figure 1 In the A area shown, real data were obtained from two open source databases, circ2Disease and circR2Disease, and medical subject (Mesh) descriptions were downloaded from the medical library. The former was used to construct the initial adjacency matrix to prepare for subsequent matrix similarity calculations, and the latter contained semantic information of different diseases to prepare for calculating disease semantic similarity and RNA functional similarity.

[0042] Specifically, as described in step 2, Figure 1 In the A region shown, multiple similarity matrices are calculated. The disease semantic similarity matrix can capture the semantic relationship between diseases and better reflect the potential connection between complex diseases. The functional similarity matrix and the Gaussian similarity matrix can smoothly process distance information and avoid the influence of possible mutations or outliers when using distance directly. The Jaccard similarity matrix is ​​particularly suitable for processing sparse discrete data. The cosine similarity matrix is ​​insensitive to the size of the vector and pays more attention to the similarity of the direction to obtain multi-directional data information.

[0043] More specifically, as described in step 2, multiple similar matrices are aggregated and concatenated as follows:

[0044] CCS(c(i))={c i x1,c i x2,c i x3,…,c i x nc}

[0045] DCS(d(i))={d i x1,d i x2,d i x3,…,d i x nd}

[0046] (CCS(c(i)),DCS(d(i)))={c i x1,c i x2,c i x3,…,c i x nc |d i x1,d i x2,d i x3,…,d i x nd}

[0047] Among them, GCS(c(i) represents the row in the circRNA Gaussian similarity matrix, GDS(d(i)) represents the i-th row in the disease Gaussian similarity matrix, nc represents the number of circRNAs, nd represents the number of diseases, and x represents the value of each column in the matrix. Similarly, the four matrices after splicing are as follows:

[0048] CDGS(c(i),d(i))=(GCS(c(i)),GDS(d(i)))

[0049] CDCS(c(i),d(i))=(CCS(c(i)),DCS(d(i)))

[0050] CDJS(c(i),d(i))=(JCS(c(i)),JDS(d(i)))

[0051] FDSS(c(i),d(i))=(CFS(c(i)),DSS(d(i)))

[0052] Among them, CDGS is the Gaussian similarity merge matrix of circRNA and disease, CDCS is the cosine similarity merge matrix of circRNA and disease, CDJS is the jaccard similarity merge matrix, and FDSS is the merge matrix of circRNA functional similarity and disease semantic similarity. The above four matrices are stacked as input features of auxiliary path R1. In order to retain the initial feature information to the greatest extent, we choose to superimpose the four generated similarity merge matrices in the form of multiple channels, so that each channel is a similarity matrix, which can maximize the retention of the original features and allow subsequent modules to obtain more information. The formula is as follows:

[0053] RX1=concat(CDGS,CDCS,CDJS,FDSS,dim=1)

[0054] Since the circRNA functional similarity matrix CFS and the disease semantic similarity matrix DSS contain sufficient original information and are relatively independent, this paper uses their merged matrix as the input feature RX2=FDSS of the main path.

[0055] Specifically, as described in step three, Figure 1 In the C region and Figure 2 As shown, the original channel is mainly composed of two large modules, namely Figure 3 The channel attention and Figure 4 Spatial attention in. Since the selected matrix merging method is the channel stacking method, in order to better read the information of each channel, we use the channel attention module to obtain the correlation information between the same channels, obtain the importance of each feature channel through autonomous learning, and then assign different weights to each channel, so as to strengthen the important features in the channel and pay more attention to them, and weaken the unimportant features. Because the original feature dimension of the matrix is ​​high, the feature dimension of each association pair is high, which will contain many redundant features. In order to improve this problem, spatial attention is then used to perform spatial transformation on each row of features and retain key information, generate weighted output of weight mask for each feature, mark important feature columns to strengthen expression, and weaken unimportant feature columns.

[0056] Specifically, as described in step three, Figure 1 In the D area and Figure 5As shown in the figure, channel 2 is also composed of two modules, namely the degree segmentation module and the Wave module. The degree of the node in the figure can represent the activity of the node. Nodes in an activity interval may be related. Therefore, the sample is divided into three parts according to the degree value of the node, and it is put into the wave with pseudo labels for multiple cycles. Finally, new sample interaction features are aggregated and weighted with the original features. The Wave module proposes a new patch fusion method based on the traditional MLP architecture, changing the expression of patches for dynamic fusion.

[0057] More specifically, as described in step 3, wave represents the feature of each sample as a wave function with amplitude and phase, which can also be regarded as a wave. These waves are dynamically aggregated to obtain enhanced features, and finally the enhanced features are combined with the original matrix to obtain a new feature representation. The amplitude part of the wave is the real-valued feature of the sample, which is used to represent the maximum intensity of the wave. The phase part is the unit complex value between the sample feature and the fixed weight, which is used to adjust the position of the wave in the cycle. In the aggregation process, the phase difference between samples will greatly affect their aggregated output results. Sample features with similar phases tend to enhance each other, so we use the above-mentioned degree segmentation method to group samples with similar information together and put them into the phase-aware feature mixing module (PATM) for dynamic aggregation and set pseudo labels with them for multiple training, so that better aggregate features can be obtained.

[0058] More specifically, as described in step 3, in the phase-aware feature mixing module (PATM), each sample feature is converted into a wave with amplitude and phase. Figure 6 The overall process of the PATM module is shown in The formula for the exponential form of a complex number is as follows:

[0059]

[0060] Among them, z j The real-valued feature representing the sample represents the amplitude of the wave, θ j is a periodic function that represents the phase of the wave, which is its position within the wave period, and ⊙ represents multiplication.

[0061] When different samples are aggregated, the phase θ j will adjust their superposition mode, assuming yes and The aggregate output, its amplitude |z r | and phase θ r The calculation formula is as follows:

[0062]

[0063] θr =θ1+atan2(|z2|⊙sin(θ2-θ1),|z1|+|z2|⊙cos(θ2-θ1))

[0064] Among them, atan2(x,y) is a two-parameter inverse tangent function. According to the above equation, the phase difference between the two samples |θ2-θ1| has an effect on the aggregate z r The impact is huge, such as Figure 7 As shown, the left side is the phasor diagram after the two waves are aggregated, and the right side shows the projection of the two waves and the aggregated wave on the real axis. It can be seen that when the two waves, that is, the tokens, have the same phase (θ2=θ1+2π*m, m∈[0,±2,±4,…]), they will enhance each other, and when their phases are opposite (θ2=θ1+π*m, m∈[±1,±3,…]), the composite wave will weaken. In other cases, their interaction is more complicated, but whether it is enhanced or weakened is still determined by the phase difference.

[0065] Specifically, as described in step 3, the overall process of wave is to first split the sample matrix and obtain the intermediate features. The sample matrix is ​​represented as X = [x1, x2, ..., x n ], where x j is a D-dimensional sample feature vector. The sample matrix is ​​cut through the convolution operation, and each sample is regarded as a patch to extract features and enhance the conversion ability of each sample feature. The obtained sample intermediate feature matrix is The formula is as follows:

[0066]

[0067] Where x is the feature of each sample, Conv(·) is a convolution block with a kernel size of (1, 10). is the intermediate feature of each sample, and n is the number of samples.

[0068] In order to satisfy the formula The wave function in needs amplitude and phase information. In order to simplify, Channel-FC is used to obtain the amplitude and phase of the token. The formula is as follows:

[0069]

[0070] in, The intermediate features of the sample are used as input, W c and W t All of them are learnable weights. According to the above equations, it can be seen that the amplitude and phase can be dynamically adjusted according to the different semantic contents of the input sample features.

[0071] In order to embed the wave function into the MLP-like architecture, Equation 1 is expanded into triangular form using Euler's formula and expressed in real and imaginary parts as follows:

[0072]

[0073] In the above formula, the complex-valued features are represented as two real-valued vectors, representing the real part and the imaginary part respectively. Then all waves are aggregated through the FC operation. The formula is as follows:

[0074]

[0075] in, Represents the set of all waves, since the output after fusion It is a complex value representation. According to the quantum measurement method, the quantum state of the complex value representation is then projected to the observable real value. The real and imaginary parts of are added to the weights to obtain the real-valued output o j , combined with equation and The following formula is obtained:

[0076]

[0077] Among them, W t and W i are all learnable weights, θ k It will still be dynamically adjusted based on the semantic content of the input data.

[0078] Specifically, as described in step 3, the step of aggregating the outputs of the two channels is to combine the output result T2 of channel one and T1 of channel two. To simplify the operation, we perform a concat operation on T1 and T2, that is, to merge the channels, so that the weight information can be retained to the maximum extent, such as Figure 8 As shown in the figure, a hole convolution layer is connected to expand the receptive field and reduce the risk of overfitting. The formula is as follows:

[0079] TF=Dlc(T1,T2)

[0080] Among them, T1 and T2 are the pre-output results of the two paths respectively, and Dlc(·) is a dilated convolution block with a dilation rate of 2.

[0081] Specifically, as described in step 4, in the five-fold cross validation, the two data sets are randomly divided into five equal parts, four of which will be used as training sets for model training, and the remaining part will be used for testing. The five parts will be used for training and testing in turn. To prevent overfitting, we average the results of the five rounds. Finally, we used five evaluation indicators to quantify the results, namely accuracy (Accuracy), F1 score (F1), Matthews correlation coefficient (Mcc), area under the precision-recall curve (AUPR), and area under the receiver operating characteristic curve (AUC). The formulas are as follows:

[0082]

[0083] Among them, TP, FP, TN and FN represent the number of true positives, false positives, true negatives and false negatives respectively. The average accuracy of the model after five-fold cross validation is as follows: the accuracy is 89.02%, the F1 score is 89.02%, the area under the receiver operating characteristic curve is 95.62%, the Matthews correlation coefficient is 78.05%, and the area under the precision-recall curve is 93.66%. In dataset 2, the indicators of the model are as follows: the accuracy is 87.48%, the F1 score is 87.43%, the area under the receiver operating characteristic curve is 94.48%, the Matthews correlation coefficient is 74.99%, and the area under the precision-recall curve is 93.27%.

[0084] As can be seen from the above, the present invention proposes a method for predicting diseases based on circular RNA of graph neural network. First, the original data in the association database of open source circular RNA and diseases is obtained, and the original data set is screened to retain only human disease information and construct an initial sparse matrix. Then, multiple similarity matrices are constructed based on the initial sparse matrix, namely, disease semantic similarity matrix, functional similarity matrix, Gaussian similarity matrix, Jaccard similarity matrix, cosine similarity matrix, and the above similarity matrix is ​​stacked and combined as the initial features of the model input. Then, a dual-channel model is built for training. Channel one is composed of a dual attention mechanism for high-order feature extraction. Channel two is first divided into piles by degree segmentation, and then the wave module is used to aggregate the potential association information between different samples, and then the output of the two channels is fused and output. Finally, in order to make the experiment more complete, 5 evaluation indicators are used to quantify the obtained results, namely, accuracy (Accuracy), F1 score (F1), Matthews correlation coefficient (Mcc), area under the precision recall curve (AUPR), and area under the receiver operating characteristic curve (AUC). And the best model is found by the five-fold cross validation method, and the results of the five rounds are averaged. This method uses graph neural networks and deep learning technology to improve the accuracy and efficiency of circular RNA in disease prediction, providing strong support for the development of bioinformatics technology.

[0085] The examples described above have been used to illustrate the purpose and technical solutions of the present application in detail. It should be understood that the above description is only an embodiment of the present application and does not limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made on the technical solution of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting diseases based on circular RNA using graph neural networks, characterized in that: The following steps are involved: Step 1: Obtain the original data from the open source circular RNA (circRNA) and disease association database, filter the original data set, retain only human disease information, and then construct the initial sparse matrix. Step 2: Based on the initial sparse matrix, multiple similarity matrices are constructed, namely disease semantic similarity matrix, functional similarity matrix, Gaussian similarity matrix, Jaccard similarity matrix, and cosine similarity matrix, and the above similarity matrices are stacked and combined as the initial features of the model input. Step 3: Build a dual-channel model for training. Channel 1 is composed of a dual attention mechanism to extract high-order features. Channel 2 first divides the samples into piles through degree segmentation, and then uses the wave module to aggregate the potential correlation information between different samples. Then it enters the channel fusion module to integrate the information from the two channels for the final output. Step 4: To make the experiment more complete, five evaluation indicators are used to quantify the results, namely accuracy, F1 score (F1), Matthews correlation coefficient (Mcc), area under the precision-recall curve (AUPR), and area under the receiver operating characteristic curve (AUC). The best model is found using a five-fold cross-validation method, and the results of the five rounds are averaged.

2. The method according to claim 1, characterized in that: In step one, the associations between circular RNA and diseases in the open source database are extracted and repeated associations are removed, and the information of animals such as mice is deleted, only the human part is retained, and then the initial matrix is ​​constructed. The horizontal axis of the matrix is ​​the name of the disease, and the vertical axis is the name of the circular RNA. If the two are associated, the matrix is ​​1, otherwise it is 0.

3. The method according to claim 1, characterized in that: In step 2, multiple similarity matrices are calculated using information such as the initial matrix, and medical subject (Mesh) descriptions are downloaded from the medical library. It contains the semantic information of different diseases to calculate the disease semantic similarity matrix, and then uses Gaussian kernel similarity calculation to obtain the Gaussian similarity matrix of circular RNA and disease. The circular RNA functional similarity matrix is ​​calculated based on the disease semantic similarity matrix and the initial sparse matrix, and the residual similarity matrix is ​​obtained based on Jaccard similarity and cosine similarity.

4. The method according to claim 1, characterized in that: In step two, the similarity matrix stacking refers to performing a column splicing operation on the similarity matrices of circular RNA and diseases, merging 8 similarity matrices into 4 matrices through column splicing, and finally using the column-spliced ​​RNA functional similarity matrix and disease semantic similarity matrix as the input features of channel two, and the remaining spliced ​​matrices as the input features of channel one.

5. The method according to claim 1, characterized in that: In step three, the dual attention mechanism in channel one refers to channel attention and spatial attention. Channel attention can better read the stacked input matrix and autonomously allocate the importance of each channel. Spatial attention can remove redundant information in features, strengthen the expression of main features, and weaken edge features.

6. The method according to claim 1, characterized in that: In step three, the degree segmentation means that the degree of a node in a graph represents the activity of this point, and different samples are divided into blocks with different activities, that is, samples with similar information are grouped together.

7. The method according to claim 1, characterized in that: In step three, the wave module proposes a new patch fusion method based on the traditional MLP architecture, which changes the patch fusion process into dynamic fusion, combines the semantic information of various patches, and obtains related information.

8. The method according to claim 1, characterized in that: In step three, the fusion module stacks the matrices output by the two channels and then feeds them into the dilated convolution to expand the receptive field and reduce the risk of overfitting.

9. The method according to claim 1, characterized in that: In step 4, the five-fold crossover refers to randomly dividing the two data sets into five equal parts, four of which will be used as training sets for model training, and the remaining part will be used for testing. The five parts will be used for training and testing in turn. It can prevent risks such as overfitting.

10. The method according to claim 1, characterized in that: In step 4, the five evaluation indicators can more accurately display the performance of each model and select the best model.

Images