circRNA-disease association prediction model construction method based on dynamic contrast sampling and linear and nonlinear feature resonance, association prediction method and related device

Abstract

The application relates to the technical field of biosciences, and is a circRNA-disease correlation prediction model construction method and correlation prediction method based on dynamic contrast sampling linear and nonlinear feature resonance and related devices, which comprises the following steps: obtaining a training set and a test set; training a multi-view contrast learning DCFR-CDA model by using the training set; introducing a loss function during the training; ending the training when the value of the loss function is stable; obtaining a circRNA-disease correlation prediction model; testing the circRNA-disease correlation prediction model by using the test set; optimizing model parameters; and outputting a circRNA-disease correlation prediction model meeting test evaluation requirements. The application jointly applies a linear decomposition method and nonlinear decomposition methods of multi-layer perception decomposition and one-dimensional convolution decomposition to generate multi-view embedding compatible with global and local modes, thereby effectively improving the prediction accuracy of the model.

Description

Technical Field

[0001] This invention relates to the field of bioscience and technology, and is a method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance of dynamic comparative sampling, an association prediction method, and related devices. Background Technology

[0002] Circular RNAs (circRNAs), due to their unique closed-loop structure and high conservation, play a crucial role in gene regulation and disease mechanisms. Research on the association between circRNAs and various diseases not only contributes to a deeper understanding of the molecular pathological processes of diseases but also provides new insights for biomarker screening and precision medicine. Currently, traditional experimental screening methods (such as qRT-PCR and high-throughput sequencing) can directly reveal circRNA-disease associations, but they suffer from limitations such as being time-consuming, costly, and requiring strict sample size requirements. Therefore, association prediction models based on bioinformatics and machine learning have emerged and show great potential in improving prediction efficiency and accuracy.

[0003] For example, Toprak et al. proposed a bidirectional imbalanced random walk based on linear neighborhood similarity, which strengthens the balance between positive and negative transition probabilities. MVGNCDA combines multi-view graph convolutional networks (GCN) and biased random walks to fuse multi-source node features to extract global topological information. GEHGAN introduces jump and stay strategies during the walk and combines multi-head graph attention to perform weighted fusion of multiple types of similarity. AMPCDA utilizes meta-path self-attention and graph attention to deeply mine path and neighborhood features under multi-source data, while adding a boundary balance factor to improve sample balance. HGECDA combines meta-path random walks, negative sampling, and CosMulformer self-attention on ternary heterogeneous graphs to achieve deep interactive representation, and completes sparse node features through joint optimization. Methods such as SVD and NMF are often coordinated with network embedding, graph neural networks (GNN), ensemble learning, and other strategies to achieve diversified predictions. MPCLCDA automatically selects meta-paths and combines GCN and contrastive learning to optimize node fusion. ACLNDA constructs a three-layer heterogeneous graph and mines cross-layer structural relationships through asynchronous contrastive learning. Wang et al. proposed a supervised contrastive knowledge graph that uniformly extracts first-order and higher-order features among multiple types of ncRNAs and diseases, effectively improving generalization ability.

[0004] Although the above methods improve prediction performance to some extent, they still have the following limitations:

[0005] While traditional matrix factorization methods can efficiently extract global linear patterns, they struggle to capture complex nonlinear interaction features, thus limiting their performance when faced with the rich local structures and high-order relationships in biological networks.

[0006] The construction of isomorphic edges based on thresholds requires manually setting thresholds and configuring dedicated message passing mechanisms for different nodes and edge types. This not only easily introduces spurious correlations and impurity edges, but also makes the model structure bloated, computationally expensive, and prone to accumulating noise in multi-hop propagation.

[0007] Traditional contrastive learning methods rely on random sampling, which often selects samples that are difficult to distinguish, thus weakening the decision boundary. Summary of the Invention

[0008] The purpose of this invention is to solve the problem that existing circRNA-disease association prediction models cannot integrate linear and nonlinear features from multiple perspectives, resulting in limited prediction performance. This invention provides a method for constructing a circRNA-disease association prediction model based on the resonance of linear and nonlinear features through dynamic comparative sampling, an association prediction method, and related devices.

[0009] One of the technical solutions of this invention discloses a method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance of dynamic contrast sampling, comprising:

[0010] Obtain training and test sets, both of which include circRNA-disease adjacency matrices. The corresponding positions of each circRNA-disease association pair in the circRNA-disease adjacency matrix store the corresponding association probabilities.

[0011] A multi-view contrastive learning model (DCFR-CDA) was trained using a training set. A loss function was introduced during training, and training ended when the loss function value stabilized, resulting in a circRNA-disease association prediction model. The DCFR-CDA model includes: a comprehensive association matrix generation unit, which processes the circRNA-disease adjacency matrix using self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix; an embedding matrix generation unit, which performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix; a reconstruction unit, which reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; and a prediction unit, which processes the composite isomorphic graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs.

[0012] The circRNA-disease association prediction model was tested using a test set, the model parameters were optimized, and a circRNA-disease association prediction model that meets the test evaluation requirements was output.

[0013] The optimized construction of the circRNA-disease adjacency matrix includes:

[0014] Human circRNAs and diseases were collected from public databases, and irrelevant or duplicate data were removed;

[0015] A circRNA-disease adjacency matrix is ​​formed by using circRNAs as rows and diseases as columns.

[0016] The association probabilities of circRNA-disease association pairs are stored in the corresponding positions of the circRNA-disease adjacency matrix, and empty spaces are stored in other positions.

[0017] The optimized, comprehensive association matrix generation unit combines self-knowledge fusion and a weighted K-nearest neighbor knowledge network to process the circRNA-disease adjacency matrix, generating a comprehensive association matrix, including:

[0018] Obtain the disease DO similarity and GIP kernel similarity of the disease part in the circRNA-disease adjacency matrix, and then weight them to obtain the comprehensive disease similarity matrix;

[0019] The functional similarity, sequence similarity, and GIP core similarity of the circRNA portion in the circRNA-disease adjacency matrix are obtained, and a weighted average is used to obtain the circRNA comprehensive similarity matrix.

[0020] The disease comprehensive similarity matrix, circRNA comprehensive similarity matrix, and circRNA-disease adjacency matrix are integrated and propagated using self-knowledge fusion and weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix.

[0021] The optimized embedding matrix generation unit performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix, including:

[0022] The composite correlation matrix is ​​approximated in low rank using singular value decomposition (SVD) to obtain the corresponding embedding matrix.

[0023] The comprehensive correlation matrix is ​​approximated by nonnegative matrix factorization (NMF) to obtain the corresponding embedding matrix.

[0024] The latent features of the comprehensive correlation matrix were extracted using the multilayer perceptron matrix factorization (MLPF), and the circRNA and disease side were encoded separately to obtain the corresponding embedding matrix.

[0025] The one-dimensional convolutional matrix factorization CNNF is used to extract two-dimensional local features from the comprehensive correlation matrix, and the circRNA and the disease side are encoded separately to obtain the corresponding embedding matrix.

[0026] Concatenate all embedding matrices row by row to obtain the circRNA-disease association pair node embedding matrix;

[0027] .

[0028] in, All are embedding matrices obtained using the linear low-rank decomposition method and the nonlinear eigenvalue decomposition method.

[0029] The optimized, reconstructed unit reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph, including:

[0030] Map each circRNA-disease association pair in the circRNA-disease association pair embedding matrix to a higher-order composite node in the isomorphism graph.

[0031] Adjacent edges are constructed based on shared circRNAs or shared diseases to form a composite isomorphic graph.

[0032] The optimized prediction unit processes the composite isomorphism graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs, including:

[0033] Multi-head attention aggregation is performed using neighborhood information from two-layer GAT high-order composite nodes;

[0034] Based on the aggregated node embedding features, the sampling probability of positive and negative samples for each anchor point is dynamically determined using feature and structural similarity, thus obtaining a set of positive samples.

[0035] The feature representation of nodes is optimized by using triplet comparison loss, and a multilayer perceptron is used to predict the association probability of circRNA and disease association pairs.

[0036] The second technical solution of the present invention discloses a circRNA-disease association prediction method, comprising:

[0037] Obtain the circRNA-disease association pairs to be predicted;

[0038] The predicted circRNA-disease association pairs are input into the circRNA-disease association prediction model to obtain the association probability of the predicted circRNA-disease association pairs. The circRNA-disease association prediction model is constructed by the circRNA-disease association prediction model construction method based on the linear and nonlinear feature resonance of dynamic contrast sampling.

[0039] The third technical solution of the present invention discloses a device for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance of dynamic contrast sampling, comprising:

[0040] The dataset acquisition unit acquires the training set and the test set. Both the training set and the test set include a circRNA-disease adjacency matrix. The corresponding position of each circRNA-disease association pair in the circRNA-disease adjacency matrix stores the corresponding association probability.

[0041] The training unit trains the multi-view contrastive learning DCFR-CDA model using the training set. A loss function is introduced during training, and training ends when the value of the loss function stabilizes, resulting in the circRNA-disease association prediction model. The multi-view contrastive learning DCFR-CDA model includes: a comprehensive association matrix generation unit, which processes the circRNA-disease adjacency matrix by combining self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix; an embedding matrix generation unit, which performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix; a reconstruction unit, which reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; and a prediction unit, which processes the composite isomorphic graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs.

[0042] The testing unit uses a test set to test the circRNA-disease association prediction model, optimizes the model parameters, and outputs a circRNA-disease association prediction model that meets the test evaluation requirements.

[0043] The fourth technical solution of the present invention discloses a circRNA-disease association prediction device, comprising:

[0044] The unit for acquiring data to be predicted acquires the circRNA-disease association pairs to be predicted.

[0045] The prediction unit inputs the circRNA-disease association pair to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease association pair to be predicted. The circRNA-disease association prediction model is constructed by the circRNA-disease association prediction model construction method based on the linear and nonlinear feature resonance of dynamic contrast sampling.

[0046] The fifth technical solution of the present invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps in the method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance of dynamic contrast sampling or the circRNA-disease association prediction method.

[0047] The beneficial effects of the present invention include:

[0048] Multi-view feature fusion: By jointly applying linear decomposition methods (SVD, NMF) with nonlinear decomposition methods such as multilayer perceptron decomposition (MLPF) and one-dimensional convolutional integral decomposition (CNNF), multi-view embeddings compatible with global and local patterns are generated.

[0049] Isomorphic composite graph construction: Each circRNA-disease association pair is mapped to a higher-order composite node in the isomorphic graph, and edges are added based on the "co-RNA" and "co-disease" relationships, and self-loop edges are added. This avoids the need to manually set thresholds or construct multi-type edge matrices in traditional methods, thereby simplifying the graph structure, reducing the number of parameters, and improving the efficiency of feature aggregation and discrimination.

[0050] Dynamic contrastive sampling: A dynamic sampling mechanism combining feature and structural similarity is proposed to accurately select highly discriminative positive and negative samples based on anchor points. This overcomes the problem of traditional random sampling that easily introduces samples that are difficult to distinguish, and enhances the feature discrimination power and generalization performance of the model.

[0051] End-to-end contrast optimization: After using a two-layer GAT to aggregate the neighborhood information of high-order composite nodes in the composite isomorphic graph through multi-head attention to capture key neighborhood information, a triple contrast loss is introduced to align and fuse multi-view embeddings. Finally, the association probability is output through a multilayer perceptron (MLP) to improve the accuracy and generalization of the model. Attached Figure Description

[0052] Appendix Figure 1 This is a schematic diagram of the implementation environment provided for one embodiment of the present invention.

[0053] Appendix Figure 2 This is a schematic diagram of a method for constructing a circRNA-disease association prediction model according to an embodiment of the present invention.

[0054] Appendix Figure 3 This is a schematic diagram of the network structure of the multi-view contrastive learning DCFR-CDA model provided in one embodiment of the present invention.

[0055] Appendix Figure 4 This is a schematic diagram of the isomorphic graph reconstruction process provided in an embodiment of the present invention, wherein (a) is a schematic diagram of the process of constructing heterogeneous bipartite graphs in cross-validation using conventional methods; and (b) is a schematic diagram of the isomorphic graph reconstruction process based on composite association pairs proposed in an embodiment of the present invention.

[0056] Appendix Figure 5 This is a schematic diagram illustrating the model prediction performance of one embodiment of the present invention.

[0057] Appendix Figure 6 This is a schematic diagram of parameter analysis provided for one embodiment of the present invention.

[0058] Appendix Figure 7 This is a schematic diagram of a circRNA-disease association prediction method provided in one embodiment of the present invention.

[0059] Appendix Figure 8 This is a schematic diagram of a circRNA-disease association prediction model construction device provided in one embodiment of the present invention.

[0060] Appendix Figure 9 This is a schematic diagram of the structure of a circRNA-disease association prediction device provided in one embodiment of the present invention. Detailed Implementation

[0061] Those skilled in the art will understand that, unless specifically stated otherwise, in the embodiments of the present invention, a "module" or "unit" refers to a computer program or part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0062] In addition, in the embodiments of the present invention, "multiple" refers to two or more, and "first" and "second" are used to distinguish descriptions and should not be construed as implying relative importance.

[0063] This invention provides a method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance using dynamic contrastive sampling, an association prediction method, and related devices. The method involves obtaining a training set and a test set, both of which include a circRNA-disease adjacency matrix. The corresponding positions of each circRNA-disease association pair in the circRNA-disease adjacency matrix store the corresponding association probability. The training set is used to train a multi-view contrastive learning DCFR-CDA model. A loss function is introduced during training, and training ends when the value of the loss function stabilizes, resulting in the circRNA-disease association prediction model.

[0064] The circRNA-disease association prediction model was tested using a test set, the model parameters were optimized, and a circRNA-disease association prediction model that meets the test evaluation requirements was output. The circRNA-disease association pairs to be predicted were input into the circRNA-disease association prediction model to obtain the corresponding association probabilities.

[0065] The method provided in this embodiment of the invention may involve artificial intelligence (AI) technology and may be implemented based on artificial intelligence technology, such as using deep learning to train a corresponding model using samples.

[0066] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0067] Deep learning (DL) specifically refers to machine learning based on deep neural network models and methods. It has developed from statistical machine learning, artificial neural network algorithms, and other algorithms, combined with the advancements in big data and computing power. The most important technical feature of deep learning is its ability to automatically extract features.

[0068] As attached Figure 1 The diagram illustrates an implementation environment provided by an embodiment of the present invention. This implementation environment may include: training equipment and usage equipment.

[0069] Both the training equipment and the equipment used are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, tablet computer, PC (Personal Computer) or other electronic devices; or, the computer device is a server, which can be a single server, a server cluster composed of multiple servers, or a cloud computing service center. This embodiment of the invention does not limit this.

[0070] Training equipment refers to computer equipment capable of training and learning a multi-view contrastive learning DCFR-CDA model. Optionally, the training equipment has the capability to acquire a multi-view contrastive learning DCFR-CDA model and train and learn it according to application requirements. For example, the training equipment acquires a multi-view contrastive learning DCFR-CDA model from other devices via a network, and then trains it using training samples according to application requirements, enabling the model to predict circRNA-disease association probabilities. Optionally, the training equipment has the capability to construct multi-view contrastive learning DCFR-CDA models. It can independently construct a multi-view contrastive learning DCFR-CDA model according to application requirements and then train and learn it. For example, to achieve circRNA-disease association probability prediction results, the training equipment independently constructs a multi-view contrastive learning DCFR-CDA model and then trains and learns it using samples according to application requirements.

[0071] The device used refers to a computer device that has the capability to use a multi-view contrastive learning DCFR-CDA model. Optionally, the device uses the multi-view contrastive learning DCFR-CDA model from other devices via the network according to application requirements. For example, if the device has the capability to predict the probability of circRNA-disease association, it can use the network to obtain a multi-view contrastive learning DCFR-CDA model that has been trained and learned to predict the probability of circRNA-disease association from other devices, and use the multi-view contrastive learning DCFR-CDA model to predict the probability of circRNA-disease association.

[0072] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0073] Example 1: As shown in the attached document Figure 2 As shown, this embodiment discloses a method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance of dynamic contrast sampling, including:

[0074] Step S110: Obtain the training set and the test set, wherein both the training set and the test set include a circRNA-disease adjacency matrix, and the corresponding position of each circRNA-disease association pair in the circRNA-disease adjacency matrix stores the corresponding association probability.

[0075] Step S120: The multi-view contrastive learning DCFR-CDA model is trained using the training set. A loss function is introduced during training. Training ends when the value of the loss function is stable, resulting in a circRNA-disease association prediction model. The multi-view contrastive learning DCFR-CDA model includes: a comprehensive association matrix generation unit, which processes the circRNA-disease adjacency matrix by combining self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix; an embedding matrix generation unit, which performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain a circRNA-disease association pair node embedding matrix; a reconstruction unit, which reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; and a prediction unit, which processes the composite isomorphic graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs.

[0076] Step S130: Test the circRNA-disease association prediction model using the test set, optimize the model parameters, and output the circRNA-disease association prediction model that meets the test evaluation requirements.

[0077] This invention discloses a method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance using dynamic contrastive sampling. It combines matrix factorization, multi-view feature fusion, isomorphic graph reconstruction, aggregation, and dynamic sampling strategies to form a multi-view contrastive learning (DCFR-CDA) model. The circRNA-disease association prediction model is obtained by training the DCFR-CDA model. This model uses linear low-rank factorization and nonlinear feature decomposition to perform multi-view fusion embedding processing on the comprehensive association matrix, obtaining a circRNA-disease association pair node embedding matrix. This matrix is ​​compatible with multi-view embedding of global and local patterns. Reconstructing the circRNA-disease association pair node embedding matrix yields a corresponding composite isomorphic graph, simplifying the graph structure, reducing the number of parameters, and improving feature aggregation and discrimination efficiency, thus providing effective support for accurate association prediction.

[0078] In step S110 above, the construction of the circRNA-disease adjacency matrix includes:

[0079] Step S111: Collect human circRNAs and diseases from public databases and remove irrelevant or duplicate data;

[0080] Step S112: Create a circRNA-disease adjacency matrix with circRNAs as rows and diseases as columns. ;

[0081] Step S113: Store the association probability of circRNA-disease association pairs in the corresponding positions of the circRNA-disease adjacency matrix, and store empty spaces in other positions. For example, an associated circRNA-disease association pair is set to 1 in the matrix, and otherwise it is set to 0.

[0082] It should also be noted that the circRNA-disease adjacency matrices in both the training and test sets can be obtained from four major public databases: circ2Disease, circR2Disease, circRNADisease, and LncRNADisease. Taking the construction of the training set as an example, it includes:

[0083] Human circRNAs and diseases (i.e., circRNA-disease association data) were collected from four major public databases: circ2Disease, circR2Disease, circRNADisease, and LncRNADisease. Irrelevant or duplicate data were removed to form the dataset shown in Table 1.

[0084] Table 1 Dataset

[0085]

[0086] (2) Form a circRNA-disease adjacency matrix with circRNA as row m and disease as column n. ;

[0087] (3) In the circRNA-disease adjacency matrix, the associated circRNA-disease pairs are set to 1, and the unassociated pairs are set to 0.

[0088] In step S130 above, the following six evaluation metrics can be selected: AUC, AUPR, Accuracy, F1, Recall, and Precision. The definitions of each metric are as follows:

[0089]

[0090] Wherein, TP is the number of true positives, i.e. the number of samples correctly predicted as positive; TN is the number of true negatives, i.e. the number of samples correctly predicted as negative; FP is the number of false positives, i.e. the number of negative samples incorrectly predicted as positive; and FN is the number of false negatives, i.e. the number of positive samples incorrectly predicted as negative.

[0091] AUC, or the area under the Receiver Operating Characteristic (ROC) curve, represents the overall measure of the model's ability to sort positive and negative samples. It ranges from 0 to 1, with higher values ​​indicating stronger model discrimination. AUPR, or the area under the Precision-Recall curve, primarily reflects the model's performance in scenarios with extremely imbalanced positive and negative samples; a higher AUPR indicates better performance when identifying sparse positive samples.

[0092] Example 2: As shown in the attached document Figure 3 As shown, based on the above embodiments, the preferred specific structure and corresponding processing of the multi-view contrastive learning DCFR-CDA model include:

[0093] (I) The comprehensive association matrix generation unit combines self-knowledge fusion and a weighted K-nearest neighbor knowledge network to process the circRNA-disease adjacency matrix and generate a comprehensive association matrix, as shown in the appendix. Figure 3 As shown in Step 1, it includes:

[0094] (1) Obtain the circRNA-disease adjacency matrix Disease DO similarity in the disease section GIP kernel similarity The weighted average of the disease similarity matrix is ​​obtained. The expression is as follows:

[0095]

[0096] (2) Obtain the functional similarity of the circRNA portion in the circRNA-disease adjacency matrix. Sequence similarity Similarity to GIP kernel The weighted average circRNA comprehensive similarity matrix is ​​obtained. The expression is as follows:

[0097]

[0098]

[0099] The two types of comprehensive similarity mentioned above reflect prior knowledge in terms of biological function and network interaction, respectively. They can be used as soft labels for subsequent self-knowledge fusion methods SKF and weighted K-nearest neighbor knowledge network WKNNK to enhance the model's ability to perceive potential associations.

[0100] (3) Using the self-knowledge fusion SKF and weighted k-nearest neighbor knowledge network WKNNK methods, respectively, for , and Through integration and dissemination, a comprehensive association matrix is ​​ultimately generated, as follows:

[0101] First, the circRNA comprehensive similarity matrix is... Disease comprehensive similarity matrix As the initial similarity matrix input between circRNA and disease (i.e., Sr) (1) =MixCS, Sd (1) =MixDS), and define a sample similarity metric based on Hamming distance, as follows:

[0102] For circRNA (row vector):

[0103]

[0104] For diseases (column vector):

[0105]

[0106] Then, SKF uses a selective update strategy to iteratively update the circRNA similarity matrix and the disease similarity matrix separately. Taking the circRNA similarity matrix update as an example, the update formula is as follows:

[0107]

[0108] in, This is the current similarity matrix. yes The similarity transition probability matrix is ​​row-normalized, where C is the number of similarity matrices. Finally, the fused similarity matrix is ​​defined as:

[0109]

[0110] Next, the similarity matrix after disease fusion is defined as Based on the fusion similarity, WKNNK is used for bidirectional propagation of association information. For any circRNA i, the propagation score for all diseases is as follows:

[0111]

[0112] in, The circRNA dimension normalization factor; As a propagation attenuation factor, the farther away the neighbor, the smaller its propagation contribution; This is the index of the t-th neighboring circRNA that is most similar to circRNA i.

[0113] For any disease j, the transmission score of all circRNAs is:

[0114]

[0115] in, Normalization factor for disease dimension; This is the index of the t-th neighboring disease that is most similar to disease j.

[0116] Finally, the comprehensive correlation matrix yes and Weighted fusion;

[0117]

[0118] (ii) Embedded matrix generation unit: The comprehensive association matrix is ​​processed using a linear low-rank decomposition method and a nonlinear eigenvalue decomposition method for multi-view fusion embedding, resulting in the circRNA-disease association pair node embedding matrix, as shown in the appendix. Figure 3 As shown in Step 2, it includes:

[0119] (1) Using the Singular Value Decomposition (SVD) method to analyze the comprehensive correlation matrix By performing a low-rank approximation (as shown in the formula below), the corresponding embedding matrix is ​​obtained. , ;

[0120]

[0121]

[0122] in, , , Let represent the first k singular values ​​and their corresponding left and right singular vectors, respectively. Here, k = min(m, n) is used to ensure the effectiveness and low rank of the decomposition.

[0123] (2) Using the Non-negative Matrix Factorization (NMF) method to analyze the comprehensive correlation matrix Perform an approximate decomposition (as shown in the formula below), preserve the non-negativity property, and obtain the corresponding embedding matrix. , ;

[0124]

[0125]

[0126] (3) Using the Multilayer Perceptron Matrix Factorization (MLPF) method to analyze the comprehensive correlation matrix Latent feature extraction is performed (formula shown below), and circRNA and disease-related features are encoded separately to obtain the corresponding embedding matrices. , ;

[0127] Specifically, the comprehensive correlation matrix will be used. As input to the neural network, latent features are extracted through MLP mapping, and circRNA and disease side are encoded respectively;

[0128]

[0129]

[0130] in, and These are the MLP coding networks for RNA and the disease side, respectively.

[0131] (4) Using the one-dimensional convolution matrix factorization method CNNF to analyze the comprehensive correlation matrix Two-dimensional local feature extraction is performed (the formula is shown below), and the circRNA and the disease side are encoded respectively to obtain the corresponding embedding matrix;

[0132]

[0133]

[0134] in, and These are convolutional coding networks for the RNA and disease sides, respectively.

[0135] (5) Concatenate all embedding matrices row by row to obtain the circRNA-disease association pair node embedding matrix;

[0136] .

[0137] in, All are embedding matrices obtained using the aforementioned linear low-rank decomposition method and nonlinear eigenvalue decomposition method.

[0138] The above-mentioned linear low-rank decomposition methods (SVD, NMF) and nonlinear eigenvalue decomposition methods (MLPF, CNNF) are used for multi-view embedding to extract the latent features of nodes. Specifically, linear decomposition mainly captures the global principal components in the data, that is, global linear information, while nonlinear decomposition aims to mine local high-order structural information, that is, nonlinear local expression, so that the node embedding in the matrix has both global linear information and nonlinear local expression, which can improve the discriminative ability of node features.

[0139] (III) Reconstruction Unit: The embedding matrix of circRNA-disease association pairs is reconstructed to obtain the corresponding composite isomorphic graph. Specifically,

[0140] In circRNA-disease association studies, due to the high sparsity of networks and the heterogeneity of node types, heterogeneous graph modeling is often used to integrate multi-source biological information and complex interactions. However, when dealing with large-scale entities and multiple types of relationships, heterogeneous graph neural networks require the design of message passing mechanisms and parameter matrices for different node and edge types, leading to a significant increase in the number of model parameters. Furthermore, information flow between heterogeneous nodes depends on type mapping or meta-path aggregation, which can easily introduce alignment errors and amplify noise during multi-hop propagation, thereby weakening the ability to model real biological associations.

[0141] As attached Figure 3 Step 3, Appendix Figure 4 As shown, this embodiment proposes an isomorphic graph reconstruction strategy based on composite association pairs. Each pair of circRNA-disease association pairs is mapped to a high-order composite node in the isomorphic graph. By splicing the multi-source attributes of circRNA and disease, node-level high-order semantic recoding is achieved. On this basis, isomorphic edges are constructed according to the relationship between "co-RNA" and "co-disease". At the same time, self-loops are added to each node to retain its own characteristics, forming a composite isomorphic graph. This eliminates the dependence on the original sparse association edges and enhances network connectivity.

[0142] (iv) Prediction Unit: The composite isomorphism graph is processed using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs, as shown in the appendix. Figure 3 As shown in Step 4, it includes:

[0143] (1) A two-layer GAT is used to perform multi-head attention aggregation on the neighborhood information of high-order composite nodes in the composite isomorphic graph in order to capture key neighborhood information. Specifically:

[0144] First-level GAT:

[0145]

[0146] Second layer GAT:

[0147]

[0148] in, Here, N represents the node features, H represents the batch size, H represents the number of attention heads, and D represents the dimension of each attention head. To Flattening, d out For output dimensions.

[0149] (2) Based on the aggregated node embedding features, the sampling probability of positive and negative samples for each anchor point is dynamically determined using feature and structural similarity, resulting in a set of positive samples, specifically including:

[0150] First, for the original graph structure, we define the trainable parameters for each edge. By using Gumbel-Softmax sampling, the existence probability of each edge is obtained, thereby constructing a differentiable adjacency matrix;

[0151]

[0152] To ensure the undirectedness and symmetry of the graph structure, the adjacency matrix is ​​symmetricized:

[0153]

[0154] Based on this, the PPR method is used to perform multi-step iterative propagation of graph structure information, initializing each node as a unit one-hot vector, i.e. Then, k iterations are performed according to the following recursive formula:

[0155]

[0156] in, It is the identity matrix. Here, deg represents the jump probability, and deg is the node degree. After k iterations, the structural similarity matrix is ​​obtained. .

[0157] To fully integrate node feature information, a node feature similarity matrix is ​​introduced. The following weighted fusion method is adopted:

[0158]

[0159] in, The weighting coefficients for feature similarity. This is the mean scaling factor, used to standardize the units of measurement. This is a temperature parameter used to improve the discriminative power of fusion similarity. The larger the Sij value, the more similar node i is to node j.

[0160] Taking positive sample sampling at anchor point i as an example, firstly, a candidate sample set C is selected based on the anchor point label. i :

[0161]

[0162] Then, for the candidate set C i For all nodes j within the node, calculate their sampling probability as follows:

[0163]

[0164] Next, from candidate set C i In the above, Npos nodes are randomly sampled without replacement according to probability Pij to obtain the set of positive samples:

[0165]

[0166] If the candidate set has fewer than Npos elements, it is padded with zeros. Similarly, for negative sample sampling at each anchor point i, a negative sample candidate set is selected from all nodes: The sampling logic is the opposite of that for positive samples; that is, the less similar a negative sample is to the anchor point, the greater the probability of sampling it.

[0167] (3) The feature representation of nodes is optimized by using triplet comparison loss, and a multilayer perceptron is used to predict the association probability between circRNA and disease association pairs. Specifically:

[0168] For each anchor point i, let its embedding be denoted as . The positive sample set is The negative sample set is The loss of a single anchor triple is defined as follows:

[0169]

[0170]

[0171] in, For anchor point embedding, This represents the mean of the embeddings of all positive sample nodes at anchor point i; For ReLU operations; The cosine similarity is calculated after L2 normalization of the embedding vectors, where m is the margin hyperparameter, typically set to 0.2.

[0172] We use MLP to map high-dimensional embeddings to the association probability space, enabling accurate prediction of each circRNA-disease association pair.

[0173]

[0174] Example 3: The model constructed in the above examples is verified as follows:

[0175] (a) Validation of model prediction capability

[0176] First, the complete model was tested on four public datasets: circ2Disease, circR2Disease, circRNADisease, and lncRNADisease. The results showed that the multi-view contrastive learning DCFR-CDA model performed well in key metrics such as AUC, Accuracy, and F1. The AUCs of the four datasets reached 0.9996, 0.9925, 0.9999, and 0.9974, respectively, demonstrating excellent robustness and generalization ability. Based on this, two ablation experiments were designed: (1) random sampling was used to replace dynamic positive and negative sample sampling to evaluate the impact of the sampling mechanism on the model's discriminative ability; (2) certain branches in the feature decomposition module were removed, retaining only SVD+CNNF, SVD+MLPF, NMF+MLPF, or NMF+CNNF to verify the effectiveness of multi-view feature fusion (SVD, NMF, MLPF, and CNNF). Specific results are attached. Figure 5 As shown in Table 2, Appendix Figure 5 Figures (a), (b), (c), and (d) show the ROC curves of ablation experiments on the four datasets: circ2Disease, circR2Disease, circRNADisease, and lncRNADisease, respectively.

[0177] Table 2 Model Prediction Performance

[0178]

[0179] The results show that removing dynamic sampling significantly reduces both the F1 score and precision, indicating that the dynamic sampling mechanism from the perspective of structure and feature fusion can effectively filter out more discriminative positive and negative samples. In feature decomposition ablation, the performance degradation of the NMF+CNNF and NMF+MLPF combinations is the most significant, demonstrating that the fusion of linear and nonlinear perspectives is crucial for improving representation capabilities. Overall, the model constructed in this invention exhibits stable performance advantages in all ablation comparison experiments, highlighting the core contributions of the dynamic sampling mechanism and multi-view matrix factorization fusion to model performance.

[0180] (II) Parameter Analysis

[0181] To systematically evaluate the impact of learning rate (lr), hidden layer dimension (hidden_dim), and GAT attention head number (heads_num) on model performance, this embodiment performed a grid search on the circ2Disease dataset for these parameters and implemented five-fold cross-validation for each parameter combination to evaluate its performance. Specific analysis results are attached. Figure 6As shown in Figures (a), (b), and (c), the performance of the multi-view contrastive learning DCFR-CDA model is demonstrated under different learning rate, GAT head number, and hidden layer dimension settings. (See attached figure.) Figure 6 It can be observed that when the learning rate is reduced from 0.5 to 0.01, the model performance improves significantly: a higher learning rate leads to excessively large update steps, causing oscillations around the optimal solution and difficulty in convergence, while a lower learning rate results in slow convergence and a tendency to get trapped in local optima; therefore, 0.01 strikes a balance between speed and stability. A hidden layer dimension of 128 can fully express complex features while avoiding the risk of overfitting; lower dimensions lack expressive power, while higher dimensions introduce redundant parameters. Furthermore, when the number of attention heads in GAT is set to 16, the multi-head mechanism enriches the aggregation of neighbor information while keeping computational overhead controllable, thus achieving optimal performance. Based on the above analysis, the optimal hyperparameter combination determined in this invention is: lr=0.01, hidden_dim=128, heads_num=16.

[0182] (III) Comparative Experiment

[0183] To evaluate the advancement of this invention under different data environments, this embodiment conducted comparative experiments with recent state-of-the-art methods on three publicly available datasets: circ2Disease, circR2Disease, and circRNADisease. Each model employed five-fold cross-validation, and F1, AUC, and Accuracy were used as evaluation metrics. Table 3 shows the performance of each method under the same conditions on the aforementioned datasets.

[0184] Table 3 Comparative Experiments

[0185]

[0186] Comparative experiments show that most existing models (such as MDGF-MCEC, MAGCDA, and MGRCDA) fail to effectively integrate linear and nonlinear features from multiple perspectives, or simply employ random sampling strategies, resulting in limited prediction performance. The multi-view contrastive learning DCFR-CDA model proposed in this invention effectively overcomes the shortcomings of the aforementioned methods in terms of single feature representation and insufficient negative sample selection by introducing dynamic contrastive sampling and a multi-view feature resonance mechanism, thus demonstrating superior performance on datasets such as circ2Disease, circRNADisease, and circR2Disease.

[0187] (iv) Case Analysis

[0188] To verify the application value of the multi-view contrastive learning DCFR-CDA model in real disease research, this embodiment conducted case analyses on three diseases: lung cancer, liver cancer, and colorectal cancer. First, all known associations with the target cancer were removed from the circR2Disease dataset, and the model was trained using the remaining data. Subsequently, the model was predicted for the cancer with potential related circRNAs, and the top 30 candidate molecules were searched in the literature to evaluate the reliability and biological significance of the recommendation results. The specific results are shown in Tables 4, 5, and 6. As shown in the table, among the top 30 candidate circRNAs for the three diseases, 23 (lung cancer), 23 (liver cancer), and 24 (colorectal cancer) have been documented in the literature. Specifically, hsa_circ_0000378 participates in lung cancer development by regulating miR-21; circPTK2 (hsa_circ_0003221) regulates the miR-942 / TRIM16 axis to inhibit lung cancer cell proliferation; circ_SLC8A1 regulates the expression of human liver CYP3A4 through hsa-miR-27a; and circ-ZNF609 promotes colorectal cancer cell migration by downregulating Gli1 through microRNA-150. Although some candidate circRNAs have not yet been confirmed in the literature, they still provide potential targets for subsequent experimental validation.

[0189] Table 4 Lung cancer

[0190]

[0191] Table 5 Liver cancer

[0192]

[0193] Table 6 Colorectal Canser

[0194]

[0195] Example 4: As shown in the appendix Figure 7 As shown, this embodiment discloses a circRNA-disease association prediction method, including:

[0196] Step S210: Obtain the circRNA-disease association pairs to be predicted;

[0197] Step S220: Input the circRNA-disease association pair to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease association pair to be predicted. The circRNA-disease association prediction model is constructed by the circRNA-disease association prediction model construction method based on the linear and nonlinear feature resonance of dynamic contrast sampling.

[0198] Example 5: As shown in the attached document Figure 8 As shown, this embodiment discloses a device for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance using dynamic contrast sampling, comprising:

[0199] The dataset acquisition unit acquires the training set and the test set. Both the training set and the test set include a circRNA-disease adjacency matrix. The corresponding position of each circRNA-disease association pair in the circRNA-disease adjacency matrix stores the corresponding association probability.

[0200] The training unit trains the multi-view contrastive learning DCFR-CDA model using the training set. A loss function is introduced during training, and training ends when the value of the loss function stabilizes, resulting in the circRNA-disease association prediction model. The multi-view contrastive learning DCFR-CDA model includes: a comprehensive association matrix generation unit, which processes the circRNA-disease adjacency matrix by combining self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix; an embedding matrix generation unit, which performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix; a reconstruction unit, which reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; and a prediction unit, which processes the composite isomorphic graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs.

[0201] The testing unit uses a test set to test the circRNA-disease association prediction model, optimizes the model parameters, and outputs a circRNA-disease association prediction model that meets the test evaluation requirements.

[0202] The specific implementation steps of each unit in this embodiment are the same as those described in Embodiments 1 and 2, and will not be repeated here.

[0203] Example 6: As shown in the appendix Figure 9 As shown, this embodiment discloses a circRNA-disease association prediction device, comprising:

[0204] The unit for acquiring data to be predicted acquires the circRNA-disease association pairs to be predicted.

[0205] The prediction unit inputs the circRNA-disease association pair to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease association pair to be predicted. The circRNA-disease association prediction model is constructed by the circRNA-disease association prediction model construction method based on the linear and nonlinear feature resonance of dynamic contrast sampling.

[0206] Example 7: This example discloses an electronic device, including a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement a method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance of dynamic contrast sampling, or a circRNA-disease association prediction method.

[0207] The processor described above can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, portable hard drives, magnetic disks, or optical disks.

[0208] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0209] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0211] The above content is only a specific embodiment of the present invention, which has strong adaptability and implementation effect. However, the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention. Therefore, equivalent changes made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using dynamic contrastive sampling, characterized in that, include: Obtain training and test sets, both of which include circRNA-disease adjacency matrices. The corresponding positions of each circRNA-disease association pair in the circRNA-disease adjacency matrix store the corresponding association probabilities. The DCFR-CDA model of multi-view contrastive learning was trained using the training set. A loss function was introduced during training, and training was terminated when the value of the loss function stabilized, resulting in a circRNA-disease association prediction model. The DCFR-CDA model of multi-view contrastive learning includes a comprehensive association matrix generation unit, which combines self-knowledge fusion and a weighted K-nearest neighbor knowledge network to process the circRNA-disease adjacency matrix and generate a comprehensive association matrix. The embedding matrix generation unit performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix; the reconstruction unit reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; the prediction unit processes the composite isomorphic graph using aggregation and positive and negative sample sampling to predict the association probability of circRNA-disease association pairs. The circRNA-disease association prediction model was tested using a test set, the model parameters were optimized, and a circRNA-disease association prediction model that meets the test evaluation requirements was output.

2. The method for constructing a circRNA-disease association prediction model based on linear and nonlinear feature resonance using dynamic contrast sampling as described in claim 1, characterized in that, The construction of the circRNA-disease adjacency matrix includes: Human circRNAs and diseases were collected from public databases, and irrelevant or duplicate data were removed; A circRNA-disease adjacency matrix is ​​formed by using circRNAs as rows and diseases as columns. The association probabilities of circRNA-disease association pairs are stored in the corresponding positions of the circRNA-disease adjacency matrix, and empty spaces are stored in other positions.

3. The method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using dynamic contrast sampling as described in claim 1 or 2, characterized in that, The comprehensive association matrix generation unit processes the circRNA-disease adjacency matrix by combining self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix, including: Obtain the disease DO similarity and GIP kernel similarity of the disease part in the circRNA-disease adjacency matrix, and then weight them to obtain the comprehensive disease similarity matrix; The functional similarity, sequence similarity, and GIP core similarity of the circRNA portion in the circRNA-disease adjacency matrix are obtained, and a weighted average is used to obtain the circRNA comprehensive similarity matrix. The disease comprehensive similarity matrix, circRNA comprehensive similarity matrix, and circRNA-disease adjacency matrix are integrated and propagated using self-knowledge fusion and weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix.

4. The method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using dynamic contrast sampling as described in claim 1 or 2, characterized in that, The embedding matrix generation unit performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix, including: The composite correlation matrix is ​​approximated in low rank using singular value decomposition (SVD) to obtain the corresponding embedding matrix. The comprehensive correlation matrix is ​​approximated by nonnegative matrix factorization (NMF) to obtain the corresponding embedding matrix. The latent features of the comprehensive correlation matrix were extracted using the multilayer perceptron matrix factorization (MLPF), and the circRNA and disease side were encoded separately to obtain the corresponding embedding matrix. The one-dimensional convolutional matrix factorization CNNF is used to extract two-dimensional local features from the comprehensive correlation matrix, and the circRNA and the disease side are encoded separately to obtain the corresponding embedding matrix. Concatenate all embedding matrices row by row to obtain the circRNA-disease association pair node embedding matrix; 。 in, All are embedding matrices obtained using the linear low-rank decomposition method and the nonlinear eigenvalue decomposition method.

5. The method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using dynamic contrast sampling as described in claim 1 or 2, characterized in that, The reconstruction unit reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph, including: Map each circRNA-disease association pair in the circRNA-disease association pair embedding matrix to a higher-order composite node in the isomorphism graph. Adjacent edges are constructed based on shared circRNAs or shared diseases to form a composite isomorphic graph.

6. The method for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using dynamic contrast sampling according to any one of claims 1 to 5, characterized in that, The prediction unit processes the composite isomorphism graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs, including: Multi-head attention aggregation is performed using neighborhood information from two-layer GAT high-order composite nodes; Based on the aggregated node embedding features, the sampling probability of positive and negative samples for each anchor point is dynamically determined using feature and structural similarity, thus obtaining a set of positive samples. The feature representation of nodes is optimized by using triplet comparison loss, and a multilayer perceptron is used to predict the association probability of circRNA and disease association pairs.

7. A circRNA-disease association prediction method, characterized in that, include: Obtain the circRNA-disease association pairs to be predicted; The circRNA-disease association pair to be predicted is input into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease association pair to be predicted, wherein the circRNA-disease association prediction model is constructed by the method described in any one of claims 1 to 6.

8. An apparatus for constructing a circRNA-disease association prediction model based on linear and nonlinear eigenre resonance using the method described in any one of claims 1 to 6, characterized in that, include: The dataset acquisition unit acquires the training set and the test set. Both the training set and the test set include a circRNA-disease adjacency matrix. The corresponding position of each circRNA-disease association pair in the circRNA-disease adjacency matrix stores the corresponding association probability. The training unit trains the multi-view contrastive learning DCFR-CDA model using the training set. A loss function is introduced during training, and training ends when the value of the loss function stabilizes, resulting in the circRNA-disease association prediction model. The multi-view contrastive learning DCFR-CDA model includes: a comprehensive association matrix generation unit, which processes the circRNA-disease adjacency matrix by combining self-knowledge fusion and a weighted K-nearest neighbor knowledge network to generate a comprehensive association matrix; an embedding matrix generation unit, which performs multi-view fusion embedding processing on the comprehensive association matrix using linear low-rank decomposition and nonlinear eigenvalue decomposition methods to obtain the circRNA-disease association pair node embedding matrix; a reconstruction unit, which reconstructs the circRNA-disease association pair node embedding matrix to obtain the corresponding composite isomorphic graph; and a prediction unit, which processes the composite isomorphic graph using aggregation and positive / negative sample sampling to predict the association probability of circRNA-disease association pairs. The testing unit uses a test set to test the circRNA-disease association prediction model, optimizes the model parameters, and outputs a circRNA-disease association prediction model that meets the test evaluation requirements.

9. A circRNA-disease association prediction device applying the method as described in claim 7, characterized in that, include: The unit for acquiring data to be predicted acquires the circRNA-disease association pairs to be predicted. The prediction unit inputs the circRNA-disease association pair to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease association pair to be predicted, wherein the circRNA-disease association prediction model is constructed by the method described in any one of claims 1 to 6.

10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the method as claimed in any one of claims 1 to 6 or to implement the steps as claimed in claim 7.

Images