A piRNA-disease association prediction method based on a feature conversion graph convolutional network

Abstract

The application discloses a piRNA-disease association prediction method based on a grouping feature conversion graph convolutional network, and comprises the following steps: firstly, mapping piRNA-disease association data, piRNA sequence similarity data, disease semantic similarity data and Gaussian interaction kernel similarity data into a two-layer heterogeneous network; secondly, taking the heterogeneous network as the input of the grouping feature conversion graph convolutional network, grouping the network to extract multi-group feature representations of piRNA and diseases, converting the features of the multi-group piRNA to the feature space of the diseases through a feature conversion network, and introducing a layer attention mechanism to solve the problem that different layers of the graph convolutional neural network have inconsistent contributions to node features; and finally, fusing the multi-group cosine similarity association scores through a feature fusion attention mechanism to obtain a piRNA-disease association reconstruction score. The grouping feature conversion network created by the application realizes more accurate piRNA-disease association prediction effect.

Description

Technical Field

[0001] This invention relates to the fields of intelligent computing and data mining, and in particular to a piRNA-disease association prediction method based on a grouped feature transformation graph convolutional network, abbreviated as iPDA-GTGCN. Background Technology

[0002] PIWI-interacting RNA (piRNA) is a class of endogenous small interfering RNAs that specifically bind to PIWI, an analogue of Argonaute protein in animal cells. Furthermore, it is involved in the development and progression of diseases such as cancer, diabetes, neurological disorders, and atherosclerosis. Therefore, identifying disease-associated piRNAs helps to elucidate the mechanisms of disease development and further advance our understanding of complex diseases.

[0003] Given the high cost and time-consuming nature of traditional biological research, various computational methods have been designed to effectively detect the potential association between piRNAs and diseases. Existing computational methods can be divided into two categories: machine learning-based methods and graph convolutional network-based methods.

[0004] Machine learning-based methods primarily extract features of piRNAs and diseases, using these features to train classifiers to predict potential piRNA-disease associations. Wei et al. proposed a predictor called iPiDA-sHN, which uses a two-step positive label-free learning strategy based on support vector machines to predict potential piRNA-disease associations. Its innovation lies in selecting high-quality negative samples from unknown piRNA-disease pairs. Furthermore, Wei et al. developed a computational predictor, iPiDi-PUL, which constructs a training set with known positive and negative piRNA-disease associations and uses ensemble learning to identify potential piRNA-disease associations. To improve stability and accuracy, Syed et al. constructed a prediction model that does not involve feature information, using two-step positive label-free learning and bootstrap techniques to predict bias, and evaluating model performance using k-fold cross-validation. To prevent the method from being affected by negative associations, Zheng et al. proposed a computational method based on structural perturbation, SPRDA, which uses a structural consistency index to obtain the predictability of links in the piRNA-disease association network, improving the model's predictive performance. Zheng et al. also developed the computational method APDA, which obtains the optimal feature subspace through stacked autoencoders for predicting potential piRNA-disease associations. Zhang et al. proposed a learning-ranked piRNA-disease association prediction method, iPiDA-LTR, which showed better prediction performance. However, the above methods suffer from low generalization ability and inability to learn the association information hidden in graph-structured data. For machine learning-based models, redundant information and noise contained in piRNA and disease features often degrade model performance. Graph convolutional network-based models can automatically extract feature representations and then use classifiers to predict piRNA-disease associations. Hou et al. proposed a prediction method, iPiDA-GCN, which uses graph convolutional networks to extract piRNA and disease features and obtains piRNA-disease association scores through fully connected networks and internal networks. To reduce the impact of network sparsity and confidence coefficients, Hou et al. developed iPiDA-SWGCN, which uses a supplementary weighting strategy to integrate various prediction factors, enrich the network structure information, and represent network nodes from different perspectives, achieving better prediction performance. Zhang et al. constructed PDA-PRGCN, integrating subgraph projection, node feature enhancement, and a double loss mechanism into a graph convolutional network. They utilized this graph convolutional network to construct a heterogeneous network link prediction task and reliably detected piRNA-disease associations. Meng et al. proposed ETGPDA, a piRNA-disease association prediction method based on embedding transformation graph convolutional networks. This method addresses the problem of inconsistent embedding spaces and develops a lightweight, high-learning-capability, and high-accuracy embedding transformation module to obtain prediction scores for piRNA-disease associations.The above method demonstrates the advantages of graph convolutional networks in the prediction process, which can obtain the feature representation of nodes by aggregating the information of nodes and edges in the graph structure. Summary of the Invention

[0005] The purpose of this invention is to provide a piRNA-disease association prediction method based on grouped feature transformation graph convolutional networks, abbreviated as iPDA-GTGCN. It aims to use graph convolutional network methods to solve, to some extent, the problem that existing methods are unable to fully handle feature representation in predicting piRNA-disease associations.

[0006] The technical solution provided by this invention is as follows.

[0007] Step 1: Map piRNA-disease association data, piRNA sequence similarity data, disease semantic similarity data, and Gaussian interaction kernel similarity data into a two-layer heterogeneous network.

[0008] Step 2: Use the heterogeneous network as input to the grouped feature transformation graph convolutional network, group the network to extract multiple sets of feature representations of piRNA and disease, and transform the features of multiple sets of piRNA to the feature space of disease through the feature transformation network. At the same time, a layer attention mechanism is introduced to solve the problem of inconsistent contributions of different layers of graph convolutional network to node features.

[0009] Step 3: The cosine similarity association scores of multiple groups are fused through the feature fusion attention mechanism to obtain the piRNA-disease association reconstruction score. Attached Figure Description

[0010] The invention will now be further described with reference to the accompanying drawings, in which...

[0011] Figure 1 This is a schematic diagram of the process for predicting potential disease associations with piRNA according to the present invention.

[0012] Figure 2 This is a directed acyclic graph representation of liver tumors and pancreatic tumors, illustrating the process of calculating the semantic similarity between the two diseases.

[0013] Figure 3 This is a comparison of the five-fold cross-validation results of the iPDA-GTGCN method with eight other methods.

[0014] Figure 4 This is the experimental result of running a single five-fold cross-validation test using the iPDA-GTGCN method. Detailed Implementation

[0015] To more clearly describe the purpose, technology, and features of this invention, the following will provide a more detailed description of the invention in conjunction with the accompanying drawings and specific embodiments.

[0016] In a specific embodiment of the present invention, the piRDisease v1.0 dataset was selected as a sample, and the technical solution of the present invention was evaluated using the five-fold cross-validation method.

[0017] Figure 1 The overall framework diagram of the model is shown, which is divided into three parts: building a heterogeneous network, grouping feature transformation graph convolutional neural network, and predicting association scores.

[0018] A method for predicting piRNA-disease association based on a grouped feature transformation graph convolutional network includes the following steps.

[0019] Step 1: Map piRNA-disease association data, piRNA sequence similarity data, disease semantic similarity data, and Gaussian interaction kernel similarity data into a two-layer heterogeneous network.

[0020] Step 1.1: 5002 experimentally validated piRNA-disease associations between 4350 piRNAs and 21 diseases were collected from the piRDisease v1.0 database. An association matrix A was constructed to represent the association between piRNAs and diseases.

[0021] ,

[0022] in, For the first One piRNA, For the first A disease, Indicates the first The first piRNA and the first The association between diseases, if the first disease... The first piRNA and the first The diseases have been experimentally verified to be related. It is 1 if it is true, otherwise it is 0.

[0023] Step 1.2: Calculate the sequence similarity of piRNAs ,

[0024] ,

[0025] in, This indicates the sequence similarity of the piRNA sequences calculated using the Smith-Waterman alignment algorithm. The first piRNA and the first Sequence alignment scores between individual piRNAs.

[0026] Step 1.3: Sequence similarity matrix of piRNA It is represented as follows,

[0027] .

[0028] Step 1.4: Calculate the semantic similarity of diseases The specific process is as follows: Figure 2 As shown, the semantic similarity matrix of the disease It is represented as follows,

[0029] .

[0030] Step 1.5: Calculate the Gaussian interaction kernel similarity of piRNA.

[0031] ,

[0032] in, Indicates the connection of piRNA With row vectors of 21 diseases, These are parameters that affect the original core bandwidth. Specifically, they are as follows:

[0033] ,

[0034] in, This indicates the number of piRNAs.

[0035] Step 1.6: Calculate the Gaussian interaction kernel similarity of the diseases.

[0036] ,

[0037] in, Indicates connection to disease With the row vector of 4350 piRNAs, These are parameters that affect the original core bandwidth. Specifically, they are as follows:

[0038] ,

[0039] in, Indicates the number of diseases.

[0040] Step 1.7: To address the limitations of individual similarity information, piRNA sequence similarity, disease semantic similarity, and Gaussian interaction kernel similarity are integrated.

[0041] ,

[0042] .

[0043] Step 1.8: Normalize the similarity of the integration.

[0044] ,

[0045] ,

[0046] in,

[0047] ,

[0048] .

[0049] Step 1.9: Map piRNA-disease association data, piRNA sequence similarity data, disease semantic similarity data, and Gaussian interaction kernel similarity data into a two-layer heterogeneous network.

[0050] Step 2: Use the heterogeneous network as input to the grouped feature transformation graph convolutional network, group the network to extract multiple sets of feature representations of piRNA and disease, and transform the features of multiple sets of piRNA to the feature space of disease through the feature transformation network. At the same time, a layer attention mechanism is introduced to solve the problem of inconsistent contributions of different layers of graph convolutional network to node features.

[0051] Step 2.1: Set the penalty factor for heterogeneous networks during the propagation process of the graph convolutional network. To balance node similarity and edge similarity. Input graph As shown below,

[0052] ,

[0053] in, It is the normalized piRNA integration similarity. It is normalized disease integration similarity. It is a piRNA-disease association matrix.

[0054] Step 2.2: Initialize the features to,

[0055] .

[0056] Step 2.3: After processing the messages between the pooling nodes and edges in the first layer of the graph convolutional network, the output node features of the first layer are obtained. ,

[0057] ,

[0058] in, It is a nonlinear function of the exponential linear unit used in each layer. It is an adjacency matrix The corresponding degree matrix, through the degree matrix The adjacency matrix is ​​updated by simultaneously normalizing both rows and columns, thus reducing the contribution of nodes and connected edges. This represents the weight matrix between the input layer and the hidden layer.

[0059] Step 2.4: Define the propagation method between layers as follows:

[0060] ,

[0061] pass The forward propagation was obtained Layer node features .

[0062] Step 2.5: Since the distance between neighboring nodes has different effects on a node, meaning that features extracted by different convolutional layers contribute differently to the fused features, an attention mechanism is introduced to address the issue of different neurons contributing differently to the task.

[0063] ,

[0064] in, These are the weights of each convolutional layer in the learned graph convolutional network, initialized to...

[0065] .

[0066] Step 2.6: Obtain the fused product after attention mechanism The dimensional feature is,

[0067] ,

[0068] in, Represents the fusion piRNA characteristics, Represents the fusion Disease characteristics are obtained by splicing piRNA and disease features. As input to the group feature transformation network.

[0069] Step 2.7: A feature transformation module is proposed to align the feature representations of piRNA and disease into a unified feature space, transforming the features of piRNA... Transform to the feature space of the disease, and denote the transformed output matrix as... This constitutes a four-layer convolutional neural network consisting of an input layer, hidden layers, and an output layer. The feature sizes of the first and second layers are set to be twice and four times the feature size of the input layer, respectively. A non-linear activation function is introduced in each layer. Introducing nonlinearity into the transformation process significantly enhances the model's learning ability.

[0070] Step 2.8: Grouping the feature representations of piRNA and disease based on dimensionality effectively solves the overfitting problem caused by too many neurons in traditional graph convolutional networks, or the phenomenon that some feature dimensions are not activated. It also improves upon the dropout technique in graph convolutional networks, which arbitrarily discards a portion of neurons in each training iteration. Since each dimension of a feature has its specific meaning, this module decomposes the feature transformation network into multiple sub-networks. Through the above process, all neurons in the sub-networks can be trained more effectively, and each neuron has a greater chance of contributing to the prediction of relevance, resulting in a more comprehensive prediction score.

[0071] Step 3: The cosine similarity association scores of multiple groups are fused through the feature fusion attention mechanism to obtain the piRNA-disease association reconstruction score.

[0072] Step 3.1: Cosine similarity operation is used to calculate Similarity between group features

[0073] .

[0074] Step 3.2: Similarity Matrix The calculation formula is as follows:

[0075] ,

[0076] in, and These represent the number of piRNAs and the number of diseases, respectively. Indicates the first The first piRNA and the first Cosine similarity scores among diseases.

[0077] Step 3.3: Similarity Matrix Used to store the predicted scores of the grouping matrix, where Indicates the number of groups.

[0078] Step 3.4: Since the features of different groups represent different emphases, and different features have different impacts on the results, a fusion attention mechanism is applied to the similarity scores obtained after grouping to achieve more comprehensive information exchange between different groups, taking into account the problem of inconsistent contributions of different feature dimensions to the features. The proposed feature fusion attention mechanism is shown below.

[0079] ,

[0080] in, This represents the learning weight for each group's predicted score, initialized to,

[0081] .

[0082] Step 3.5: The fused similarity prediction score is used as the reconstruction score, as shown below.

[0083] .

[0084] To verify the effectiveness of this invention, this embodiment compares the association matrix predicted by the prediction model of the present invention with the actual association labels, uses five-fold cross-validation for prediction evaluation, calculates the AUC value, and compares it with existing methods. Figure 3 As shown, this invention compared eight methods, including PDA-PRGCN, ETGPDA, DFL-PiDA, piRDA, iPiDA-sHN, iPiDi-PUL, GAPDA, and SPRDA. The prediction results show that the method proposed in this invention has a significantly higher AUC value than the compared methods, effectively improving the performance of piRNA-disease association prediction and demonstrating high practicality.

[0085] like Figure 4 As shown, the iPDA-GTGCN method achieved a mean AUC of 0.9656, with an AUC exceeding 96% per fold. The mean accuracy of iPDA-GTGCN reached 0.9572, reflecting the accuracy of the latent association score calculated by this method. Furthermore, the mean recall and specificity results were 0.8662 and 0.9481, respectively. These experimental results confirm the reliability of the iPDA-GTGCN method in predicting potential piRNA-disease associations. This invention demonstrates robust and balanced performance across multiple evaluation metrics, enhancing its applicability for predicting potential piRNA-disease associations.

[0086] It should be emphasized that the embodiments described in this invention are illustrative and not limiting. Therefore, this invention is not limited to the embodiments described in the specific implementation. Any other implementation methods derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1.A piRNA-disease association prediction method based on a graph convolutional network with feature transformation of packets, characterized in that, Includes the following steps: Step 1: A grouped feature transformation graph convolutional network was created. The piRNA-disease two-layer heterogeneous network was grouped to extract the feature representations of piRNA and disease. The feature transformation network was used to transform the features of multiple groups of piRNA to the feature space of the disease. Step 2: A feature fusion attention mechanism was proposed to fuse multiple sets of cosine similarity association scores to obtain the piRNA-disease association reconstruction score; The feature fusion attention mechanism described in step 2 fuses multiple sets of cosine similarity association scores to obtain the piRNA-disease association reconstruction score. The specific steps are as follows: Step 2.1: Cosine similarity operation is used to compute Similarity between group features: ； Step 2.2: Similarity matrix The formula is as follows: ， wherein, and represent the number of piRNAs and diseases, respectively, represents the cosine similarity score between the th piRNA and the th disease; Step 2.3: Similarity matrix used as a prediction score to hold the group matrix, wherein denotes the number of groups; Step 2.4: Use a feature fusion attention mechanism to fuse the similarity scores of each group obtained after grouping, as shown below: ， in, The learning weights for each predicted score are initialized as follows: ； Step 2.5: The fused similarity prediction score is used as the reconstruction score, as shown below: 。 2. The method according to claim 1, characterized in that, The grouped feature transformation graph convolutional network described in step 1 uses heterogeneous networks as input to group the network to extract multiple sets of feature representations of piRNA and disease. The feature transformation network transforms the features of multiple sets of piRNA to the feature space of the disease, and introduces a layer attention mechanism. The specific steps are as follows: Step 1.1: Set the penalty factor for heterogeneous networks during the propagation process of the graph convolutional network. To balance node similarity and edge similarity, input graph As shown below: ， in, It is the normalized piRNA integration similarity. It is normalized disease integration similarity. It is a piRNA-disease association matrix; Step 1.2: Initialize the features as follows: ； Step 1.3: After processing the messages between the pooling nodes and edges in the first layer of the graph convolutional network, the output node features of the first layer are obtained. : ， in, It is a nonlinear function of the exponential linear unit used in each layer. It is an adjacency matrix The corresponding degree matrix, through the degree matrix The adjacency matrix is ​​updated by simultaneously normalizing both rows and columns, thus reducing the contribution of nodes and connected edges. The weight matrix represents the relationship between the input layer and the hidden layer; Step 1.4: Define the propagation method between layers as follows: ， pass The forward propagation was obtained Layer node features ; Step 1.5: Introduce an attention mechanism to address the issue of different neurons contributing differently to the task: ， in, These are the weights of each convolutional layer in the learned graph convolutional network, initialized as follows: ； Step 1.6: Obtain the fused product after attention mechanism The dimensional features are: ， in, Represents the fusion piRNA characteristics, Represents the fusion Disease characteristics are obtained by splicing piRNA and disease features. As input to the group feature transformation network; Step 1.7: A feature transformation module is proposed to align the feature representations of piRNA and disease into a unified feature space, transforming the features of piRNA... Transform to the feature space of the disease, and denote the transformed output matrix as... This constitutes a four-layer convolutional neural network consisting of an input layer, hidden layers, and an output layer. The feature sizes of the first and second layers are set to be twice and four times the feature size of the input layer, respectively. A non-linear activation function is introduced in each layer. Introducing nonlinearity into the conversion process; Step 1.8: Group the feature representations of piRNA and disease based on dimensionality, and decompose the feature transformation network into multiple sub-networks to better train all neurons in the sub-networks.

Images