Method, device and equipment for predicting piRNA-disease relevance based on multilateral type graph convolutional network of heterogeneous graph
By constructing heterogeneous graph and multilateral type graph convolution networks, combined with attention mechanism, the problem of low prediction accuracy of piRNA-disease association in the existing technology is solved, and higher prediction accuracy and effect are achieved.
Patent Information
- Application Number
- CN202411188060.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-17
AI Technical Summary
The existing piRNA-disease association prediction methods fail to make full use of various similarity data and neighbor node information, resulting in unsatisfactory prediction results.
Using a multilateral type graph convolution network based on heterogeneous graphs, we learn the isomorphic embedding features and heterogeneous embedding features of piRNA and disease by constructing piRNA-like isomers and disease-like isomers, and use attention mechanisms to fuse the features, and finally use the Multilayer perceptron predicts piRNA-disease association.
It effectively improves the accuracy of piRNA-disease association prediction, and achieves an AUC of 0.9292 and an AUPR of 0.5006, which is better than the previous method.
Smart Images
Figure CN119153111A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bioinformatics, and in particular to a method, device and equipment for predicting piRNA-disease association based on a heterogeneous graph-based multi-edge graph convolutional network. Background Art
[0002] piRNAs are a unique class of small non-coding RNAs that are primarily expressed in germ cells. Because they play an important role in various biological processes and are associated with numerous diseases, including cancer, neurodegenerative diseases, and other geriatric diseases. piRNAs are approximately 24-35 nucleotides in length and bind to Piwi-subfamily Argonaute proteins, playing a key role in transposable element silencing, genome defense, and histone modification. Therefore, discovering the relationship between piRNAs and diseases is crucial for medical researchers to gain a deeper understanding of the complex pathological processes of diseases. In traditional prediction methods, the correlation between piRNAs and diseases is inferred through biological experiments, which have the disadvantages of special equipment requirements, complex operations, long experimental cycles, and high costs. Therefore, an efficient and accurate computational method can be designed to provide guidance for biological experiments, making the discovery of piRNA-disease associations more efficient.
[0003] For example, Zhang et al. disclosed a piRNA-disease association prediction method iPiDA-LTR in PLoS Computational Biology in 2022. iPiDA-LTR combines the component method and the learning ranking method to identify the association between known piRNAs and diseases. Zhang et al. disclosed a piRNA-disease association prediction method iPiDA-GCN in PLoS Computational Biology in 2022. iPiDA-GCN uses two GCNs to extract features of piRNAs and diseases from the piRNA-disease interaction network and two similarity networks, and uses a fully connected network and internal production as output modules to predict piRNA-disease association scores. Chen et al. disclosed a piRNA-disease association prediction method PUTransGCN in Briefings in Bioinformatics in 2024. PUTransGCN uses an attention mechanism to automatically adjust the weight parameters of aggregated heterogeneous node features. To address the imbalanced dataset problem, a combined positive unlabeled learning (PUL) method consisting of PU bagging, two-step and spy techniques is adopted to select reliable negative associations, and a heterogeneous graph convolutional network is used to discover potential piRNA-disease associations.
[0004] Although the above methods are effective in studying piRNA-disease associations, there are still many shortcomings. On the one hand, some methods do not fully utilize the information useful for piRNA-disease prediction and do not comprehensively consider various similarity data; on the other hand, some methods fail to aggregate neighbor node information well when learning piRNA and disease embedding features, resulting in unsatisfactory prediction results.
[0005] In view of this, it is necessary to provide a new method to solve the above technical problems. Summary of the invention
[0006] The technical problem to be solved by the present invention is to provide a method for predicting piRNA-disease association based on a multi-edge graph convolutional network of heterogeneous graphs, which can effectively improve the accuracy of piRNA-disease association prediction, thereby solving the technical problem of low prediction accuracy in the prior art. The technical solution is as follows:
[0007] The present invention provides a method for predicting piRNA-disease association based on a heterogeneous graph-based multi-edge graph convolutional network, comprising the following steps:
[0008] S1, using piRNA similarity matrix and disease similarity matrix to construct piRNA similarity isomorphism graph and disease similarity isomorphism graph, and using piRNA-disease association relationship data to construct piRNA-disease association heterogeneity graph;
[0009] S2, using a graph convolutional neural network to learn the isomorphic embedding features of piRNAs and the isomorphic embedding features of diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and using a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and the heterogeneous embedding features of diseases from the piRNA-disease association heterogeneous graph;
[0010] S3, using an attention mechanism to fuse the isomorphic embedding features of the piRNA and the isomorphic embedding features of the disease, as well as the heterogeneous embedding features of the piRNA and the heterogeneous embedding features of the disease, to obtain a fused embedding feature of the piRNA and a fused embedding feature of the disease;
[0011] S4, horizontally splicing the fusion embedding feature of the piRNA and the fusion embedding feature of the disease to form a piRNA-disease pair feature, and using a multi-layer perceptron to predict the association of the piRNA-disease pair.
[0012] Furthermore, in step S1, data on the association between known human piRNAs and diseases are downloaded from the MNDR 4.0 database, a piRNA-disease association matrix is constructed and converted into the piRNA-disease association heterogeneous graph.
[0013] Furthermore, in step S1, the piRNA similarity isomorphism graph is constructed using piRNA sequence similarity and piRNA Gaussian kernel similarity; and the disease similarity isomorphism graph is constructed using disease semantic similarity and disease Gaussian kernel similarity.
[0014] Furthermore, constructing the piRNA similarity isomorphism graph comprises the following steps:
[0015] First, the piRNA p was calculated using the Smith–Waterman algorithm. i and p j The sequence similarity between them was calculated and the piRNA sequence similarity matrix was obtained. The formula is as follows:
[0016]
[0017] Among them, SW(p i ,p j ) represents the sequence alignment score between the i-th and j-th piRNAs based on the Smith-Waterman algorithm;
[0018] Second, calculate piRNA p i and p j The Gaussian kernel similarity between them is calculated and the piRNA Gaussian kernel similarity matrix is obtained. The formula is as follows:
[0019] PGS(p i ,p j )=exp(-γ p ||IP(p i )-IP(p j )|| 2 ),
[0020]
[0021] Among them, IP(p i ) represents the piRNA p in the piRNA-disease association matrix i The corresponding row, parameter γ p Used to control kernel bandwidth; parameter γ, p It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N p represents the number of piRNAs in the piRNA-disease association matrix;
[0022] Third, the piRNA sequence similarity matrix and the piRNA Gaussian kernel similarity matrix were averaged to calculate the piRNA m i and m jThe integrated similarity between them is then used to obtain the piRNA integrated similarity matrix The calculation formula is as follows:
[0023]
[0024] Fourth, the k most similar neighbors of each piRNA were selected in the piRNA ensemble similarity matrix, where k = 2; based on this, the final piRNA ensemble similarity matrix was obtained and converted into a piRNA similarity isomorphism graph, which was represented by a binary matrix The formula is as follows:
[0025]
[0026] in, Represents the matrix A p The element in row i and column j, is piRNA p i Including p i The set of k nearest neighbors, N k (p i ) is p i k nearest neighbors.
[0027] Furthermore, constructing the disease similarity isomorphism graph comprises the following steps:
[0028] First, we obtain the directed acyclic graph (DAG) of diseases from the MeSH database, where nodes represent diseases and connections between diseases describe the relationships between them; and calculate the disease semantic similarity matrix based on DAG.
[0029]
[0030] in, Indicates disease i The semantic value of T(d i ) indicates that node d i and the node set consisting of its ancestor nodes; Indicates disease i For disease i The semantic contribution of is calculated as follows:
[0031]
[0032] Where Δ is the semantic contribution attenuation factor, which is set to 0.5;
[0033] Second, calculate the disease d i and d j The Gaussian kernel similarity between them is calculated, and the disease Gaussian kernel similarity matrix is obtained. The formula is as follows:
[0034] DGS(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 ),
[0035]
[0036] Among them, IP(d i ) represents the disease d in the piRNA-disease association matrix i The corresponding column, parameter γ d Used to control kernel bandwidth; parameter γ' d It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N d represents the number of diseases in the piRNA-disease association matrix;
[0037] Third, the disease semantic similarity matrix and the disease Gaussian kernel similarity matrix are averaged to calculate the disease d i and d j The integrated similarity between them is then used to obtain the disease integrated similarity matrix The calculation formula is as follows:
[0038]
[0039] Fourth, the k most similar neighbors of each disease are selected in the disease integrated similarity matrix, where k = 2; based on this, the final disease integrated similarity matrix is obtained and converted into a disease similarity isomorphism graph, which is represented by a binary matrix The formula is as follows:
[0040]
[0041] in, Represents the matrix A d The element in row i and column j, It is a disease i Included i The set of k nearest neighbors, N k (d i ) is d i k nearest neighbors.
[0042] Further, step S2 includes the following steps:
[0043] S21, GCN is used to learn the isomorphic embedding features of piRNA and disease from the piRNA similarity isomorphism graph and the disease similarity isomorphism graph respectively, the formula is as follows:
[0044]
[0045]
[0046] in, and are the information of piRNA and disease at layer l, and are the weight matrices of the GCN layer l where piRNA and disease are located, ReLU is the activation function, D p and D d A p and A d The diagonal matrix of piRNA and disease is embedded in the last layer as and
[0047] S22, using a multi-edge-type graph convolutional network based on a heterogeneous graph to learn heterogeneous embedding features of piRNAs and diseases from a piRNA-disease association heterogeneous graph; in the piRNA-disease association heterogeneous graph, known and unknown piRNA-disease associations are treated as different edge types, and a separate processing channel is assigned to each edge type t∈{0,1}; specifically, each edge type of the graph convolution can be viewed as a form of message passing, in which vector-valued messages are passed and transformed on the edges of the graph; in the above model, a specific transformation is assigned to each edge type, thereby obtaining a disease d j to piRNA p i Edge type message m j→i,t The format is as follows:
[0048]
[0049] Among them, N(p i ) represents node p i The neighbor set, N(d j ) represents node d j The neighbor set, W t is the edge type specific parameter matrix, x j is the disease node d j The feature vector of piRNA p i To disease j Message i→j,t is processed in a similar way; after the message passing step, the edge type t is computed by t (d j) and accumulate the results for each edge type into a single vector representation to accumulate the incoming messages for each node:
[0050]
[0051] Where sum represents the accumulation operation, σ represents the tanh activation function, and in order to obtain the final representation of piRNA, a linear operator is used to operate on the intermediate output z i To transform:
[0052] z i =Wh i ,
[0053] Disease Embedded j A similar method is used for calculation; in the linear operator, the parameter matrix W of piRNA is the same as that of disease. Applying the above transformation to all nodes of the piRNA-disease association heterogeneous graph, we can obtain piRNA and diseases Final representation in a heterogeneous graph.
[0054] Further, step S3 includes the following steps:
[0055] S31, the corresponding importance of drug embedding and disease embedding is adaptively learned using the attention mechanism as follows:
[0056]
[0057]
[0058] Among them, att is a neural network that performs attention operation, and Represents piRNA and disease embedding, respectively and The attention value;
[0059] S32, combine these embeddings according to these attention values to get the final piRNA embedding Z p and disease embedding Z d :
[0060]
[0061]
[0062] Further, step S4 includes the following steps:
[0063] S41, concatenate the two obtained embeddings to represent the piRNA-disease pair, and feed the embeddings of the piRNA-disease pair into an MLP for prediction:
[0064]
[0065] in, indicates the embedding of piRNA i, represents the embedding of disease j, MLP is a three-layer MLP neural network, represents the probability of association between piRNA i and disease j pairs;
[0066] S42, calculate the binary cross entropy loss between the predicted results and the true results:
[0067]
[0068] Where (i,j) represents the piRNA i-disease j pair, y ij represents the true label; considering the common semantics between homogeneous and heterogeneous graphs, consistency constraints are used to enhance their universality; for piRNA embedding, in order to capture the similarity of piRNA between different graphs, the following constraints are given:
[0069]
[0070] Similarly, calculate the disease embedding constraint L Cd , the cross entropy loss L BCE , consistency constraint L Cp and L Cd The weighted combination gives the final loss L:
[0071] L=L BCE +λL Cp +λL Cd
[0072] where λ is a hyperparameter that balances these three terms.
[0073] The present invention provides a device for predicting piRNA-disease association based on a heterogeneous graph-based multi-edge graph convolutional network, comprising:
[0074] The graph construction module uses the piRNA similarity matrix and the disease similarity matrix to construct the piRNA similarity isomorphism graph and the disease similarity isomorphism graph, and uses the piRNA-disease association relationship data to construct the piRNA-disease association heterogeneity graph;
[0075] The feature learning module uses graph convolutional neural networks to learn the isomorphic embedding features of piRNAs and diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and uses a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and diseases from the piRNA-disease association heterogeneous graph;
[0076] The attention fusion module uses the attention mechanism to fuse the homogeneous embedding features and heterogeneous embedding features of piRNA and disease respectively to obtain the fused embedding features of piRNA and disease;
[0077] The association prediction module horizontally splices the fusion embedding features of piRNA and disease to form piRNA-disease pair features, and uses a multi-layer perceptron to predict the association of piRNA-disease pairs.
[0078] The present invention discloses a device for predicting piRNA-disease associations based on a heterogeneous graph-based multi-edge graph convolutional network, characterized in that it comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the method steps described in any one of claims 1 to 8 when executing the computer program.
[0079] The method for predicting piRNA-disease association using a heterogeneous graph-based multi-edge graph convolutional network provided by the present invention has the following beneficial effects:
[0080] 1. The method for predicting piRNA-disease association based on a heterogeneous graph-based multi-edge graph convolutional network provided by the present invention makes full use of various similar data of piRNA and diseases, and fully considers the neighbor node information, effectively improving the accuracy of piRNA-disease association prediction.
[0081] 2. The method for predicting piRNA-disease association using a heterogeneous graph-based multilateral graph convolutional network provided by the present invention is evaluated using five-fold cross validation and case validation, and an AUC of 0.9292 and an AUPR of 0.5006 are achieved in the five-fold cross validation. BRIEF DESCRIPTION OF THE DRAWINGS
[0082] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0083] Figure 1 A flowchart of a method for predicting piRNA-disease association using a heterogeneous graph-based multi-edge graph convolutional network provided by the present invention. DETAILED DESCRIPTION
[0084] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention and to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the specific implementation modes of the present invention are further described below in conjunction with the accompanying drawings.
[0085] It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as there is no conflict between them.
[0086] Embodiment 1
[0087] Please refer to Figure 1 In this embodiment, a method for predicting piRNA-disease association based on a heterogeneous graph multi-type graph convolutional network comprises the following steps:
[0088] S1, using piRNA similarity matrix and disease similarity matrix to construct piRNA similarity isomorphism graph and disease similarity isomorphism graph, and using piRNA-disease association relationship data to construct piRNA-disease association heterogeneity graph;
[0089] S2, using a graph convolutional neural network to learn the isomorphic embedding features of piRNAs and the isomorphic embedding features of diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and using a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and the heterogeneous embedding features of diseases from the piRNA-disease association heterogeneous graph;
[0090] S3, using an attention mechanism to fuse the isomorphic embedding features of the piRNA and the isomorphic embedding features of the disease, as well as the heterogeneous embedding features of the piRNA and the heterogeneous embedding features of the disease, to obtain a fused embedding feature of the piRNA and a fused embedding feature of the disease;
[0091] S4, horizontally splicing the fusion embedding feature of the piRNA and the fusion embedding feature of the disease to form a piRNA-disease pair feature, and using a multi-layer perceptron to predict the association of the piRNA-disease pair.
[0092] Preferably, in step S1, data on the association between known human piRNAs and diseases are downloaded from the MNDR4.0 database, a piRNA-disease association matrix is constructed and converted into the piRNA-disease association heterogeneous graph.
[0093] Preferably, in step S1, the piRNA similarity isomorphism graph is constructed using piRNA sequence similarity and piRNA Gaussian kernel similarity; and the disease similarity isomorphism graph is constructed using disease semantic similarity and disease Gaussian kernel similarity.
[0094] Preferably, constructing the piRNA similarity isomorphism graph comprises the following steps:
[0095] First, the piRNA p was calculated using the Smith–Waterman algorithm. iand p j The sequence similarity between them was calculated and the piRNA sequence similarity matrix was obtained. The formula is as follows:
[0096]
[0097] Among them, SW(p i ,p j ) represents the sequence alignment score between the i-th and j-th piRNAs based on the Smith-Waterman algorithm;
[0098] Second, calculate piRNA p i and p j The Gaussian kernel similarity between them is calculated and the piRNA Gaussian kernel similarity matrix is obtained. The formula is as follows:
[0099] PGS(p i ,p j )=exp(-γ p ||IP(p i )-IP(p j )|| 2 ),
[0100]
[0101] Among them, IP(p i ) represents the piRNA p in the piRNA-disease association matrix i The corresponding row, parameter γ p Used to control kernel bandwidth; parameter γ, p It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N p represents the number of piRNAs in the piRNA-disease association matrix;
[0102] Third, the piRNA sequence similarity matrix and the piRNA Gaussian kernel similarity matrix were averaged to calculate the piRNA m i and m j The integrated similarity between them is then used to obtain the piRNA integrated similarity matrix The calculation formula is as follows:
[0103]
[0104] Fourth, the k most similar neighbors of each piRNA were selected in the piRNA ensemble similarity matrix, where k = 2; based on this, the final piRNA ensemble similarity matrix was obtained and converted into a piRNA similarity isomorphism graph, which was represented by a binary matrix The formula is as follows:
[0105]
[0106] in, Represents the matrix A p The element in row i and column j, is piRNA p i Including p i The set of k nearest neighbors, N k (p i ) is p i k nearest neighbors.
[0107] Preferably, constructing the disease similarity isomorphism graph comprises the following steps:
[0108] First, we obtain the directed acyclic graph (DAG) of diseases from the MeSH database, where nodes represent diseases and connections between diseases describe the relationships between them; and calculate the disease semantic similarity matrix based on DAG.
[0109]
[0110] in, Indicates disease i The semantic value of T(d i ) indicates that node d i and the node set consisting of its ancestor nodes; Indicates disease i For disease i The semantic contribution of is calculated as follows:
[0111]
[0112] Where Δ is the semantic contribution attenuation factor, which is set to 0.5;
[0113] Second, calculate the disease d i and d j The Gaussian kernel similarity between them is calculated, and the disease Gaussian kernel similarity matrix is obtained. The formula is as follows:
[0114] DGS(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 ),
[0115]
[0116] Among them, IP(d i) represents the disease d in the piRNA-disease association matrix i The corresponding column, parameter γ d Used to control kernel bandwidth; parameter γ, d It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N d represents the number of diseases in the piRNA-disease association matrix;
[0117] Third, the disease semantic similarity matrix and the disease Gaussian kernel similarity matrix are averaged to calculate the disease d i and d j The integrated similarity between them is then used to obtain the disease integrated similarity matrix The calculation formula is as follows:
[0118]
[0119] Fourth, the k most similar neighbors of each disease are selected in the disease integrated similarity matrix, where k = 2; based on this, the final disease integrated similarity matrix is obtained and converted into a disease similarity isomorphism graph, which is represented by a binary matrix The formula is as follows:
[0120]
[0121] in, Represents the matrix A d The element in row i and column j, It is a disease i Included i The set of k nearest neighbors, N k (d i ) is d i k nearest neighbors.
[0122] Preferably, step S2 comprises the following steps:
[0123] S21, GCN is used to learn the isomorphic embedding features of piRNA and disease from the piRNA similarity isomorphism graph and the disease similarity isomorphism graph respectively, the formula is as follows:
[0124]
[0125]
[0126] in, and are the information of piRNA and disease at layer l, and are the weight matrices of the GCN layer l where piRNA and disease are located, ReLU is the activation function, D p and D d A p and A d The diagonal matrix of piRNA and disease is embedded in the last layer as and
[0127] S22, using a multi-edge-type graph convolutional network based on a heterogeneous graph to learn heterogeneous embedding features of piRNAs and diseases from a piRNA-disease association heterogeneous graph; in the piRNA-disease association heterogeneous graph, known and unknown piRNA-disease associations are treated as different edge types, and a separate processing channel is assigned to each edge type t∈{0,1}; specifically, each edge type of the graph convolution can be viewed as a form of message passing, in which vector-valued messages are passed and transformed on the edges of the graph; in the above model, a specific transformation is assigned to each edge type, thereby obtaining a disease d j to piRNA p i Edge type message m j→i,t The format is as follows:
[0128]
[0129] Among them, N(p i ) represents node p i The neighbor set, N(d j ) represents node d j The neighbor set, W t is the edge type specific parameter matrix, x j is the disease node d j The feature vector of piRNA p i To disease j Message i→j,t is processed in a similar way; after the message passing step, the edge type t is computed by t (d j ) and accumulate the results for each edge type into a single vector representation to accumulate the incoming messages for each node:
[0130]
[0131] Where sum represents the accumulation operation, σ represents the tanh activation function, and in order to obtain the final representation of piRNA, a linear operator is used to operate on the intermediate output z i To transform:
[0132] z i =Wh i ,
[0133] Disease Embedded j A similar method is used for calculation; in the linear operator, the parameter matrix W of piRNA is the same as that of disease. Applying the above transformation to all nodes of the piRNA-disease association heterogeneous graph, we can obtain piRNA and diseases Final representation in a heterogeneous graph.
[0134] Preferably, step S3 comprises the following steps:
[0135] S31, the corresponding importance of drug embedding and disease embedding is adaptively learned using the attention mechanism as follows:
[0136]
[0137]
[0138] Among them, att is a neural network that performs attention operation, and Represents piRNA and disease embedding, respectively and The attention value;
[0139] S32, combine these embeddings according to these attention values to get the final piRNA embedding Z p and disease embedding Z d :
[0140]
[0141]
[0142] Preferably, step S4 comprises the following steps:
[0143] S41, concatenate the two obtained embeddings to represent the piRNA-disease pair, and feed the embeddings of the piRNA-disease pair into an MLP for prediction:
[0144]
[0145] in, indicates the embedding of piRNA i, represents the embedding of disease j, MLP is a three-layer MLP neural network, represents the probability of association between piRNA i and disease j pairs;
[0146] S42, calculate the binary cross entropy loss between the predicted results and the true results:
[0147]
[0148] Where (i,j) represents the piRNA i-disease j pair, y ij represents the true label; considering the common semantics between homogeneous and heterogeneous graphs, consistency constraints are used to enhance their universality; for piRNA embedding, in order to capture the similarity of piRNA between different graphs, the following constraints are given:
[0149]
[0150] Similarly, calculate the disease embedding constraint L Cd , the cross entropy loss L BCE , consistency constraint L Cp and L Cd The weighted combination gives the final loss L:
[0151] L=L BCE +λL Cp +λL Cd
[0152] where λ is a hyperparameter that balances these three terms.
[0153] Example 2
[0154] A device for predicting piRNA-disease association based on a heterogeneous graph multi-edge graph convolutional network in this embodiment includes:
[0155] The graph construction module uses the piRNA similarity matrix and the disease similarity matrix to construct the piRNA similarity isomorphism graph and the disease similarity isomorphism graph, and uses the piRNA-disease association relationship data to construct the piRNA-disease association heterogeneity graph;
[0156] The feature learning module uses graph convolutional neural networks to learn the isomorphic embedding features of piRNAs and diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and uses a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and diseases from the piRNA-disease association heterogeneous graph;
[0157] The attention fusion module uses the attention mechanism to fuse the homogeneous embedding features and heterogeneous embedding features of piRNA and disease respectively to obtain the fused embedding features of piRNA and disease;
[0158] The association prediction module horizontally splices the fusion embedding features of piRNA and disease to form piRNA-disease pair features, and uses a multi-layer perceptron to predict the association of piRNA-disease pairs.
[0159] Example 3
[0160] The present embodiment provides a device for predicting piRNA-disease associations based on a heterogeneous graph-based multi-edge graph convolutional network, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the method described in Example 1 when executing the computer program.
[0161] Example 4
[0162] Five-fold cross validation was used to evaluate the prediction accuracy of the piRNA-disease association prediction method. The specific method is as follows:
[0163] First, the known piRNA-disease pairs were used as positive samples, and the unknown associated pairs were used as negative samples; then, the positive samples and negative samples were combined and divided into 5 balanced subsets of equal number, each subset was used as the test sample set in turn without repetition, and the remaining four subsets were used as training sample sets; finally, in order to prevent the bias caused by random segmentation of the data set, 10 five-fold cross validations were repeated and the average values of all evaluation indicators were calculated. The experimental results are shown in Table 1:
[0164] Table 1: Five-fold cross validation experimental results of PPDAMEGCN and other methods under MNDR 4.0
[0165]
[0166] As can be seen from Table 1, the piRNA-disease association prediction method of the present invention achieved an AUC of 0.9292 and an AUPR of 0.5006 in the five-fold cross validation, both of which are better than all the comparative methods.
[0167] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0168] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations made to these embodiments without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for predicting piRNA-disease association based on a multi-edge graph convolutional network based on a heterogeneous graph, characterized in that: The steps include: S1, using piRNA similarity matrix and disease similarity matrix to construct piRNA similarity isomorphism graph and disease similarity isomorphism graph, and using piRNA-disease association relationship data to construct piRNA-disease association heterogeneity graph; S2, using a graph convolutional neural network to learn the isomorphic embedding features of piRNAs and the isomorphic embedding features of diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and using a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and the heterogeneous embedding features of diseases from the piRNA-disease association heterogeneous graph; S3, using an attention mechanism to fuse the isomorphic embedding features of the piRNA and the isomorphic embedding features of the disease, as well as the heterogeneous embedding features of the piRNA and the heterogeneous embedding features of the disease, to obtain a fused embedding feature of the piRNA and a fused embedding feature of the disease; S4, horizontally splicing the fusion embedding feature of the piRNA and the fusion embedding feature of the disease to form a piRNA-disease pair feature, and using a multi-layer perceptron to predict the association of the piRNA-disease pair.
2. The method according to claim 1, characterized in that In step S1, data on the association between known human piRNAs and diseases are downloaded from the MNDR 4.0 database, a piRNA-disease association matrix is constructed and converted into the piRNA-disease association heterogeneous graph.
3. The method according to claim 2, characterized in that In step S1, the piRNA similarity isomorphism graph is constructed using piRNA sequence similarity and piRNA Gaussian kernel similarity; and the disease similarity isomorphism graph is constructed using disease semantic similarity and disease Gaussian kernel similarity.
4. The method according to claim 3, characterized in that Constructing the piRNA similarity isomorphism graph comprises the following steps: First, the piRNA p was calculated using the Smith–Waterman algorithm. i and p j The sequence similarity between them was calculated and the piRNA sequence similarity matrix was obtained. The formula is as follows: Among them, SW(p i ,p j ) represents the sequence alignment score between the i-th and j-th piRNAs based on the Smith-Waterman algorithm; Second, calculate piRNA p i and p j The Gaussian kernel similarity between them is calculated and the piRNA Gaussian kernel similarity matrix is obtained. The formula is as follows: PGS(p i ,p j )=exp(-γ p ||IP(p i )-IP(p j )|| 2 ), Among them, IP(p i ) represents the piRNA p in the piRNA-disease association matrix i The corresponding row, parameter γ p Used to control kernel bandwidth; parameter γ ， p It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N p represents the number of piRNAs in the piRNA-disease association matrix; Third, the piRNA sequence similarity matrix and the piRNA Gaussian kernel similarity matrix were averaged to calculate the piRNA m i and m j The integrated similarity between them is then used to obtain the piRNA integrated similarity matrix The calculation formula is as follows: Fourth, the k most similar neighbors of each piRNA were selected in the piRNA ensemble similarity matrix, where k = 2; based on this, the final piRNA ensemble similarity matrix was obtained and converted into a piRNA similarity isomorphism graph, which was represented by a binary matrix The formula is as follows: in, Represents the matrix A p The element in row i and column j, is piRNA p i Including p i The set of k nearest neighbors, N k (p i ) is p i k nearest neighbors.
5. The method according to claim 4, characterized in that Constructing the disease similarity isomorphism graph includes the following steps: First, we obtain the directed acyclic graph (DAG) of diseases from the MeSH database, where nodes represent diseases and connections between diseases describe the relationships between them; and calculate the disease semantic similarity matrix based on DAG. in, Indicates disease i The semantic value of T(d i ) indicates that node d i and the node set consisting of its ancestor nodes; Indicates disease i For disease i The semantic contribution of is calculated as follows: Where Δ is the semantic contribution attenuation factor, which is set to 0.5; Second, calculate the disease d i and d j The Gaussian kernel similarity between them is calculated, and the disease Gaussian kernel similarity matrix is obtained. The formula is as follows: DGS(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 ), Among them, IP(d i ) represents the disease d in the piRNA-disease association matrix i The corresponding column, parameter γ d Used to control kernel bandwidth; parameter γ ， d It is measured by the average number of associations between all diseases and piRNAs, with a value of 1; N d represents the number of diseases in the piRNA-disease association matrix; Third, the disease semantic similarity matrix and the disease Gaussian kernel similarity matrix are averaged to calculate the disease d i and d j The integrated similarity between them is then used to obtain the disease integrated similarity matrix The calculation formula is as follows: Fourth, the k most similar neighbors of each disease are selected in the disease integrated similarity matrix, where k = 2; based on this, the final disease integrated similarity matrix is obtained and converted into a disease similarity isomorphism graph, which is represented by a binary matrix The formula is as follows: in, Represents the matrix A d The element in row i and column j, It is a disease i Included i The set of k nearest neighbors, N k (d i ) is d i k nearest neighbors.
6. The method according to claim 1, characterized in that Step S2 includes the following steps: S21, GCN is used to learn the isomorphic embedding features of piRNA and disease from the piRNA similarity isomorphism graph and the disease similarity isomorphism graph respectively, the formula is as follows: in, and are the information of piRNA and disease at layer l, and are the weight matrices of the GCN layer l where piRNA and disease are located, ReLU is the activation function, D p and D d A p and A d The diagonal matrix of piRNA and disease is embedded in the last layer as and S22, using a multi-edge-type graph convolutional network based on a heterogeneous graph to learn heterogeneous embedding features of piRNAs and diseases from a piRNA-disease association heterogeneous graph; in the piRNA-disease association heterogeneous graph, known and unknown piRNA-disease associations are treated as different edge types, and a separate processing channel is assigned to each edge type t∈{0,1}; specifically, each edge type of the graph convolution can be viewed as a form of message passing, in which vector-valued messages are passed and transformed on the edges of the graph; in the above model, a specific transformation is assigned to each edge type, thereby obtaining a disease d j to piRNA p i Edge type message m j→i,t The format is as follows: Among them, N(p i ) represents node p i The neighbor set, N(d j ) represents node d j The neighbor set, W t is the edge type specific parameter matrix, x j is the disease node d j The feature vector of piRNA p i To disease j Message i→j,t is processed in a similar way; after the message passing step, the edge type t is t (d j ) and accumulate the results for each edge type into a single vector representation to accumulate the incoming messages for each node: Where sum represents the accumulation operation, σ represents the tanh activation function, and in order to obtain the final representation of piRNA, a linear operator is used to operate on the intermediate output z i To transform: z i =Wh i , Disease Embedded j A similar method is used for calculation; in the linear operator, the parameter matrix W of piRNA is the same as that of disease. Applying the above transformation to all nodes of the piRNA-disease association heterogeneous graph, we can obtain piRNA and diseases Final representation in a heterogeneous graph.
7. The method according to claim 1, characterized in that Step S3 includes the following steps: S31, the corresponding importance of drug embedding and disease embedding is adaptively learned using the attention mechanism as follows: Among them, att is a neural network that performs attention operation, and Represents piRNA and disease embedding, respectively and The attention value; S32, combine these embeddings according to these attention values to get the final piRNA embedding Z p and disease embedded in Z d :
8. The method according to claim 1, characterized in that Step S4 includes the following steps: S41, concatenate the two obtained embeddings to represent the piRNA-disease pair, and feed the embeddings of the piRNA-disease pair into an MLP for prediction: in, indicates the embedding of piRNA i, represents the embedding of disease j, MLP is a three-layer MLP neural network, represents the probability of association between piRNA i and disease j pairs; S42, calculate the binary cross entropy loss between the predicted results and the true results: Where (i,j) represents the piRNA i-disease j pair, y ij represents the true label; considering the common semantics between homogeneous and heterogeneous graphs, consistency constraints are used to enhance their universality; for piRNA embedding, in order to capture the similarity of piRNA between different graphs, the following constraints are given: Similarly, calculate the disease embedding constraint L Cd , the cross entropy loss L BCE , consistency constraint L Cp and L Cd The weighted combination gives the final loss L: L=L BCE +λL Cp +λL Cd where λ is a hyperparameter that balances these three terms.
9. A device for predicting piRNA-disease association based on a multi-edge graph convolutional network of heterogeneous graphs, characterized in that: include: The graph construction module uses the piRNA similarity matrix and the disease similarity matrix to construct the piRNA similarity isomorphism graph and the disease similarity isomorphism graph, and uses the piRNA-disease association relationship data to construct the piRNA-disease association heterogeneity graph; The feature learning module uses graph convolutional neural networks to learn the isomorphic embedding features of piRNAs and diseases from the piRNA similarity isomorphic graph and the disease similarity isomorphic graph, respectively, and uses a multi-edge type graph convolutional network based on a heterogeneous graph to learn the heterogeneous embedding features of piRNAs and diseases from the piRNA-disease association heterogeneous graph; The attention fusion module uses the attention mechanism to fuse the homogeneous embedding features and heterogeneous embedding features of piRNA and disease respectively to obtain the fused embedding features of piRNA and disease; The association prediction module horizontally splices the fusion embedding features of piRNA and disease to form piRNA-disease pair features, and uses a multi-layer perceptron to predict the association of piRNA-disease pairs.
10. A device for predicting piRNA-disease association based on a multi-edge graph convolutional network of heterogeneous graphs, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the method according to any one of claims 1 to 8 is executed when the processor executes the computer program.