Association prediction method based on hypergraph convolutional neural network and graph transformer
Through the hypergraph convolutional neural network and graph transformer method integrating multiple associated information, the problem of insufficient information in miRNA-disease prediction is solved, and more efficient feature extraction and prediction accuracy is achieved.
Patent Information
- Application Number
- CN202510120199.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-25
- Publication Date
- 2025-07-25
AI Technical Summary
The existing miRNA-disease association prediction methods lack sufficient prior information, resulting in insufficient prediction accuracy and traditional methods fail to effectively utilize multiple association relationships in complex biological systems.
A method of combining hypergraph convolutional neural network with graph transformer is adopted to integrate a variety of key information, such as miRNA-disease association, functional similarity, sequence similarity, semantic similarity, etc., attribute features are extracted through hypergraph convolutional neural network, and biometric features are obtained using graph transformer, and finally prediction is made in a multi-layer perception machine.
It improves the accuracy and reliability of miRNA-disease association prediction, significantly improves prediction performance, and can effectively identify known and potential miRNA-disease associations.
Smart Images
Figure CN120373349A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of neural network models, and particularly relates to an association prediction method based on a hypergraph convolutional neural network and a graph transformer Background Art
[0002] MicroRNAs (miRNAs) are non-coding RNAs consisting of approximately 22 nucleotides. They participate in the processes of biological cell growth, differentiation, and apoptosis by regulating target genes. In recent years, many studies have shown that the abnormal expression of miRNAs is closely related to the occurrence of diseases. In-depth study of the association between miRNAs and diseases can reveal the potential role of miRNAs in the disease pathogenesis, contribute to exploring potential biomarkers, and also provide new perspectives and methods for the early diagnosis and personalized treatment of diseases. In addition, miRNA can also be used as a drug target to open up new ways for disease treatment
[0003] The interaction mode between miRNA and diseases has become a breakthrough point for researchers to understand the disease pathogenesis at the molecular level. Currently, there are many biological methods available for detecting disease-related miRNAs, such as quantitative real-time polymerase chain reaction, high-throughput sequencing, etc. Although traditional biological methods can accurately determine the relationship between miRNAs and diseases, they require a large amount of time and effort. Computational model methods can discover potential miRNA-disease associations quickly at low cost. Therefore, it is necessary to develop effective computational models to predict miRNA-disease associations. Currently, most computational methods for predicting disease-related miRNAs rely on the assumption that functionally related miRNAs tend to be associated with similar diseases. This process includes constructing an association network based on various association information and then inferring new miRNA-disease associations using known relationships. According to the inference methods adopted, these methods can be divided into methods based on similarity network learning and methods based on machine learning. Methods based on similarity network learning can measure the similarity between miRNAs and diseases, but their computational models are relatively simple. More importantly, they rarely pay attention to using the complex relationships between miRNAs and diseases in heterogeneous networks, which may limit their accuracy in similarity calculation. In contrast, miRNA-disease association prediction models based on machine learning have significantly reduced the computational cost. However, the quality of feature extraction of the models has a significant impact on the prediction results, and using manually crafted features severely limits their representation ability, thereby affecting the prediction accuracy. Existing machine learning computational methods mainly focus on biological-related molecules directly related to miRNAs or diseases, ignoring that the human body is a highly complex system, and miRNAs or diseases may be indirectly related to various types of biomolecules, such as proteins, circRNAs, LncRNAs, etc
[0004] With the emergence of deep learning, many end-to-end computing methods have been developed. These new prediction techniques are superior to traditional machine learning techniques. Among them, the graph neural network GNN (Graph Neural Network) has received extensive attention. In particular, due to its superior ability in processing graph-structured data, graph convolutional networks GCNs (Graph Convolutional Networks) have received extensive attention. Commonly used GNN models can effectively aggregate task context information related to graph data, including graph convolutional networks GCNs, graph attention networks GATs (Graph Attention Networks), and graph autoencoders GAEs (GraphAutoencoders). However, there are still some problems to be solved in these methods, such as less preprocessing information, insufficient effective utilization of information, information sparsity, and ineffective extraction of information features. Therefore, the main direction for improving the prediction model lies in increasing auxiliary information and more effectively extracting feature information. Summary of the Invention
[0005] The main problem addressed by the present invention is the lack of sufficient prior information, and an association prediction method based on hypergraph convolutional neural network and graph transformer is proposed. To solve the problem of insufficient similarity information, multi-source similarity information is integrated, and a variety of key and effective information is integrated, mainly including known miRNA-disease associations, miRNA functional similarity, miRNA sequence similarity, and miRNA's GIP (Gaussian Interaction Profile) kernel similarity, disease semantic similarity, target-based disease similarity, and disease GIP kernel similarity, protein information datasets, LncRNA information datasets, drug node information datasets, circRNA information datasets, and microbial information datasets, to obtain biological entity information. The hypergraph convolutional neural network algorithm is used to process relevant data to improve the prediction accuracy of the algorithm.
[0006] To obtain more effective information, the hypergraph convolutional neural network extracts features from the association matrix to obtain the attribute features of the nodes. At the same time, a biological entity graph is constructed, and the graph transformer is used to obtain biological features. To solve the non-linearity problem of the data, a multi-layer perceptron (MLP) is introduced into the method. Finally, the fused data is input into the MLP to obtain the predicted correlation score. The specific steps of the present invention are as follows:
[0007] (1) Obtain the miRNA-disease association information and construct the association matrix A;
[0008] (2) Download the functional similarity of miRNAs, the sequence similarity of miRNAs, and the disease similarity based on targets to construct the miRNA functional similarity matrix MF; the miRNA sequence similarity MS and the disease similarity DT based on targets; prepare the protein information dataset, the LncRNA information dataset, the drug node information dataset, the circRNA information dataset, and the microbial information dataset to obtain the biological entity information;
[0009] (3) Calculate the semantic similarity of diseases Construct the semantic similarity network of diseases, and the semantic contribution of disease d to its ancestor D Let T(d) be the set of nodes of disease d and its ancestors, Δ represent the semantic contribution factor and be set to 0.5, n represent the number of disease types, and d i ' represent the child node of d i The GIP kernel similarity DG of diseases = exp(-γ d ||A(d i ) - A(d j )|| 2 ), where γ d is the parameter used to adjust the kernel bandwidth, and the calculation formula is: γ' d = 1, n d is the number of diseases, and the GIP kernel similarity MG of miRNAs = exp(-γ m ||A(m i ) - A(m j )|| 2 ), where γ m is the parameter used to adjust the kernel bandwidth γ′ m = 1, n r is the number of miRNAs;
[0010] (4) Perform data fusion on the association information, and the obtained miRNA dataset is: C m = {MF, MS, MG}; the obtained disease dataset is C d = {DS, DT, DG}; at the same time, use the Torch Geometric tool to construct a biological entity graph, and input the collected biological entities as nodes into the HeteroData graph;
[0011] (5) Use the hypergraph convolutional neural network to aggregate and extract features from the associated data, and finally obtain the attribute features on the miRNA side as the attribute features on the disease side as
[0012] (6) Use the graph transformer to extract features from the biological entity data, and obtain the biological entity features. The miRNA-side biological features are The disease-side biological features are
[0013] (7) Combine the attribute features extracted using the hypergraph convolutional neural network and weight-assigned by the attention mechanism with the biological features extracted from the biological entity graph to obtain the fusion feature matrix on the miRNA side The fusion feature matrix on the disease side Embed the miRNA And the embedding of the disease Connect them to obtain the complete dataset: X ij = concatenating(V m , V d ), X ij Represents the feature correlation pair between miRNA m i And disease d j Use a multi-layer perceptron to score the final miRNA-disease association prediction as follows: X l = ReLU(X l-1 W l + b l ), Represents the number of hidden layers, X l Represents the output of the l-th hidden layer, W l Represents the learning parameter matrix of the l-th layer, b l Represents the bias term of the l-th layer, and minimize the binary cross-entropy loss function as
[0014] Represents the final prediction score, (i, j) represents the association pair between disease d j And miRNA m i Y + Represents the positive sample set, y - Represents the negative sample set, and N represents the number of all miRNA-disease association pairs.
[0015] Specifically, in step (5), after constructing the interaction networks of miRNAs and diseases respectively, use the hypergraph convolutional neural network to extract various similarity information and update the embedding of each node. The specific steps are as follows:
[0016] 5i) Normalize the adjacency matrix of the miRNA-miRNA interaction network to obtain the normalized matrix: Represents the adjacency matrix of miRNA. E m ∈R m×m Represents Cm ∈R m×m The degree matrix of, where m is the size of the matrix. The present invention constructs the initial incidence matrix of the hypergraph. The calculation definition of the initial incidence matrix on the miRNA side is: represents the node embedding of the l-th layer, and Y(:, j) represents all elements of the j-th column of the incidence matrix. To explore the high-order connectivity in the miRNA-disease bipartite graph, if disease (i) is a k-th order reachable neighbor of disease (j), there exists a path between i and j, and the number of diseases within this path is less than k. The k-th order reachable matrix of diseases is: where power(·, k) is the k-th power of the matrix. Using this k-th order reachable adjacency matrix of diseases, the present invention constructs a hypergraph reflecting the k-th order neighbor relationship of miRNAs. The k-th order hypergraph incidence matrix of miRNAs is:
[0017] 5ii) The present invention performs a connection operation on miRNAs, fusing the hypergraph based on the initial incidence matrix and the hypergraph based on the second-order reachability rule. The incidence matrix H of the obtained miRNA hypergraph m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the calculation representation of miRNA nodes is as follows: represents the node at the l th -th embedding in the hypergraph convolutional neural network layer. represents the learnable weight matrix, and σ(·) is the non-linear activation function. Finally, the attribute features on the miRNA side are obtained as
[0018] 5iii) Normalize the adjacency matrix of the interaction network between diseases to obtain the normalized matrix: C d ∈R d×d represents the adjacency matrix of miRNAs. E d ∈R d×d represents C d ∈R d×d 's degree matrix. The present invention constructs the initial incidence matrix of the hypergraph. The calculation definition of the initial incidence matrix on the disease side is: $h_{l}$ represents the node embedding of the $l$-th layer, and $Y(i, :)$ represents all elements of the $i$-th row of the association matrix. To explore the high-order connectivity in the miRNA-disease bipartite graph, the present invention defines the $k$-th order reachable neighbors of miRNAs and diseases according to previous studies. In the miRNA-disease bipartite graph, if miRNA(i) is the $k$-th order reachable neighbor of miRNA(j), there exists a path between $i$ and $j$, and the number of miRNAs within this path is less than $k$. The $k$-th order reachable matrix of miRNAs is: where $power(·, k)$ is the $k$-th power of the matrix. Using this $k$-th order reachable adjacency matrix of miRNAs, the present invention constructs a disease hypergraph reflecting the $k$-th order neighborhood relationship of diseases. The $k$-th order hypergraph incidence matrix of diseases is:
[0019] 5iiii) The present invention performs a connection operation on the disease matrix, fusing the hypergraph based on the primary incidence matrix and the hypergraph based on the second-order reachability rule. The incidence matrix $H$ of the obtained disease hypergraph m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the node calculation representation of diseases is as follows: $h_{l}$ represents the embedding of the node in the $l$-th th hypergraph convolutional neural network layer. $W$ represents the learnable weight matrix, and $\sigma(·)$ represents the non-linear activation function. Finally, the attribute features on the miRNA side are The attribute features on the disease side are
[0020] Specifically, in step (6), in the biological entity graph part, it is divided into three modules: heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation. The specific calculation steps are as follows:
[0021] 6a) In the heterogeneous mutual attention part: Different from the traditional Transformer model, the present invention assigns a unique set of projection weights $W$ to each pair of meta-relations, while in the traditional Transformer, all words share the same set of weights. The present invention applies linear projection technology to represent the node $t$ as a $Q - Linear_i$ vector. To convert the target node $t$ into the $i$-th query vector, the present invention performs the following calculation: $Q$ i $(t)$ represents the $i$-th query vector to which the target node $t$ is projected. Similarly, the present invention performs a linear projection on the source node, representing the node $s$ as a $K - Linear_i$ vector. To convert the source node $s$ into the $i$-th vector, the present invention performs the following calculation: $K$ i(s) represents the i-th Key vector to which the source node is projected. The present invention uses a multi-head attention mechanism to perform weight distribution, and the calculation method is as follows: ATT-head i represents the i-th attention head, e represents the relationship between source node s and target node t, and μ(·) represents the general meaning of each relationship triple as a means of adaptively scaling attention. The operation of the Attention(·) vector mainly consists of connecting h ATT-head connections to obtain the attention vector for each node pair A softmax function is used in the computation to generate a probability distribution specific to each target node t, which is formed using the attention vectors accumulated from its neighboring nodes N(t).
[0022] 6b) Message passing part: This module introduces the meta-relation of edges in the message passing mechanism to alleviate the distribution differences of different types of nodes and edges and pass information from the source node to the target node. Each pair of nodes e = (s, t) calculates its multi-head message through the following formula: Project the source node s into the i-th message vector. Finally, connect all h message headers to get the complete information: Represents the weight in the message passing process.
[0023] 6c) Target aggregation part: In this module, the above heterogeneous mutual attention module and heterogeneous message passing module are first aggregated from the source node to the target node. The above steps are as follows: Then, the above results are input into the Relu activation layer, and the output is linearly transformed and residually connected. The activation function part can be expressed as:
[0024] Finally, the embedding extraction of the biological entity graph is completed to obtain the biological entity features. The biological features on the miRNA side are Disease-side biological characteristics are
[0025] The present invention has the following advantages compared with the prior art: The present invention adds various association similarity information, solves the problem of sparse prediction prior information, improves the reliability of prediction, constructs a hypergraph using these association information, makes full use of the marginal association between multiple nodes, uses a hypergraph convolutional neural network for feature extraction, and ensures the extraction of the required key features. In order to improve the reliability of the algorithm model, the present invention adds protein information datasets, LncRNA information datasets, drug node information datasets, circRNA information datasets, and microbial information datasets to construct a biological entity graph, uses graph transformers to extract features from the biological entity graph, and finally obtains the biological entity features of the nodes, enriching the biological entity data. To solve the non-linear problem of data, a multi-layer perceptron (MLP) is introduced in the method. Finally, the biological features and attribute features are input into the MLP to obtain the predicted correlation scores. Compared with the existing methods, the prediction performance has been significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 is a flowchart of the method of the present invention;
[0027] Figure 2 is a display of the ROC performance curve of the five-fold cross-validation of the method of the present invention;
[0028] Figure 3 is a display of the ROC performance curve of the ten-fold cross-validation of the method of the present invention; SPECIFIC EMBODIMENTS
[0029] The present invention will be clearly and completely described below in conjunction with specific datasets. The datasets applied in the present invention are only used to illustrate the present invention and are not used to limit the scope of use of the present invention.
[0030] Example 1
[0031] An association prediction method based on a hypergraph convolutional neural network and a graph transformer, characterized in that the method comprises the following steps:
[0032] (1) Prepare the miRNA-disease association dataset and obtain the association information between miRNA and disease; prepare protein information datasets, LncRNA information datasets, drug node information datasets, cirRNA information datasets, and microbial information datasets to obtain biological entity information;
[0033] (2) Download the miRNA functional similarity matrix MF, miRNA sequence similarity MS, and target-based disease similarity DT;
[0034] (3) Calculate the semantic similarity DS of diseases. According to the constructed miRNA-disease association matrix A, calculate the GIP kernel similarity DG of diseases and the GIP kernel similarity MG of miRNAs respectively;
[0035] (4) Concatenate the obtained similarity matrices to get the miRNA association dataset: C m = {MF, MS, MG}, and the obtained disease association dataset is C d = {DS, DT, DG}; for the biological entity graph part, construct a HeteroData graph;
[0036] (5) Use the hypergraph convolutional neural network to aggregate and extract features from the associated data and Use the graph transformer to extract features from the biological entity data to obtain biological entity features and
[0037] Use the hypergraph convolutional neural network to extract various similarity information and update the embedding of each node. The specific steps for miRNA side attribute feature extraction are as follows:
[0038] 5i) Process the miRNA adjacency matrix C m to obtain the normalized matrix L m , represents the miRNA adjacency matrix. E m ∈ R m×m is denoted as C m ∈ R m×m is the degree matrix of, and m is the size of the matrix. Construct the initial association matrix of the hypergraph. The calculation of the initial association matrix on the miRNA side is defined as: L m (i,:) represents all columns of the i-th row of the L m matrix, represents the node embedding of the l-th layer, and Y(:,j) represents all elements of the j-th column of the association matrix.
[0039] 5ii) To explore the high-order connectivity in the miRNA-disease bipartite graph, if disease i is a k-order reachable neighbor of disease j, there exists a path between i and j, and the number of diseases within this path is less than k. The k-order reachable matrix of diseases is: where power(·,k) is the k-th power of the matrix, represents the initial association matrix on the disease side, represents the transpose of the initial association matrix on the disease side; use the k-order reachable adjacency matrix of diseases Construct a hypergraph that reflects the k-order neighbor relationship of miRNAs, and the incidence matrix of the k-order hypergraph of miRNAs The calculation method is as follows: Denote the initial incidence matrix on the miRNA side, Denote the (k - 1)-order reachable adjacency matrix of diseases.
[0040] 5iii) Perform a connection operation on miRNAs to fuse the hypergraph based on the primary incidence matrix and the hypergraph based on the second-order reachability rule. The incidence matrix H of the resulting miRNA hypergraph m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the incidence matrix obtained after the miRNA node undergoes l + 1 iterations is calculated as follows: L m Denote the normalized incidence matrix on the miRNA side, Denote the embedding of the node in the th l-th hypergraph neural network layer, Denote the learnable weight matrix, σ(·) is the non-linear activation function, and the final attribute feature on the miRNA side is
[0041] In the part of using the graph transformer to extract biological entity features, it is divided into three modules: heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation. The specific calculation steps are as follows:
[0042] 5i) In the heterogeneous mutual attention part: Assign a unique set of projection weights W to each pair of meta-relations, apply the linear projection technique, and represent the node t as a Q-Lineari vector. In order to convert the target node t into the i-th query vector, this invention performs the calculation, and the calculation formula is as follows: Q i (t) represents that the target node t is projected onto the i-th query vector, Denote the linear projection process, H (l-1) [t] represents the iterative result of the target node t in the (l - 1)-th layer. Similarly, perform a linear projection on the source node and represent the node s as a K-Lineari vector. In order to convert the source node s into the i-th vector, this invention performs the calculation, and the calculation formula is as follows: K-linear(·) represents the linear projection process, K i (s) represents the i-th Key vector that the source node is projected onto, H (l-1) [s] represents the (l-1) iterative result of the source node s in the
[0043] Weight distribution is performed using the multi-head attention mechanism, and the calculation method is as follows: ATT-head i represents the i-th attention head. As a way to adaptively scale attention, e represents the relationship between the source node s and the target node t, and μ(·) represents the general meaning between each pair of triple relationships. K i (s) represents the linear projection vector of the source node s, represents the weight, Q i (t) T represents the query vector of the target node t, represents the scaling factor, τ(s), τ(t) represents the dot product multiplication. The operation of the Attention(·) vector mainly includes connecting h ATT-heads to obtain the attention vector for each pair of nodes The softmax function is used in the solution process, normalizes ATT-head i (s,e,t) to generate a probability distribution specific to each target node t, which is formed using the attention vectors accumulated from adjacent nodes N(t);
[0044] 5ii) Message passing part: This module introduces the meta-relationship of edges in the message passing mechanism to alleviate the distribution differences of different types of nodes and edges, and passes information from the source node to the target node. For each pair of nodes e=(s,t), its multi-head message is calculated through the following formula: represents projecting the source node s onto the i-th message vector, H (l-1) [s] represents the result of the iterative process, represents the weight in the message passing process, Message HGT (s,e,t) represents connecting all h message heads to obtain the complete information, and the calculation process is as follows:
[0045] 5iii) Target aggregation part: In this module, first, the above heterogeneous cross-attention module Attention HGT (s,e,t) and the heterogeneous message passing module Message HGT (s,e,t) are aggregated from the source node to the target node through The above steps are shown as follows: Then, the result obtained above is input into the Relu activation layer. σ(·) represents the activation function, and a linear transformation is performed on the output and the residual H (l-1) [t] are connected, and the activation function part can be expressed as:
[0046] Finally, through i iterations of the activation function, the miRNA-side biometric matrix is calculated as The disease-side biometric matrix is
[0047] (6) Using the hypergraph convolutional neural network to extract and the attention mechanism to assign weights, the attribute features are combined with the biometric features extracted from the biological entity graph: concatenating(·) represents matrix concatenation;
[0048] (7) Using a multi-layer perceptron to process the data and obtain the final prediction score.
[0049] Example 2
[0050] As Figure 1 shown, a correlation prediction method based on a hypergraph convolutional neural network and a graph transformer, the specific implementation steps are as follows:
[0051] Step 1, Input and output data description
[0052] Download the miRNA-disease association dataset from a public database and preprocess it. After removing some redundant data, the present invention selects 12,446 associations between m miRNAs and d diseases. Among them, m = 853 and d = 591. The number of protein-disease association pairs is 2,042, the number of LncRNA-disease association pairs is 874, the number of drug-disease association pairs is 17,414, the number of circRNA nodes is 421, and the number of microbial nodes is 100.
[0053] Step 2, Construct initialization data
[0054] (1) Using the known information to construct the association matrix A, A(i,j) = 0 indicates that there is no association between miRNA m i and disease d j and A(i,j) = 1 indicates that there is an association between miRNA m i and disease d j exists.
[0055] (2) Using the known association information to calculate the semantic similarity of diseases. The relationship between diseases is described as a directed acyclic graph DAG. Then, the present invention uses the calculation method proposed by predecessors to calculate the semantic value of each disease. The calculation method of the semantic contribution of a disease to its ancestors is: The calculation method of the disease semantic value is as follows: Among them, T(d) is the set of nodes of disease d and its ancestors, Δ represents the semantic contribution factor and is set to 0.5, n represents the number of disease types which is 591, and d i ' represents the child node of d i . According to the previous view, if there are more directed acyclic graphs shared between two diseases, the two diseases are more similar. The semantic similarity between disease d i and d j is defined as follows:
[0056] (3) According to the hypothesis that similar miRNAs are more likely to be related to similar diseases, the GIP Gaussian kernel similarity of diseases is calculated as follows: DG = exp(-γ d ||A(d i ) - A(d j )|| 2 ), where A(d i ) corresponds to the i-th row vector in the association matrix A. The parameter γ is used to adjust the kernel bandwidth, and the calculation method is as follows: γ' d is set to 1, n d is the number of diseases and is set to 591.
[0057] (4) Calculate the Gaussian kernel similarity of miRNAs, which is calculated as follows: MG = exp(-γ m ||A(m i ) - A(m j )|| 2 ), where A(m i ) corresponds to the i-th column vector in the association matrix A. The parameter γ is used to adjust the kernel bandwidth, and the calculation method is: γ′ m is set to 1, n m is the number of miRNAs and is set to 853.
[0058] (5) Integrate miRNA functional similarity MF, miRNA sequence similarity MS, disease semantic similarity DS, target-based disease similarity DT, the calculated miRNA Gaussian kernel similarity MG and disease Gaussian kernel similarity DG. Respectively obtain the miRNA-side dataset, which is calculated as follows: C m = {MF, MS, MG}, and the disease-side dataset, which is calculated as follows: C d = {DS, DT, DG}.
[0059] The present invention constructs a biological entity graph part. First, the biological entities collected by the present invention are input into a heterogeneous graph neural network (HeteroData is a built-in data structure for representing heterogeneous graphs) as nodes. Second, the present invention constructs a node mapping according to different node types, and each node is mapped to a unified number, thereby constructing an edge index in HeteroData. The present invention constructs node type labels to represent the types of each node in HeteroData, and the feature dimension of each node type is 901.
[0060] Step 3: Use a hypergraph convolutional neural network for feature extraction
[0061] (1) First, normalize the obtained miRNA-side adjacency matrix C m The calculation is as follows: E m represents the degree matrix of C m The calculation definition of the initial hypergraph incidence matrix on the miRNA side is as follows: represents the node embedding of the l-th layer, and Y(:,j) represents all elements of the j-th column of the incidence matrix. The calculation of the k-th order reachability matrix of the disease is as follows: where power(·,k) is the k-th power of the matrix. The calculation of the k-th order hypergraph incidence matrix of miRNAs is as follows: The miRNA incidence matrix H after fusing the initial hypergraph incidence matrix and the second-order hypergraph incidence matrix m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the calculation of the miRNA node is as follows: represents the node at the l th -th hypergraph convolutional neural network layer. represents the learnable weight matrix, and σ(·) is the non-linear activation function. Finally, the attribute features on the miRNA side are obtained as where the number of layers of the hypergraph convolutional neural network is set to 2.
[0062] (2) Normalize the obtained disease-side adjacency matrix C d The calculation is as follows: E m represents the degree matrix of C m The calculation definition of the initial incidence matrix on the disease side is: represents the node embedding of the l-th layer, and Y(i,:) represents all elements of the i-th row of the incidence matrix. The calculation of the k-th order reachability matrix of miRNAs is: Among them, power(·, k) represents the k-th power of a matrix. The incidence matrix of the k-th order hypergraph of a disease is as follows: The hypergraph based on the primary incidence matrix and the hypergraph based on the second-order reachability rule are fused, and the incidence matrix H of the obtained disease hypergraph m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the node calculation representation of the disease is as follows: represents the embedding of the node in the l th -th hypergraph convolutional neural network layer. represents the learnable weight matrix, σ(·) represents the non-linear activation function, and finally the attribute features on the disease side are Similarly, the number of layers of the hypergraph convolutional neural network is defined as 2.
[0063] Step 4, use the graph transformer to extract features from the constructed biological entity graph.
[0064] (1) In the heterogeneous mutual attention part: The present invention applies the linear projection technology to transform the target node t into the i-th query vector, and the calculation is as follows: Q i (t) represents the i-th query vector to which the target node t is projected. The source node s is transformed into the i-th K-Lineari vector, and the calculation is as follows: K i (s) represents the i-th Key vector to which the source node is projected. The multi-head attention mechanism is used for weight allocation, and the calculation is as follows: ATT-head i represents the i-th attention head, e represents the relationship between the source node s and the target node t, and μ(·) represents the general meaning of each relationship triple. The operation of the Attention(·) vector connects h attention heads, uses the softmax function to generate the probability distribution for each specific target node t, and this distribution forms the distribution using the attention vectors accumulated from the adjacent nodes N(t).
[0065] (2) Message passing part: This module introduces the meta-relationship of the edge in the message passing mechanism to alleviate the distribution differences of different types of nodes and edges, and transmits the information from the source node to the target node. Specifically, for a pair of nodes e = (s, t), its single message passing head is calculated by the following formula: The source node s is projected into the i-th message vector. Finally, h message heads are connected and weighted to obtain the final message passing vector represents the weight in the message passing process.
[0066] (3) Target aggregation part: Aggregate the above heterogeneous cross-attention module and heterogeneous message passing module from the source node to the target node, and the calculation is as follows: Input the result into the Relu activation layer, and the activation function part is expressed as:
[0067] Finally, obtain the biological entity features. The biological features on the miRNA side are The biological features on the disease side are
[0068] Step 5, calculate the final association prediction score
[0069] (1) Combine the attribute features extracted by the hypergraph convolutional neural network and the biological features extracted by the graph transformer, and the calculation is as follows:
[0070] (2) Concatenate the embedding of each miRNA and the embedding of each disease as follows: X ij = concatenating(V m , V d ), X ij represents the feature association pair between miRNA m i and disease d j . Then use a multi-layer perceptron to score the final miRNA-disease association prediction as follows: X l = Relu(X l-1 W l + b l ), represents the number of hidden layers, X l represents the output of the l hidden layer, W l represents the learning parameter matrix of the l layer, b l represents the bias term of the l layer. Finally, train the model by minimizing the binary cross-entropy loss function error as follows:
[0071]
[0072] represents the final prediction score, (i, j) represents the association pair between disease d j and miRNA m i , y + represents the positive sample set, y -Denoted as the negative sample set, N represents the number of all miRNA-disease association pairs, which is set to 12446. Among them, the Adam optimizer is used to minimize the loss function, and the optimal parameters are selected, including the number of hidden layers L, the number of epochs e, the learning rate lr, and the weight decay wd. In the present invention, the optimal hyperparameters are set as lr = 10^-5, e = 100, wd = 10^-3, T = 4, and L = 4. The number of layers of the multi-layer perceptron l is set to 2. The dimension of the first layer is set to 256, the dimension of the second layer is set to 128, and the output data is the predicted association probability matrix
[0073] Step 6, Experimental description
[0074] The effect of the present invention is further illustrated by experiments, using five-fold cross-validation and leave-one-out cross-validation experiments.
[0075] (1) The experiment uses python simulation processing. In the five-fold cross-validation, the maximum number of iterations is 100 times.
[0076] (2) Analyze the prediction performance of the present invention. And it is compared with four existing prediction methods (Note the multi-view similarity network and hypergraph learning for miRNA-disease association recognition AMHMDA, predicting miRNA-disease associations through node-level graph attention autoencoders PGCNMDA, predicting the correlation between miRNA and disease based on lncRNA-miRNA interactions and graph convolutional networks SGAEMDA). Figure 2 is the ROC performance curve of the five-fold cross-validation for the method of the present invention Figure 3 is the ROC performance curve of the leave-one-out cross-validation for the method of the present invention. Table 1 shows the comparison of the evaluation indexes between the method of the present invention and other methods after the five-fold cross-validation experiment.
[0077] Table 1 Comparison of the five-fold cross-validation between the method of the present invention and four other methods
[0078]
[0079] (3) Case analysis of the prediction results of the present invention is carried out through three common cancers (colon cancer, breast cancer, gastric cancer). The prediction scores are sorted in descending order, and the top 10 miRNAs are selected for ranking. Verification is carried out through two databases (HMDD 3.2 and dbDEMC) and published literature. The specific results can be seen in Table 2, Table 3 and Table 4. It can be seen that the present invention can effectively predict the confirmed miRNA-disease associations and can predict potential miRNA-disease associations.
[0080] Table 2 The top 10 miRNAs predicted to be related to colon cancer
[0081]
[0082] Table 3 The top 10 miRNAs predicted to be related to breast cancer
[0083]
[0084] Table 4 The top 10 miRNAs predicted to be related to gastric cancer
[0085]
[0086]
[0087] The above simulation results show that the present invention can effectively predict miRNA-disease associations and can significantly improve the prediction accuracy compared with existing methods.
Claims
1. An association prediction method based on a hypergraph convolutional neural network and a graph transformer, characterized in that The method includes the following steps: (1) Prepare the miRNA-disease association dataset and obtain the miRNA-disease association information; prepare the protein information dataset, LncRNA information dataset, drug node information dataset, cirRNA information dataset, and microorganism information dataset, and obtain the biological entity information; (2) Download the miRNA functional similarity matrix MF, miRNA sequence similarity MS, and target-based disease similarity DT; (3) Calculate the semantic similarity DS of diseases, and calculate the GIP kernel similarity DG of diseases and the GIP kernel similarity MG of miRNAs respectively according to the constructed miRNA-disease association matrix A; (4) Perform matrix splicing to obtain the miRNA association dataset: C m = {MF, MS, MG}, and the obtained disease association dataset is C d = {DS, DT, DG}; for the biological entity graph part, construct a HeteroData graph; (5) Aggregate and extract features from associated data using a hypergraph convolutional neural network and Extract features from biological entity data using a graph transformer to obtain biological entity features and (6) Combine the attribute features after extraction using the hypergraph convolutional neural network and weight assignment by the attention mechanism with the biological features extracted from the biological entity graph: Indicates the concatenation of the feature matrices; (7) Use a multi-layer perceptron to process the data and obtain the final prediction score.
2. The correlation prediction method based on the hypergraph convolutional neural network and the graph transformer according to claim 1, wherein: In step (5), use the hypergraph convolutional neural network to extract various similarity information and update the embedding of each node. The specific steps for extracting miRNA-side attribute features are as follows: 5i) Process the miRNA adjacency matrix C m to obtain the normalized matrix L m , and construct the initial incidence matrix of the hypergraph. The calculation definition of the initial incidence matrix on the miRNA side is as follows: L m (i,:) represents all columns of the i-th row of the normalized matrix L m and represents the node embedding of the l-th layer, and Y(:,j) represents all elements of the j-th column of the incidence matrix; 5ii) To explore the high-order connectivity in the miRNA-disease bipartite graph, if disease i is a k-order reachable neighbor of disease j, there exists a path between i and j, and the number of diseases within this path is less than k. The k-order reachable matrix of diseases is as follows: represents the node embedding of the l-th layer on the disease side, represents the transpose matrix of the node embedding matrix of the l-th layer on the disease side, where power(·, k) is the k-th power of the matrix; using the k-order reachable adjacency matrix of diseases construct a hypergraph reflecting the k-order neighborhood relationship of miRNAs. The incidence matrix of the k-order hypergraph of miRNAs is calculated as follows: represents the incidence matrix of the l-th layer, the (k - 1)-order reachable adjacency matrix of diseases; 5iii) Perform a connection operation on the miRNA, fuse the hypergraph based on the primary association matrix and the hypergraph based on the second-order reachability rule, and the association matrix H of the obtained miRNA hypergraph m is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the miRNA node is obtained after l + 1 iterations The calculation is expressed as follows: represents the node at the l th th embedding in the hypergraph neural network layer, represents the learnable weight matrix, σ(·) is the non-linear activation function, and the final attribute feature on the miRNA side is 3. The correlation prediction method based on a hypergraph convolutional neural network and a graph transformer according to claim 1, wherein: In step (5), use the hypergraph convolutional neural network to extract various similarity information and update the embedding of each node. The specific steps for extracting miRNA-side attribute features are as follows: 5i) For the disease adjacency matrix C d Normalize it to obtain L d , and construct the initial incidence matrix of the hypergraph. The calculation definition of the initial incidence matrix on the disease side is as follows: L d (:,j) represents all rows of the j-th column of the normalized matrix L d The j-th column, represents the node embedding of the l-th layer, and Y(i,:) represents all elements of the i-th row of the incidence matrix; 5ii) To explore the high-order connectivity in the miRNA-disease bipartite graph, define the k-order reachable neighbors of miRNAs and diseases in the miRNA-disease bipartite graph. If miRNA (i) is the k-order reachable neighbor of miRNA (j), then there exists a path between i and j, and the number of miRNAs within this path is less than k. The k-order reachable matrix of miRNAs is as follows: where power(·, k) is the k-th power of the matrix, denotes the initial association matrix on the miRNA side, denotes the transpose of the initial association matrix on the miRNA side; using this k - 1 order reachable adjacency matrix of miRNAs construct a disease hypergraph matrix reflecting the k-order neighborhood relationship of diseases The k-order hypergraph association matrix of diseases is: 5iii) Perform a connection operation on the disease-side matrix, fuse the hypergraph based on the primary association matrix and the hypergraph based on the second-order reachability rule, and obtain the disease hypergraph association matrix H d is defined as: According to the inter-layer propagation rule of the hypergraph convolutional neural network, the node of the disease undergoes l+1 iterative calculations to obtain the node embedding The calculation is expressed as follows: represents the node at the l th th hypergraph convolutional neural network layer, represents the learnable weight matrix, L d represents the normalized disease-side association matrix, σ(·) represents the non-linear activation function, and finally the attribute features on the disease side are obtained as 4. The correlation prediction method based on the hypergraph convolutional neural network and the graph transformer according to claim 1, characterized in that: In step (5), in the part of using the graph transformer to extract biological entity features, it is divided into three modules: heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation. The specific calculation steps are as follows: 5i) In the heterogeneous mutual attention part: A unique set of projection weights W is assigned to each pair of meta-relations. The linear projection technique is applied to represent node t as a Q-Lineari vector. In order to convert the target node t into the i-th query vector, this step performs calculations, and the calculation formula is as follows: Q i (t) represents that the target node t is projected onto the i-th query vector, represents the linear projection process, H (l-1) [t] represents the iterative result of the target node t at the (l - 1)-th layer. Similarly, the source node is linearly projected, and node s is represented as a K-Lineari vector. In order to convert the source node s into the i-th vector, this invention performs calculations, and the calculation formula is as follows: K-linear(·) represents the linear projection process, K i (s) represents the i-th Key vector onto which the source node is projected, H (l-1) [s] represents the iterative result of the source node s at the (l - 1)-th layer; Weight allocation is performed using the multi-head attention mechanism, and the calculation method is as follows: ATT-head i represents the i-th attention head, e represents the relationship between the source node s and the target node t, and μ(·) represents the general meaning between each pair of triple relationships. K i (s) represents the linear projection vector of the source node s. represents the weight, Q i (t) T represents the query vector of the target node t. represents the scaling factor, τ(s), τ(t) represents the dot product multiplication, ATT-head i represents the i-th attention head. As a way to adaptively scale attention, e represents the relationship between the source node s and the target node t, and μ(·) represents the general meaning between each pair of triple relationships. The operation of the Attention(·) vector includes connecting h ATT-head connections to obtain the attention vector for each node pair. The softmax function is used in the solution process. Normalize ATT-head i (s, e, t) to generate a probability distribution specific to each target node t, which is formed using the attention vectors accumulated from adjacent nodes N(t); 5ii) Message passing part: This module introduces the meta-relation of edges in the message passing mechanism to alleviate the distribution differences of different types of nodes and edges, and transmits information from the source node to the target node. For each pair of nodes e = (s, t), its multi-head message is calculated by the following formula: Represents the projection of the source node s to the i-th message vector, H (l-1) [s] represents the result of the iterative process, Represents the weight in the message transmission process, Message HGT (s,e,t) means connecting all h message headers to get complete information. The calculation process is as follows: 5iii) Target Aggregation Part: In this module, first, the above heterogeneous mutual attention module Attention HGT (s, e, t) and the heterogeneous message passing module Message HGT (s, e, t) are aggregated from the source node to the target node through as follows: Then, the result obtained above is input into the Relu activation layer, where σ(·) represents the activation function, and a linear transformation is performed on the output and concatenated with the residual H (l-1) [t]. The activation function part is expressed as: The miRNA-side biometric matrix obtained through i iterations of calculation by the activation function is The disease-side biometric matrix is 5. The association prediction method based on the hypergraph convolutional neural network and graph transformer according to claim 1, characterized in that: The steps of step (7) are as follows: Embed each miRNA and embed each disease and concatenate them as follows: X ij = concatenating(V m , V d ), where X ij represents the feature association pair between miRNA m i and disease d j . Then use a multi-layer perceptron to score the final miRNA-disease association prediction as follows: X l = Relu(X l-1 W l + b l ), where l ∈ [1, 2] represents the number of hidden layers, X l represents the output of the l-th hidden layer, W l represents the learning parameter matrix of the l-th layer, b l represents the bias term of the l-th layer, sigmoid is the activation function, X 2 represents the output of the second hidden layer, W 2 represents the learning parameter matrix of the second layer, b 3 represents the bias term of the third layer. Finally, train the model by minimizing the binary cross-entropy loss function error as follows: Denote the final prediction score, and (i, j) denote disease d j and miRNA m i The association pair between them, y + Denote as the positive sample set, y - Denote as the negative sample set, N denotes the number of all miRNA-disease association pairs, set to 12446. Among them, the Adam optimizer is used to minimize the loss function, and the optimal parameters are selected, including the number of hidden layers L, the number of epochs e, the learning rate lr, and the weight decay wd. The output data is the predicted association probability matrix