MiRNA and disease association prediction method based on graph convolution network

Abstract

The application relates to the technical field of graph convolution network, in particular to a miRNA and disease correlation prediction method based on a graph convolution network, which comprises the following steps: constructing a heterogeneous network by using disease integrated similarity, miRNA integrated similarity and miRNA-disease correlation information; learning the features of diseases and miRNAs by using GCN; inserting an active mask convolution layer in the embedding layer of the GCN; inserting a CRF layer in the embedding layer of the GCN; integrating the active mask convolution layer and the CRF layer; introducing an attention mechanism to reconstruct a miRNA-disease score matrix; and reconstructing an adjacency matrix of the miRNA and disease correlation by using a bilinear decoder. The application solves the problems that the complex network algorithm model and the deep learning model do not consider the complexity of model data and the large amount of calculation, and the node information in the graph data is not effectively aggregated in numerous GCN models.

Description

Technical Field

[0001] This invention relates to the field of graph convolutional network technology, and in particular to a method for predicting miRNA-disease associations based on graph convolutional networks. 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.

[0003] Existing computational methods can be divided into models based on complex network algorithms and prediction models based on deep learning.

[0004] Models based on complex network algorithms primarily utilize network algorithms on various constructed complex networks to predict potential disease-miRNA associations. For example, Li et al. proposed a novel multi-kernel fusion framework—Multi-View Kernel Fusion Network (MvKFN)—to efficiently fuse similar kernels from different data sources using a highly nonlinear approach. Yu et al. proposed a meta-path-based miRNA-disease association prediction, constructing a heterogeneous miRNA-disease gene information network. Huang et al., to fully explore the potential relationship between miRNAs and diseases, collected multi-source data and constructed a heterogeneous network. After network construction, they used a designed heterogeneous network framework (MEAHNE) to mine potential associations within the network.

[0005] Deep learning-based models are currently favored by many researchers due to their superior feature extraction and predictive performance. Han et al. proposed a novel deep learning model based on hierarchical graph attention networks for predicting miRNA-disease associations. Su et al. improved the accuracy of miRNA-disease association prediction by integrating similarity information and deep autoencoders. Ji et al. proposed a semi-supervised learning method for miRNA-disease association prediction based on variational autoencoders. SVAEMDA trains a predictor based on variational autoencoders using known miRNA-disease associations as concatenated dense vectors.

[0006] Many studies have used graph neural networks (Graph Neural Networks) to learn feature representations of miRNA-disease association pairs, achieving high performance in predicting these pairs. Unlike standard convolutional neural networks (CNNs), graph convolutional neural networks perform convolution operations on graph data. Generally, the more graph data available, the more significant its contribution to the model. However, large amounts of graph data can introduce redundancy and increase computational complexity. Furthermore, graph data, compared to general data, possesses similarity information between nodes. Therefore, preserving this similarity information in the hidden layers of graph convolutional neural networks is crucial.

[0007] However, neither complex network-based algorithms nor deep learning-based models take into account the complexity of the model data and the high computational cost. Furthermore, many GCN models fail to effectively aggregate node information from graph data. Summary of the Invention

[0008] To address the shortcomings of existing methods, the technical solution adopted in this invention is: a method for predicting miRNA-disease association based on graph convolutional networks, comprising the following steps:

[0009] Step 1: Construct heterogeneous networks using disease integration similarity, miRNA integration similarity, and miRNA-disease association information;

[0010] Furthermore, the formula for heterogeneous networks is:

[0011]

[0012] Among them, M m It is the standardization of SR, M d This is the standardization of SD, where A is the correlation matrix.

[0013] Step 2: Use GCN to learn the feature representations of diseases and miRNAs. Insert an active masking convolutional layer into the embedding layer of GCN to actively hide some associated nodes; and insert a CRF layer to enhance the representation aggregation of the neighborhood; integrate the results of the active masking convolutional layer and the CRF layer.

[0014] Furthermore, GCN is used to learn the features of diseases and miRNAs. Active masking convolutional layers are inserted into the embedding layers of the GCN to actively hide a portion of the associated nodes. Specifically:

[0015] Deploy GCN on the constructed heterogeneous network graph to combine node similarity and direct link association information;

[0016] Set the input graph G as:

[0017] G = A H (5)

[0018] Initialize the embedding as follows:

[0019]

[0020] The first-level formula for the GCN encoder is:

[0021]

[0022] in, It is the input to the hidden layer weight matrix. This is the first layer of embedding, and K is the dimension of the embedding layer.

[0023] Furthermore, an active mask convolutional layer is inserted to output the first layer of the GCN encoder, as shown in the formula:

[0024]

[0025] Among them, U (0) It is H (0) The masked part indicates that the masked part is a masked part of the ...

[0026] Furthermore, a CRF layer is inserted to enhance the aggregation of neighborhood representations; the results of integrating the active mask convolutional layer and the CRF layer specifically include:

[0027] Define a CRF layer using the following formula:

[0028]

[0029] Where Z(·) is the normalization factor and E(·) is the energy function;

[0030] Furthermore, energy functions are divided into univariate energy functions and paired energy functions. The definition of a univariate energy function is as follows:

[0031]

[0032] Among them, H i This represents the node i, Q updated in the CRF layer. i This represents the initial embedding of node i obtained from the GCN;

[0033] Furthermore, the pairwise energy function is defined as follows:

[0034]

[0035] Where, λ ij This represents the similarity between nodes i and j.

[0036] The CRF function is defined as follows:

[0037]

[0038] Where, N i α and β are the neighbors of node i, and α and β are weight factors;

[0039] Update H embedded in CRF layer i Nodes;

[0040]

[0041] Wherein, the initial embedding Set to Q i ,and It is the embedding that is updated in the k-th iteration.

[0042] Step 3: Introduce an attention mechanism. Based on the learned embeddings, reconstruct the miRNA-disease score matrix; use a bilinear decoder to reconstruct the adjacency matrix of miRNA-disease association.

[0043] Furthermore, the formula for the miRNA-disease score matrix is:

[0044] H = a1H1 + a2H2 + a3H3 (15)

[0045] in, It is the final miRNA embedding. The final disease embedding is defined by H1 and H3, which are active mask convolutional layers, and H2, which is a CRF layer. A1, A2, and A3 are attention weights automatically learned by the graph convolutional neural network.

[0046] Furthermore, the adjacency matrix formula for miRNA-disease association is as follows:

[0047]

[0048] in, It is a trainable matrix; miRNAm i and disease d j The correlation prediction score is related to the numerical value in the corresponding A' matrix A'(i,j).

[0049] The beneficial effects of this invention are:

[0050] 1. By inserting active mask convolutional layers into the hidden layers, more accurate prediction results are obtained by using a relatively simple data network to represent the miRNA-disease association while greatly reducing data redundancy; this can reduce the impact of redundant data on the model, improve the model's computation speed, and increase the model's generalization ability.

[0051] 2. To maximize the aggregation and preservation of node information of miRNAs and diseases, a conditional random field convolutional layer is used to ensure that similar nodes (i.e., miRNAs or diseases) have similar representations.

[0052] 3. An attention mechanism is used to assign weights to the results of different convolutional layers and fuse them to score unobserved miRNA-disease associations. In 5-fold cross-validation (CV) and 10-fold cross-validation (CV), the IACGCN model of this invention achieved AUCs of 0.9351 and 0.9401 on HMDD v2.0. Attached Figure Description

[0053] Figure 1(a), (b), and (c) are respectively the block diagrams of the miRNA-disease association prediction method based on graph convolutional networks, the active mask convolutional layer, and the conditional random field convolutional layer of the present invention.

[0054] Figure 2 This is a schematic diagram of the masked layer of the present invention;

[0055] Figure 3 This is a schematic diagram of the prediction time for different active mask convolutional layer adjustment parameter P pairs according to the present invention;

[0056] Figure 4 This invention relates to the impact of the Conditional Random Field (CRF) convolutional layer on model performance (IACGCN indicates that CRF is introduced, and IAGCN indicates that CRF is not introduced).

[0057] Figure 5 These are the attention weights of the three convolutional layers in the IACGCCN of this invention;

[0058] Figure 6 This is a comparison chart of the IACGCCN model of this invention and existing models using ten-fold cross-validation.

[0059] Figure 7 This is a comparison chart of the AUC of the IACGCCN model of this invention and existing models. Detailed Implementation

[0060] 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.

[0061] like Figure 1 As shown, a method for predicting miRNA associations with diseases based on graph convolutional networks includes the following steps:

[0062] 5430 known associations between 383 diseases and 495 miRNAs were extracted from the HMDDV2.0 database. An association matrix A was constructed to store the known disease-miRNA association information; in association matrix A, the number of rows M and columns N represent the number of miRNAs and the number of diseases, respectively. If the m of the miRNAs... i With disease d j If there is a known relationship between them, then A(m) i ,d j ) is set to 1 otherwise; a directed acyclic graph (DAG) is used to obtain the semantic information of the disease and two different disease semantic similarities are constructed.

[0063] Disease Semantic Similarity 1: Specifically, based on the assumption that the greater the shared portion of the DAGs of two diseases, the greater their semantic similarity value; calculate Disease Semantic Similarity 1 (SS1) and Disease Semantic Similarity 2: Clearly, for a given disease in the DAG of disease D, the fewer other diseases' DAGs disease D appears in, the greater its semantic contribution to the current disease D. Therefore, in Disease Semantic Similarity 1, it is inaccurate to consider diseases at the same level in the DAG of disease D as having the same semantic contribution value.

[0064] Disease semantic similarity 2 (SS2) was constructed to calculate the functional similarity (MS) between miRNA pairs. Functional similarity information for miRNAs can be obtained from the website http: / / www.cuilab.cn / files / images / cuilab / misim.zip. Based on the assumption that functionally similar miRNAs are usually associated with similar diseases, Gaussian interaction kernel similarity can be used as another measure to calculate the similarity between diseases or miRNA pairs. The Gaussian interaction kernel similarity (GD and GR) between diseases and miRNAs was calculated using the association matrix A. Integrating the disease semantic similarity with the disease Gaussian interaction kernel similarity yields the integrated disease similarity matrix SD. If disease d... i and d jIf a corresponding DAG exists, then the final disease ensemble similarity is the average of the disease semantic similarity and the disease Gaussian interaction spectrum kernel similarity; otherwise, the disease ensemble similarity is filled in using the Gaussian interaction spectrum kernel similarity. Similarly, integrating miRNA functional similarity with miRNA Gaussian interaction spectrum kernel similarity yields the miRNA ensemble similarity matrix SR. Constructing Heterogeneous Networks: Heterogeneous graph inference can effectively utilize the implicit topological information in heterogeneous graphs (obtained by fusing new association and similarity information), and GCNs can effectively learn the hidden topological information of each node. IACGCCN (Prediction of miRNA-disease association prediction based on integrated active mask layer and conditional random field layer Graph convolutional network) uses a graph convolutional network with an active masked layer and a conditional random field (CRF) layer to predict miRNA-disease associations, integrating the association matrix A with disease integration similarity information and miRNA integration similarity information into the heterogeneous network. In the process of constructing the heterogeneous network, IACGCCN standardizes SR and SD according to formulas and constructs the heterogeneous network A. H As shown in the formula:

[0065]

[0066]

[0067] Where m = 495, n = 383, M m It is the standardization of SR and M d It is the standardization of SD, SR is the miRNA integration similarity matrix, and SD is the disease integration similarity matrix.

[0068] The formula for constructing a heterogeneous network AH is:

[0069]

[0070] GCN is a multi-layered neural network architecture used to learn low-dimensional representations of nodes from graph-structured data. Each layer of GCN aggregates information about neighbors through direct links in the graph to reconstruct the embedding as input to the next layer; in a graph, GCN can aggregate information about neighbors, capture dependencies between data, and generate useful node representations.

[0071] The propagation method between layers of GCN is as follows:

[0072]

[0073] Among them, H l It is the embedding of the l-th layer node, D = diag(∑ j 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.

[0074] 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 G is set as:

[0075] G = A H (5)

[0076] Then, initialize the embedding as follows:

[0077]

[0078] The first-level formula for the GCN encoder is:

[0079]

[0080] in, It is the input to the hidden layer weight matrix. This is the first layer of embedding, where K is the dimension of the embedding layer. The definition of graph G is shown in the formula; for l = 1, 2, ..., L and And G as defined by the formula and H as defined by the formula (0) The subsequent layers of the GCN encoder follow the formula. After L iterations, L K-dimensional embeddings can be obtained from different graph convolutional layers. All graph convolutional layers use exponential linear units as non-linear activation functions, which not only accelerates the learning speed but also significantly improves the generalization performance.

[0081] In the training of GCN, heterogeneous networks are used to generate training data. Heterogeneous networks can bring a lot of miRNA-disease association information (known association information and potential association information) to GCN; however, they can also bring redundant information, thereby increasing the computational complexity of the model.

[0082] To address these issues, an active masking convolutional layer is inserted into the model. In the active mask, "masked" represents random miRNA-miRNA association nodes, disease-disease association nodes, and miRNA-disease association nodes in the hidden graph G. The presence of masked data causes related nodes to cluster together. By constructing unknown node associations through known node associations, the GCN is forced to learn better features and patterns.

[0083] The masked layer is defined as follows:

[0084]

[0085] Where P is the adjustment parameter of the masked layer, P∈(0.0...1.0); (GP) represents the association of the corresponding nodes masked from 1%...99% of the entire heterogeneous graph G. Figure 2 As shown, in the original embedding, the out-degree and in-degree of the disease 4 node are 7, and after masking, its out-degree and in-degree are 4. By operating in this way, many redundant association methods can be reduced between the disease 4 node and the miRNA 4 node, thereby reducing redundant information in the graph data and improving the computational efficiency of the model.

[0086] When l = 0, the first layer formula of the GCN encoder is:

[0087]

[0088] Among them, U (0) It is H (0) The masked representation; because in the first layer of embedding, both the feature matrix and the heterogeneous graph G need to be masked.

[0089] Intuitively, the GCN in the equation is first obtained by using The features of adjacent nodes are aggregated based on the adjacency matrix, and then a residual transformation is performed. For graph data, nodes are connected by edges, and these edges carry similarity relationships between different nodes. Therefore, in order to learn an efficient representation H from GCN... (l+1) It is necessary to force H (l+2) To maintain a similarity relationship. Otherwise, H (l+1) The input graph data cannot be fully represented, causing downstream tasks to degrade. To address this issue, a CRF is used to limit the H... (l+1) The behavior of [the study] is as follows: After deriving the initial embeddings, pairwise relationships between different nodes are captured using CRF to ensure that similar miRNAs (or diseases) are also similar in the feature space, i.e., they have similar embeddings. Simultaneously, the embeddings need to be smoothly updated.

[0090] Specifically, the node representation H is considered as a random variable {H} i}, where H i This represents node i being updated in the CRF layer, where these random variables are expressed in {Q}. i} as a condition, Q i The initial embedding of node i obtained from GCN is represented by the CRF layer as follows:

[0091]

[0092] Where Z(·) is the normalization factor, and E(·) is the energy function; the energy function is divided into univariate energy function and pairwise energy function; the definition of the univariate energy function is as follows:

[0093]

[0094] The univariate energy function measures the distance between the two, and minimizing this function as much as possible makes H i Approaching Q i In order to capture the similarity relationships between different nodes;

[0095] The pairwise energy function is defined as follows:

[0096]

[0097] Where, λ ij This represents the similarity between nodes i and j. Intuitively, when λ ij When it is large, minimize ψ P (H i H j Q i Q j ) will make H i Approaching H j Otherwise, it will H i Push away H j .

[0098] Based on the formulas, the function definition of CRF is as follows:

[0099]

[0100] Where, N i α and β are the neighbors of node i, and α and β are weighting factors.

[0101] At the same time, update the embedded H in the CRF layer according to the following rules i Nodes;

[0102]

[0103] Wherein, the initial embedding Set to Q i ,and The embedding is updated in the k-th iteration. Let k be the final representation of node i, and in the experiment, k is set to 2. Note that the first layer considers all neighbors equally, while the CRF layers focus on similar / important neighbors. Furthermore, as the number of iterations k in the CRF layers increases, the node will gradually obtain more and more information from its higher-order neighbors.

[0104] Considering that the active masking convolutional layer and the conditional random field convolutional layer constructed above each have their own characteristics and their contributions to the model are inconsistent, the graph convolutional network does not know the degree of contribution of the active masking convolutional layer and the conditional random field convolutional layer to the model. If we simply distribute their weights equally, this operation may cause the model to ignore the characteristics of different convolutional layers. In order to make reasonable use of their different characteristics, an attention mechanism is introduced to integrate these different embedding results and obtain the final miRNA-disease embedding H, the formula is:

[0105] H = a1H1 + a2H2 + a3H3 (15)

[0106] in, It is the final miRNA embedding. The final disease embedding is defined by H1 and H3, which are active mask convolutional layers, and H2, which is a conditional random field convolutional layer. A1, A2, and A3 are attention weights automatically learned by the graph convolutional neural network.

[0107] Using a bilinear decoder A' = f(H) M H D The adjacency matrix of miRNAs associated with diseases is reconstructed using the following formula:

[0108]

[0109] in, It is a trainable matrix; miRNAm i and disease d j The correlation prediction score between them is related to the values ​​in the corresponding A' matrix A'(i,j), and A'(i,j) can be regarded as a' ij .

[0110] Based on experimental data, there were 5430 miRNA-disease association pairs and 184155 unknown miRNA-disease association pairs. The miRNA-disease association pairs were considered positive samples, and the unknown miRNA-disease association pairs were considered negative samples, denoted as P0. + and P - Distinguishing between miRNAs and diseases is a binary classification problem. Clearly, the number of known miRNA-disease associations is far fewer than the number of miRNA-disease pairs for which no association has been observed. Therefore, a weighted cross-entropy loss function is used to train the model, calculated as follows:

[0111]

[0112] Where N and M are constants, N = 383, M = 495; (i,j) represents miRNAm i and disease d j Yes, weighting factor |P + | and | P - | represents the number of positive and negative samples, respectively; λ emphasizes the importance of the observed associations and reduces the impact of data imbalance.

[0113] Parameter settings

[0114] During model training, the Xaiver initialization method is used to randomly initialize all trainable weight matrices, and the Adam optimizer is used to optimize the loss function with a recurrent learning rate during optimization. To reduce the impact of overfitting on the model's predictive ability, IACGCCN introduces node dropout in the graph convolutional layers. This method discards some data in the embedding layers during training to prevent the model from treating unrelated miRNA-disease pairs as related ones.

[0115] In this experiment, we used 5-fold and 10-fold cross-validation to evaluate the performance of IACGCCN in predicting potential disease-related miRNAs. In the 5-fold cross-validation, known miRNA-disease associations were treated as positive samples and randomly divided into five identical 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 a similar manner to the 5-fold cross-validation. In each fold, a predictive model was built on the known associations in the training set and used to predict associations in the test set.

[0116] IACGCN has several parameters: embedding dimension k, number of layers L, learning rate lr, total training time α, node dropout rate β, and adjustment parameter P for the active mask convolutional layer. Through experimentation, we continuously adjusted these parameters, setting them to k = 256, L = 3, lr = 0.01, α = 4000, β = 0.4, and P = 0.66. Ablation experiments of the model were then conducted.

[0117] The effects of active masked convolutional layers:

[0118] The impact of active masking convolutional layers on model performance will be analyzed, and the feasibility of the method will be verified from both spatial and temporal perspectives.

[0119] First, considering the spatial aspect: In Table 1, when P = 0.1, the active masking convolutional layer hides 10% of the nodes in the graph data; IACGCCN achieves an AUC of 0.8859 in the five-fold cross-validation evaluation. As the adjustment coefficient in the active masking convolutional layer gradually increases, the AUC of IACGCCN in the five-fold cross-validation evaluation also increases accordingly; at P = 0.66, the AUC of IACGCCN reaches its highest value; further increasing the adjustment coefficient reveals that the AUC of IACGCCN begins to decrease.

[0120] From a time perspective, such as Figure 3 A diagram illustrating the prediction time for different active mask convolutional layer adjustment parameters P; when P = 0.1 or without inserting a random mask convolutional layer, model training takes 50 minutes to complete, but the training time decreases as P increases. Considering all factors, P = 0.66 is used as the adjustment coefficient for the random mask convolutional layer.

[0121] These two aspects demonstrate that inserting active mask convolutional layers can reduce the impact of redundant data on the model to some extent, improve the model's computation speed, and increase its generalization ability.

[0122] Table 1: Impact of different active masked convolutional layers on IACGCCN

[0123]

[0124] The effects of Conditional Random Field convolutional layers:

[0125] IACGCN has three convolutional layers, two of which are active mask convolutional layers and the other is a CRF layer. Three different models were reconstructed to compare with IACGCN, and their results are shown in Table 2.

[0126] Among them, GCN-I: an integrated convolutional layer model based on attention mechanism; GCN-A: a convolutional layer model based on active masking; and GCN-C: a convolutional layer model based on conditional random field and active masking.

[0127] In Table 2, the GCN-I, GCN-A, and GCN-C models all achieved good performance in five-fold cross-validation. Furthermore, when IACGCN integrated the different characteristics of these three models, it achieved the best AUC score.

[0128] Among them, GCN-C has a higher AUC than GCN-S and GCN-A; Figure 4 Among the models, the one that incorporates a Conditional Random Field (CRF) convolutional layer achieves the highest AUC score. This demonstrates the effectiveness of inserting a CRF convolutional layer into the GCN model.

[0129] Table 2: The impact of different convolutional layers on the model

[0130]

[0131] The impact of ensemble convolutional layers:

[0132] Active mask convolutional layers and conditional random field convolutional layers are inserted into the hidden layers of the model. They each have their own characteristics, and an ensemble convolutional layer needs to be built to fully utilize their characteristics. The ensemble convolutional layer is a component of the IACGCCN network architecture and is responsible for managing and quantifying the interdependencies of different convolutional layers.

[0133] IACGCN-L1 uses a first embedding layer to construct the IACGCN model, and the same applies to IACGCN-L2 and IACGCN-L3. Table 3 shows the results of all models evaluated by 5-CV on the main dataset; IACGCN-L1 and IACGCN-L2 produce better results than IACGCN-L3, indicating that the first and second embedding layers contain more information than the third embedding layer. This result may be due to the over-smoothing of GCNs. IACGCN with three ensemble embedding layers produces better results than IACGCN-L1, IACGCN-L2, and IACGCN-L3.

[0134] Perform 15 5-CV runs for IACGCCN, and in Figure 5 The attention weights of the three convolutional layers are visualized. The three layers have different attention weights, with the second layer > the first layer > the third layer. This is because the first and third layers are active masking convolutional layers to improve computational efficiency and reduce the impact of redundant data, while the second layer inserts a CRF layer to avoid the loss of miRNA-disease node information. The results help explain the performance of IACGCCN-L1, IACGCCN-L2, and IACGCCN-L3 in Table 3. Table 4 also shows the similarity of the first, second, and third convolutional layers, revealing distinct differences. Therefore, integrating different convolutional layers using an attention mechanism is essential. In Table 3, IACGCCN-ave assigns uniform weights to different embeddings; IACGCCN-add directly fuses different embeddings. When uniform weights are assigned to different embeddings (IACGCN-ave), its AUC is lower than IACGCCN-L1 and IACGCCN-L2 because it ignores the different characteristics of these three convolutional layers. When directly fusing different embeddings (IACGCN-add), its AUC performance is good. This operation integrates information from all convolutional layers, but it ignores the contribution of each convolutional layer to the model. Embedding layers that contribute more to IACGCN receive more attention, while those that contribute less receive less attention, allowing for a more reasonable allocation of weights and utilization of the characteristics of each embedding layer.

[0135] Table 3: Performance of IACCGCN based on different convolutional layers

[0136]

[0137] Table 4: Similarity between different convolutional layers in IACGCCN

[0138]

[0139] IACGCN was compared with four state-of-the-art miRNA-based disease prediction methods and two baseline methods; the comparison results are shown in Table 5.

[0140] VGAMF constructs two subnetworks for miRNA-disease association. Deep autoencoders are used to train the miRNA-disease subnetwork data, and non-negative matrix factorization is used to obtain the features of miRNA and disease.

[0141] NIMCGCN constructs graph data from miRNA similarity networks and disease similarity networks, then feeds them into the GCN network for learning, and introduces neural induction matrix completion to predict the association between miRNAs and diseases.

[0142] VAEMDA constructs two matrices by integrating disease-miRNA association, disease similarity, and miRNA similarity, which are used to train variational autoencoders.

[0143] NCMCMDA is used to fill in the neighborhood constraint matrix for miRNA-disease association prediction to predict potential miRNA-disease associations.

[0144] The MDHGI model uses matrix factorization to obtain low-rank information of the miRNA-disease association matrix and utilizes the integration similarity between diseases and miRNAs to construct a heterogeneous network.

[0145] LAGCN uses graph convolutional networks to predict drug-disease associations, adopts the graph embeddings contributed by each convolutional layer, and uses an attention mechanism to reasonably distribute the graph embeddings of each graph convolutional layer.

[0146] According to Table 5, IACGCN's AUC score shows impressive performance compared to state-of-the-art models. MDHGI and NCMCMDA, based on matrix factorization and heterogeneous network prediction models, have AUC scores of 0.8805 and 0.8942, respectively. VGAMF and VAEMDA use deep autoencoder models to calculate miRNA-disease association scores, with AUC scores of 0.9243 and 0.9091, respectively. IACGCN, LAGCN, and NCMCMGCN are GCN-based methods for predicting miRNA-disease association scores; compared to these two baseline methods, IACGCN achieves the best results. IACGCN introduces masked convolutional layers and conditional random field convolutional layers on top of GCN, reducing the impact of redundant data and the loss of miRNA-disease node information. Therefore, under the same training data, IACGCN is more advanced than LAGCN and NCMCMGCN.

[0147] The results based on the 10-fold cross-validation are as follows: Figure 6As shown, ROC curves were plotted at different thresholds with the true positive rate (TPR, sensitivity) on the X-axis and the false positive rate (FPR, 1-specificity) on the Y-axis. Sensitivity refers to the proportion of test samples that rank before the threshold, while specificity refers to the percentage of negative miRNA-disease pairs that rank after the threshold. AUC was calculated as the evaluation metric; a higher AUC indicates more accurate predictive performance of the model.

[0148] Known miRNA-disease associations are crucial for predicting unknown miRNA-disease associations; the number of known associations significantly impacts method performance. To test the robustness of IACGCN with MDHGI, VAEMDA, NIMCGCN, and LAGCN, a subset of known associations were randomly removed from the main dataset, and 5-CV was implemented to evaluate the method. Figure 7 As shown, the number of miRNA-disease associations is an important factor in miRNA-disease association prediction, and more associations can produce better prediction models; compared with the four existing models, IACGCN is the most robust and performs best on data with different richness levels.

[0149] Table 5: Performance of comparison methods on HDMM V2.0

[0150] Model 5-fold of AUC Standard deviation IACGCN 0.9351 + / -0.0020 VGAMF 0.9243 + / -0.0019 NIMCGCN 0.9091 + / -0.0065 VAEMDA 0.8312 + / -0.0138 NCMCMDA 0.8942 + / -0.0015 MDHGI 0.8805 + / -0.0021 LAGCN 0.9081 + / -0.0039

[0151] Case Study:

[0152] To further demonstrate the predictive performance of IACGCCN, case studies of three different types of colorectal cancer, esophageal cancer, and kidney cancer were conducted. In the first type of case study, to verify the adaptability of the model to the old data HMDD v1.0, known disease-miRNA associations in HMDD were used as the training dataset.

[0153] In the first case study, known miRNA-disease correlations in HMDD v1.0 were used as training samples for the model to evaluate whether IACGCN could demonstrate stable performance on different databases.

[0154] Researchers have discovered an increasing number of miRNAs linked to colorectal cancer. They used IACGCCN to predict potential miRNAs associated with colorectal cancer and validated the top 50 miRNAs associated with colorectal cancer using the HMDD v2.0, dbDEMC, and miR2Disease databases. Table 6 shows that 43 of the top 50 miRNAs associated with colorectal cancer were confirmed in the dbDEMC, HMDD v2.0, and miR2Disease databases.

[0155] Table 6: Validation of the top 50 miRNAs predicted by the model and their association with colon cancer.

[0156]

[0157] Among them, dbDEMC is abbreviated as ①, miR2Disease as ②, and HMDD as ③.

[0158] The second case study used the IACGCCN model to validate the association between esophageal cancer and 50 other miRNAs in the disease-miRNA association databases dbDEMC and miR2Disease.

[0159] Studies have shown the impact of miRNAs on esophageal tumors. Therefore, detecting miRNA expression in esophageal tumor patients is crucial for analyzing their condition. IACGCCN was used to predict potential miRNAs associated with esophageal cancer, and the top 50 miRNAs associated with esophageal cancer were validated using the HMDDv2.0, dbDEMC, and miR2Disease databases. Table 7 shows that 47 of the top 50 miRNAs associated with esophageal cancer were validated in the dbDEMC, HMDD v2.0, and miRDisease databases.

[0160] Table 7: Validation of the top 50 miRNAs predicted by the model and their association with esophageal cancer.

[0161]

[0162] Among them, dbDEMC is abbreviated as ①, miR2Disease as ②, and HMDD as ③.

[0163] In the third case study, the aim was to evaluate whether IACGCN could be applied to diseases for which no known associated miRNAs were found. Given an investigated disease, known associations in HMDD v2.0 were removed, and it was treated as a new disease. Therefore, IACGCN predicted the studied disease solely based on information about known associated miRNAs from other diseases and similarity information between all diseases and miRNAs.

[0164] CDIGGCN was used to predict potential miRNAs associated with kidney tumors; Table 8 shows that 46 of the top 50 miRNAs associated with kidney tumors were validated by the dbDEMC, HMDD v2.0 and miR2Disease databases.

[0165] Table 8: Validation of the top 50 miRNAs predicted by the model and their association with kidney tumors.

[0166]

[0167] Among them, dbDEMC is abbreviated as ①, miR2Disease as ②, and HMDD as ③.

[0168] 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 predicting miRNA-disease association based on graph convolutional networks, characterized in that, Includes the following steps: Step 1: Construct heterogeneous networks using disease integration similarity, miRNA integration similarity, and miRNA-disease association information; Step 2: Use GCN to learn the characteristics of diseases and miRNAs; Active mask convolutional layers are inserted into the first and third convolutional layers of the GCN embedding layer; include: Insert an active masked convolutional layer to output the first layer of the GCN encoder, as shown in the formula: (9)； in, yes The masked part indicates that the masked part is used to indicate the masked part. These are the adjustment parameters for the masked layer; Represented as a non-linear activation function; It is the input to the hidden layer weight matrix; It is the first layer of embedding; K It is the dimension of the embedding layer; D For the image G The degree matrix; To initialize the embedding; M The number of miRNAs; N The number of diseases; Insert a CRF layer into the second convolutional layer of the GCN embedding layer; include: Define the CRF layer, given the initial embedding Q At that time, the updated node H The formula for conditional probability is: (10) ； in, As a normalization factor, It is an energy function; Embedded in CRF functions The loss function is defined as follows: (13)； in, It is a node i The neighbors, and and It is a weighting factor. For nodes obtained from GCN i Initial embedding; This indicates the node being updated in the CRF layer. i ; This indicates the node being updated in the CRF layer. j ; For nodes i , j Similarity between them; Update the CRF layer k +1 iterations of the node i Embedded The formula is: (14)； in, and It is a weighting factor; For the first k The node of the next iteration j Embedding; It also integrates active mask convolutional layers and CRF layers; Step 3: Introduce an attention mechanism to reconstruct the miRNA-disease embedding; Reconstructing the adjacency matrix of miRNAs and disease associations using a bilinear decoder. ; (16)； in, It is a trainable matrix. It is the final miRNA embedding. It is the final disease embedding; Using the i-th miRNA and the jth disease Adjacency matrix between medium elements Calculate the predicted score.

2. The method for predicting miRNA-disease association based on graph convolutional networks according to claim 1, characterized in that, The formula for heterogeneous networks is: （3）； in, It is a miRNA integration similarity matrix SR Standardization It is a disease integrated similarity matrix SD Standardization A This is the correlation matrix.

3. The method for predicting miRNA-disease association based on graph convolutional networks according to claim 1, characterized in that, Energy functions are divided into univariate energy functions and paired energy functions. The definition of a univariate energy function is as follows: （11）。 4. The method for predicting miRNA-disease association based on graph convolutional networks according to claim 3, characterized in that, The pairwise energy function is defined as follows: (12)； in, Q j For nodes obtained from GCN j Initial embedding.

5. The method for predicting miRNA-disease association based on graph convolutional networks according to claim 1, characterized in that, The formula for miRNA-disease embedding is: (15)； in, , It is the final miRNA embedding. It is the final disease embedding. H 1 , H 3 For active masked convolutional layers, H 2 For CRF layer, These are attention weights that are automatically learned by a graph convolutional neural network.

Images