MiRNA and disease prediction method based on cross-modal and graph convolution

Abstract

The present application relates to the field of image technology, especially to a miRNA and disease prediction method based on cross-modal and graph convolution, comprising: obtaining a filled correlation matrix by using a cross-modal data filling algorithm; constructing miRNA integrated similarity and disease integrated similarity; constructing a heterogeneous network graph G and a feature matrix H(0) as inputs of a graph convolution neural network by using the miRNA integrated similarity, the filled correlation matrix and the disease integrated similarity; comparing the cosine similarity between three node embedding gates with a similarity threshold ST to determine whether to output an integrated convolution layer or a single convolution layer; obtaining a miRNA-disease score matrix by using a bilinear decoder; and training a CIGGNET model by using a weighted cross-entropy loss function. The present application solves the problems that the existing biological information prediction model cannot fill the correlation matrix and does not utilize unknown information of the correlation data when using a matrix decomposition method to process the correlation matrix.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to miRNA and disease prediction methods based on cross-modal and graph convolution. Background Technology

[0002] Today, a growing body of research indicates that miRNAs play a vital role in various biological processes. Understanding the molecular mechanisms of disease is a key objective of biomedical research. Extensive evidence suggests that miRNAs are crucial in the development and progression of human diseases; numerous studies have demonstrated their association with cancer and other disease types, and miRNAs are emerging as novel potential biomarkers and diagnostic / therapeutic tools.

[0003] In recent years, significant progress has been made in predicting potential miRNA-disease associations. With the development of machine learning and continuous improvements in systems biology, numerous computational methods have been developed to predict novel miRNA-disease associations. Existing computational methods can be categorized into two types: models based on complex network algorithms and predictive models based on machine learning.

[0004] Current research utilizes graph neural networks (Graph Neural Networks) to learn feature representations of miRNA-disease association pairs, such as MMGCN, LAGCN, GCN-MF, IDDkin, and FCGCNMDA, achieving high performance in predicting miRNA-disease association pairs. However, most models do not process miRNA-disease association information, and effectively utilizing this information remains a challenge. For example, VGAMF and NIMCGCN incorporate matrix factorization into graph convolutional networks. Based on the sparsity of the miRNA-disease association matrix, they obtain the required low-rank matrix and remove unnecessary sparse terms through matrix factorization algorithms, thus learning the necessary feature values ​​through graph convolution.

[0005] Existing bioinformatics prediction models use matrix factorization to process the association matrix. While this method effectively removes low-rank information (which can be understood as noisy data) from the association matrix, it cannot fill in the gaps in the association matrix. Furthermore, existing methods do not fully utilize the unknown information in the association data. Summary of the Invention

[0006] To address the shortcomings of existing methods, this invention utilizes known data to evaluate and predict unknown data, effectively leveraging potentially valuable information within the unknown data.

[0007] The technical solution adopted in this invention is: a miRNA and disease prediction method based on cross-modal and graph convolution, comprising the following steps:

[0008] Step 1: Construct miRNA-disease association matrix A, and use a cross-modal data imputation algorithm to obtain the imputed association matrix A. * ;

[0009] Furthermore, step one specifically includes:

[0010] Step 11: Convert the correlation matrix A to grayscale and then into a first-order tensor. Finally, fold the first-order tensor into a third-order tensor χ.

[0011] Step 12: Define the objective function for cross-modal data imputation and set the initial estimation tensor. satisfy:

[0012]

[0013] Where, χ ijk The set of indices of the corresponding elements is (i,j,k)∈Ω;

[0014] Step 13: Set additional variables q = 1, 2, 3;

[0015] Step 14: Set cross-modal data imputation constraints and update the intermediate tensor β according to the maximum number of iterations. q Estimating tensors and additional variable y q ;

[0016]

[0017]

[0018]

[0019] Where ρ is the iteration parameter, fold q (.) represents the transformation of a matrix into a tensor function, S is a third-order tensor consisting of elements 0 and 1, and χ is a third-order sparse tensor of the incidence matrix A. Let y be the estimator tensor of χ. q(q) For the additional tensor y q Expansion under mode q Let β be a singular value decomposition function. q This is the q-th intermediate tensor;

[0020] Step 15: Obtain the filled third-order estimation tensor Will Expanding yields a filled first-order tensor image, which is then converted into a filled correlation matrix A. * .

[0021] Furthermore, the formula for the cross-modal data imputation objective function is as follows:

[0022]

[0023] Where β1, β2, β3 are intermediate tensors, β 1(1) This is the expansion of the intermediate tensor β1 in mode 1; β 2(2) For the mode 2 expansion of β2, β 3(3) α1, α2, and α3 are the modal 3 expansions of β3; α1, α2, and α3 are the weight coefficients.

[0024] Furthermore, the formula for cross-modal data completion constraints is as follows:

[0025]

[0026] Where S is a third-order tensor consisting of elements 0 and 1, and χ is a third-order sparse tensor of the incidence matrix A. Let β be the estimated tensor of χ. q Let q be the q-th intermediate tensor.

[0027] Step 2: Construct miRNA integration similarity (SM) and disease integration similarity (SD) using miRNA similarity and disease similarity, respectively;

[0028] Furthermore, if miRNA i and miRNA j have functional similarity, sequence similarity, or semantic similarity, the miRNA integration similarity SM is the average of the miRNA sequence similarity matrix SM1, the miRNA functional similarity matrix SM2, the semantic similarity matrix SM3, and the miRNA GIP similarity SM4; otherwise, the miRNA integration similarity SM is simply the miRNA GIP similarity SM4.

[0029] Step 3: Utilize SM and A * SD constructs a heterogeneous network graph G and feature matrix H(0), and uses the heterogeneous network graph G and feature matrix H(0) as input to the graph convolutional neural network;

[0030] Furthermore, step three specifically includes:

[0031] Step 31: Construct a heterogeneous network G H G H =M m (m i ,m j )×A * ×M d (d i ,d j M m (m i ,m j ) is the miRNA normalization matrix, M d (d i ,dj () represents the disease standardization matrix;

[0032] Step 32: Use the double random walk algorithm to perform random walks on the miRNA ensemble similarity network and the disease ensemble similarity network respectively to obtain initial association scores; calculate the random walks on the disease feature matrix network and the miRNA feature matrix network based on the initial association scores and the normalization matrix; the formula for the feature matrix network is:

[0033] L_R t = (1-α)*M d *R t-1 +α*Y d ;

[0034] R_R t = (1-α)*M m *R t-1 +α*Y m (5)

[0035] Where α is the decay factor of the random walk, M m M d These are miRNA, disease normalization matrix, and R. t-1 Y represents the probability of disease-miRNA association. m Y d These represent miRNA and initial disease association score, respectively.

[0036] Step 4: Embed the H of GCN layers 1-3 nodes 1 -H 3 The input is fed into the control gate; based on the cosine similarity between the three nodes embedded in the gate and the similarity threshold ST, it is determined whether the gate output is an integrated convolutional layer or a single convolutional layer; the miRNA-disease score matrix is ​​obtained using a bilinear decoder, resulting in the CIGGNET model;

[0037] Furthermore, the condition for determining whether a gate outputs an integrated convolutional layer or a single convolutional layer is as follows: if the cosine similarity CS between the three node embedding gates is greater than a given similarity threshold ST, the gate outputs a single convolutional layer; conversely, if the similarity CS between the three convolutional layers is less than a given similarity threshold ST, the gate outputs an integrated convolutional layer.

[0038] Furthermore, the formula for determining whether the gate output is an integrated convolutional layer or a single convolutional layer is:

[0039]

[0040] in, H M It is the miRNA embedding matrix, H DThe embedding matrix for the disease, where σ1, σ2, σ3 represent the weights of the three-layer GCN node embedding, and H1, H2, H3 represent the three-layer GCN node embedding.

[0041] Furthermore, the formula for the miRNA-disease score matrix is:

[0042]

[0043] Where W′∈R k×k For the training matrix, H M The final embedding matrix of miRNAs, H D The final embedding matrix for the disease.

[0044] Step 5: Train the CIGGNET model using the weighted cross-entropy loss function.

[0045] Furthermore, the formula for the weighted cross-entropy loss function is:

[0046]

[0047] Where M represents known miRNA association pairs, N represents known disease association pairs, and (i,j) represents disease d. i and miRNA m j Yes, λ is the weighting factor, y + and y - α′ represents the number of positive and negative samples, respectively. ij The values ​​are in the miRNA-disease score matrix A'.

[0048] The beneficial effects of this invention are:

[0049] 1. This invention proposes a cross-modal imputation algorithm to complete missing miRNA-disease association information. In order to obtain the relationship between missing information and existing information more intuitively and effectively, the association matrix is ​​transformed to the image domain, and tensor repair is performed on the constructed sparse image to finally obtain the cross-modal repair association matrix.

[0050] 2. A gated graph convolutional neural network algorithm is proposed. In the gating operation, each convolutional layer is regarded as the convergence of circuit information into the control gate. In order to make reasonable and full use of the characteristics of different convolutional layers, the control gate adaptively determines the fusion operation between layers based on the similarity of the convolutional layers.

[0051] 3. By integrating cross-modal imputation algorithms and gated GCNs, and combining the miRNA-disease feature matrix constructed by BRWMDA with a heterogeneous network based on miRNA and disease similarity networks, the task of predicting miRNA-disease associations is achieved. Attached Figure Description

[0052] Figure 1 This is a logic block diagram of the miRNA and disease prediction method based on cross-modal and graph convolution of the present invention;

[0053] Figure 2 The graph shows a performance comparison between CIGGNET and five comparative models using 5-fold cross-validation on the HMDD v3.2 database.

[0054] Figure 3 The graph shows a performance comparison of CIGGNET with five comparative models using 10-fold cross-validation on the HMDD v3.2 database. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0056] CIGGNET (Cross-modal Imputation and Gated GCN for Predicting miRNA-Disease Associations).

[0057] like Figure 1 As shown, the miRNA and disease prediction method based on cross-modal and graph convolution includes the following steps:

[0058] Step 1: Obtain the imputed correlation matrix A using a cross-modal data imputation algorithm. * ;

[0059] Multiple datasets need to be fused for the CIGGNET model. First, miRNAs and diseases that are not present in the miRBase, mirTarBase, and HMDD v3.2 mesh are removed, leaving 788 miRNAs, 374 diseases, and 8968 known association pairs. Then, an association matrix A is constructed to represent the association information between miRNAs and diseases.

[0060] The association matrix contains a large amount of unknown data; the unknown associations contain a lot of noisy data, which can have an adverse effect on the prediction results. Cross-modal data imputation algorithms are used to impute the disease-miRNA association matrix to reduce the impact of noise on the prediction results.

[0061] Cross-domian data imputation algorithms include: to effectively observe the correlation matrix A 788*374The structural characteristics and more intuitive and effective acquisition of the relationship between missing and existing information will enable the association matrix A to... 788*374 Convert to grayscale; make the correlation matrix A 788*374 The transition from the data domain to the image domain allows us to apply image restoration techniques to repair the correlation matrix.

[0062] The grayscale correlation matrix A 788*374 As a first-order tensor (grayscale image), the correlation matrix A is used. 788*374 Construct a third-order tensor χ (color image) of size 788×374×3; the index set of the elements corresponding to the third-order tensor χ is (i,j,k)∈Ω. At the same time, construct a tensor S of size 788×374×3, which is a tensor composed of elements 0 and 1, as shown in formula (1):

[0063]

[0064] Where i = 788, j = 374, k = 3, and χ is a third-order sparse tensor composed of the correlation matrix A.

[0065] The objective function of the cross-domian completion algorithm can be written in the following form:

[0066]

[0067] in, This represents the estimate of the tensor χ, where the intermediate tensors β1, β2, and β3 are all of size η1×η2×η3, and the matrix β 1(1) This represents the expansion of the intermediate tensor β1 in mode-1 (model-1 unfolding); similarly, the matrix β 2(2) This represents the model-2 unfolding of β2; matrix β 3(3) This represents the model-3 unfolding of β3, where α1, α2, and α3 are the weight coefficients.

[0068] There are two constraints for cross-modal data imputation: the first is to ensure the estimation tensor The first is that the elements of the tensor χ are equal on the set Ω; the second is that the intermediate tensors β1, β2, β3 are all equal to the estimated tensor. Specifically as follows:

[0069]

[0070] In the objective function, the weight coefficients α1=α2=α3 need to satisfy: α1+α2+α3=1. Using the ADMM (alternating direction method of multipliers) framework, we can obtain tensors β1, β2, β3 and estimate the tensors. The iterative update formula.

[0071] The specific process of the cross-modal data imputation method is as follows:

[0072]

[0073] In cross-modal data imputation methods There is a special definition, with the correlation matrix A 788*374 For example, U and V are the singular value decomposition matrices of A; where... The symbol "fold" q "(.)" indicates that the matrix is ​​reduced to a tensor, which is the opposite of the unfolding process; σ i Let be the i-th largest singular value in the correlation matrix A.

[0074] Cross-modal data imputation can reduce noise or outliers in the association matrix by filling in information with known association pairs; after imputation, the disease-miRNA association matrix has 8968 known association pairs and 285744 unknown association pairs.

[0075] Step 2: Construct miRNA integration similarity (SM) and disease integration similarity (SD) using miRNA similarity and disease similarity, respectively;

[0076] Specifically, if miRNAm i With disease d j There is a known association between them. =1; otherwise, The first step is to calculate the semantic similarity between the two diseases, which is then represented as SD1. The Gaussian Interaction Spectrum Kernel (GIP) similarity matrix SD2 is calculated using the association matrix. Then, the semantic similarity and GIP similarity are integrated to obtain the disease ensemble similarity SD. If there is semantic similarity between disease i and disease j, the overall similarity is the average of the semantic similarity and GIP similarity; otherwise, the overall similarity is the GIP similarity. Finally, a miRNA sequence similarity matrix SM1 and a miRNA functional similarity matrix SM2 are constructed. Referring to the semantic similarity of the diseases, miRNAs also have corresponding semantic similarities; a miRNA semantic similarity matrix SM3 is constructed. Similar to the GIP kernel similarity, miRNA GIP similarity is calculated, and the formula for calculating the miRNA GIP similarity matrix is ​​SM4. Likewise, miRNA sequence similarity, miRNA functional similarity, miRNA semantic similarity, and miRNA GIP similarity are integrated into the miRNA ensemble similarity SM.

[0077] If miRNA i and miRNA j have functional similarity, sequence similarity, or semantic similarity, the integration similarity of miRNAs is the average of SM1, SM2, SM3, and SM4; otherwise, the integration similarity of miRNAs is the miRNA GIP similarity SM4.

[0078] Step 3: Utilize SM and A * SD constructs a heterogeneous network graph G and feature matrix H(0), and uses the heterogeneous network graph G and feature matrix H(0) as input to the graph convolutional neural network;

[0079] The correlation matrix A * Similarity information between diseases and miRNAs were integrated into heterogeneous networks; an association matrix A was used. * miRNA normalization matrix M m and disease standardization matrix M d Building heterogeneous networks G H .

[0080]

[0081]

[0082] Where m = 788, n = 374; M m It is the standardized form of SM, M d It is a standardized form of SD, heterogeneous network G H The formula is:

[0083] G H =M m (mi ,m j )×A * ×M d (d i ,d j (6)

[0084] A dual random walk algorithm is used to perform random walks on both the miRNA and disease ensemble similarity networks. This operation makes reasonable use of the similarity between miRNAs and diseases; M is obtained. d and M m Following the matrix calculation, the initial association probability of miRNA-disease with no known association is further calculated; and the initial association score Y is calculated using the completed adjacency matrix; for miRNAm i The calculation process is as follows:

[0085] Y m (i,:)=A * (:,j)*M m (m i ,m j (7)

[0086] Among them, miRNAm i With miRNAm j There is associated similarity, i,j∈(788×788), and Y is the initial association score of miRNA.

[0087] Similarly, for disease d i Its calculation process is as follows:

[0088] Y d (i,:)=A * (:,j)*M d (d i ,d j (8)

[0089] Among them, disease d i and disease d j There is a correlation similarity, i,j∈(383×383), and Y is the initial correlation score of the disease.

[0090] Define matrix R as having dimension N. d ×N m The value of matrix R(i,j) represents the association probability between disease i and miRNA j; the iterative random walk process of the disease feature network and miRNA feature network is defined as follows:

[0091] Disease Feature Matrix Network:

[0092] L_R t = (1-α)*Md *R t-1 +α*Y d (9)

[0093] miRNA feature matrix network:

[0094] R_R t = (1-α)*M m *R t-1 +α*Y m (10)

[0095] Where t represents the number of iterations, and α is the decay factor of the random walk, with a value ranging from 0 to 1; L_R t and R_R t The disease-miRNA feature association matrix, represented by random walks on the disease feature matrix network and the miRNA feature matrix network respectively, R t-1 This represents the probability of disease-miRNA association.

[0096] Step 4: H 1 H 2 H 3 Treating these as three circuit signals input to the control gate, the control gate will respond according to H. 1 H 2 H 3 Choose an appropriate convolutional layer based on the degree of similarity.

[0097] Gate-Graph Convolution

[0098] The idea behind GCN is to propose the Gate-GCN algorithm, which aggregates information about neighbors, captures dependencies between data, and generates useful node representations in graph convolution. Moreover, Gate-GCN can select appropriate convolutional layers based on gating operations.

[0099] The propagation method between layers of GCN is as follows:

[0100]

[0101] Among them, H l It is the embedding of the i-th layer node in GCN, D = diag(∑ i G ij W is the degree matrix of graph G. (l) It is a trainable weight matrix, and σ(.) represents a non-linear activation function.

[0102] To build a GCN model capable of learning disease-miRNA associations, a GCN is deployed on a constructed heterogeneous network graph to combine node similarity and direct link association information. The input graph is set as G = G H Secondly, constructing a GCN model requires initializing the feature matrix H. 0 The double random walk algorithm is used to construct the feature matrix H. 0 This can fully represent the attribute characteristics of miRNA-disease pairs, thus enriching their initial feature expression capabilities; the disease-miRNA feature matrix H is inferred from the double random walk algorithm. 0 Set the input feature matrix H 0 for:

[0103]

[0104] The first layer of the GCN encoder is formulated as follows:

[0105]

[0106] Among them, W (0) ∈R (788*374)×K It is the input to the hidden layer weight matrix, H (0) ∈R (788*374)×K It is the initial feature matrix constructed by a double-restart random walk, and K is the dimension of the embedding layer.

[0107] The subsequent layers of the GCN encoder follow formula (13). After L iterations, L K-dimensional embeddings can be obtained from different graph convolutional layers. All graph convolutional layers use exponential linear units as nonlinear activation functions, which not only speeds up the learning process but also significantly improves the generalization performance.

[0108] Using different convolutional layers in GCN produces different prediction results. However, too many convolutional layers in GCN can lead to oversmoothing. Generally, GCN uses 2-3 convolutional layers in its hidden layers. This invention's model uses three convolutional layers. Extensive experiments have shown that if the similarity between convolutional layers is very close, there is no need to integrate them; simply select the best-performing convolutional layer. Conversely, if there are significant differences between convolutional layers, it is necessary to combine their different characteristics. Borrowing ideas from LSTM and GRU, a decision gate operation is added to the hidden layers. CIGGNET integrates the three convolutional layers (i.e., the node embedding gate H)... 1 -H 3 The three circuit signals are transmitted to the decision gate, which outputs two signals: one for the integrated convolutional layer and the other for a single convolutional layer. If the cosine similarity CS between the three convolutional layers is greater than a given similarity threshold ST, the decision gate outputs the signal for a single convolutional layer (output H). 1 -H 3(Any one); Conversely, if the cosine similarity CS between the three convolutional layers is less than a given similarity threshold ST, the judgment gate outputs an integrated convolutional layer, with the specific formula as follows:

[0109]

[0110] in H M It is the miRNA embedding matrix, H D The embedding matrix of the disease, σ1, σ2, σ3 represent three different weights, and H1, H2, H3 represent the embedding of three layers of GCN nodes.

[0111] The essence of graph convolution is data. To facilitate calculation, different convolutional layers are transformed into matrix form. The cosine similarity between convolutional layers is shown in formula (15):

[0112]

[0113] Where i∈(1,2…), j∈(1,2,…), and M is the matrix form of the convolutional layer H. Representing vectors sum vector dot product, It represents the magnitude of the vector.

[0114] To reconstruct the adjacency matrix of miRNAs and disease associations, the bilinear decoder A′=f(H M H D The formula for ) is:

[0115]

[0116] Where W′∈R k×k For the training matrix, H M This is the final embedding matrix of the miRNA, H D The final embedding matrix of the disease, miRNA m i and disease d j The correlation prediction score between them is related to the value of A′(i,j) in the corresponding A′ matrix. A′(i,j) can be regarded as a i ′ j .

[0117] Step 5: Train the CIGGNET model using the weighted cross-entropy loss function;

[0118] In a dataset containing N diseases and M miRNAs, miRNA-disease association pairs are considered positive samples, and those without association are considered negative samples; positive and negative samples are denoted as Y. + and Y -Distinguishing between miRNAs and diseases is a binary classification problem because the number of associated miRNA-disease pairs is far fewer than the number of miRNA-disease pairs where no association was observed. The model is trained using a weighted cross-entropy loss function, calculated as follows:

[0119]

[0120] Where M represents known miRNA association pairs, N represents known disease association pairs, and (i,j) represents disease d. i and miRNA m j Yes, λ is the weighting factor. y + and y - λ represents the number of positive and negative samples, respectively, and the weighting factor λ emphasizes the importance of the observed associations and reduces the impact of data imbalance.

[0121] All trainable weight matrices are initialized using the Xaiver initialization method, and the loss function is optimized using the Adam optimizer. The Adam optimizer iteratively updates the neural network weights based on the training data. To prevent overfitting, CIGGNET introduces node loss and rule loss mechanisms into the graph convolutional layers. Node loss can be considered as training different models on various subnets. Furthermore, by integrating different models on the subnets, unknown miRNA-disease pairs can be predicted. In addition, a cyclic learning rate is used during optimization. The simple cyclic learning rate allows the learning rate to vary between the maximum and minimum learning rates, helping to balance training speed and accuracy.

[0122] Parameter settings

[0123] In this experiment, 5-fold and 10-fold cross-validation were used to evaluate the performance of CIGGNET in predicting potential disease-associated miRNAs. In each round, known miRNA-disease associations were treated as positive samples and randomly divided into five equal subsets. The cross-validation process was repeated five times, with one subset used as the test set and the remaining four subsets used as the training set. Similarly, the 10-fold cross-validation was performed in the same manner as the 5-fold cross-validation. In each round, a predictive model was built on the known associations in the training set and used to predict associations in the test set. AUPR and AUC were used as the primary metrics to measure the performance of the method without any specific thresholds.

[0124] CIGGNET has several hyperparameters, such as the similarity threshold ST, embedding dimension k, number of layers L, learning rate lr, total training time α, and two dropout rates: node dropout rate β and regular dropout rate γ. Different ranges of parameters were set: α∈{500,1000,2000,4000}, β,γ∈{2,4,6,8,10}, k∈{64,128,256}. Through experiments, these parameters were continuously adjusted, and they were set to k=256, L=3, lr=0.01, α=4000, β=0.6, γ=0.4, and ST=0.95.

[0125] Impact of cross-modal data in filling the correlation matrix

[0126] CIGGNET uses a cross-modal imputation algorithm to impute the correlation matrix A to obtain A. * More unknown miRNA-disease association pairs were obtained, and the accuracy comparison results of CIGGNET-filled and CIGGNET-unfilled pairs are shown in Table 1.

[0127] Table 1: Impact of cross-modal data imputation on CIGGNET

[0128]

[0129] Similarity between different convolutional layers

[0130] Different convolutional layers produce different results. Different models are built based on different convolutional layers: CIGGNET-L1 (using one convolutional layer), CIGGNET-L2 (using two convolutional layers), CIGGNET-L3 (using three convolutional layers), CIGGNET-ave (averaging the three convolutional layers), and CIGGNET-atn (using an attention mechanism to distribute the convolutional layers). Table 2 shows that CIGGNET-L3 and CIGGNET-L2 have relatively weak performance, possibly because the first convolutional layer embeds more information than the second and third convolutional layers, or because of excessive smoothing in GCN. Among them, CIGGNET-L1 has slightly higher accuracy than CIGGNET-ave and CIGGNET-atn.

[0131] Graph convolution is essentially data. We transformed the convolutions of these five models into matrix form to calculate their cosine similarity. Referring to Table 3, we found that CIGGNET-L1 and CIGGNET-L2 have a similarity of 98%, and CIGGNET-L1 and CIGGNET-L3 have a similarity of 97%, indicating very high similarity between the first, second, and third convolutional layers. Similarly, CIGGNET-L1 and CIGGNET-atn have a similarity of 99%, and CIGGNET-L1 and CIGGNET-ave have a similarity of 98%, also showing very high similarity. Furthermore, the data for CIGGNET-L1, CIGGNET-atn, and CIGGNET-ave in Table 2 are very close.

[0132] Table 2 shows that CIGGNET suffers from data loss and decreased prediction performance as the number of convolutional layers increases. Table 3 shows the similarity of each convolutional layer. Each convolutional layer in CIGGNET-L1, L2, and L3 has a similarity of at least 97%, meaning there's no need to consider integrating them. Considering the CIGGNET-L1 model as the prediction model, the specific convolutional layer to use is determined by the similarity between each layer. The condition for setting the first convolutional layer is that the similarity between CIGGNET-L1, L2, and L3 cannot be lower than 95%. If the similarity of these three convolutional layers is lower than 95%, integration should be considered. This different approach greatly enhances the model's generalization ability.

[0133] Where L1, L2, and L3 represent the first, second, and third convolutional layers, respectively.

[0134] Table 2 shows the impact of different convolutional layers on CIGGNET.

[0135]

[0136] Table 3 shows the different similarities between the convolutional layers of CIGGNET.

[0137]

[0138] Experimental comparison

[0139] To demonstrate the effectiveness of the CIGGNET model, the model of this invention is compared with the following baseline models. In this study, all comparison models and the CIGGNET model were predicted and trained using data from the HMDD v3.2 database.

[0140] The comparison models include:

[0141] The LAGCN method uses graph convolutional networks to build a prediction model, adopts the graph embeddings contributed by each convolutional layer, and uses an attention mechanism to reasonably allocate the graph embeddings of each graph convolutional layer.

[0142] The BRWMDA method constructs microbial and disease networks separately. In addition, the disease network is calculated using the similarity network fusion (SNF) method. Random walks are then performed on these two networks to predict new potential microbial and disease associations.

[0143] The NIMCGCN method constructs graphs of miRNA similarity networks and disease similarity networks, then feeds them into the GCN network for learning, and introduces the neural induction matrix completion method to predict the association between miRNAs and diseases.

[0144] The MDHGI method uses matrix factorization to separate noise to obtain the required low-rank matrix, and then constructs disease ensemble similarity matrix and miRNA ensemble similarity matrix separately, and integrates them into a single network.

[0145] The VAEMDA method constructs two matrices: the overall disease-miRNA association and disease similarity, as well as miRNA similarity. These matrices are used to train the variational autoencoder.

[0146] CIGGNET was compared with five other methods in 5-fold and 10-fold cross-validation experiments. All these comparative experiments were conducted on the HMDD v3.2 database. In the 5-fold cross-validation, the mean and standard deviation of the AUCs for the five experiments (LAGCN, BRWMDA, NIMCGCN, MDHGI, and VAEMDA) were 0.9375 + / - 0.0039, 0.9127 + / - 0.0043, 0.8588 + / - 0.0052, 0.8801 + / - 0.0021, and 0.8799 + / - 0.0049.

[0147] In the 10-fold cross-validation, the mean and standard deviation of the AUCs for the five experiments (LAGCN, BRWMDA, NIMCGCN, MDHGI, and VAEMDA) were 0.9371 + / - 0.0025, 0.9184 + / - 0.0029, 0.8508 + / - 0.0060, 0.8910 + / - 0.0024, and 0.8783 + / - 0.0077, respectively. The results show that in 100 5-fold cross-validations, the mean and standard deviation of the AUC for CIGGNET were 0.9423 + / - 0.000322. In 10-fold cross-validation, the mean and standard deviation of the AUC for CIGGNET were 0.9422 + / - 0.0038. Detailed comparison results are shown in Table 4. Figure 2 and Figure 3 .

[0148] Table 4 compares the performance of CIGGNET with other methods on the HDMM V3.2 database.

[0149]

[0150]

[0151] Case Studies

[0152] To further demonstrate the predictive performance of CIGGNET, three case studies were conducted. In the first case study, to verify the model's adaptability to the older HMDD v2.0 data, known disease-miRNA associations from HMDD v2.0 were used as the training dataset. CIGGNET predicted associations between breast cancer and 50 other miRNAs, and then ranked these candidate miRNAs according to their prediction scores. Finally, two important disease-miRNA association databases, dbDEMC and miR2Disease, were used to validate the top 50 predictions for breast cancer. In the second case study, the CIGGNET model was used to validate associations between esophageal cancer and 50 other miRNAs in the dbDEMC and miR2Disease disease miRNA association databases. In the third case study, the ability of the CIGGNET model to predict unknown disease-miRNA association pairs was tested. Known disease-miRNA association pairs in the HMDD v3.2 database were set as unknown association pairs. In this experiment, associations between liver cancer and lung cancer and 50 other miRNAs were predicted. dbDEMC is abbreviated as ①, miR2Disease as ②, and HMDD... v3.2 is abbreviated as ③.

[0153] Breast cancer is one of the most common malignant tumors in women. In recent years, a large number of studies have found that many microRNAs play a regulatory role in promoting or inhibiting the invasion and metastasis of breast cancer. miR-21 is located in the fragile region 17q23.3 of the chromosome. miR-21 is significantly upregulated in various malignant tumors such as breast cancer and liver cancer, and is positively correlated with the malignancy grade of breast cancer. Abnormal expression of miR-145 may be involved in the occurrence and development of breast cancer and is a potential molecular marker for breast cancer. Upregulation of miR-27a-3p can enhance the proliferation, invasion and migration of MCF-7 cells, and its mechanism may be related to promoting the expression of PCNA, MMP-2 and MMP-9 proteins. Furthermore, researchers have discovered an increasing number of miRNAs linked to breast cancer; CIGGNET was used to predict breast cancer-related miRNAs, and the latest database HMDDv3.2 was used; the dbDEMC and miR2Disease databases were used to validate the top 50 miRNAs that may be associated with breast cancer; Table 5 shows that all of the top 10 miRNAs associated with breast cancer were validated; 44 of the top 50 individual miRNAs associated with breast cancer were validated by at least one dbDEMC and miRDisease database.

[0154] Table 5 shows the validation of the top 50 miRNAs predicted by the model and their association with breast cancer.

[0155]

[0156] Esophageal cancer is one of the most common malignant tumors in my country. In recent years, with the development of high-throughput sequencing technology, more and more miRNAs have been found to be related to esophageal cancer. miR-451 is lowly expressed in esophageal squamous cell carcinoma tissues, while miR-21 is highly expressed in these tissues. miR-185 can inhibit the proliferation and migration of esophageal cancer through its target gene Six1's downstream genes, cell cycle A1 and E-cadherin. miRNA-375 is lowly expressed in esophageal cancer tissues, which is related to its promoter region methylation, and miRNA-375 is involved in the development and progression of esophageal cancer.

[0157] Studies have shown the impact of miRNAs on esophageal cancer. Therefore, detecting miRNA expression in esophageal cancer patients is crucial for analyzing their condition. CIGGNET was used to predict potential miRNAs associated with esophageal cancer. The dbDEMC and miR2Disease databases validated the top 50 miRNAs potentially associated with esophageal cancer. Table 6 shows that all of the top 9 miRNAs associated with esophageal cancer were validated. Of the top 50 human miRNAs associated with esophageal cancer, 40 were validated by at least one dbDEMC and miRDisease database.

[0158] Table 6 shows the validation of the top 50 miRNAs predicted by the model and their association with esophageal tumors.

[0159]

[0160] Liver cancer (HCC) refers to malignant tumors that occur in or originate in the liver. In 2021, 782,000 people were diagnosed with liver cancer, resulting in 746,000 deaths. In 2013, 300,000 people died from liver cancer caused by hepatitis B, 343,000 from hepatitis C, and 92,000 from alcohol-induced liver cancer. Regions with prevalent hepatitis B and C, including Asia and sub-Saharan Africa, also have higher rates of liver cancer. Men have a higher incidence of hepatocellular carcinoma than women. Resveratrol (Res) may inhibit the proliferation of HepG2 liver cancer cells and induce apoptosis by downregulating miR-151 expression; serum miR-122-5p and miR-486-5p may serve as serological markers for the auxiliary diagnosis of liver cancer; differential miRNA expression exists between liver cancer tissue and adjacent non-cancerous tissue, with the degree of miR-338-3p downregulation being associated with malignant liver behavior.

[0161] All studies have demonstrated the inseparable relationship between miRNAs and liver cancer; moreover, regulating different miRNAs can become an important mediator for the treatment of liver cancer; CIGGNET was used to predict potential miRNAs associated with liver cancer; Table 7 shows that all of the top 10 miRNAs associated with liver cancer were validated by the HMDD v3.2 database; 48 of the top 50 miRNAs associated with liver cancer were validated by the HMDD v3.2 database.

[0162] Table 7 shows the validation of the top 50 miRNAs predicted by the model and their association with liver cancer.

[0163]

[0164]

[0165] Lung cancer is a malignant tumor originating from the bronchial mucosa or glands of the lungs. It is one of the malignant tumors with the fastest increasing incidence and mortality rates, posing the greatest threat to human health and life. miR-145 can inhibit the proliferation of stem cells in the A549 lung adenocarcinoma cell line by downregulating OCT4 gene expression, and is a potential "protective" miRNA for lung cancer. The plasma miRNA23-a and miRNA-451 have potential biomarkers for the early diagnosis of non-small cell lung cancer (NSCLC). miRNA23-a and miRNA-451 are candidate biomarkers for the early diagnosis of NSCLC, and the combined detection of these two biomarkers is more effective for the diagnosis of NSCLC. The effect of microRNA (miRNA)-126 on the function of lung cancer A549 cells and its related mechanism of action: miRNA-126 can significantly reduce the protein levels of p-EGFR, p-AKT and p-mTOR in lung cancer A549 cells, which may inhibit their cell proliferation and migration and invasion capabilities.

[0166] Studies have shown that controlling miRNAs can influence lung cancer. CIGGNET was used to predict potential miRNAs associated with lung cancer; these potential miRNAs were then categorized based on the model's prediction scores. Table 8 shows that all of the top 10 miRNAs associated with lung cancer were validated using one of the following databases: HMDD v3.2, dbDEMC, or miR2Disease. Of the top 50 miRNAs associated with liver cancer, 48 were validated using at least one of the following databases: HMDD v3.2, dbDEMC, or miR2Disease.

[0167] Table 8 shows the validation of the top 50 miRNAs predicted by the model and their association with lung cancer.

[0168]

[0169]

[0170] The CIGGNET model proposed in this invention is used to predict the association between miRNAs and diseases. To more intuitively and effectively obtain the correlation between missing information (unknown miRNA-disease association) and existing information (known miRNA-disease association), CIGGNET performs cross-modal data estimation on the miRNA-disease association matrix. CIGGNET integrates the calculated miRNA-disease association matrix with miRNA similarity networks and disease similarity networks to construct a heterogeneous network, and uses a random walk algorithm to construct a miRNA-disease feature matrix. In addition, CIGGNET uses GCN to process the heterogeneous network and feature matrix to learn the miRNA-disease embedding.

[0171] Meanwhile, to improve the model's generalization performance, gating was employed in the selection of convolutional layers. In 100 five-fold cross-validations (5-fold CV), CIGGNET's mean AUC and standard deviation on HMDD v3.2 were 0.9423 + / - 0.00322, respectively. To further validate the predictive performance of the CIGGNET model, three case studies were conducted. The results showed that CIGGNET was highly effective in predicting the association between miRNAs and diseases in three important human diseases: liver cancer, breast cancer, and esophageal cancer. The third case study demonstrated that the CIGGNET model can predict the association between unknown miRNAs and diseases.

[0172] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for miRNA and disease prediction based on cross-modal and graph convolution, characterized in that, Includes the following steps: Step 1: Construct the miRNA-disease association matrix and use a cross-modal data imputation algorithm to obtain the imputed association matrix; Step one specifically includes: Step 11: Assemble the correlation matrix. The grayscale process converts the data into a first-order tensor, and then folds the first-order tensor into a third-order tensor. ; Step 12: Define the objective function for cross-modal data imputation and set the initial estimation tensor. ,satisfy: in, The set of indices of the corresponding elements is ; The formula for the objective function of cross-modal data imputation is: in, For intermediate tensors, intermediate tensor Expansion in mode 1; for Mode 2 unfolding, for Modal 3 expansion; , , These are the weighting coefficients; Step 13: Set additional variables , q =1,2,3; Step 14: Set cross-modal data imputation constraints and update the intermediate tensor based on the maximum number of iterations. Estimating tensors and additional variables ; ； ； ； in, For iteration parameters, To convert a matrix back into a tensor function, A third-order tensor composed of elements 0 and 1, For the correlation matrix A The third-order sparse tensor, for The estimated tensor, For additional tensors In modality q The following unfolds, () is the singular value decomposition function. For the first An intermediate tensor; The formula for the constraint conditions of cross-modal data completion is: in, A third-order tensor composed of elements 0 and 1, For the correlation matrix A The third-order sparse tensor, for The estimated tensor, For the first An intermediate tensor; Step 15: Obtain the filled third-order estimation tensor ,Will Expanding yields a filled first-order tensor image, which is then converted into a filled correlation matrix. ; Step 2: Construct miRNA integration similarity and disease integration similarity using miRNA similarity and disease similarity, respectively; Step 3: Construct a heterogeneous network graph using miRNA integration similarity, association matrix completion, and disease integration similarity. G and characteristic matrix and heterogeneous network graph G and characteristic matrix As input to a graph convolutional neural network; Step three specifically includes: Step 31: Construct a heterogeneous network , , For miRNA normalization matrix, A standardized matrix for diseases; Step 32: Use the double random walk algorithm to perform random walks on the miRNA ensemble similarity network and the disease ensemble similarity network respectively to obtain initial association scores; calculate the probability of random walks on the disease feature matrix network and the miRNA feature matrix network based on the initial association scores and the normalized matrix; the probability formula for random walks on the feature matrix network is: ； ； in, The decay factor for random walks, , These are miRNA and disease normalization matrix, respectively. The probability of disease-miRNA association. , These are miRNA and initial disease association score, respectively. Step 4: Embedding GCN Layer 1-3 nodes H 1 - H 3 The input is fed into the control gate; based on the cosine similarity between the three nodes embedded in the gate and the similarity threshold ST, it is determined whether the gate output is an integrated convolutional layer or a single convolutional layer; the miRNA-disease score matrix is ​​obtained using a bilinear decoder, resulting in the CIGGNET model; The condition for determining whether a gate outputs an ensemble convolutional layer or a single convolutional layer is as follows: if the cosine similarity between the three node embedding gates is greater than a given similarity threshold ST, the gate outputs a single convolutional layer; conversely, if the similarity between the three convolutional layers is less than the given similarity threshold ST, the gate outputs an ensemble convolutional layer. The formula for determining whether a gate outputs an integrated convolutional layer or a single convolutional layer is: ； in, , It is the miRNA embedding matrix. Disease embedding matrix This indicates the weight of the embedded three-layer GCN nodes. This indicates the embedding of a three-layer GCN node; Step 5: Train the CIGGNET model using the weighted cross-entropy loss function.

2. The miRNA and disease prediction method based on cross-modal and graph convolution according to claim 1, characterized in that, If miRNA i and miRNA j If functional similarity, sequence similarity, or semantic similarity is present, the miRNA ensemble similarity is the average of the miRNA sequence similarity matrix, miRNA functional similarity matrix, semantic similarity matrix, and miRNA GIP similarity; otherwise, the miRNA ensemble similarity is... SM It refers to miRNA GIP similarity.

3. The miRNA and disease prediction method based on cross-modal and graph convolution according to claim 1, characterized in that, The formula for the miRNA-disease score matrix is: ； in, For the training matrix, This is the final embedding matrix of miRNAs. The final embedding matrix for the disease.

4. The miRNA and disease prediction method based on cross-modal and graph convolution according to claim 1, characterized in that, The formula for the weighted cross-entropy loss function is: ； in, M For miRNAs, known association pairs N For known disease association pairs, For disease and miRNA Yes, λ is the weighting factor. and These represent the number of positive samples and the number of negative samples, respectively. miRNA-disease score matrix The value in.

Images