A miRNA-disease association prediction method based on a graph neural network

Abstract

The application provides a miRNA-disease association prediction method based on a graph neural network, and relates to the technical field of the graph neural network.The method comprises the following steps: acquiring miRNA sequences, disease description texts, miRNA family information, disease parent-child relationships and miRNA-disease association relationship data, and constructing a heterogeneous graph after data alignment; based on the miRNA sequences and the disease description texts, using a pre-trained language model to acquire miRNA representation and disease representation as node features, and updating the heterogeneous graph; constructing a heterogeneous graph MDA association prediction model, inputting the updated heterogeneous graph and a to-be-predicted miRNA-disease pair, and obtaining an association prediction score of the to-be-predicted miRNA-disease pair.The method can fully represent biological entities on entity representation, can effectively aggregate information of different types of nodes on model construction, and can perform message passing of multiple depths on different types of edges, so that the integration of heterogeneous information of nodes and edges is realized, and the performance of miRNA-disease association prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology, and in particular to a miRNA-disease association prediction method based on graph neural network. Background Art

[0002] MicroRNAs (miRNAs) are a class of endogenous noncoding RNAs (ribonucleic acids) that bind to protein-coding mRNAs through partial complementary pairing, directing post-transcriptional repression and blocking mRNA-to-protein translation. This leads to regulation of diverse biological processes, including cell differentiation, proliferation, and apoptosis, and influences disease development and progression. For example, miR-203 expression inhibits primary ovarian tumor growth and the development of metastatic tumors in multiple peritoneal organs, including the liver and spleen. Silencing miR-93-5p prevents granulosa cell dysfunction, making miR-93-5p a novel target for improving granulosa cell function in patients with polycystic ovary syndrome. In patients with certain infectious diseases, mRNA-16-2-3p is significantly upregulated, while miR-21-5p is significantly downregulated. Therefore, identifying potential miRNA-disease associations (MDAs) is crucial, providing new insights into the biological functions of miRNAs and the pathogenesis of complex diseases, and holds important implications for clinical prevention, diagnosis, and treatment.

[0003] Techniques such as reverse transcription polymerase chain reaction (RT-PCR) and microarray analysis can effectively validate MDAs, but these methods are costly and time-consuming. During this process, a large amount of experimentally validated miRNA-disease association data has been integrated into databases such as the Human MicroRNA Disease Database (HMDD), the RNADisease (RNA and Disease Database), and the dbDEMC (Database of Differentially Expressed MiRNAs in Cancers). This authoritative data has facilitated the development of data-driven computational methods for MDA prediction. Computational methods aim to use models or algorithms to extract features of miRNAs and drugs and predict whether there is an association between the two. Computational methods can predict potential MDAs on a large scale, allowing researchers to use relatively accurate predictions to guide experimental validation of miRNA-disease associations and reduce the cost of biological experiments.

[0004] Currently, the calculation methods in the existing technology mainly have the following technical problems:

[0005] 1. Insufficient representation of biological entities. Most existing studies use similarity features and integration, or indirectly obtain features from other biological entities to represent biological entities. However, they still cannot well represent miRNAs and disease entities, which have been less studied.

[0006] 2. The model fails to effectively aggregate information from different types of nodes and cannot perform necessary message passing on specific types of edges. The prediction effect still has room for improvement. Summary of the Invention

[0007] To solve the above technical problems, the present invention aims to provide a computational method for predicting miRNA-disease associations that is based on sufficient representation of biological entities, can effectively aggregate information of different types of nodes, and can perform necessary message passing on specific edges.

[0008] Specifically, the present invention provides a miRNA-disease association prediction method based on graph neural network, and the technical solution includes the following steps:

[0009] Step S1: Obtain miRNA sequences, disease description text, miRNA family information, disease parent-child relationship, and miRNA-disease association data, align the data, and construct a heterogeneous graph;

[0010] The heterogeneous graph includes miRNA nodes, disease nodes, miRNA-miRNA edges, disease-disease edges and miRNA-disease edges; wherein a pair of miRNA nodes and disease nodes connected by a miRNA-disease edge is a miRNA-disease pair;

[0011] Step S2: Input the miRNA sequence and disease description text into the pre-trained language model respectively, obtain the miRNA representation and disease representation as node features, and update the heterogeneous graph;

[0012] Step S3: Build and train a heterogeneous graph MDA prediction model based on graph neural networks with multi-level attention and multi-depth message passing;

[0013] Step S4: Input the updated heterogeneous graph and the miRNA-disease pair to be predicted into the heterogeneous graph MDA prediction model to obtain the association prediction score of the miRNA-disease pair to be predicted.

[0014] Preferably, step S1 specifically includes:

[0015] Step S11: obtaining miRNA family information and two mature sequences of miRNA from the miRBase database, and constructing miRNA-miRNA association relationships between miRNAs based on the miRNA family information;

[0016] Step S12: Obtain disease description text and parent-child relationship information between diseases from the MeSH database; wherein the disease description text includes the disease name and annotation information;

[0017] Step S13: Obtain miRNA-disease association data from the HMDD database and the RNADisease database;

[0018] Step S14: Based on the miRNA family information, the two mature sequences of miRNA, the disease description text and the parent-child relationship information between diseases, the miRNA-disease data is aligned to remove the miRNA-disease relationship with missing entity features;

[0019] Step S15: Construct a heterogeneous graph based on miRNA family information, miRNA-disease association data, and disease-disease parent-child relationships.

[0020] Preferably, step S2 specifically includes:

[0021] Step S21: For each miRNA,

[0022] Step S21: For each miRNA, its two mature sequences and Input RNA-FM model, corresponding to the output of two miRNA representations, namely and ;

[0023] For each disease, the disease name and annotation information are input into the PubMedBERT model, and the disease name representation is output. and annotation information representation ;

[0024] Step S22: miRNA expression and 、Disease name and annotation information representation As node features of heterogeneous graphs.

[0025] Preferably, step S3 specifically includes:

[0026] Step S31: embed the node features of the heterogeneous graph through MLP to obtain the corresponding embedded features.

[0027] ;

[0028] ;

[0029] Where, represents the corresponding feature embedding for miRNA, Indicates two mature miRNA sequences, Represent the corresponding feature embedding for the disease;

[0030] Step S32: including steps S321 to S325;

[0031] Step S321: Through the learnable matrix The node Embedding features Projected into a unified feature space,

[0032] ;

[0033] Where, For nodes Feature Embedding The projection result of ; among them, the node Including miRNA nodes and disease nodes;

[0034] Step S322: Normalize the node-level attention using the softmax function to obtain the dynamic attention weight value of each neighbor node; where node-level attention refers to the contribution value of the node to the target node through a specific edge;

[0035] Step S323: Aggregate the dynamic attention weight values ​​and the feature embeddings of the corresponding neighbor nodes to obtain the node-level attention embeddings of the current node corresponding to different types of edges;

[0036] Step S324: Perform a nonlinear transformation on each set of node-level attention embeddings, and use the semantic-level attention mechanism to determine the importance of each set of node-level attention embeddings to obtain the importance value of the corresponding type of edge to the node; normalize the importance value using the softmax function to obtain the attention weight value of the edge;

[0037] Step S325: Fuse each set of node-level attention embeddings and corresponding attention weight values ​​to obtain the final embedding of the current node;

[0038] Step S33: For the miRNA node and disease node of the miRNA-disease pair to be predicted, the corresponding multi-level attention is finally embedded and input into the classifier to obtain the association prediction score.

[0039] Preferably, step S323 is specifically as follows:

[0040] use Head attention mechanism, calculation Secondary Node The node-level attention embedding of Embedded features are concatenated to obtain nodes exist Semantic embedding under type edge ,

[0041] ;

[0042] wherein, is a node a group ordinal of the node-level attention embedding, , is a node a group number of the node-level attention embedding, is a head ordinal in the head attention mechanism, , is a node a set of neighbor nodes on type edges, is an element of the set , is a dynamic attention weight value of the neighbor node, is a node a projection result of the feature embedding.

[0043] Preferably, the step S324 is specifically:

[0044] ;

[0045] wherein, denotes a node type, denotes an importance value of a pair of type edges for a type node, denotes a transpose, is a weight matrix, is a bias vector;

[0046] normalizing the importance value using a softmax function , obtaining an attention weight value of the pair of type edges ;

[0047] The step S325 is specifically:

[0048] ;

[0049] wherein, is a multi-level attention embedding of the type node, is a semantic embedding of the type node under the type edge.

[0050] Further, the method further comprises:

[0051] For each node,

[0052] In the first L-1 layers of the heterogeneous graph MDA prediction model, according to step S32, the node obtains the information of all nodes in the heterogeneous graph and aggregates them;

[0053] In the L layer of the heterogeneous graph MDA prediction model, according to step S32, the node obtains the information of the heterogeneous nodes in the heterogeneous graph and aggregates them to obtain the final embedding of the miRNA and disease nodes;

[0054] Among them, the heterogeneous graph MDA prediction model is an L-layer structure.

[0055] Preferably, step S33 is specifically as follows:

[0056] Perform vector concatenation on the final embeddings of miRNA nodes and disease nodes,

[0057] ;

[0058] Where, is the prediction score, is the final embedding of miRNA node, is the final embedding of the disease node.

[0059] Furthermore, after step S33, the following steps are further included:

[0060] The binary cross entropy loss function is used to calculate the score loss Loss.

[0061] ;

[0062] Where, is a supervised label. If the input miRNA-disease pair belongs to miRNA-disease association data, then ,otherwise ;

[0063] The obtained score loss Loss is used for gradient backpropagation to optimize the heterogeneous graph MDA prediction model.

[0064] Furthermore, the training method of the heterogeneous graph MDA prediction model includes:

[0065] In the miRNA-disease dataset, verified miRNA-disease pairs were selected as positive samples. Among all miRNA-disease pairs, the same number of unverified miRNA-disease pairs were randomly selected as negative samples. Among them, the unverified miRNA-disease pairs did not belong to the miRNA-disease dataset.

[0066] The positive and negative samples are randomly shuffled and divided into K parts, and K-fold cross validation is performed;

[0067] The model parameters with the highest score in K-fold cross validation are selected as the final model parameters;

[0068] Among them, the scoring indicators include the area under the operating characteristic curve, the area under the precision-recall curve, accuracy and F1 score.

[0069] The technical solution provided by the present invention can fully represent biological entities, achieve better representation of miRNA and disease entities that have been less studied, and achieve highly accurate prediction of diseases that lack miRNA-disease associations. In addition, the present invention adopts a multi-level attention mechanism to aggregate the features of the neighbors of each node in the constructed heterogeneous graph, and performs multi-depth message passing on different types of edges, which can realize the integration of heterogeneous information of nodes and edges, and help improve the miRNA-disease association prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Schematic diagram of the overall process framework of the miRNA-disease association prediction method of the present invention.

[0071] Figure 2 Schematic diagram of the structure of the multi-level attention mechanism in the present invention. DETAILED DESCRIPTION

[0072] Hereinafter, the technical solution provided by the present invention will be further elaborated in conjunction with the accompanying drawings.

[0073] This paper processes raw miRNA and disease data, combines it with a pre-trained language model to represent the raw features of miRNAs and diseases, and constructs a heterogeneous graph with rich node information. Applying multi-level attention and multi-depth message passing mechanisms, this paper proposes a graph neural network-based miRNA-disease association prediction method, helping researchers predict potential MDAs, particularly for diseases without known MDAs. The technical solution includes: data acquisition and processing, miRNA and disease representation, model construction, and model training.

[0074] The specific steps are as follows:

[0075] Step 1: Data acquisition and processing

[0076] MiRNA sequences, disease description text (including disease name and annotation information), miRNA family information, disease parent-child relationship and miRNA-disease association data were obtained from public databases, and heterogeneous graphs were constructed after data alignment.

[0077] 1. For miRNA, the standard name, two mature sequences and family information of miRNA were obtained from miRBase, a database for naming miRNA. The mature sequence was used as the original feature of miRNA, and the family information was used to construct the miRNA-miRNA association relationship.

[0078] 2. For diseases, the disease name and annotation information are obtained from the MeSH database as the original features of the disease, and the parent-child relationship between diseases is obtained to construct the disease-disease association relationship.

[0079] 3. Obtain data on miRNA-disease associations from the HMDD and RNADisease databases, unify the names of miRNAs and diseases involved in the two datasets, merge the datasets and remove duplicates based on the entity names, and use the processed data as both edges in the subsequent heterogeneous graph and as prediction targets.

[0080] 4. Based on the miRNA and disease data obtained in 1 and 2, data alignment was performed on the three miRNA-disease association datasets, and MDA with missing entity features was removed.

[0081] 5. Use the miRNA-disease association data, as well as miRNA family association and disease parent-child relationship data from the dataset to construct a heterogeneous graph. In this heterogeneous graph, nodes include miRNAs and diseases, and edges include connections between nodes of the same type (miRNA-miRNA and disease-disease) and between nodes of different types (miRNA-disease). All edges are undirected.

[0082] Step 2: miRNA and disease expression

[0083] After data acquisition and processing, the original features of miRNA and disease are further represented to construct a heterogeneous graph with rich node semantics, which is used as the input of subsequent models together with the miRNA-disease pairs to be predicted.

[0084] 1. Input the two mature miRNA sequences obtained in step 1 into the pre-trained RNA-FM model to obtain the miRNA representations corresponding to the two mature sequences. and ;

[0085] 2. Input the disease name and annotation information obtained in step 1 into the pre-trained PubMedBERT model, and the obtained disease name representation and annotation information representation .

[0086] 3. Use the miRNA representation (representation of two miRNA sequences) and disease representation (disease name representation and annotation information representation) output by the pre-trained model as the node features of the heterogeneous graph obtained in step 1 to obtain a heterogeneous graph with rich node information.

[0087] Step 3: Model construction

[0088] like Figure 1 As shown in the figure, a heterogeneous graph MDA prediction model with multi-level attention and multi-depth message passing is constructed. The model input is the heterogeneous graph with rich node semantic information obtained in step 2, and the output is the prediction score of the miRNA-disease pair.

[0089] 1. Node feature encoding: The representations of miRNA and disease nodes are embedded into appropriate dimensions through MLP (Multilayer Perceptron). The formula is as follows:

[0090] ;

[0091] ;

[0092] Where, is the feature embedding corresponding to the mature sequence of miRNA, Feature embedding for diseases.

[0093] 2. Heterogeneous Graph Neural Network: Design and implement heterogeneous graph neural network structures with multi-level attention and multi-depth message passing.

[0094] (1) Multi-level attention mechanism: such as Figure 2 As shown in Figure 3, the multi-level attention mechanism consists of node-level attention and semantic-level attention.

[0095] ①For Type of node, first use the learnable matrix Set miRNA and disease nodes (set as nodes ) characteristics Projected into a unified feature space, as follows:

[0096] ;

[0097] ②Node-level attention mechanism: For nodes , node-level attention Representation node Through a specific type of edge contribution.

[0098] ③ Normalize the node-level attention through the softmax function to obtain the dynamic attention weight value of the neighboring nodes .

[0099] ④ Obtain the node by aggregating the attention weight value and the features of the corresponding neighbor nodes In order to enhance the expressiveness of the model and enable it to adaptively learn different features, a multi-head attention mechanism is adopted. Secondary Node The node-level attention embedding of The embedding splicing obtained at the beginning is used to get the node exist Semantic embedding under type edge ,

[0100] ;

[0101] Where, For nodes The group ordinal number of node-level attention embeddings, , For nodes The number of groups of node-level attention embeddings, for The head order in the head attention mechanism, , For nodes exist The set of neighbor nodes on the type edge, For collection Elements, is the dynamic attention weight value of the neighbor node, For nodes Projection result of feature embedding.

[0102] After embedding using the node-level attention mechanism on all edges, two sets of embeddings can be obtained for the miRNA node: and ; For disease nodes, we can obtain and The semantically specific node embeddings obtained from different types of edges contain the features of different types of neighboring nodes.

[0103] ⑤Semantic-level attention mechanism: Use the semantic-level attention mechanism to balance the importance of different edges and further aggregate the node embeddings of these specific semantics. Nodes of type Node embeddings for group-specific semantics are first applied to these embeddings by applying a nonlinear transformation and then using a semantic-level attention vector To measure the importance of embedding. The importance of node embedding for each group of specific semantics is averaged to obtain the type The edge of type The importance value of the node is expressed as , the calculation process is as follows:

[0104] ;

[0105] Where, For nodes The group ordinal number of node-level attention embeddings, , represents transpose, represents the weight matrix, represents the bias vector.

[0106] ⑥Use the softmax function to normalize the above importance values ​​to obtain the attention weight value of the edge .

[0107] ⑦Similar to the calculation of node embedding of specific semantics, the learned attention weight value is fused to obtain the embedding of specific semantics of type Multi-level attention embedding of nodes .

[0108] The calculation process is as follows:

[0109] ;

[0110] Where, for Type node in Semantic embedding under type edge.

[0111] ⑧ Perform residual connection on the multi-level attention embedding of miRNA and disease with the input features to obtain the final embedding of miRNA and disease, which are and .

[0112] (2) Multi-depth message passing mechanism: At different embedding stages, different depths of message passing on different types of edges enable the model to focus on specific information. Therefore, a multi-depth message passing mechanism is introduced. For each node, according to the method of steps ① to ⑧, the model The first layer obtains and aggregates information from both homogeneous and heterogeneous nodes; in the last layer, the model only obtains and aggregates information from heterogeneous nodes, thereby enhancing the learning of heterogeneous information and obtaining the final embedding of miRNA and disease nodes.

[0113] Each layer of the model is structured as follows Figure 2 shown.

[0114] 3. Classifier: The final embedding vectors of the calculated candidate miRNA-disease pairs are concatenated and passed through the classifier to output the prediction score. The loss is calculated based on the label and used for gradient backpropagation.

[0115] (1) Process it using two linear layers and use the sigmoid activation function to obtain the prediction score as follows:

[0116] ;

[0117] ;

[0118] Where, is the prediction score, is the multi-level attention embedding of miRNA nodes, Multi-level attention embedding for disease nodes.

[0119] (2) Calculate the binary cross entropy loss function Loss for gradient backpropagation to optimize the model:

[0120] ;

[0121] in, Represents the supervision label. If the associated record of the input miRNA-disease pair is included in the positive example of the dataset, ,otherwise, .

[0122] Step 4: Model training

[0123] Select appropriate metrics for quantitative model evaluation. Collect negative samples of miRNA-disease associations and split the data into training, validation, and test sets. Use five-fold cross-validation during training to improve model generalization. Test the selected model on the test set to obtain the final prediction results.

[0124] 1. Indicator selection:

[0125] The area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), accuracy (Acc), and F1 score (F1) were selected as quantitative evaluation metrics. The experimental results were categorized into four categories based on the correctness of the judgments: TP, TN, FP, and FN. The resulting evaluation metrics are shown in Table 1.

[0126] Among them, with TPR and FPR as the vertical and horizontal axes, the area enclosed by the curve and the coordinate axes is called AUC; with recall as the horizontal axis and precision as the vertical axis, the area enclosed by the curve and the coordinate axes is called AUPR.

[0127] Table 1 Evaluation indicators of confusion matrix

[0128] .

[0129] 2. Negative sample sampling:

[0130] The miRNA-disease associations that were verified in the dataset were considered positive samples, the miRNA-disease associations that did not appear in the dataset were considered unrelated, and the same number of unverified miRNA-disease pairs randomly selected from them were considered negative samples.

[0131] 3. Randomly shuffle the positive and negative samples and select the first 10% as the test set. The remaining 90% of the positive and negative samples are equally divided into five groups for five-fold cross-validation. For the k-th fold, use the k%5, (k+1)%5, (k+2)%5, and (k+3)%5 positive and negative samples as the training set, and the (k+4)%5 positive and negative samples as the validation set. To prevent data leakage, the MDA in the positive samples used for validation and testing must not appear on the edges of the heterogeneous graph in the training and testing folds.

[0132] Among them, the last 90% positive and negative samples can also be divided into K parts respectively and used for K-fold cross-validation, where K is a natural number greater than 1.

[0133] 4. Select the model parameters with the highest score in the five-fold cross-validation as the final model parameters, and test them on the test set to obtain the final prediction results of the model.

[0134] Among them, the last 90% positive and negative samples can also be divided into K parts respectively and used for K-fold cross-validation, where K is a natural number greater than 1.

[0135] The technical solution provided by the present invention can fully represent biological entities, achieve better representation of miRNA and disease entities that have been less studied, and achieve highly accurate prediction of diseases that lack miRNA-disease associations. In addition, the present invention adopts a multi-level attention mechanism to aggregate the features of the neighbors of each node in the constructed heterogeneous graph, and performs multi-depth message passing on different types of edges, which can realize the integration of heterogeneous information of nodes and edges, and help improve the miRNA-disease association prediction performance.

[0136] Furthermore, the optimization method and training method of the heterogeneous graph MDA prediction model can further improve the prediction accuracy of the miRNA-disease association prediction method provided by the present invention.

Claims

1. A miRNA-disease association prediction method based on graph neural network, characterized in that: The following steps are involved: Step S1: Obtain miRNA sequences, disease description text, miRNA family information, disease parent-child relationship, and miRNA-disease association data, align the data, and construct a heterogeneous graph; The heterogeneous graph includes miRNA nodes, disease nodes, miRNA-miRNA edges, disease-disease edges and miRNA-disease edges; wherein a pair of miRNA nodes and disease nodes connected by a miRNA-disease edge is a miRNA-disease pair; Step S2: Input the miRNA sequence and disease description text into the pre-trained language model respectively, obtain the miRNA representation and disease representation as node features, and update the heterogeneous graph; Step S3: Build and train a heterogeneous graph MDA prediction model based on graph neural networks with multi-level attention and multi-depth message passing; Step S4: inputting the updated heterogeneous graph and the miRNA-disease pair to be predicted into the heterogeneous graph MDA prediction model to obtain the association prediction score of the miRNA-disease pair to be predicted; Step S1 specifically includes: Step S11: obtaining miRNA family information and two mature sequences of miRNA from the miRBase database, and constructing miRNA-miRNA association relationships between miRNAs based on the miRNA family information; Step S12: Obtain disease description text and parent-child relationship information between diseases from the MeSH database; wherein the disease description text includes the disease name and annotation information; Step S13: Obtain miRNA-disease association data from the HMDD database and the RNADisease database; Step S14: Based on the miRNA family information, the two mature sequences of miRNA, the disease description text and the parent-child relationship information between diseases, the miRNA-disease data is aligned to remove the miRNA-disease relationship with missing entity features; Step S15: constructing a heterogeneous graph based on miRNA family information, miRNA-disease association data, and disease-disease parent-child relationships; Step S2 specifically includes: Step S21: For each miRNA, its two mature sequences and Input RNA-FM model, corresponding to the output of two miRNA representations, namely and ; For each disease, the disease name and annotation information are input into the PubMedBERT model, and the disease name representation is output. and annotation information representation ; Step S22: miRNA expression and 、Disease name and annotation information representation As node features of heterogeneous graphs; Step S3 specifically includes: Step S31: embed the node features of the heterogeneous graph through MLP to obtain the corresponding embedded features. ； ； Where, represents the corresponding feature embedding for miRNA, Represent the corresponding feature embedding for the disease; Step S32: including steps S321 to S325; Step S321: Through the learnable matrix The node Embedding features Projected into a unified feature space, ； Where, For nodes Feature Embedding The projection result of ; among them, the node Including miRNA nodes and disease nodes; Step S322: Normalize the node-level attention using the softmax function to obtain the dynamic attention weight value of each neighbor node; where node-level attention refers to the contribution value of the node to the target node through a specific edge; Step S323: Aggregate the dynamic attention weight values ​​and the feature embeddings of the corresponding neighbor nodes to obtain the node-level attention embeddings of the current node corresponding to different types of edges; Step S324: Perform a nonlinear transformation on each set of node-level attention embeddings, and use the semantic-level attention mechanism to determine the importance of each set of node-level attention embeddings to obtain the importance value of the corresponding type of edge to the node; normalize the importance value using the softmax function to obtain the attention weight value of the edge; Step S325: Fuse each set of node-level attention embeddings and corresponding attention weight values ​​to obtain the final embedding of the current node; Step S33: For the miRNA node and disease node of the miRNA-disease pair to be predicted, the corresponding final embeddings are spliced ​​and input into the classifier to obtain the association prediction score.

2. The miRNA-disease association prediction method based on graph neural network according to claim 1, characterized in that: Step S323 is specifically as follows: use Head attention mechanism, calculation Secondary Node The node-level attention embedding of Embedded features are concatenated to obtain nodes exist Semantic embedding under type edge , ； Where, For nodes The group ordinal number of node-level attention embeddings, , For nodes The number of groups of node-level attention embeddings, for The head order in the head attention mechanism, , For nodes exist The set of neighbor nodes on the type edge, For collection Elements, is the dynamic attention weight value of the neighbor node, For nodes Projection result of feature embedding.

3. The miRNA-disease association prediction method based on graph neural network according to claim 2, characterized in that: Step S324 is specifically as follows: ； Where, Indicates the node type, express Type edge pair The importance value of the type node, is the semantic level attention vector, represents transpose, is the weight matrix, is the deviation vector; Normalize the importance value using the softmax function ,get Attention weight value of type edge ; Step S325 is specifically as follows: ； Where, for Multi-level attention embedding of type nodes, for Type node in Semantic embedding under type edge.

4. A method for predicting miRNA-disease association based on graph neural network according to any one of claims 1 to 3, characterized in that: Also includes: For each node, In the first L-1 layers of the heterogeneous graph MDA prediction model, according to step S32, the node obtains the information of all nodes in the heterogeneous graph and aggregates them; In the L layer of the heterogeneous graph MDA prediction model, according to step S32, the node obtains the information of the heterogeneous nodes in the heterogeneous graph and aggregates them to obtain the final embedding of the miRNA and disease nodes; Among them, the heterogeneous graph MDA prediction model is an L-layer structure.

5. The miRNA-disease association prediction method based on graph neural network according to claim 1, characterized in that: Step S33 specifically includes: Perform vector concatenation on the final embeddings of miRNA nodes and disease nodes, ； Where, is the prediction score, is the final embedding of miRNA node, is the final embedding of the disease node.

6. The miRNA-disease association prediction method based on graph neural network according to claim 1, characterized in that: After step S33, the following steps are also included: The binary cross entropy loss function is used to calculate the score loss Loss. ； Where, is a supervised label. If the input miRNA-disease pair belongs to miRNA-disease association data, then ,otherwise ; The obtained score loss Loss is used for gradient backpropagation to optimize the heterogeneous graph MDA prediction model.

7. The miRNA-disease association prediction method based on graph neural network according to claim 1, characterized in that: The training methods of heterogeneous graph MDA prediction models include: In the miRNA-disease dataset, verified miRNA-disease pairs were selected as positive samples. Among all miRNA-disease pairs, the same number of unverified miRNA-disease pairs were randomly selected as negative samples. Among them, the unverified miRNA-disease pairs did not belong to the miRNA-disease dataset. The positive and negative samples are randomly shuffled and divided into K parts, and K-fold cross validation is performed; The model parameters with the highest score in K-fold cross validation are selected as the final model parameters; Among them, the scoring indicators include the area under the operating characteristic curve, the area under the precision-recall curve, accuracy and F1 score.

Images