Identification of piRNA-disease associations based on multimodal collaborative representation learning

Abstract

The present invention belongs to the field of computer-aided disease diagnosis, and is specifically a method for identifying piRNA-disease associations based on multimodal collaborative representation learning. The method constructs a comprehensive heterogeneous network by fusing piRNA-disease association information and multi-source similarity information of piRNA and disease; the features of the heterogeneous network representation learning module, the association network representation learning module and the multi-view representation learning module are fused to construct an integrated multimodal piRNA-disease association pair feature representation, and the association confidence between piRNA and disease is predicted based on a fully connected neural network. The present invention fully exploits five types of multi-source similarity information from piRNA and disease; three types of collaborative representation learning modules are introduced from different perspectives, and through the interaction and complementation between the modules, MCRL-PDA can more accurately capture the complex relationship between piRNA and disease. The method of multi-source similarity fusion and multi-module representation learning of the present invention can effectively improve the ability of MCRL-PDA to identify complex relationships between different entities.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer-aided disease diagnosis, and in particular relates to a method for identifying piRNA-disease association based on multimodal collaborative representation learning. Background Art

[0002] Piwi-interacting RNA (piRNA) is a class of non-coding RNA composed of approximately 23-36 nucleotides. They mainly interact with the germline-specific PIWI proteins in the Argonaute protein family, ensuring the stability of the germ cell genome and the accurate transmission of genetic information. In addition, increasing evidence shows that abnormal expression of piRNA is related to the pathogenesis of various diseases.

[0003] For example, Fu et al. found that abnormal expression of piR-021285 can promote the methylation of the gene ARHGAP11A at the 5'-UTR and the first exon CpG sites, leading to mRNA apoptosis and then inhibiting the apoptosis of breast cancer cells.

[0004] Liu et al. found that the interaction between piR-30188 and OIP5-AS1 can inhibit the proliferation of glioma cells. Downregulation of the expression level of the gene OIP5-AS1 leads to a decrease in CEBPA levels, thereby promoting the malignant phenotype of glioma cells.

[0005] As the role of piRNA in disease epigenetics continues to be discovered, more and more studies have shown that piRNA can be used as a potential biomarker or therapeutic target for diagnosing and predicting diseases. Therefore, accurate identification of disease-related piRNA is of great significance for diagnosing diseases and revealing the pathogenesis of diseases.

[0006] In recent years, graph neural networks (GNNs) have shown excellent performance in association prediction problems, and several GNN-based piRNA-disease association prediction methods have been proposed. For example, Zheng et al. proposed a graph neural network framework GAPDA with piRNA-disease associations as nodes, which aims to capture topological information and use neighbor node information to calculate the hidden representation of associations in the network.

[0007] iPiDA-GCN is based on two graph convolutional networks (Asso-GCN and Sim-GCN) to capture features from the piRNA-disease association network and two similarity networks, respectively. iPiDA-SWGCN enriches the structural information of the association network by integrating the confidence calculated by the base predictor for supplementary weighting, and then uses GCN to learn the network topology information.

[0008] However, although the existing calculation methods have achieved certain results, they also have certain limitations:

[0009] First, existing methods mainly extract node features based on piRNA sequence similarity and disease semantic similarity information, but do not fully tap the multi-source feature information of piRNA and disease; second, due to the transition smoothing problem of GCN, the node features gradually lose valuable distinguishing information during the message propagation process. Therefore, using only a single model based on GNN to identify complex associations between different entities may not be the best method. Summary of the invention

[0010] The purpose of the embodiments of the present invention is to provide a method for identifying piRNA-disease associations based on multimodal collaborative representation learning, aiming to solve the problem that existing computational methods do not fully exploit the multi-source characteristics of piRNA and diseases, and the problem that the performance of the model based on a single GNN is limited due to excessive smoothing.

[0011] To achieve the above object, the present invention provides the following technical solutions.

[0012] A method for identifying piRNA-disease associations based on multimodal collaborative representation learning, the method comprising the following steps:

[0013] Step S101: integrating piRNA-disease association information and multi-source similarity information of piRNA and disease to construct a comprehensive heterogeneous network;

[0014] Step S102: design a heterogeneous network representation learning module based on a multi-head attention mechanism to capture diverse features from piRNA and disease nodes;

[0015] Step S103: construct a GCN-based association network representation learning module, taking the piRNA-disease association network as input, the integrated piRNA similarity network PS, and the integrated disease similarity network DS as features, and iteratively updating the feature representation of the nodes respectively;

[0016] Step S104: designing a multi-view representation learning module based on dilated convolution;

[0017] Step S105: The features of the heterogeneous network representation learning module, the association network representation learning module and the multi-view representation learning module are integrated to construct an integrated multimodal piRNA-disease association pair feature representation, and the association confidence between piRNA and disease is predicted based on a fully connected neural network.

[0018] Furthermore, in the step of constructing a comprehensive heterogeneous network, known experimentally verified piRNA-disease associations were collected from the MNDRv3.0 and piRDisease v1.0 databases, respectively, and an m × n association network was used. To represent the association information of piRNA-disease, the association network is defined as follows: ; Mining five types of multi-source similarity information from piRNA and disease: piRNA sequence similarity PSS, functional similarity PFS, Gaussian interaction profile kernel similarity GIPKp, piRNA similarity network and disease similarity networks ;

[0019] Constructing a comprehensive heterogeneous network by integrating piRNA-disease association information and multi-source similarity information of piRNAs and diseases , the construction is expressed as follows: ,in, represents the adjacency matrix of piRNA-disease, yes The transpose of Represents the similarity network of piRNAs, represents the disease similarity network. Furthermore, the heterogeneous network representation learning module based on the multi-head attention mechanism includes a total of 3 encoding layers. In the hth attention head of one of the encoding layers, the query matrix 、key matrix And the value matrix The linear transformation of is as follows:

[0020] in, , , represents the weight matrix, It is Layer heterogeneous network represents the output of learning module; computing and The dot product of ,matrix The attention scores of all piRNA and disease nodes are recorded, reflecting the degree of association between all piRNA and disease nodes; the attention score matrix Normalized and multiplied by To obtain the A heterogeneous network representation of 4 attention heads: ;

[0021] Where d is and The vector dimension of ; by concatenating the results of all attention heads to obtain the Output of the layer :

[0022] ;in, Represents a splicing operation, Indicates the number of attention heads;

[0023] After representation learning of 3 encoding layers, the heterogeneous network representation matrix is ​​obtained ;

[0024] Rows i and j represent the learned diverse feature representations of piRNA and disease, respectively; , The feature vector captures the interaction information between piRNA molecules and disease entities, and between disease molecules and piRNA entities. and Represent the number of piRNA nodes and disease nodes respectively.

[0025] Further, in step S103, the first The message propagation mechanism of the layer is as follows:

[0026] ;

[0027] ;

[0028] Where D is the degree matrix of the associated network A. By multiplying A on the left and right

[0029] Normalize the rows and columns of the association network respectively, It is The embedding features of layer nodes, It is The trainable weight matrices in the layer, It is a non-linear activation function, and the non-linear activation function uses the Relu function.

[0030] Furthermore, in step S104:

[0031] Combine the ith row of the piRNA similarity network PS and the ith row of A to construct The characteristic representation of:

[0032] in, Represents a splicing operation, Represents piRNA The similarity score with all other piRNAs, Represents all Interactive disease information;

[0033] Combine the j-th row of the disease similarity network DS and the j-th column of A to construct The characteristic representation of: ;

[0034] Represents disease The similarity score with all other diseases, Represents all diseases Information on interacting piRNAs;

[0035] Then, construct Feature representation of node pairs: . Further, in step S104, using dilated convolution, the receptive field of the convolution kernel is expanded to a variable size according to the dilation rate; by integrating the multi-view features revealed by different receptive fields, the comprehensive feature representation T of the complex relationship between piRNA-disease node pairs is captured. Further, zeros are filled around the comprehensive feature representation T and the zero-filled matrix is ​​represented as: ; The convolutional layer contains The height and width are and The filter, given a dilation rate d, when the qth filter moves to When the mth row and nth column are The definition is as follows: ;in, , , ; Apply the qth filter to , the element value in the affected area of ​​the qth feature map is updated through the convolution operation : ;in, is the Relu function, * represents the convolution operation, and Represents the learnable weight matrix and bias vector of the qth filter; introduces the maximum pooling layer to extract features on the feature map, and sets the height and width of the filter to and , the element value of the affected area on the qth feature map after the maximum pooling The definition is as follows: Given a dilation rate d, after the above dilated convolution-pooling layer processing, the feature representation of view d is obtained , set d=[1,2,3], and obtain , and , reflecting the different views The feature representation of node pairs; , and splicing, and then input into another convolution-pooling layer for fusion to form a multi-view Feature representation of node pairs Furthermore, in step S105, a linear transformation layer is introduced after the heterogeneous network representation learning module and the multi-view representation learning module to perform dimension mapping on the features extracted by each representation learning module so that they meet the dimensional consistency requirements of subsequent fusion and analysis. Furthermore, in step S105, the output of the heterogeneous network representation learning module is mapped to the multi-view representation learning module. middle The i-th row and The jth row of is concatenated to form a heterogeneous network representation learning module Feature representation of associated pairs: ; Accordingly, the association network represents the learning module middle The i-th row and The jth row of is concatenated to form an association network representation learning module Feature representation of associated pairs:

[0036] ; Each module after the linear transformation layer dimension mapping The feature representations of the associated pairs are fused as follows: Afterwards, Multi-module feature representation of association pairs As input to the fully connected neural network, the confidence scores of piRNA-disease association pairs were calculated. : ; Where w and b represent the learnable weight matrix and bias respectively. The loss function uses the cross entropy function to measure the difference between the predicted value and the actual value. The loss function is defined as follows: ; Among them, N and They represent the number of piRNA-disease association pairs and their true labels respectively;

[0037] Subsequently, the Adam optimizer is used to minimize the loss function.

[0038] Compared with the prior art, the method for identifying piRNA-disease association based on multimodal collaborative representation learning of the present invention has the following beneficial effects:

[0039] First, the computational method of multi-module collaborative representation learning MCRL-PDA of the present invention is used to identify potential piRNA-disease associations, which fully mines five multi-source similarity information from piRNAs and diseases;

[0040] Second, the present invention introduces three collaborative representation learning modules from different perspectives, which respectively capture the diverse characteristics of nodes in heterogeneous networks, the topological structure information implicit in the association network, and the multi-view features revealed by different receptive fields. Through the interaction and complementation between the modules, MCRL-PDA can more accurately capture the complex relationship between piRNA and disease.

[0041] Third, the MCRL-PDA of the present invention achieved AUC values ​​of 0.9824 and 0.9657 respectively under five-fold cross validation on two data sets, surpassing the other seven common models, verifying the effectiveness of the present invention; in addition, the ablation experiment results of the present invention show that the method of multi-source similarity fusion and multi-module representation learning can effectively improve the ability of MCRL-PDA to identify complex relationships between different entities. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0043] Figure 1 It is a schematic diagram of the implementation process of the method for identifying piRNA-disease association based on multimodal collaborative representation learning of the present invention;

[0044] Figure 2 Schematic diagram of ROC curves of different calculation methods on the MNDRv3.0 database under five-fold cross validation of the present invention;

[0045] Figure 3 It is a schematic diagram of the PR curve of different calculation methods on the MNDRv3.0 database under five-fold cross validation of the present invention;

[0046] Figure 4 Schematic diagram of ROC curves of different calculation methods on the piRDisease v1.0 database under five-fold cross validation of the present invention;

[0047] Figure 5 It is a schematic diagram of PR curves of different calculation methods on the piRDisease v1.0 database under five-fold cross validation of the present invention;

[0048] Figure 6 A schematic diagram showing the visual comparison of the associations predicted by MCRL-PDA of the present invention;

[0049] Figure 7 This is a schematic diagram of the association visualization comparison predicted by iPiDA-SWGCN of the present invention;

[0050] Figure 8 This is a schematic diagram of the correlation visualization comparison predicted by PUTransGCN of the present invention;

[0051] Fig. 9 This is a schematic diagram of the correlation visualization comparison of ETGPDA prediction of the present invention;

[0052] Fig.10 This is a system architecture diagram of the piRNA-disease association identification method based on multimodal collaborative representation learning of the present invention;

[0053] Fig.11 Schematic diagram of the multi-head attention mechanism in the system architecture of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] Currently, in the existing technology, iPiDA-GCN is based on two graph convolutional networks (Asso-GCN and Sim-GCN) to capture features from the piRNA-disease association network and two similarity networks respectively. iPiDA-SWGCN enriches the structural information of the association network by integrating the confidence calculated by the base predictor for supplementary weighting, and then uses GCN to learn the network topology information.

[0056] However, although the existing computational methods have achieved certain results, they also have certain limitations: first, the existing methods mainly extract node features based on piRNA sequence similarity and disease semantic similarity information, but do not fully mine the multi-source feature information of piRNA and disease; second, due to the transition smoothing problem of GCN, the node features gradually lose valuable distinguishing information during the message propagation process. Therefore, using only a single model based on GNN to identify complex associations between different entities may not be the best method.

[0057] To solve the above problems, the present invention provides a method for identifying piRNA-disease associations based on multimodal collaborative representation learning. The method's multi-module collaborative representation learning computational method MCRL-PDA is used to identify potential piRNA-disease associations. The method fully mines five multi-source similarity information from piRNA and disease; the method introduces three collaborative representation learning modules from different perspectives, which respectively capture the diverse characteristics of nodes in heterogeneous networks, the topological structure information implicit in the association network, and the multi-view features revealed by different receptive fields. Through the interaction and complementation between the modules, MCRL-PDA can more accurately capture the complex relationship between piRNA and disease.

[0058] The specific implementation of the piRNA-disease association identification method based on multimodal collaborative representation learning of the present invention is described in detail below in conjunction with specific embodiments.

[0059] like Figure 1 , Fig.10 and Fig.11 As shown, in one embodiment of the present invention, a method for identifying piRNA-disease associations based on multimodal collaborative representation learning is provided, the method comprising: constructing a piRNA-disease heterogeneous network, heterogeneous network representation learning, association network representation learning, multi-view feature representation learning, and integration and association score prediction.

[0060] Specifically, the method comprises the following steps:

[0061] Step S101: integrating piRNA-disease association information and multi-source similarity information of piRNA and disease to construct a comprehensive heterogeneous network;

[0062] Step S102: design a heterogeneous network representation learning module based on a multi-head attention mechanism to capture diverse features from piRNA and disease nodes;

[0063] Step S103: construct a GCN-based association network representation learning module, taking the piRNA-disease association network as input, the integrated piRNA similarity network PS, and the integrated disease similarity network DS as features, and iteratively updating the feature representation of the nodes respectively;

[0064] Step S104: designing a multi-view representation learning module based on dilated convolution;

[0065] Step S105: The features of the heterogeneous network representation learning module, the association network representation learning module and the multi-view representation learning module are integrated to construct an integrated multimodal piRNA-disease association pair feature representation, and the association confidence between piRNA and disease is predicted based on a fully connected neural network.

[0066] Furthermore, in the step of constructing a comprehensive heterogeneous network, known experimentally validated piRNA-disease associations were collected from the MNDRv3.0 and piRDisease v1.0 databases, respectively;

[0067] Among them, MNDR v3.0 is a comprehensive ncRNA-disease association database that stores the latest and most complete piRNA-disease association data among all piRNA-disease databases. After removing piRNAs that have not been verified in the piRbase database and duplicate associations, we collected a total of 11,981 experimentally verified piRNA-disease associations, covering 10,149 piRNAs and 19 diseases. Due to equipment performance limitations, we randomly selected 5,000 piRNA-disease associations from all collected associations, covering 4,300 piRNAs and 19 diseases.

[0068] In addition, piRDisease v1.0 contains 7939 validated associations involving 4796 piRNAs and 28 different diseases. After removing duplicate data, we collected 4981 experimentally validated piRNA-disease associations, including 4349 piRNAs and 21 diseases. Subsequently, the present invention uses an m×n association network To represent the association information of piRNA-disease, where m represents the number of piRNAs and n represents the number of diseases. and disease When relevant, the corresponding position is 1, otherwise is 0, the associated network is defined as follows:

[0069] ; In addition, we fully mined five types of multi-source similarity information from piRNAs and diseases.

[0070] The five types of multi-source similarity information include: piRNA sequence similarity PSS, functional similarity PFS, Gaussian interaction spectrum kernel similarity GIPKp, piRNA similarity network and disease similarity networks ; Among them, the disease similarity network (DS): We describe and calculate the semantic similarity DSS of diseases based on the directed acyclic graph (DAG) information of diseases in MeSH; based on the assumption that similar piRNAs in the interaction network show similar modes of action in diseases, we can measure the similarity information of diseases GIPKd through the Gaussian interaction spectrum kernel. Then, we use the advanced similarity network fusion technology SNF to continuously iteratively update and fuse the above two similarities to form an integrated disease similarity network (DS); piRNA similarity network (PS): Chaotic game representation (CGR) technology can uniquely map piRNA sequences to a two-dimensional plane by combining positional and nonlinear information. Subsequently, we mapped the relative position of each nucleotide on the two-dimensional plane to a frequency network with N×N, where N is set to 8. Finally, a vector representation of piRNA is constructed based on the information of the frequency network and the Pearson correlation coefficient is used to measure the piRNA sequence similarity PSS; previous studies have shown that the functional similarity of piRNA can be calculated and measured by evaluating the similarity information between the diseases associated with it. Therefore, we need to consider the contribution of diseases associated with piRNAs. First, we define the semantic similarity between a disease and a group of diseases: , where d is a disease and D is a group of diseases, that is Subsequently, piRNA and The functional similarity between is defined as: ;in, Is related to disease Related disease collections, yes An element in , m and n represent the piRNA and Based on the assumption that similar piRNAs in the interaction network show similar modes of action in diseases, we can measure the similarity information of piRNAs by Gaussian interaction spectrum kernel GIPKp, and then use SNF technology to iteratively fuse the above three similarities to form an integrated piRNA similarity network (PS).

[0071] Finally, the present invention constructs a comprehensive heterogeneous network by integrating piRNA-disease association information and multi-source similarity information of piRNA and disease , the construction is expressed as follows: ,in, represents the adjacency matrix of piRNA-disease, yes The transpose of Represents the similarity network of piRNAs, Representing a disease similarity network. Further, in step S102 of the present invention, the present invention designs a heterogeneous network representation learning module based on a multi-head attention mechanism, aiming to capture the diverse features from piRNA and disease nodes. By introducing a multi-head attention mechanism, the variance that may be generated in representation learning can be reduced because it allows the model to learn multiple different representations in parallel, reducing the dependence on a single representation, making the learning process more robust;

[0072] The heterogeneous network representation learning module based on the multi-head attention mechanism of the present invention includes a total of 3 encoding layers. In the hth attention head of one encoding layer, the query matrix 、key matrix And the value matrix The linear transformation of is as follows: ;

[0073] in, , , represents the weight matrix, It is Layer heterogeneous network represents the output of learning module; computing and The dot product of ,matrix The attention scores of all piRNA and disease nodes were recorded, reflecting the degree of association between all piRNA and disease nodes;

[0074] The attention score matrix Normalized and multiplied by To obtain the A heterogeneous network representation of 4 attention heads: ; where d is and The vector dimension of , the normalization process helps to improve the stability during training;

[0075] By concatenating the results of all attention heads, we obtain Output of the layer :

[0076] ;in, Represents a splicing operation, Indicates the number of attention heads;

[0077] After representation learning of 3 encoding layers, the heterogeneous network representation matrix is ​​obtained ;

[0078] Row i and row j represent the learned diversified feature representations of piRNA and disease, respectively;

[0079] in, , The feature vector captures the interaction information between piRNA molecules and disease entities, and between disease molecules and piRNA entities. and Represent the number of piRNA nodes and disease nodes respectively. Furthermore, GCN can effectively extract implicit topological information from graph structure data and learn effective representations of nodes. It is based on the message propagation mechanism of neighbor aggregation, in which the feature vector of each node is iteratively updated and calculated through the characteristic information of its directly adjacent nodes. In step S103, we constructed a GCN-based association network representation learning module. We take the piRNA-disease association network as input, the integrated piRNA similarity network PS, and the integrated disease similarity network DS as features, and iteratively update the feature representation of the node respectively. The message propagation mechanism of the lth layer in GCN is as follows: ;in, It is The embedding features of layer nodes, is the piRNA-disease association network, It is The trainable weight matrices in the layer, is a nonlinear activation function, and here we use the Relu function. In the associative network, since the number of connections (degrees) of different nodes are different, in the information aggregation process of GCN, nodes with higher degrees tend to produce larger feature vectors, while the feature vectors of nodes with smaller degrees are relatively small. This may cause gradient explosion or gradient disappearance problems during the learning process, thus affecting the stability of network training and the performance of the model. Therefore, we use the Laplacian matrix to normalize the associative network A. Therefore, the first The message propagation mechanism of the layer is updated as follows: ;

[0080] ; Where D is the degree matrix of the associated network A, by multiplying A on the left and right The rows and columns of the association network are normalized separately. In addition, GCNs with different numbers of layers can capture different structural information of the association network. For example, the first layer captures direct link information, and higher layers capture multi-hop neighbor information (high-order proximity) by iteratively updating the embedding. Here, we use a 3-layer GCN to capture the feature information of piRNA and disease nodes.

[0081] Furthermore, in step S104: before introducing the multi-view feature representation learning module, we first introduce a new fusion strategy to provide Given the association network, the integrated piRNA similarity network PS, and the integrated disease similarity network DS, we first combine the i-th row of the piRNA similarity network PS and the i-th row of A to construct The characteristic representation of:

[0082] in, Represents a splicing operation, Represents piRNA The similarity score with all other piRNAs, Representing all Interactive disease information; combined with the j-th row of the disease similarity network DS and the j-th column of A to construct The characteristic representation of:

[0083] ; Represents disease The similarity score with all other diseases, Represents all diseases The interacting piRNA information was then constructed by Feature representation of node pairs:

[0084] Through this construction method, the attribute embedding of the node pair not only contains the similarity information between piRNA (disease) nodes, but also the interaction information between piRNA and each disease. This rich similarity information and interaction information enables the model to have a deeper understanding of the connection between complex biological molecules and diseases.

[0085] Since traditional CNN can only extract features from a single view due to its fixed receptive field, in this section, we design a multi-view representation learning module based on dilated convolution. With dilated convolution, the receptive field of the convolution kernel can be expanded to a variable size according to the dilation rate. Receptive fields of different sizes can extract features from different views, while larger receptive fields can capture more abstract and global features, and smaller receptive fields mean more local and specific features. By integrating multi-view features revealed by different receptive fields, we can effectively capture a comprehensive feature representation T of the complex relationship between piRNA-disease node pairs.

[0086] Furthermore, in order to effectively capture the edge information of the feature representation T, the present invention fills zeros around the overall feature representation T and expresses the zero-filled matrix as: ; The convolutional layer contains The height and width are and The filter, given a dilation rate d, when the qth filter moves to When the mth row and nth column are The definitions are as follows:; ;

[0087] in, , , ; Apply the qth filter to , the element value in the affected area of ​​the qth feature map is updated through the convolution operation : ;

[0088] in, is the Relu function, * represents the convolution operation, and Represents the learnable weight matrix and bias vector of the qth filter; introduces the maximum pooling layer to extract more important features on the feature map, and sets the height and width of the filter to and , the element value of the affected area on the qth feature map after the maximum pooling The definition is as follows: ;

[0089] Given a dilation rate d, after the above dilated convolution-pooling layer processing, the feature representation of view d is obtained , set d=[1,2,3], and obtain , and , which reflect the The feature representation of node pairs; in order to effectively fuse features from different views, , and splicing, and then input into another convolution-pooling layer for fusion to form a multi-view Feature representation of node pairs .

[0090] Furthermore, in step S105, a linear transformation layer is introduced after the heterogeneous network representation learning module and the multi-view representation learning module to perform dimension mapping on the features extracted by each representation learning module so that the features meet the dimensional consistency requirements of subsequent fusion and analysis. It is worth noting that in the association network representation learning module, since the feature dimension remains unchanged before and after, the module does not introduce a linear transformation layer; specifically, the output of the heterogeneous network representation learning module is mapped to the linear transformation layer. middle The i-th row and The jth row of is concatenated to form a heterogeneous network representation learning module Feature representation of associated pairs:

[0091] ; Accordingly, the association network represents the learning module middle The i-th row and The jth row of is concatenated to form an association network representation learning module Feature representation of associated pairs:

[0092] ; In the multi-view representation learning module, the present invention has constructed a feature representation of the associated pair ;

[0093] Each module after the linear transformation layer dimension mapping The feature representations of the associated pairs are fused as follows: Afterwards, Multi-module feature representation of association pairs As input to the fully connected neural network, the confidence scores of the piRNA-disease association pairs were calculated. :

[0094] ; Where w and b represent the learnable weight matrix and bias respectively. The loss function uses the cross entropy function to measure the difference between the predicted value and the actual value. The loss function is defined as follows:

[0095] ;

[0096] Among them, N and represent the number of piRNA-disease association pairs and their true labels, respectively; subsequently, the Adam optimizer was used to minimize the loss function.

[0097] Furthermore, in verifying the effectiveness of the embodiment of the present invention, the following steps are included:

[0098] (1) Evaluation indicators:

[0099] Five-fold cross validation was used to evaluate the performance of the model because it helps to comprehensively evaluate the generalization ability of the model and can effectively prevent the model from overfitting. We considered all known piRNA-disease associations as positive samples, and unknown piRNA-disease associations meant that there were no relevant biological experiments to verify their association with each other. Therefore, all unknown piRNA-disease associations were considered negative samples. To ensure the balance of positive and negative samples, we randomly selected the same number of negative samples as positive samples from all negative samples.

[0100] Since AUPR and AUC can represent the performance of the model at different thresholds, we choose them as the main evaluation indicators. In addition, we also use other threshold-based evaluation indicators, such as F1 score, precision, and recall. The relevant mathematical formulas are as follows: ; TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively. AUC refers to the area under the receiver operating characteristic curve, which quantitatively reflects the performance of the model measured based on the ROC curve. AUPR stands for the area under the Precision-Recall (PR) curve. The larger the area under the PR curve, the better the performance of the model.

[0101] (2) Comparison with existing methods:

[0102] In order to evaluate the prediction performance of the model, the present invention compares the prediction performance of our method MCRL-PDA with seven existing advanced methods: GAPDA, iPiDA-GCN, iPiDA-SWGCN, PUtransGCN, ETGPDA, iPiDA-GBNN and piRDA on two datasets: MNDRv3.0 and piRDiseasev1.0. The experimental results of MNDR3.0 are shown in Table 1. Figure 2 , Figure 3 As shown;

[0103] Figure 2 Schematic diagram of ROC curves of different calculation methods on the MNDRv3.0 database under five-fold cross validation of the present invention; Figure 3 Schematic diagram of PR curves of different calculation methods on the MNDRv3.0 database under five-fold cross validation of the present invention.

[0104] MCRL-PDA achieved the highest AUC of 0.9824 and AUPR of 0.9747, which were 4.31% and 3.09% higher than PUtransGCN, the best performing of the seven advanced methods. In addition, MCRL-PDA achieved the highest results in all evaluation indicators.

[0105] Table 1: Performance comparison of different methods under 5-fold cross validation under MNDRv3.0 . The experimental results of piRDiseasev1.0 are shown in Table 2. Figure 4 , Figure 5 As shown;

[0106] Figure 4 Schematic diagram of ROC curves of different calculation methods on the piRDisease v1.0 database under five-fold cross validation of the present invention; Figure 5 Schematic diagram of PR curves of different calculation methods on the piRDisease v1.0 database under five-fold cross validation of the present invention.

[0107] MCRL-PDA achieved the highest AUC of 0.9657, which was 7.38%, 4.89%, 4.14%, 1.19%, 3.69%, 13.73% and 4.96% higher than GAPDA, iPiDA-GCN, iPiDA-SWGCN, PUtransGCN, ETGPDA, iPiDA-GBNN and piRDA, respectively. The AUPR of MCRL-PDA was 0.9612, which was 10.62%, 6.63%, 4.95%, 0.99%, 6.9%, 9.89% and 8.39% higher than the above methods, respectively. In addition, MCRL-PDA also achieved the highest results in all evaluation indicators. The experimental results on the two datasets show that MCRL-PDA has better performance in predicting piRNA-disease association.

[0108] Table 2: Performance comparison of different methods under piRDisasev1.0 with five-fold cross validation . (3) Visual comparison of predicted associations:

[0109] To clearly demonstrate the excellent performance of MCRL-PDA in predicting piRNA-disease associations, we made an intuitive visual comparison of the prediction results of MCRL-PDA with the three best-performing methods on two datasets. Specifically, we selected three representative piRNA-disease associations on the MNDRv3.0 dataset for in-depth analysis, namely <piR-hsa-10004, Parkinson disease>, <piR-hsa-11988, Alzheimer disease> and <piR-hsa-11604, Cardiovascular disease>.

[0110] The visualization results are as Figure 6-Figure 9 shown. iPiDA-SWGCN only successfully predicted one association, <piR-hsa-11604, Cardiovascular disease>. PUTransGCN introduced an attention mechanism and an advanced negative sampling technique, achieving the successful prediction of two associations. At the same time, ETGPDA also achieved the successful prediction of two associations. In contrast, through the representation learning and fusion of multiple modules, as well as the interaction and complementarity between modules, MCRL-PDA achieved the best performance of successfully predicting all three associations.

[0111] (4) Ablation experiment:

[0112] (a) Similarity network:

[0113] In this section, we conducted ablation experiments on two datasets to deeply explore the specific contributions of each similarity network to the MCRL-PDA prediction framework. Specifically, we designed five groups of experiments, deleting the piRNA sequence similarity PSS, piRNA functional similarity PFS, GIPK similarity GIPKp, disease semantic similarity DSS, and disease GIPK similarity GIPKd one by one. Based on these 5 different similarity network fusion strategies, we constructed the inputs of three representation learning modules respectively, and used five-fold cross-validation to evaluate the impact of each similarity network on the prediction performance of MCRL-PDA.

[0114] The experimental results are shown in Tables 3 and 4. We can observe that different similarity network information contributes differently to the model. Among them, the GIPK similarity of piRNA and the semantic similarity of diseases on the MNDRv3.0 dataset contribute the most to the performance improvement of MCRL-PDA. In the experiments of removing these two similarity networks, the performance of MCRL-PDA decreased by 5.34% and 3.51% in AUC, and decreased by 6.33% and 5.26% in AUPR. The experimental results on the piRDiseasev1.0 dataset show that the GIPK similarity information of piRNA and disease contributes the most to the model. After removing these two similarity networks, the model's AUC decreased by 2.62% and 4.36%, and AUPR decreased by 4.3% and 3.07%.

[0115] Table 3: Similarity network ablation experiment results under MNDRv3.0 five-fold cross validation . Table 4: Similarity network ablation experiment results under piRDiseasev1.0 five-fold cross validation . (b) represents the learning module:

[0116] In this section, we will further explore the contribution and impact of the collaborative representation learning module on the prediction performance of MCRL-PDA. Specifically, we constructed three sets of experiments on two datasets, removing the heterogeneous network representation learning module, the associated network representation learning module, and the multi-view feature representation learning module one by one, aiming to explore the contribution and impact of each representation learning module on the prediction performance of MCRL-PDA one by one. All experiments use five-fold cross validation to ensure the accuracy of the evaluation results.

[0117] The experimental results are shown in Table 5 and Table 6. The results of the two datasets show that each representation learning module can significantly improve the prediction performance of the model, among which the heterogeneous network representation learning and multi-view feature representation learning modules contribute the most to the model. In addition, the collaborative learning of the three representation learning modules enables our model to achieve the highest prediction performance, which proves the effectiveness of the collaborative representation learning method.

[0118] In addition, we found that the removal of the similarity network has a more significant impact on the prediction ability of MCRL-PDA than the case of removing the representation learning module. One possible reason is that after removing a certain similarity network, the three collaborative representation learning modules of MCRL-PDA all lost the learning and representation of the similarity network information. In contrast, when a certain representation learning module is removed, the similarity network information can still be learned and integrated by the other two representation learning modules.

[0119] Table 5: Ablation experiment results of representation learning module under MNDRv3.0 five-fold cross validation . Table 6: Ablation experiment results of the representation learning module under piRDiseasev1.0 five-fold cross validation . In summary, the computational method of multi-module collaborative representation learning MCRL-PDA of the present invention is used to identify potential piRNA-disease associations. This method fully exploits five types of multi-source similarity information from piRNA and disease. The present invention introduces three collaborative representation learning modules from different perspectives, which respectively capture the diverse characteristics of nodes in heterogeneous networks, the topological structure information implicit in the association network, and the multi-view features revealed by different receptive fields. Through the interaction and complementation between the modules, MCRL-PDA can more accurately capture the complex relationship between piRNA and disease. Under five-fold cross validation on two data sets, MCRL-PDA of the present invention achieved AUC values ​​of 0.9824 and 0.9657, respectively, surpassing the other seven most advanced models. In addition, the ablation experiment results show that the method of multi-source similarity fusion and multi-module representation learning can effectively improve the ability of MCRL-PDA to identify complex relationships between different entities.

[0120] The above solutions are only an illustration of a preferred example, but are not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0121] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A method for identifying piRNA-disease associations based on multimodal collaborative representation learning, characterized in that: The method comprises the following steps: Step S101: integrating piRNA-disease association information and multi-source similarity information of piRNA and disease to construct a comprehensive heterogeneous network; Step S102: design a heterogeneous network representation learning module based on a multi-head attention mechanism to capture diverse features from piRNA and disease nodes; Step S103: construct a GCN-based association network representation learning module, taking the piRNA-disease association network as input, the integrated piRNA similarity network PS, and the integrated disease similarity network DS as features, and iteratively updating the feature representation of the nodes respectively; Step S104: designing a multi-view representation learning module based on dilated convolution; In step S104: combine the ith row of the piRNA similarity network PS and the ith row of A to construct The characteristic representation of: in, Represents a splicing operation, Represents piRNA The similarity score with all other piRNAs, Represents all Interactive disease information; Combine the j-th row of the disease similarity network DS and the j-th column of A to construct The characteristic representation of: ； Represents disease The similarity score with all other diseases, Represents all diseases Information on interacting piRNAs; Then, construct Feature representation of node pairs: ； Using dilated convolution, the receptive field of the convolution kernel is expanded to a variable size according to the dilation rate; by integrating multi-view features revealed by different receptive fields, a comprehensive feature representation T of the complex relationship between piRNA-disease node pairs is captured; Fill zeros around the overall feature representation T and express the zero-filled matrix as: ; The convolutional layer contains The height and width are and The filter, given a dilation rate d, when the qth filter moves to When the mth row and nth column are The definition is as follows: ； in, , , ; Apply the qth filter to , the element value in the affected area of ​​the qth feature map is updated through the convolution operation : ； in, is the Relu function, * represents the convolution operation, and Represents the learnable weight matrix and bias vector of the qth filter; The maximum pooling layer is introduced to extract features on the feature map, and the height and width of the filter are set to and , the element value of the affected area on the qth feature map after the maximum pooling The definition is as follows: ； Given a dilation rate d, after the above dilated convolution-pooling layer processing, the feature representation of view d is obtained , set d=[1,2,3], and obtain , and , reflecting the different views Feature representation of node pairs; Will , and splicing, and then input into another convolution-pooling layer for fusion to form a multi-view Feature representation of node pairs ; Step S105: integrating the features of the heterogeneous network representation learning module, the association network representation learning module and the multi-view representation learning module to construct an integrated multimodal piRNA-disease association pair feature representation, and predicting the association confidence between piRNA and disease based on a fully connected neural network; In step S105, a linear transformation layer is introduced after the heterogeneous network representation learning module and the multi-view representation learning module to perform dimension mapping on the features extracted by each representation learning module; The heterogeneous network represents the output of the learning module middle The i-th row and The jth row of is concatenated to form a heterogeneous network representation learning module Feature representation of associated pairs: ； The association network represents the learning module middle The i-th row and The jth row of is concatenated to form an association network representation learning module Feature representation of associated pairs: ； Each module after the linear transformation layer dimension mapping The feature representations of the associated pairs are fused: ;Will Multi-module feature representation of association pairs As input to the fully connected neural network, the confidence scores of piRNA-disease association pairs were calculated. : ； Among them, w and b represent the learnable weight matrix and bias respectively. The loss function uses the cross entropy function to measure the difference between the predicted value and the actual value. The loss function is defined as follows: ； Among them, N and They represent the number of piRNA-disease association pairs and their true labels respectively; Finally, the Adam optimizer is used to minimize the loss function.

2. The method for identifying piRNA-disease associations based on multimodal collaborative representation learning according to claim 1, characterized in that: In the step of constructing a comprehensive heterogeneous network, validated piRNA-disease associations were collected from the MNDRv3.0 and piRDisease v1.0 databases, respectively. The associated network To represent the association information of piRNA-disease, the association network is defined as follows: ； Mining five types of multi-source similarity information from piRNA and disease: piRNA sequence similarity PSS, functional similarity PFS, Gaussian interaction profile kernel similarity GIPKp, and piRNA similarity network and disease similarity networks ; Constructing a comprehensive heterogeneous network by integrating piRNA-disease association information and multi-source similarity information of piRNAs and diseases , the construction is expressed as follows: ,in, represents the adjacency matrix of piRNA-disease, yes The transpose of Represents the similarity network of piRNAs, Represents the disease similarity network.

3. The method for identifying piRNA-disease associations based on multimodal collaborative representation learning according to claim 2, characterized in that: The heterogeneous network representation learning module based on the multi-head attention mechanism includes three encoding layers. In the hth attention head of one encoding layer, the query matrix 、key matrix And the value matrix The linear transformation of is as follows: ;in, , , represents the weight matrix, It is The layer-heterogeneous network represents the output of the learning module; calculate and The dot product of , to obtain the attention score matrix ; The attention score matrix Normalized and multiplied by To obtain the A heterogeneous network representation of 4 attention heads: ； Where d is and The vector dimension of By concatenating the results of all attention heads, we get Output of the layer : ； in, Represents a splicing operation, Indicates the number of attention heads; After representation learning of 3 encoding layers, the heterogeneous network representation matrix is ​​obtained , row i and row j represent the learned diversified feature representations of piRNA and disease respectively; among them, , The feature vector captures the interaction information between piRNA molecules and disease entities, and between disease molecules and piRNA entities. and Represent the number of piRNA nodes and disease nodes, respectively.

4. The method for identifying piRNA-disease associations based on multimodal collaborative representation learning according to claim 3, characterized in that: In step S103, the first The message propagation mechanism of the layer is as follows: ； ； Where D is the degree matrix of the associated network A. By multiplying A on the left and right Normalize the rows and columns of the association network respectively, It is The embedding features of layer nodes, It is The trainable weight matrices in the layer, It is a non-linear activation function, and the non-linear activation function uses the Relu function.

Images