Phosphorylation site and disease association prediction method based on graph neural network

Abstract

The invention is suitable for the technical field of bioinformatics, and provides a phosphorylation site and disease association prediction method based on a graph neural network, and the method comprises the steps: constructing multiple views of phosphorylation sites and diseases based on multi-source biological data; constructing a multi-view heterogeneous graph based on the constructed phosphorylation site sequence similar matrix and disease semantic similar matrix in combination with site-disease associated information; coding the constructed heterogeneous view through a layer attention mechanism to generate representations of sites and diseases; and optimizing the locus and the disease representation by using a comparative learning method, and predicting the locus and the disease representation. According to the method, the prediction accuracy is remarkably improved, and the problems of insufficient information utilization, weak feature expression ability, deep map neural network over-smoothness and the like in an existing method are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bioinformatics, and particularly relates to a phosphorylation site and disease association prediction method based on a graph neural network. BACKGROUND

[0002] Protein phosphorylation is a key post-translational modification in cell signaling and metabolic regulation, and its abnormal changes are closely related to many diseases including Alzheimer's disease, tumors and metabolic diseases. Existing research mainly focuses on phosphorylation site prediction, and there are few systematic studies on "site-disease" association. In the prior art, multi-modal features are usually directly spliced or simple similarity networks are constructed, which has the following disadvantages: insufficient fusion of heterogeneous biological data, serious over-smoothing problem of deep graph model, and noise of negative samples affecting model performance. Therefore, a new method is needed that has multi-view fusion, inter-layer attention and contrast learning capabilities to significantly improve the performance of site-disease association prediction. SUMMARY

[0003] The purpose of the embodiment of the application is to provide a phosphorylation site and disease association prediction method based on a graph neural network, which aims to solve the problems raised in the background.

[0004] The embodiment of the application is implemented as follows: a phosphorylation site and disease association prediction method based on a graph neural network, comprising the following steps:

[0005] (1) Constructing a multi-view of phosphorylation sites and diseases based on multi-source biological data, including a phosphorylation site sequence similarity matrix, a disease semantic similarity matrix and site-disease association information;

[0006] (2) Constructing a multi-view heterogeneous graph based on the constructed phosphorylation site sequence similarity matrix and disease semantic similarity matrix combined with site-disease association information;

[0007] (3) Encoding the constructed multi-view heterogeneous graph through a graph layer attention mechanism to generate representations of sites and diseases;

[0008] (4) Optimizing the representations of sites and diseases using a contrast learning method and making predictions.

[0009] The embodiment of the application uses multi-source heterogeneous data of protein sequences, disease semantics and known site-disease associations, adopts a multi-view graph network combined with a graph layer attention mechanism and a contrast learning strategy, significantly improves the prediction accuracy, and solves the problems of insufficient information utilization, weak feature expression capability and over-smoothing of deep graph neural networks in existing methods. Through accurate multi-view information integration, node feature enhancement and similarity measurement optimization, the method provided by the embodiment of the application can provide strong prediction support in multiple biological fields and has important application value. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A flowchart of a phosphorylation site and disease association prediction method based on a graph neural network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0012] The specific implementation of the present application is described in detail below in combination with specific embodiments.

[0013] As shown in the figure, a flowchart of a phosphorylation site and disease association prediction method based on a graph neural network provided by an embodiment of the present application, including the following steps: Figure 1

[0014] Step 1, construction and processing of multi-source biological data, including:

[0015] Phosphorylation site sequence similarity network: using a pre-trained protein language model ProtT5 to encode the 21 amino acid sequences around the site, and using cosine similarity to calculate the similarity between sites;

[0016] Disease semantic similarity network: using MeSH (Medical Subject Headings) to construct a directed acyclic graph (DAG) of diseases, combining semantic contribution factors and rarity characteristics to generate a disease semantic similarity matrix;

[0017] GIP (Gaussian Interaction Profile) similarity network: based on the site-disease association matrix, the GIP similarity between sites and diseases is constructed to supplement the information in the sparse area;

[0018] By fusing semantic / sequence similarity and GIP similarity, a complete phosphorylation site sequence similarity matrix (PS) and disease semantic similarity matrix (DS) are obtained.

[0019] Step 2, construction of multi-view graph network (multi-view heterogeneous graph), specifically including the following contents:

[0020] Using the constructed phosphorylation site sequence similarity matrix (PS) and disease semantic similarity matrix (DS), they are combined into a similarity graph and k-nearest neighbor sparsification is performed to obtain a site similarity graph and a disease similarity graph, and Pg and Dg are obtained through GLA module (an innovative module adding layer attention based on graph attention network GAT), and combined with site-disease association network S, a heterogeneous graph adjacency matrix in the following form is constructed​ :

[0021] ;

[0022] It includes three different data views: similarity information between phosphorylation sites, similarity information between diseases, and known association information between sites and diseases. A transpose matrix representing site-disease association information;

[0023] Such heterogeneous graphs can integrate multi-level information about loci and diseases.

[0024] Step 3: Layer Attention Mechanism (GLA) encoding, specifically including:

[0025] Using GLA encoding of the heterogeneous graph, Ph and Dh are obtained:

[0026] First, features of each node are extracted using a neighborhood aggregation method. Then, using a layer attention mechanism, cross-layer feature fusion is performed by adaptively weighting the output features of each layer to generate a site representation Ph and a disease representation Dh.

[0027] By fusing the node features output by each layer using layer attention weights, the oversmoothing problem that may occur in traditional deep graph neural networks is avoided, and the discriminative ability of node features is enhanced.

[0028] Step 4: Comparative learning and multi-view information fusion:

[0029] To improve discriminative power and effectively integrate multi-view information, a contrastive learning and information fusion mechanism is introduced in this step. Negative cosine similarity is used as the loss function. Contrastive learning forces the model to maintain the consistency of the representation of the same node in different views and ensures the representation difference between different nodes. Specifically, the distance between positive and negative sample pairs is compared to optimize the model so that the representation distance of the same node from different views is minimized, while the representation distance of different nodes is maximized.

[0030] The embedding representations obtained from different views (site similarity view and disease similarity view) are weighted and fused to obtain the final site representation P and disease representation D. This fusion process is adaptively adjusted by learned gating parameters. The final representations of site and disease are fused using the following formulas:

[0031] ;

[0032] ;

[0033] in, and It is a learnable gating parameter that dynamically balances view-specific embeddings (Pg, Dg) and heterogeneous graph embeddings (Ph, Dh);

[0034] Calculate the final loss, including site contrast loss. Disease contrast loss ;

[0035] Using the fused P and D to represent phosphorylation sites and disease characteristics, the prediction scores for phosphorylation site i and disease j are calculated as follows. :

[0036] ;

[0037] in, Let be the transpose matrix of the eigenvectors of node i. This represents the feature vector of node j;

[0038] Cross-entropy loss As the primary loss, combined with the comparative loss, it constitutes the total loss. :

[0039] ;

[0040] in, and This is a hyperparameter used to balance the importance of contrastive loss and cross-entropy loss, and to optimize the accuracy of association prediction.

[0041] Once training is complete, it can predict disease associations at unknown sites.

[0042] The method provided in the embodiments of the present invention will be specifically implemented, and the detailed operation process and technical details are as follows:

[0043] (1) Data preparation: Known phosphorylation site information from the PhosphoSitePlus and PTMD databases was used. These data include the sequence, function and related protein information of each phosphorylation site. Semantic tags and related information of various diseases were extracted from MeSH. Site-disease association data in PhosphoSitePlus were used to construct the association matrix between phosphorylation sites and diseases and fill in the corresponding association relationships. Relevant data were extracted from the original database and standardized to remove missing data and noise. All phosphorylation site sequences and disease names were standardized. The amino acid sequence of each phosphorylation site was encoded using the ProtT5 model to generate a 1024-dimensional embedding vector corresponding to each site. The semantic similarity between diseases was calculated using the MeSH tags of the diseases. The diseases were encoded using DAG and the semantic similarity values ​​between the diseases were calculated.

[0044] (2) Construction of the multi-view graph network: The cosine similarity method is used to calculate the similarity between each pair of phosphorylation sites, and a similarity matrix PS of phosphorylation sites is constructed. Each element of the similarity matrix represents the sequence similarity between sites. The semantic similarity DS between diseases is calculated to construct a disease similarity matrix. This matrix is ​​based on the hierarchical relationship of DAG in MeSH. The similarity score of each pair of diseases is calculated using the calculation formula. A binary association matrix S is constructed through the known association data between phosphorylation sites and diseases. The site similarity matrix PS, the disease similarity matrix DS and the site-disease association matrix S are merged to construct an adjacency matrix of a heterogeneous graph. Then, the layer attention network (GAT) is used to extract features from the heterogeneous graph. The features of each layer are calculated by the neighborhood aggregation method. An adaptive attention mechanism is used to adjust the contribution weight of each layer to alleviate the oversmoothing problem (see the next step for details).

[0045] (3) Model Training and Optimization: A layered attention network (GAT) architecture was used for graph data processing, and the feature expression capability of the deep model was enhanced through inter-layer attention mechanism. The specific process included: input was the constructed heterogeneous graph and the embedding vectors of phosphorylation sites and diseases; information fusion was performed on different views (sites, diseases) through layered attention mechanism, and the neighborhood of nodes was weighted and aggregated using graph attention network to ensure that information was effectively fused from different views; loss functions were designed including cross-entropy loss and contrastive learning loss, and the performance of the model was improved by optimizing these losses; the dataset was divided using the 10-fold cross-validation method; each fold of the training set contained 90% of the data, and the test set contained 10% of the data; the learning rate was set to 0.001, the batch size was 128, and the number of training epochs was set to 500; the optimizer was Adam, and Dropout was used to prevent overfitting; the main evaluation metrics included AUC and AUPR; in addition, common metrics such as accuracy, precision, and recall of the model were also evaluated.

[0046] (4) Performance evaluation and application: The model was tested on a benchmark dataset and evaluated using 10-fold cross-validation. The final model achieved AUC and AUPR of 0.9961 and 0.9924, respectively, which far exceeded other existing phosphorylation site-disease association prediction methods (0.9786 and 0.9836). It also achieved excellent performance in other indicators, with an accuracy of 0.9844, an improvement of 4.96%; sensitivity of 0.9931, an improvement of 9.16%; and MCC of 0.9526, an improvement of 8.07%. In addition, case studies were conducted. In the Alzheimer's disease analysis, 29 of the top 30 predicted candidate phosphorylation sites were validated in the PhosphoSitePlus and PTMD databases, and the remaining one was confirmed in PTMD 2.0. In the tests on breast cancer and neuroblastoma, the sites predicted by the method proposed in this embodiment were validated in databases and literature, showing good prediction accuracy.

[0047] In summary, this invention, through a multi-view graph network combined with layer attention mechanisms and contrastive learning, can simultaneously process information from multiple views (such as site similarity, disease semantics, and known associations). Furthermore, the layer attention mechanism effectively avoids oversmoothing, enhancing the model's expressive and discriminative capabilities. Therefore, this method significantly outperforms traditional methods in prediction accuracy, particularly in metrics such as AUC and AUPR. Experimental results demonstrate that this method has significant application potential in multiple biological fields, providing strong support for the precise diagnosis and treatment of diseases.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting phosphorylation sites and disease associations based on graph neural networks, characterized in that, Includes the following steps: (1) Construct a multi-view of phosphorylation sites and diseases based on multi-source biological data, including a phosphorylation site sequence similarity matrix, a disease semantic similarity matrix, and site-disease association information; (2) Based on the constructed phosphorylation site sequence similarity matrix and disease semantic similarity matrix, combined with site-disease association information, a multi-view heterogeneous map is constructed; (3) Encode the constructed multi-view heterogeneous map through the layer attention mechanism to generate representations of loci and diseases; (4) Use contrastive learning methods to optimize the representation of loci and diseases and make predictions.

2. The method for predicting phosphorylation sites and disease associations based on graph neural networks according to claim 1, characterized in that, The method for constructing a multi-view of phosphorylation sites and diseases based on multi-source biological data includes steps such as a phosphorylation site sequence similarity matrix, a disease semantic similarity matrix, and site-disease association information. The process includes the following: By using the pre-trained protein language model ProtT5 to encode the amino acid sequences of phosphorylation sites, embedding vectors of site sequences are generated, and the similarity between phosphorylation sites is calculated using cosine similarity to construct a phosphorylation site sequence similarity matrix. We construct a directed acyclic graph of diseases using MeSH, and calculate the semantic similarity between diseases by combining semantic contribution factors and rarity features, thus obtaining a disease semantic similarity matrix. Based on known site-disease association data, a site-disease association matrix is ​​constructed and a GIP similarity network is calculated.

3. The method for predicting phosphorylation sites and disease associations based on graph neural networks according to claim 2, characterized in that, The steps for constructing a multi-view heterogeneous map based on the constructed phosphorylation site sequence similarity matrix and disease semantic similarity matrix, combined with site-disease association information, are as follows: The process includes the following: Based on the constructed phosphorylation site sequence similarity matrix PS and disease semantic similarity matrix DS, a similarity graph is merged. Through k-nearest neighbor sparsification, site similarity graphs and disease similarity graphs are obtained. Pg and Dg are then obtained through GLA encoding. Combined with site-disease association information S, a heterogeneous graph adjacency matrix is ​​constructed as shown below: ； It includes information on the similarity between phosphorylation sites, the similarity between diseases, and the association between sites and diseases. A transpose matrix representing site-disease association information.

4. The method for predicting phosphorylation sites and disease associations based on graph neural networks according to claim 3, characterized in that, The step of encoding the constructed multi-view heterogeneous map through a layer attention mechanism to generate representations of sites and diseases is specifically... The process includes the following: Features of each node in the heterogeneous graph are extracted using a neighborhood aggregation method. By utilizing the layer attention mechanism, cross-layer feature fusion is performed by adaptively weighting the output features of each layer to generate a site representation Ph and a disease representation Dh.

5. The method for predicting phosphorylation sites and disease associations based on graph neural networks according to claim 4, characterized in that, The steps of optimizing site and disease representations using contrastive learning methods and making predictions are as follows: The process includes the following: Using negative cosine similarity as the loss function, the method optimizes the representation distance by comparing the distance between positive and negative sample pairs, so that the representation distance of the same node from different views is minimized, while the representation distance of different nodes is maximized. The embedding representations obtained from the site similarity view and the disease similarity view are weighted and fused to obtain the final site representation P and disease representation D: ； ； in, and It is a learnable gating parameter that dynamically balances view-specific embeddings (Pg, Dg) and heterogeneous graph embeddings (Ph, Dh); Calculate the final loss, including site contrast loss. Disease contrast loss ; Using the fused P and D to represent phosphorylation sites and disease characteristics, the prediction scores for phosphorylation site i and disease j are calculated as follows. : ； in, Let be the transpose matrix of the eigenvectors of node i. This represents the feature vector of node j; Cross-entropy loss As the primary loss, combined with the comparative loss, it constitutes the total loss. : ； in, and This is a hyperparameter used to balance the importance of contrastive loss and cross-entropy loss, and to optimize the accuracy of association prediction. Predicting disease associations at unknown sites.

Images