A disease circRNA correlation prediction method for a multi-source heterogeneous network

Abstract

The application discloses a disease circRNA correlation prediction method of a multi-source heterogeneous network. The method comprises the following steps: integrating multi-omics data to construct a multi-source heterogeneous molecular network of diseases and circRNAs; based on a gate mechanism, designing a variational graph autoencoder to extract topological structure information and context features of nodes, and using a multi-head self-attention model to fuse to obtain low-dimensional embedding representation of circRNAs and diseases; and using a support vector machine to predict unknown circRNA-disease correlation relationships. The application is modeled based on a multi-source biological heterogeneous molecular network, combined with an embedding representation learning algorithm, and can quantitatively infer and calculate the relationship between nodes in the network, support in-depth understanding of the correlation analysis mechanism between circRNAs and diseases in the heterogeneous network, and has good scalability, practicability and application prospect.

Description

Technical Field

[0001] This invention belongs to the field of biological data mining and knowledge engineering, and relates to a method for predicting disease circRNA associations in multi-source heterogeneous networks. Background Technology

[0002] Circular RNAs (circRNAs) play crucial roles in gene expression, splicing, and transcription. Increasing research indicates a significant link between circRNAs and disease development, their interaction with multiple target genes, and their important role in the occurrence, metastasis, proliferation, and differentiation of cancer cells. These interactions form a complex network of biomolecular interactions. Artificial intelligence technologies for identifying disease-related circRNAs can contribute to understanding the molecular pathology and pathogenesis of diseases, promoting the development of smart and personalized medicine.

[0003] With the development of electronic information and sequencing technologies, massive amounts of omics data have been accumulated in fields such as biomedicine. Predicting the association between miRNAs and diseases in biological experiments requires significant financial and time investment. Therefore, using intelligent computing algorithms to predict circRNA-disease association information has become an urgent problem to be solved. In 2021, Lei Xiujuan et al. published an article in *Science in China: Information Sciences* entitled "Prediction of circRNA-Disease Association Based on Multi-Data Fusion," which studied an improved non-negative matrix factorization algorithm to predict circRNA-disease association information based on circRNA expression profile data and disease similarity data. In 2020, He Xin et al. published an article in *Life Sciences* entitled "Research Progress on the Regulatory Role of circRNA-miRNA-mRNA Networks in Cardiovascular Diseases," which elucidated that circRNAs, acting as a "sponge" for microRNAs (miRNAs), competitively inhibit the binding of miRNAs to mRNAs and play an important regulatory role through circRNA-miRNA-mRNA networks, providing support for circRNAs to become biological targets for cardiovascular diseases. In addition, Kuang Zhufang et al. disclosed an efficient prediction method for the association between circRNA and disease based on the restart random walk algorithm and graph convolutional neural network in patent application No. 202110492443.7; Ji Cunmei et al. disclosed a prediction method for the association between circRNA and disease based on graph attention mechanism in patent application No. 202111154802.4; and Fan Yongxian et al. disclosed a prediction method for the association between circRNA and disease based on network model in patent application No. 201811580662.5.

[0004] The aforementioned methods typically use the similarity information between diseases and circRNAs as input to the algorithm, ignoring the interactions between molecules such as miRNAs, lncRNAs, and proteins closely related to disease miRNAs, thus limiting the predictive performance of the model. This invention is based on a multi-source heterogeneous molecular network composed of diseases, circRNAs, miRNAs, and lncRNAs. It utilizes gating mechanisms and graph convolutional networks to automatically learn the multi-scale topological features of circRNAs and diseases, and then uses support vector machines to predict the association information between diseases and circRNAs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a disease circRNA association prediction method based on multi-source heterogeneous networks. First, it constructs a multi-source heterogeneous molecular network between diseases and circRNAs using biomolecules such as circRNAs, miRNAs, and lncRNAs from multi-omics data of complex diseases. Then, based on the constructed heterogeneous network, a deep neural network is used to automatically capture the topological features and multi-level semantic information of circRNAs and diseases. Finally, a support vector machine is used to predict the association relationship. Specifically, to achieve the above-mentioned objectives, the specific technical solution of this invention includes:

[0006] 1. A method for predicting disease circRNA associations in multi-source heterogeneous networks, characterized by:

[0007] Step 1. Obtain datasets and interaction association datasets of proteins, circRNAs, miRNAs, lncRNAs, and diseases, and construct a heterogeneous biomolecular network of disease miRNAs. The set of nodes in this heterogeneous network Including proteins, diseases, circRNAs, miRNAs, and lncRNAs; edge objects. This includes association edges between proteins, diseases, circRNAs, miRNAs, and lncRNAs; where node type mapping functions are used. Edge type mapping function ; and These represent the number of node type sets and edge type sets on the heterogeneous molecular network, respectively.

[0008] Step 2. Protein, circRNA, miRNA, lncRNA, and disease node feature extraction: Calculate the sequence features of proteins respectively. Disease semantic similarity Similarity to Gaussian kernel function Sequence characteristics of circRNA, miRNA, and lncRNA ( , , ) and Gaussian kernel function similarity features ( , , ).

[0009] Step 3. Constructing a heterogeneous graph embedding representation model for disease circRNAs: Based on the heterogeneous network from Step 1 and the node features from Step 2, a deep representation model for extracting features from multi-source heterogeneous molecular networks is designed using gate mechanisms and graph convolutional networks. This deep model is used to automatically aggregate the topological semantic features of proteins, diseases, circRNAs, miRNAs, and lncRNAs in the heterogeneous molecular network. Based on the multi-source heterogeneous molecular networks for diseases and circRNAs designed in Step 1, a multi-layer graph convolutional network based on gate mechanisms is used to extract the topological features of proteins, diseases, circRNAs, miRNAs, and lncRNAs. The process is as follows: , It is the embedding feature of the node. These are the weights learned by the model. This is a characteristic of heterogeneous networks; then, a multi-head self-attention model is used to fuse disease and circRNA to construct node-level and semantic-level topological information in a multi-source heterogeneous molecular network.

[0010] Step 4. Construct a disease circRNA association predictor: Based on the embedding information of heterogeneous molecular networks, use support vector machines to predict the association relationships of disease circRNAs.

[0011] Step 5. Design the loss function for the disease circRNA association prediction task.

[0012] Step 6. Deep variational autoencoder model training. The model is trained using protein, disease, circRNA, miRNA, and lncRNA datasets and their interactions. The training process is then optimized using gradient descent, Dropout, and regularization methods to obtain the optimal parameters of the model. Attached Figure Description

[0013] Figure 1 Example diagram of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method may include the following steps:

[0016] 1. A method for predicting disease circRNA associations in multi-source heterogeneous networks, characterized by:

[0017] Step 1. Obtain datasets and interaction association datasets of proteins, circRNAs, miRNAs, lncRNAs, and diseases, and construct a heterogeneous biomolecular network of disease miRNAs. The set of nodes in this heterogeneous network Including proteins, diseases, circRNAs, miRNAs, and lncRNAs; edge objects. This includes association edges between proteins, diseases, circRNAs, miRNAs, and lncRNAs; where node type mapping functions are used. Edge type mapping function ; and These represent the number of node type sets and edge type sets on the heterogeneous molecular network, respectively.

[0018] Step 2. Protein, circRNA, miRNA, lncRNA, and disease node feature extraction: Calculate the sequence features of proteins respectively. Disease semantic similarity Similarity to Gaussian kernel function Sequence characteristics of circRNA, miRNA, and lncRNA ( , , ) and Gaussian kernel function similarity features ( , , Specifically,

[0019] semantic similarity of diseases The relationships between diseases are described as a directed acyclic graph (G), where nodes represent diseases and edges represent the relationships between diseases; in a DAG, diseases... Represented as ,in This includes diseases Including diseases The ancestor node set, It is the edge set connecting these diseases. Therefore, for a certain disease s in G, there is a disease pair with the disease... The semantic contribution value is:

[0020]

[0021] ε is a semantic contribution factor, representing the disease. semantic value ,disease and semantic similarity between for: .

[0022] Gaussian kernel similarity of diseases : Disease and Gaussian kernel similarity is defined as The calculation method is as follows:

[0023]

[0024] in and It is a disease and The interaction spectrum, as the eigenvectors of the Gaussian kernel. and These represent the vectors in rows i and j, respectively. It is a kernel bandwidth parameter, calculated as follows: Gaussian kernel function similarity of circRNA, miRNA, and lncRNA , , Similarity to the Gaussian kernel function of the disease The calculation method is the same.

[0025] Molecular sequences of length L (circRNA, miRNA, and lncRNA) were divided into groups containing... The k-mer string of 1 base is used, and then the k-mer sequence data of circRNA, miRNA, and lncRNA are trained using the word embedding algorithm GloVe to obtain the k-mer embedding features rich in contextual patterns. The k-mer embedding features of circRNA, miRNA, and lncRNA are denoted as follows: , , .

[0026] Therefore, the characteristics of a protein are represented as The characteristics of the disease are represented as The characteristics of circRNA are represented as follows The characteristics of miRNA are represented as follows The characteristics of lncRNA are represented as follows: , This represents a connection operation along a feature channel dimension.

[0027] Step 3. Construct a disease circRNA heterogeneous graph embedding representation model: Based on the heterogeneous molecular network in Step 1 and the protein feature representations learned in Step 2, the model is... Disease node characteristics circRNA node characteristics miRNA node characteristics and lncRNA node characteristics A deep representation learning model for extracting features from multi-source heterogeneous molecular networks was designed using gate mechanisms and graph-gated convolutional networks. The designed multi-layer graph-gated convolutional network was then used to extract topological embedding features of diseases and circRNAs. These are the weight parameters learned by the model. It is the verified correlation feature matrix in heterogeneous molecular networks; the feature learning process of graph-gated convolutional networks is as follows: , This model can automatically aggregate the topological semantics of lncRNAs and miRNAs in heterogeneous molecular networks to obtain discriminative features. and Then, by utilizing information from proteins, diseases, lncRNAs, and miRNAs that are adjacent to the disease and circRNA, the topological semantic features of the disease and circRNA in the heterogeneous molecular network are aggregated to obtain the topological semantic features of the disease and circRNA. and Then, the topological embedding features of the disease are obtained by fusing them using a variational graph autoencoder. Topological embedding features of lncRNA The process is denoted as , Follows a standard normal distribution. These are the weights learned by the model. It is a verified correlation feature matrix in heterogeneous molecular networks. :

[0028]

[0029] In the decoder, through the hidden variables or An inner product reconstruction between them yields the adjacency matrix of lncRNAs and diseases. and The inner product calculation process is defined as follows: and From a probabilistic sampling perspective, the feature learning process of the decoder is defined as follows: and During training, the loss function of the variational graph autoencoder is changed by... and Cross-entropy between and and The Kullback-Leibler (KL) divergence between the components, along with a gradient descent-based optimizer, was used to optimize the variational graph autoencoder for extracting embedding information of disease and lncRNA. The loss function... and Defined as: , It is the expected log-likelihood, responsible for reconstructing the penalty term; The KL divergence of q and p is the penalty term for regularization; the topological information of the disease and circRNA extracted by the variational autoencoder are respectively and .

[0030] By utilizing multi-self-attention fusion to integrate the topological information of nodes and semantics in a multi-source heterogeneous molecular network composed of diseases and miRNAs, dependency information between disease, circRNA, miRNA, protein, and lncRNA nodes associated with the disease (circRNA) is extracted. Both represent the input matrix of the attention model. The calculation process for dot product attention is as follows: D represents the dimension of the vectors in the query, keyword, and content matrices; the feature extraction process of the self-attention model is as follows: Using N self-attention models, the feature output of the multi-self-attention model is: The features extracted by the multi-attention model for the disease are denoted as... The characteristics of circRNA are denoted as .

[0031] Step 4. Construct a disease circRNA association predictor. Based on the topological embedding features of the disease extracted in Step 3. and dependency information circRNA topological embedding information and dependency information The possibility of using support vector machines to predict the relationships between disease circRNAs .

[0032] Step 5. Design the loss function for the disease circRNA association prediction task: Train and optimize the model based on the cross-entropy loss function, introducing a term into the entropy loss. The cross-entropy loss calculation is used to assess the loss between the output of the disease circRNA association prediction model and the actual data. The calculation process of the loss function is as follows: , It is an adaptive hyperparameter. It is circRNA With disease Related tags between them.

[0033] Step 6. Deep Variational Autoencoder Model Training: The model is trained based on protein, disease, circRNA, miRNA, and lncRNA datasets and interaction datasets to perform nonlinear transformations on the input heterogeneous molecular network data to fit the output. Gradient descent, Dropout, and regularization methods are used to optimize the model training and obtain the optimal parameters.

Claims

1. A method for predicting disease circRNA associations in multi-source heterogeneous networks, characterized by: Step 1. Construct a heterogeneous biological information network of disease circRNAs. The set of nodes in this heterogeneous network Including diseases, proteins, circRNAs, miRNAs, lncRNAs, and edge objects. This includes association edges between diseases, proteins, circRNAs, miRNAs, and lncRNAs; and calculates the sequence features, semantic similarity, and Gaussian kernel function similarity of proteins, as well as the sequence features and Gaussian kernel function similarity of circRNAs, miRNAs, and lncRNAs. Step 2. Construct a heterogeneous graph embedding representation model: Based on the heterogeneous network and node features from Step 1, a deep variational autoencoder model is designed using gate mechanisms and graph convolutional networks to extract features from multi-source heterogeneous molecular networks; the topological semantic features of diseases, proteins, circRNAs, miRNAs, and lncRNAs at different levels in the heterogeneous network are automatically aggregated, and then the multi-head self-attention model is used to aggregate the multi-layer topological semantic features of circRNAs and diseases to obtain the embedding features of diseases and circRNAs; Step 3. Construct a disease circRNA association predictor. Based on the embedding information of heterogeneous molecular networks, use support vector machines to predict the association relationships of disease circRNAs and design a loss function for the disease circRNA association prediction task. Train the model using protein, disease, circRNA, miRNA, and lncRNA datasets, and optimize the model training process using Dropout and regularization methods to obtain the optimal parameter settings of the model.

2. The method for predicting disease circRNA associations in a multi-source heterogeneous network according to claim 1, characterized in step 1: obtaining datasets and interaction association datasets of proteins, circRNAs, miRNAs, lncRNAs, and diseases, and constructing a multi-source heterogeneous molecular network. Then, the sequence features, semantic similarity, and Gaussian kernel function similarity of proteins, as well as the sequence features and Gaussian kernel function similarity of circRNA, miRNA, and lncRNA, are calculated. These features are then fused to obtain the characteristics of proteins, diseases, circRNA, miRNA, and lncRNA, and are then used respectively... and express.

3. The method for predicting disease circRNA associations in a multi-source heterogeneous network according to claim 1, step 2 is characterized by: designing a multi-layer graph convolutional network based on a gate mechanism to extract topological embedding features of proteins, diseases, circRNAs, miRNAs, and lncRNAs, the main process of which is as follows: ,in It is the embedding feature of the node. These are the weights learned by the model. This is a characteristic of heterogeneous networks; a multi-head self-attention model is used to adaptively extract and fuse topological information in multi-source heterogeneous molecular networks.

4. The method for predicting disease circRNA associations in a multi-source heterogeneous network according to claim 1, step 3 is characterized in that: based on the topological embedding information extracted by a deep neural network, a support vector machine is used to predict the association relationship of disease circRNAs; an effective loss function optimization model is designed for the disease miRNA association prediction task, the model is trained using disease, circRNA, miRNA and lncRNA datasets, and the optimal parameter settings of the model are obtained by optimizing the model using the gradient descent algorithm, Dropout and regularization methods.

Images