Circular RNA-disease association prediction method, device and medium based on weighted graph attention and heterogeneous graph neural network

Abstract

The application discloses a circular RNA-disease association prediction method and device based on a weighted graph attention and a heterogeneous graph neural network and a medium, and the method comprises the following steps: acquiring data to determine similarity and constructing a circular RNA network, an miRNA network and a disease network respectively; based on initial node features and connection weights between nodes in the network, a weighted graph attention mechanism is used to extract node features from the circular RNA network, the miRNA network and the disease network respectively; the circular RNA, the miRNA and the disease are taken as three kinds of heterogeneous nodes respectively, three kinds of heterogeneous edges are established according to the association between different types of nodes, and a multi-source heterogeneous network is constructed; a heterogeneous graph neural network model is used to extract the aggregated features of each node in the multi-source heterogeneous network; and the circular RNA-disease association hidden in the multi-source heterogeneous network is predicted based on the aggregated features. The application effectively predicts the meta-path hidden in multi-source data, improves the accuracy of the model, and has stable prediction performance.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and specifically to a method for predicting circular RNA-disease associations based on weighted graph attention and heterogeneous graph neural networks. Background Technology

[0002] The ENCODE project discovered that 98% of the transcripts in the human genome are non-coding RNAs, and circular RNAs (RRNAs) are a unique class of endogenous non-coding RNAs. RRNAs play important biological functions in many life activities, and their detection in different tissues and diseases has shown a close correlation with the occurrence and development of various human diseases, such as neurological diseases, cardiovascular diseases, and diabetes. For example, Parkinson's disease is characterized by a high concentration of α-synuclein, and CDR1as can be downregulated to inhibit miR-7 of α-synuclein. The more stable structure, tissue-specific expression, and high conservation of RRNAs compared to linear RNA give them great potential as biomarkers for disease diagnosis. For example, RRNAs can be detected in exosomes and body fluids, specifically in peripheral blood and urine. Research on the correlation between RRNAs and diseases can not only provide a better understanding of the fundamental mechanisms of life activities at the RRNA level but also contribute to disease diagnosis, treatment, and prognosis.

[0003] As precision medicine research deepens, increasing evidence suggests that single-source genomic analysis has significant limitations in elucidating the function of circular RNAs and their correlation with diseases. For example, circular RNAs perform their biological functions by regulating multiple biomolecules and are associated with diseases through multi-level molecular regulation. Therefore, it is insufficient to solely mine the unique biological functions of circular RNAs from a single source of data to discover their potential links to diseases. However, the rapid development of sequencing and high-throughput technologies has facilitated the generation of biological data at multiple levels, such as genomics, epigenetics, transcriptomics, and metabolomics. Multi-source data integration analysis is far more comprehensive than the simple summation of single-source data analysis. Multi-source data integration effectively helps data scientists discover common mechanisms and patterns behind heterogeneous data, thereby providing a more complete and accurate analysis of life science laws, deepening our understanding of life processes, and building more accurate quantitative models.

[0004] Traditional biovalidation methods are extremely time-consuming and labor-intensive; therefore, computational methods can efficiently predict the association between circular RNA and disease. Based on the computational models used to predict the association between circular RNA and disease, these methods can be broadly categorized into three types:

[0005] 1) Network propagation algorithm. Fan et al. calculated the Gaussian interaction spectrum similarity of circular RNA (circRNA) and the Gaussian spectrum similarity of disease based on known associations between circRNA and disease. Then, combining circRNA expression similarity and disease phenotypic similarity, they fused circRNA networks, disease networks, and circRNA-disease association networks, and used the KATZ method to calculate the probability values ​​of circRNA-disease association in heterogeneous networks. Lei et al. calculated the Gaussian interaction spectrum similarity between circRNA and disease based on known circRNA-disease associations, calculated disease functional similarity based on disease gene relationships, and calculated circRNA semantic similarity based on circRNA gene associations and gene ontology information. They fused disease similarity networks, circRNA similarity networks, and circRNA-disease association networks, and finally used a graph-based path algorithm to predict the probability values ​​of circRNA-disease association.

[0006] 2) Machine Learning Algorithms. Yan et al. calculated the Gaussian interaction spectrum similarity between circular RNAs and diseases based on known associations between circular RNAs and diseases, integrated disease semantic similarity, and finally used the Kronecker inner product kernel method to predict circular RNA-disease associations. This method uses the Kronecker inner product to linearly connect the kernels of circular RNAs and diseases, but its model's expressive power is insufficient. Wang et al. considered circular RNA expression similarity, circular RNA functional similarity, and disease similarity, and designed a graph-based recommendation algorithm. Zheng et al. proposed an SVM-based classification model, using circular RNA genes, circular RNA sequences, and known circular RNA-disease associations. Li et al. used a matrix filling model to predict candidate associations based on circular RNA sequences, disease semantic information, and the Gaussian interaction spectrum kernels between circular RNAs and diseases. These models rely excessively on manually designed features using relevant professional knowledge, resulting in significant performance differences and instability.

[0007] 3) Neural Network-Based Models. Lu et al. used nonlinear neural networks to replace the linear approximation of matrix factorization methods based on the association patterns in the Gaussian interaction spectrum of circular RNA and diseases, further exploring the model's fitting ability. Wang et al. fused disease semantic similarity, disease Gaussian interaction spectrum kernel similarity, and circular RNA Gaussian interaction spectrum kernel similarity, then used convolutional neural networks to extract latent features, and employed extreme learning as a separator to classify potential circular RNA-disease associations. Lu et al. encoded circular RNA sequences using k-mer, used an unsupervised model to learn the k-mer features of the sequences, learned disease ontology embeddings from the semantic structure tree of diseases, and used BiLSTM to connect features with contextual relevance to complete the circular RNA-disease association prediction. However, these models cannot effectively fuse features from multiple data sources when dealing with various types of data.

[0008] 4) Graph Neural Network-Based Models. Mudiyanselage et al. constructed a heterogeneous network using circular RNA sequence similarity and Gaussian spectrum similarity between circular RNA and diseases. On this network, a graph convolutional network was used to learn node features and predict potential circular RNA-disease associations. Lan et al. integrated multi-source associations among circular RNA, miRNA, and lncRNA to construct a heterogeneous network for multiple data types, employing a graph attention neural network to learn corresponding node features and predict potential circular RNA-disease associations. This type of model integrates data from multiple different sources to construct a heterogeneous multi-source network, but it does not distinguish between different data types.

[0009] In summary, when discovering hidden circular RNA-disease associations, effectively exploring the essential characteristics of different types of nodes in multi-source data and fully utilizing the consistency of multi-source data are crucial. However, current methods for predicting circular RNA-disease associations rely excessively on expert knowledge and fail to effectively utilize the biological knowledge contained in multi-source data. Summary of the Invention

[0010] To address the shortcomings of current methods for predicting the association between circular RNA and diseases, such as the lack of extraction of features specific to multi-source data and the neglect of differences in pathogenicity among different source data, this invention provides a circular RNA-disease association prediction method based on weighted graph attention and heterogeneous graph neural networks. This method effectively predicts the meta-paths hidden in multi-source data, significantly improves the accuracy of the model, and exhibits more stable predictive performance.

[0011] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0012] A method for predicting circular RNA-disease associations based on weighted graph attention and heterogeneous graph neural networks includes:

[0013] Step 1: Based on known circular RNA-disease associations, miRNA-miRNA similarities, and disease-gene data, construct weighted circular RNA networks, miRNA networks, and disease networks, respectively.

[0014] Step 2: Based on the initial features of circular RNA, miRNA, and disease, as well as the connection weights between nodes in the circular RNA network, miRNA network, and disease network, a weighted graph attention mechanism is used to extract the features of each circular RNA, miRNA, and disease node from the circular RNA network, miRNA network, and disease network, respectively.

[0015] Step 3: Circular RNA, miRNA, and disease are treated as three different types of nodes. The extracted node features are used as node representations. Based on the known circular RNA-miRNA association, miRNA-disease association, and circular RNA-disease association, three different types of edges are established to construct a multi-source heterogeneous network.

[0016] Step 4: Using a heterogeneous graph neural network model, different attention weights are assigned to different types of edges in the multi-source heterogeneous network, and the contextual information of the meta-paths in the multi-source heterogeneous network is learned to obtain the aggregation features of each node in the multi-source heterogeneous network.

[0017] Step 5: Using the aggregation features of known circular RNAs and diseases as samples, train an association prediction classifier. Based on the aggregation features of circular RNAs and disease nodes, use the association prediction classifier network to predict the unknown association between circular RNAs and diseases in the multi-source heterogeneous network.

[0018] Furthermore, the method for constructing a circular RNA network based on known circular RNA-disease associations is as follows: obtain circular RNA-disease association data from the database, calculate the Gaussian interaction spectrum kernel similarity between different circular RNAs; construct a weighted circular RNA network with circular RNAs as nodes and the Gaussian interaction spectrum kernel similarity between different circular RNAs as the connection weights between nodes;

[0019] The method for constructing a miRNA network based on known miRNA-miRNA similarity is as follows: download the functional relationships of miRNAs from the database, normalize the functional relationship data using Z-score normalization and use it as the similarity between miRNAs; construct a weighted miRNA network with miRNAs as nodes and the similarity between different miRNAs as the connection weights between nodes;

[0020] The method for constructing a disease network based on known disease-gene pairs is as follows: download gene-disease data from the database to obtain the association probability between genes and diseases; then, calculate the similarity between different diseases based on the association probability; and finally, construct a weighted disease network with diseases as nodes and the similarity between different diseases as the connection weight between nodes.

[0021] Furthermore, the formula for calculating Gaussian interactive spectral kernel similarity is:

[0022]

[0023]

[0024] Among them, c i and c j Representing circular RNAi and circular RNAj respectively; Sim(ci ,c j ) represents the Gaussian interaction spectrum nuclear similarity between circular RNAi and circular RNAj; nc represents the total number of circular RNAs; GIP(c i ) represents c i The interaction spectrum, i.e., the row vector of the i-th row in the circular RNA-disease association matrix A; The mean of all circular RNA interaction profiles is used to control the width of the Gaussian nucleus.

[0025] Furthermore, the similarity calculation method between different diseases is as follows:

[0026]

[0027] Among them, Sim(d i ,d j ) indicates disease d i and disease d j Similarity between them; It is related to disease d i There are related gene sets. It is related to disease d j There exists a set of related genes, where v and u represent genes that meet the criteria, and P(·,·) represents the probability of association between the gene and the disease.

[0028] Furthermore, the method for extracting the initial features of circular RNA is as follows: the sequence of each circular RNA is encoded using k-mer, and then the shallow model Doc2Vec is used to learn the k-mer continuous word vectors in the encoded sequence as the initial features of the circular RNA.

[0029] The method for extracting initial miRNA features is as follows: extract the miRNA association vector from known miRNA-disease association data, and use it as the initial miRNA feature;

[0030] The method for extracting the initial features of a disease is as follows: obtain the linguistic definition of the disease from the medical lexicon, then use Doc2Vec to obtain the linguistic description vector contained in the definition, and then concatenate the obtained ontology vector and description vector as the initial features of the disease.

[0031] Furthermore, a weighted graph attention mechanism is used to aggregate the features of adjacent nodes in the topology of the circular RNA network, while considering the edge weight information of adjacent nodes, to obtain the unique features of each circular RNA node in the circular RNA network; the same method is used to obtain the unique features of each miRNA node in the miRNA network and the unique features of each disease node in the disease network; specifically:

[0032] h i =Wx i

[0033]

[0034]

[0035]

[0036] In the formula, x i Let h represent the feature of the i-th node in the input of the weighted graph attention layer, W represent the linear transformation matrix of the weighted graph attention layer, and h represent the feature of the i-th node in the input of the weighted graph attention layer. i h j This indicates that the weighted graph attention mechanism layer focuses on node features x. i x j Features obtained by linear transformation W represents the learnable weights shared by the weighted graph attention mechanism layers. e This represents the learnable parameters shared by all layers, where LeakyReLU is the activation function, and e ij ω represents the attention score of neighboring node j to node i generated when the weighted graph attention mechanism layer aggregates the features of node i; ij In a circular RNA network, miRNA network, or disease network, α represents the connection weight between node j and node i; N(i) represents the set of neighboring nodes of node i; ij e ij The attention coefficients obtained after normalization; M represents the number of multi-heads used in the weighted graph attention mechanism layer. W represents the attention coefficient obtained from the attention of the m-th head. m Let z denote the linear transformation matrix under the m-th head attention, and σ() denote the ReLU nonlinear activation function; i This indicates that the weighted graph attention mechanism layer aggregates the features of node i with the features of its neighboring nodes and simultaneously considers the edge weight information of the neighboring nodes to obtain the features.

[0037] Further, step 4 specifically involves: the architecture of the heterogeneous graph neural network includes a Graph Transformer layer, a convolutional layer, and a multi-head attention mechanism layer. The Graph Transformer layer aggregates the adjacency matrices representing various edge types in the multi-source heterogeneous network to obtain a new adjacency matrix, i.e., generating new heterogeneous edges. Then, for the nodes in the multi-source heterogeneous network, the multi-head attention mechanism layer aggregates the features of its neighboring nodes and considers the attention weights of the connecting edges to obtain the aggregated features of the nodes.

[0038] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to implement the circular RNA-disease association prediction method based on weighted graph attention and heterogeneous graph neural networks as described above.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the circular RNA-disease association prediction method based on weighted graph attention and heterogeneous graph neural networks as described above.

[0040] Beneficial effects

[0041] This invention establishes separate networks for circular RNA, miRNA, and disease similarity. A weighted graph attention mechanism is designed to extract unique features of circular RNA, miRNA, and disease. Based on multi-source association information, a multi-source heterogeneous network is constructed. A heterogeneous graph neural network is then used to assign different attention weights to different types of edges, and the contextual information of meta-paths in the heterogeneous network is learned to predict circular RNA-disease associations. This method effectively improves the model's accuracy and exhibits more stable predictive performance. Furthermore, the introduction of multi-source data allows for further exploration of the circular RNA-miRNA-disease regulatory pathway, contributing to the revelation of more specific and detailed biological regulatory pathways. Attached Figure Description

[0042] Figure 1 Flowchart of the circular RNA-disease association prediction method (CDHGNN) based on weighted graph attention and heterogeneous graph neural network described in this invention;

[0043] Figure 2 Network node dimension testing;

[0044] Figure 3 Changes in various network attention scores

[0045] Figure 4 Network analysis of heterogeneous networks on dataset 1

[0046] Figure 5 Network analysis of heterogeneous networks on dataset 2

[0047] Figure 6 AUC curves comparing CDHGNN with other methods;

[0048] Figure 7 Comparison of the accuracy of the prediction results from Top 10 to Top 40;

[0049] Figure 8 KEGG regulation analysis in acute kidney injury. Detailed Implementation

[0050] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0051] This embodiment provides a method for predicting circular RNA-disease associations based on weighted graph attention and heterogeneous graph neural networks. The specific implementation process is as follows:

[0052] I. Construction of circular RNA networks, miRNA networks, and disease networks

[0053] 1. Construction of circular RNA networks

[0054] Circular RNA-disease association data were obtained from two standard databases, MNDR 3.0 and CircFunBase, respectively. The nuclear similarity of the Gaussian interaction spectrum of circular RNA was calculated, and a circular RNA network was established based on this.

[0055] Known circular RNA-disease associations can be represented by an association matrix A∈R nc×nd If the vector in the i-th row is represented as a matrix, then the vector can be expressed as the association pattern between circular RNA i and the disease. The Gaussian interaction kernel similarity between any two circular RNAs is calculated using the interaction patterns between any two circular RNAs and the disease, as defined by the following formula:

[0056]

[0057]

[0058] Among them, c i and c j represent circular RNAi and circular RNAj respectively; nc represents the total number of circular RNAs; GIP(c i ) represents c i The interaction spectrum, i.e., the row vector of the i-th row in the correlation matrix A; Used to control the width of the Gaussian kernel.

[0059] From the formula definition, it can be seen that the more similar the interaction patterns of two circular RNAs, the greater the similarity of their Gaussian interaction spectra. A weighted circular RNA network is constructed using the calculated circular RNA similarity.

[0060] 2. Construction of miRNA network

[0061] To construct a miRNA functional similarity network, functional relationships of miRNAs were downloaded from the MISIM v2.0 database. To ensure that the similarity values ​​had the same scale, the data were normalized using Z-score normalization and used as the functional similarity between miRNAs. A weighted miRNA network was then constructed using the normalized miRNA similarity.

[0062] 3. Construction of disease networks

[0063] To construct a disease network, gene-disease data was downloaded from the DisGeNET database to obtain the association probabilities between genes and diseases. Then, the disease similarity between disease i and disease j was calculated and used as weights between disease nodes to construct the disease network. The similarity calculation between diseases is as follows:

[0064]

[0065] Among them, Sim(d i ,d j ) indicates disease d i and disease d j Disease similarity between them; It is related to disease d i There are related gene sets. It is related to disease d j There exists a set of related genes, where P(·,·) represents the probability of association between a gene and a disease.

[0066] II. Extraction of circular RNA, miRNA, and disease characteristics

[0067] Data on the association between circular RNA and disease is non-Euclidean structural data. Graph neural networks, which have been successfully applied, were used to extract node features. Each biomolecule plays a different role in its function, thus requiring different weights to be assigned to different molecules. Therefore, Graph Attention Networks (GAT) have become an important tool for automatically assigning node weights and thus aggregating and acquiring node features.

[0068] Generic Artificial Attention (GAT) is a neural network architecture that operates on graph-structured data. In each layer, nodes participate in the generation of features from their neighbors. These layers are stacked and then used with masked self-attention layers to solve graph-based convolution problems. The input layer of GAT uses the initial features of the nodes. Where n is the number of nodes in the network, and f represents the feature dimension of each node. Let represent the features of the i-th node. To obtain sufficient expressive power to transform the input features into higher-level features, at least one learnable linear transformation is required. As an initial step, a shared linear transformation is used:

[0069]

[0070] Where W∈R f'×f This W acts on every point. After connecting the features of the l-th layer node i and its neighboring j nodes together, a shared, learnable weight a:R is applied. f′ ×R f′ →R represents the pairwise attention score between node i and node j, and the process is as follows:

[0071]

[0072] The association between node i and node j is determined by the weight ω of the edge between them. ij This is reflected in the fact that it affects the features generated by the aggregation nodes. Therefore, the design incorporates edge weights into the generation process of pairwise attention scores, and the calculation formula is as follows:

[0073]

[0074] Among them W e These are learnable parameters shared by all layers.

[0075] The graph structure is inserted into the model using a masked attention mechanism, and the attention coefficients are obtained by normalizing all neighboring nodes j of node i using the softmax function.

[0076]

[0077] Multi-head attention aggregation is used to aggregate the embeddings of neighbors simultaneously, as shown in the following formula:

[0078]

[0079] Using the same method, a graph attention neural network was used to extract node features of circular RNA, miRNA, and disease from the constructed circular RNA network, miRNA network, and disease network, respectively.

[0080] In addition, since graph attention neural networks require initial features of input nodes, corresponding initial node features are generated for each type of node.

[0081] Non-coding RNA does not express function as individual bases, but rather through continuous RNA fragments. A common method for sequence analysis is to divide them into so-called k-mers, which are subsequences of length k. For example, the tetramer of an RNA sequence is AAAA, AAAC, ..., UUUU. K-mer representation is widely used to predict RNA-protein binding sites. For circular RNA, we first encode all circRNA sequences in the database using k-mers to uncover the functional fragments present in the circular RNA sequences. Specifically, we use a continuous encoding method to preserve the contextual relationships between k-mers. Since the variance in the length of circular RNA sequences is relatively large, this embodiment uses the shallow model Doc2Vec to learn the continuous word vectors of k-mers for each sequence. The details are as follows:

[0082] Context(w,seq)=[w(seq),w(t-2),w(t-1),w(t+1),w(t+2)] (9)

[0083]

[0084] Where w represents the word in each sequence, i.e., k-mer, seq represents the sequence information of each circular RNA, corpus represents the vocabulary composed of all circular RNA sequences, and L represents the loss function.

[0085] Therefore, the k-mer embeddings of the circular RNA sequence are obtained and used as the initial features of the weighted graph attention neural network to extract features of the circular RNA.

[0086] For initial miRNA features, known miRNA-disease association data are obtained from the HMDD database. This association data reflects the association patterns between miRNAs and diseases. The miRNA association vectors are extracted from these data and used as the initial vectors for a weighted graphical attention neural network to extract miRNA features.

[0087] The Disease Ontology aims to provide the biomedical community with a consistent, reusable, and sustainable vocabulary of human disease terms, phenotypic features, and related medical disease concepts. The Disease Ontology draws from a wealth of sources, containing significant hidden disease-related information. It stores the relationships between diseases in the form of a Directed Acyclic Graph (DAG). Each disease is represented by a DO (Domain) term, and its structure in the tree represents its corresponding system description, which can be viewed as a statement describing the state of the disease. By serializing the semantic tree path of disease i, the embedding of the disease ontology is obtained using natural language processing methods. Simultaneously, to obtain more accurate disease feature vectors, the linguistic definition of the disease is retrieved from the medical lexicon of the UMLS Metathesaurus Browser. Then, Doc2Vec is used to extract the linguistic description vectors contained in the definition. The obtained ontology vector and description vector are concatenated as the initial disease vector, and a weighted graph attention neural network is used to extract the disease features.

[0088] III. Constructing Multi-Source Heterogeneous Networks

[0089] Heterogeneous networks contain three different types of nodes: circular RNA, miRNA, and disease. Correspondingly, there are three different types of edges: circular RNA-miRNA, circular RNA-disease, and miRNA-disease. A heterogeneous network is defined as G = (V, E), where V represents the three types of network nodes (circular RNA, miRNA, and disease), and E represents the three different types of edges (circular RNA-miRNA, circular RNA-disease, and miRNA-disease). There is a network node type mapping function ψ in a heterogeneous network. v :V→T v A mapping function of the type of an edge. node v i It uniquely maps to one node type, such as: Similarly, edge e i It uniquely maps to one edge type, such as: When the type of an edge in the network is greater than 1, such as: If the network is heterogeneous, then it is a heterogeneous network. Various types of neighbor matrices can be used. To construct a heterogeneous network, where A i ∈R N×N N = nc + nm + nd. A meta-path in a heterogeneous network. It is a path that connects multiple types, such as: in Metapath Adjacency matrix can be used To represent it using multiplication, in This represents the adjacency matrix of the type of the i-th edge in the metapath.

[0090] IV. Using Heterogeneous Graph Neural Networks to Aggregate Node Features in Multi-Source Heterogeneous Networks

[0091] Multi-source data networks contain various types of edges, and these edges can be represented by graphs to depict their complex data structures. Graph neural networks have been successfully applied to this type of data. However, most graph neural network models treat such complex networks containing multiple data types as homogeneous networks. However, different types of data have varying degrees of influence on the pathogenic process of circular RNAs (circRNAs), so the homogeneous network approach loses relevant information generated by different data sources regarding circRNA-disease associations. Alternatively, manual specification of meta-paths is required, but this necessitates domain-specific expertise for filtering. Therefore, heterogeneous graph neural networks are used to learn the potential associations between circRNAs, miRNAs, and diseases hidden in multi-source data by automatically selecting meta-paths. Meta-paths are formed by connecting nodes of various types, i.e., multiplying different adjacency matrices. Therefore, a GT (Graph Transformer) layer is constructed. Through this layer, two adjacency matrices are selected from three to obtain a new adjacency matrix, thus generating new heterogeneous edges. The formula for calculating Q is:

[0092] Q = F(A; W) φ )=φ(A;softmax(W φ (11)

[0093] Where φ represents the convolutional layer in the GT layer. The parameters representing the convolutional layer φ, Q is the sum of A and the weight parameters W. φ After convolution, each Q is obtained. i use To represent, where Represents the type set of heterogeneous edges. Indicates the t-th l The weight of each type of edge in the l-th layer.

[0094] Therefore, the metapath can be represented as:

[0095]

[0096] The GraphTransformer layer generalizes the adjacency matrix obtained by multiplying two adjacency matrices, so that the values ​​are no longer limited to 0 and 1, but also include decimals. It represents the probability that two types of edges are connected. Furthermore, the ground truth (GT) layer assigns weights, meaning different weights are given to different types of edges. The GT layer retains the probability of the two different types of edges being connected.

[0097] A multi-head attention mechanism is used to simultaneously aggregate features from multiple nodes, as shown in the following formula:

[0098]

[0099] Where C represents the number of channels in the multi-head attention mechanism, in order to incorporate the influence of its own node features. yes The degree matrix, I is the identity matrix, W gt ∈R F×F This represents the shared, trainable weight matrix. V. Association Prediction Classification

[0100] The nodes learned by the heterogeneous graph neural network are aggregated with features and then classified. The classification prediction training method is as follows:

[0101]

[0102]

[0103] In the formula, f c ,f d These are circular RNAs extracted from multi-source heterogeneous networks and aggregated features of diseases, respectively; W3 and b3 are the weights and bias parameters of the association prediction classifier network. It is an association prediction classifier network based on aggregated features f c ,f d The predicted output, It is the aggregate feature f c ,f d The corresponding true label, Loss is the loss function value of the association prediction classifier network.

[0104] VI. Experimental Verification

[0105] 1. Experimental Data

[0106] The following data were mainly used in the experiment: (1) Circular RNA sequence data. Since the function of circular RNA is specifically implemented by continuous RNA fragments, i.e., k-mers, in order to explore the embedding of k-mers in circular RNA, circular RNA sequence data was retrieved from the CircBase database and approximately 140,732 circular RNA sequences were obtained; (2) miRNA functional similarity data. In order to construct the miRNA network, the functional similarity of miRNAs was obtained from the MISIM v2.0 database, which included the functional relationships between 664 miRNAs. This similarity was based on the positive and negative correlation between miRNAs and diseases, disease semantics, and the upregulation and downregulation relationships between miRNAs and diseases, etc., and a more comprehensive miRNA functional similarity was generated by integrating multiple aspects of information related to miRNA function. (3) Gene-disease association data. Gene-disease associations were downloaded from DisGeNET, which included 262,989 associations between 13,705 genes and 1,977 diseases. This data was integrated by normalizing genes, variants, diseases, and block-driven ontology and controllable vocabulary. (4) Disease ontology data. (5) Disease definition description. The UMLS Metathesaurus Browser is a large biomedical lexicon organized by concept or meaning, containing synonyms from more than 200 different sources. Language descriptions of relevant diseases were obtained from this database. (6) Circular RNA-miRNA association data. Known circular RNA-miRNA associations were collected from the starBase database, which contained 18,320 circular RNA-miRNA associations between 886 circular RNAs and 638 miRNAs. (7) miRNA-disease association data. Biologically validated miRNA-disease associations were collected from the HMDD 3.0 database, which contained 27,872 miRNA-disease associations between 1,054 miRNAs and 226 diseases. (8) Circular RNA-disease associations. Download circular RNA-disease association data from the MNDR 3.0 database, which includes 3,206 circular RNA-disease associations between 2,396 circular RNAs and 165 diseases. 9) Circular RNA-disease association dataset 2 contains 2,900 circular RNA-disease associations between 2,537 circular RNAs and 67 diseases, from the CircFunBase database.

[0107] 2. Evaluation Indicators

[0108] To verify the effectiveness of this method, we used 10x cross-validation and De novo prediction to test the prediction performance of the MBiRW method.

[0109] (1) Cross-validation method

[0110] Cross-validation is a statistical method for estimating the performance of machine learning models, exhibiting lower bias than other methods. This method divides the samples into training and test sets, trains the model using the training data, and then evaluates the model's performance using the test set data. This method employs a 5-fold cross-validation approach, dividing the data into five equal parts: four parts serve as the training set, and the remaining part as the test set. This process is repeated five times until each part has been used as the test set in turn. The advantage of this method is that it effectively avoids overfitting and underfitting.

[0111] The receiver operating characteristic (ROC) curve is used to evaluate the classification ability of a binary classifier as its discrimination threshold changes. This metric describes the relationship between the true positive rate (TPR, sensitivity) and the false positive rate (FPR, 1-specificity) based on different discrimination thresholds. TPR is the proportion of correctly predicted positive samples out of all positive samples. FPR is the proportion of incorrectly predicted positive samples out of all negative samples. A predicted value higher than the discrimination threshold is classified as a positive sample, and vice versa. TP refers to the number of samples classified as positive that are also positive, FP refers to the number of samples classified as positive but also negative, TN refers to the number of samples classified as negative that are also negative, and FN refers to the number of samples classified as negative but also positive.

[0112]

[0113]

[0114] The area under the ROC curve (AUC) is used as a metric to measure model performance.

[0115] (2) Other evaluation indicators

[0116] To more comprehensively evaluate model performance, this paper also uses several other evaluation metrics, primarily precision, accuracy, and F1-score. Precision refers to the proportion of correctly classified positive samples. Accuracy refers to the proportion of correctly classified samples in the total sample. The F1-score is the harmonic mean of recall and precision. The specific formula is as follows:

[0117]

[0118]

[0119]

[0120] 3. Network Node Dimension Testing

[0121] To learn meta-paths within the constructed heterogeneous network, node representations are needed for the heterogeneous graph neural network. CDHGNN uses a designed edge-weighted graph attention neural network to learn node features and concatenates three types of node features as node representations in the heterogeneous network. Node dimension is a key factor in model performance. Cross-validation is performed on datasets 1 and 2 to evaluate the impact of node dimension. Figure 2 As shown, the model's performance gradually improved as the node dimension of circular RNA increased from 16 to 126, the miRNA dimension increased from 16 to 64, and the disease dimension increased from 16 to 128. The model performed best when the circular RNA dimension was 128, the miRNA dimension was 64, and the disease dimension was 128. With increasing dimensions, the noise introduced by this feature leads to a performance decrease. The dimensions of circular RNA, miRNA, and disease were set to 128, 64, and 128, respectively.

[0122] 4. Changes in attention scores across various online platforms

[0123] To add contextual information about different edge types in the metapath, an attention mechanism is used to assign attention weights to each edge type. During training, the changes in the weights of each edge type reflect the importance of that type of data to the pathogenicity. The test results on datasets 1 and 2 are as follows: Figure 3 As shown in the figure, CM represents circular RNA-miRNA association, CD represents circular RNA-disease association, and MD represents miRNA-disease association. In heterogeneous graph neural network aggregation, IM represents the identity matrix to account for the influence of its own nodes. As can be seen from the figure, the CD association is assigned the highest weight, meaning it has the greatest impact on the final prediction accuracy. CM and MD are also important for accuracy.

[0124] 5. Network Analysis of Heterogeneous Networks

[0125] To analyze the heterogeneous networks constructed on datasets 1 and 2, statistical analysis was performed on their degree distribution and node connectivity. Figure 4 (a) shows the degree distribution of the network in dataset 1. It can be seen that the distribution follows a power law, indicating that most nodes in the network are linked to a small number of nodes, while a small number of nodes are linked to a large number of nodes. Figure 4(b) shows the network node connectivity on dataset 1. Disease nodes have higher connectivity than the other two types of nodes, while circRNA nodes have the lowest connectivity, indicating that disease nodes have more connections with other nodes, while circRNA nodes have the fewest connections with other nodes. The heterogeneous network constructed on dataset 2 is as follows: Figure 5 As shown.

[0126] 6. Comparison with other methods

[0127] To evaluate the performance of CDHGNN, it was compared with five state-of-the-art computational methods: KGANCDA, MGRCDA, CDASOR, GCNCDA, and NSL2CD. For a fair and reasonable comparison, all models were evaluated using a 5x crossover on the same benchmark dataset, and the parameters were the recommended optimal parameters. On Dataset 1, CDHGNN achieved the best performance with an AUC of 0.886, representing improvements of 4.5%, 4.1%, 7.2%, 9.8%, and 8.3% compared to KGANCDA, MGRCDA, CDASOR, GCNCDA, and NSL2CD, respectively. Other metrics also showed improvement: AUCR of 0.817, Accuracy of 0.824, Precision of 0.808, Recall of 0.814, and F1-score of 0.804. On dataset 2, CDHGNN outperforms other methods, with an AUC of 0.896, an AUPR of 0.823, an accuracy of 0.831, a precision of 0.819, a recall of 0.823, and an F1-score of 0.811. The results demonstrate that CDHGNN improves both accuracy and stability. Detailed AUC results for both datasets are shown below. Figure 4 As shown, a higher TPR value with the same FPR indicates higher accuracy for CDHGNN. Furthermore, the correct retrieval ratio of Top 10 to Top 40 associations in the prediction results was compared, such as... Figure 5 As shown, CDHGNN can retrieve more experimentally validated circRNA-disease associations.

[0128] 7. Case Analysis

[0129] Based on the current database and literature, the model's performance was evaluated through case studies of the prediction results. The model was trained on known and validated circular RNA-disease associations to uncover potential circular RNA-disease associations, and the accuracy of the top 20 predicted values ​​was analyzed, as shown in Table 1. Hsa_circ_0091702 originates from the vacuolar ATPse assembly factor and alleviates sepsis-related acute kidney injury by regulating miR-9-3p / SMG1 / inflammation and oxidative stress. Circ-AKT3 (hsa_circ_0000199) is associated with acute kidney injury via the miR-144-5p / Wnt / β-catenin pathway. CircFUT8 (hsa_circ_0003028) sponges miR-570-3p and regulates the miR-570-3p / KLF10 axis as a tumor suppressor in bladder cancer. Hsa_circ_0000515, by binding to miR-296-5p, can activate bladder cancer cell growth. Immobilized miR-200a-3p and exosome-mediated circ_UBE2D2 (hsa_circ_0005728) metastasis enhanced tamoxifen resistance in breast cancer. Spongy miR-532-3p and circRNA_103809 (hsa_circ_0072088) inhibited breast cancer cell proliferation and metastasis. CircRNA_0001946 (hsa_circ_0001946) participates in the miR-135a-5p / EMT axis, acting as a tumor promoter in rectal cancer cells. Knockdown and modulation of the miR-296-5p / RUNX1 axis inhibited colorectal cancer cell proliferation, and circ_0000512 (hsa_circ_0000512) is a promising therapeutic target for colorectal cancer. Circ-RanGAP1 (hsa_circ_0063526) promotes gastric cell development by mediating the miR-877-3p / VEGFA axis. In gastric cancer, hsa_circ_000487, associated with the negative regulatory loop, forms hsa_circ_0004872 / miR-224 / Smad44 / ADAR1, which functions as a tumor suppressor. As an endogenous RNA of miR-654-3p, circRHOBTB3 (hsa_circ_0006404) inhibits gastric cancer cell growth by promoting the p21 signaling pathway. Isolating miR-17, circ-ITCH (hsa_circ_0001141) functions as a tumor suppressor in gastric cancer via the Wnt / β-catenin signaling pathway. Spongy miR-134-5p activates BTG-2 expression, and circZNF609 (hsa_circ_0000615) inhibits glioma cell proliferation and migration.Targeting the miR-520a-5p / CDK4 regulatory axis, exosome-mediated hsa_circ_0014235 activates the malignant development of non-small cell lung cancer.

[0130] Table 1. Case Analysis Results

[0131]

[0132] Furthermore, identifying potential circular RNA-disease associations based on multi-source data is beneficial for exploring the relationships between circRNA, miRNA, and diseases. Among the top 20 predicted circular RNA-disease associations, the biological regulatory pathways of circular RNAs associated with acute kidney injury in the predicted outcomes were further explored based on the microRNAs in cancer identified in KEGG, specifically as follows: Figure 8 Show.

Claims

1. A method for predicting circular RNA-disease associations based on weighted graph attention and heterogeneous graph neural networks, characterized in that, include: Step 1: Based on known circular RNA-disease associations, miRNA-miRNA similarities, and disease-gene data, construct weighted circular RNA networks, miRNA networks, and disease networks, respectively. Step 2: Based on the initial features of circular RNA, miRNA, and disease, as well as the connection weights between nodes in the circular RNA network, miRNA network, and disease network, a weighted graph attention mechanism is used to extract the features of each circular RNA, miRNA, and disease node from the circular RNA network, miRNA network, and disease network, respectively. A weighted graph attention mechanism is used to aggregate the features of adjacent nodes in the topology of a circular RNA network, while also considering the edge weight information of adjacent nodes, to obtain the unique features of each circular RNA node in the circular RNA network. The same method is used to obtain the unique features of each miRNA node in the miRNA network and the unique features of each disease node in the disease network; specifically: ； ； ； ； In the formula, The input to the weighted graph attention mechanism layer is the first... The characteristics of each node This represents the linear transformation matrix of the weighted graph attention mechanism layer. , This indicates that the weighted graph attention mechanism layer focuses on node features. , Features obtained by linear transformation This represents the learnable weights shared by the weighted graph attention mechanism layers. This represents the learnable parameters shared by all layers. For activation function, This indicates that the weighted graph attention mechanism layer is applied to nodes. Neighbor nodes generated when features are aggregated For nodes Attention score; This refers to nodes in circular RNA networks, miRNA networks, or disease networks. With nodes Connection weights between them; Represents a node The set of neighboring nodes; express The attention coefficient obtained after normalization; This indicates the number of multi-heads used in the weighted graph attention mechanism layer. Indicates the first Attention coefficient obtained from head attention Let m be the linear transformation matrix under the m-th head attention. Represents the ReLU nonlinear activation function; This indicates that the weighted graph attention mechanism layer is applied to nodes. The features are obtained by aggregating the features of its neighboring nodes and simultaneously considering the edge weight information of the neighboring nodes. Step 3: Circular RNA, miRNA, and disease are treated as three different types of nodes. The extracted node features are used as node representations. Based on the known circular RNA-miRNA association, miRNA-disease association, and circular RNA-disease association, three different types of edges are established to construct a multi-source heterogeneous network. Step 4: Using a heterogeneous graph neural network model, different attention weights are assigned to different types of edges in the multi-source heterogeneous network, and the contextual information of the meta-paths in the multi-source heterogeneous network is learned to obtain the aggregation features of each node in the multi-source heterogeneous network. Step 5: Using the aggregation features of known circular RNAs and diseases as samples, train an association prediction classifier. Based on the aggregation features of circular RNAs and disease nodes, use the association prediction classifier network to predict the unknown association between circular RNAs and diseases in the multi-source heterogeneous network.

2. The method according to claim 1, characterized in that, The method for constructing a circular RNA network based on known circular RNA-disease associations is as follows: obtain circular RNA-disease association data from the database, calculate the Gaussian interaction spectrum kernel similarity between different circular RNAs; construct a weighted circular RNA network with circular RNAs as nodes and the Gaussian interaction spectrum kernel similarity between different circular RNAs as the connection weights between nodes. The method for constructing a miRNA network based on known miRNA-miRNA similarity is as follows: download the functional relationships of miRNAs from the database, normalize the functional relationship data using Z-score normalization, and use it as the similarity between miRNAs; A weighted miRNA network is constructed, with miRNAs as nodes and the similarity between different miRNAs as the connection weights between nodes. The method for constructing a disease network based on known disease-gene pairs is as follows: download gene-disease data from the database to obtain the association probability between genes and diseases; then, calculate the similarity between different diseases based on the association probability; and finally, construct a weighted disease network with diseases as nodes and the similarity between different diseases as the connection weight between nodes.

3. The method according to claim 2, characterized in that, The formula for calculating Gaussian interactive spectral kernel similarity is: ； ； in, and These represent circular RNA i and circular RNA j, respectively. The Gaussian interaction spectrum kernel similarity between circular RNA i and circular RNA j is represented; nc represents the total number of circular RNAs. express The interaction spectrum, i.e., the row vector of the i-th row in the circular RNA-disease association matrix A; The mean of all circular RNA interaction profiles is used to control the width of the Gaussian nucleus.

4. The method according to claim 1, characterized in that, The similarity calculation method between different diseases is as follows: ； in, Indicates disease and disease Similarity between them; Is with disease There are related gene sets. Is with disease There are related gene sets. and This indicates genes that meet the criteria. This indicates the probability of an association between a gene and a disease.

5. The method according to claim 1, characterized in that, The method for extracting the initial features of circular RNA is as follows: the sequence of each circular RNA is encoded using k-mer, and then the shallow model Doc2Vec is used to learn the k-mer continuous word vectors in the encoded sequence as the initial features of the circular RNA. The method for extracting initial miRNA features is as follows: extract the miRNA association vector from known miRNA-disease association data, and use it as the initial miRNA feature; The method for extracting the initial features of a disease is as follows: obtain the linguistic definition of the disease from the medical lexicon, then use Doc2Vec to obtain the linguistic description vector contained in the definition, and then concatenate the obtained ontology vector and description vector as the initial features of the disease.

6. The method according to claim 1, characterized in that, The process of constructing a multi-source heterogeneous network is as follows: Circular RNA-disease association data are obtained from the MNDR 3.0 and CircFunBase databases; miRNA-disease association data are obtained from the HMDD 2.0 database; and circular RNA-miRNA association data are obtained from the starBase database. The obtained association data are then fused to construct the multi-source heterogeneous network. ; among them This represents a set of nodes of the following three types: circular RNA, miRNA, and disease, through a node type mapping function. Determine the node type corresponding to each node; The set of edges representing the following three association types: circular RNA-miRNA, circular RNA-disease, and miRNA-disease, is mapped by the edge type function. Determine the edge type corresponding to each edge.

7. The method according to claim 1, characterized in that, Step 4 specifically involves the following: The architecture of the heterogeneous graph neural network includes a Graph Transformer layer, a convolutional layer, and a multi-head attention mechanism layer. The Graph Transformer layer aggregates the adjacency matrices representing various edge types in the multi-source heterogeneous network to obtain a new adjacency matrix, i.e., generating new heterogeneous edges. Then, for the nodes in the multi-source heterogeneous network, the multi-head attention mechanism layer aggregates the features of its neighboring nodes and considers the attention weights of the connecting edges to obtain the aggregated features of the nodes.

8. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to implement the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images