Tag-guided methods for predicting protein interactions

Abstract

This invention specifically relates to a label-guided method for predicting protein-protein interactions, comprising: acquiring a protein to be predicted; inputting the protein to be predicted into a trained prediction model, outputting the corresponding predicted relationship; firstly, performing graph data augmentation on the protein to obtain a multi-scale graph representation; secondly, inputting the multi-scale graph representation into a graph neural network to obtain a multi-scale protein feature representation, and introducing contrastive learning to eliminate differences in protein feature representations at different scales; then constructing a self-learning label relationship graph and learning the relationships between labels to obtain label feature representations; finally, correcting the protein feature representations using the label feature representations and guiding the prediction of protein-protein interactions; and using the predicted relationship of a protein to be predicted as its protein-protein interaction prediction result. This invention can improve the generalization ability of protein feature representations and the classification accuracy of the prediction model, thereby improving the prediction effect of protein-protein interactions.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and natural language processing, specifically to a label-guided method for predicting protein-protein interactions. Background Technology

[0002] Protein-protein interactions play a crucial role in a wide range of biological processes, such as DNA replication, transcription, translation, and transmembrane signal transduction. Therefore, detecting protein-protein interactions (PPIs) and their types is essential for understanding cellular biological processes in normal and disease states. This type of research also contributes to the identification of therapeutic targets and the design of new drugs. Early research on protein-protein interactions relied on laboratory-based methods, primarily including yeast two-hybrid screening, protein microarrays, and mass spectrometry-based protein complex identification. Laboratory experiments are typically time-consuming and labor-intensive, leading to low efficiency in identifying protein-protein interactions. Furthermore, the limitations of laboratory experiments result in incomplete data on protein interactions generated by these methods.

[0003] Existing research on protein interaction prediction using deep learning algorithms primarily employs Convolutional Neural Networks (CNNs) to extract local protein features or Recurrent Neural Networks (RNNs) to preserve long-range dependencies in the context. However, these deep learning algorithms still suffer from several limitations, such as the inability to effectively filter and aggregate local protein features, difficulty in simultaneously preserving important contextual and sequence amino acid information, and a lack of utilization of the interactions between protein pairs. With the development of Graph Neural Networks (GNNs), current techniques have begun to construct protein interaction network graphs and incorporate them for prediction. This approach not only considers the influence between protein pairs but also enhances the feature representation through the relationships between protein pairs, further improving the accuracy of protein interaction prediction.

[0004] However, the applicant found in their actual research that existing methods for predicting protein interactions based on graph neural networks only construct protein interaction network graphs and protein feature representations based on the original dataset, without fully exploring the original dataset. This results in insufficient generalization ability of the protein feature representations, leading to poor performance in protein interaction prediction. Furthermore, proteins often have multiple interactions, which may contain interrelated information. Existing methods for predicting protein interactions based on graph neural networks do not consider this interrelationship information, resulting in insufficient classification accuracy of the protein interaction prediction models. Therefore, designing a method that can improve the generalization ability of protein feature representations and the classification accuracy of prediction models is an urgent technical problem to be solved. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a label-guided protein interaction prediction method that can improve the generalization ability of protein feature representation and the classification accuracy of the prediction model, thereby improving the prediction effect of protein interaction and enabling better analysis of the cell biological processes of the protein's parent in normal and disease states.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] Tag-guided protein interaction prediction methods include the following steps:

[0008] S1: Obtain a test protein to be predicted;

[0009] S2: Input a protein to be tested into the trained prediction model and output the corresponding prediction relationship;

[0010] The prediction model first performs graph data augmentation on the protein to be tested to obtain a multi-scale graph representation; then, the multi-scale graph representation is input into a graph neural network to obtain a multi-scale protein feature representation, and contrastive learning is introduced to eliminate the differences in protein feature representations at different scales; then, a self-learning label relationship graph is constructed and the relationships between labels are learned to obtain label feature representations; finally, the label feature representations are used to correct the protein feature representations and guide the prediction of protein interaction relationships, thereby outputting the corresponding predicted relationships.

[0011] S3: The predicted relationship of a protein to be tested is used as the prediction result of its protein interaction relationship. Then, based on the prediction result of the protein interaction relationship, the cell biological processes of the subject to which the protein to be tested belongs in normal and disease states are analyzed.

[0012] Preferably, in step S2, the prediction model includes a protein feature encoder module for extracting local and global features of proteins, a multi-scale graph neural network module for data augmentation, graph neural network processing, and contrastive learning, a self-learning label relationship graph module for learning the relationship between labels, and a multi-label loss calculation module for performing self-supervised learning tasks and supervised learning tasks.

[0013] Preferably, in step S2, the prediction model is trained through the following steps:

[0014] S201: Obtain a pair of proteins for training and input them into the prediction model;

[0015] S202: Extract local and global features of proteins through the protein feature encoder module to obtain protein feature representations with local and global information;

[0016] S203: Construct an original graph of protein interactions based on protein feature representations; perturb the original graph using a multi-scale graph neural network module to obtain a corresponding perturbed graph; then input the original graph and the perturbed graph into the graph neural network, outputting the original node feature representation and the perturbed node feature representation, i.e., the multi-scale protein feature representation; then fuse the original node feature representation and the perturbed node feature representation through contrastive learning to obtain the fused node feature representation; finally, calculate the fused edge feature representation using the fused node feature representation.

[0017] S204: Obtain the tag name embedding representation through the self-learning tag relationship graph module and construct the tag relationship graph; then input the tag relationship graph into the graph convolutional neural network and output the tag node feature representation;

[0018] S205: Correct the fusion edge feature representation by using the label node feature representation to obtain the protein relationship graph edge feature representation;

[0019] S206: The multi-label loss calculation module performs self-supervised learning through the original node feature representation and the perturbed node feature representation to obtain the self-supervised learning loss function; then it performs supervised learning through the edge feature representation of the protein relationship graph to obtain the supervised learning loss function; finally, it calculates the training loss function based on the self-supervised learning loss function and the supervised learning loss function, and then optimizes and updates the parameters of the prediction model through the training loss function.

[0020] S207: Repeat steps S201 to S206 until the prediction model converges.

[0021] Preferably, in step S202, the protein feature encoder module includes a local feature encoder and a global feature encoder;

[0022] The local feature encoder, consisting of a convolutional neural network and a max-pooling layer, extracts the input protein using the following formula. Local feature representation h in i ;

[0023] h i =f GMP (f CNN (p i ;θ CNN ));

[0024]

[0025] In the formula: f CNN Indicates the convolution operation; f GMP This represents the max pooling layer operation; Represents a collection of proteins; This refers to a defined amino acid vocabulary; a j This refers to amino acids in the amino acid vocabulary; θ CNN Represents the training parameters for the convolution operation;

[0026] The global feature encoder consists of a bidirectional gated recurrent unit and a global average pooling layer, and extracts the local feature representation h of the input using the following formula. i The global feature representation in the model yields a protein feature representation x that contains both local and global information. i ∈X;

[0027] x i =f GAP (f BiGRU (h i ;θ BiGRU ));

[0028] In the formula: f BiGRU Indicates a bidirectional gated loop operation; f GAP Represents the global average pooling layer operation; θ BiGRU represents the training parameters for the bidirectional gated loop operation; X represents the protein feature representation obtained based on the protein feature encoder module.

[0029] Preferably, in step S203, the original graph G = (X, A) is defined, and the node features are... and adjacency matrix

[0030] 1) The multi-scale graph neural network module first applies random graph data augmentation functions from two different perspectives. and The edges and node features of the original graph G=(X,A) are perturbed respectively to obtain the first perturbed graph G1=(X,A1) and the second perturbed graph G2=(X2,A);

[0031]

[0032]

[0033]

[0034] ∈1~Bernoulli(N,1-δ1);

[0035]

[0036]

[0037] In the formula: ∈1 represents the N Bernoulli distribution result obtained based on the hyperparameter δ1∈(0,1); This indicates that the graph data augmentation function is applied based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; Let represent the set of edges in the original graph; Bernoulli represents the Bernoulli distribution; δ1∈(0,1) is a hyperparameter representing the ratio of edges removed. X1 represents the uniform distribution result obtained based on the hyperparameter δ2∈(0,1); X2 represents the graph data augmentation function based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; X represents the node features of the original graph; Unifrom represents a uniform distribution; δ2∈(0,1) is a hyperparameter representing the ratio of node features set to 0;

[0038] 2) Input the original graph G = (X,A), the first perturbation graph G1 = (X,A1), and the second perturbation graph G2 = (X2,A) into the graph neural network, and output the original node feature representations. First perturbation node feature representation Second perturbation node feature representation

[0039] A graph neural network with k iterations is represented as:

[0040]

[0041]

[0042] In the formula: AGG represents the representation of node v after aggregating the features of its neighboring nodes; AGG represents the function that aggregates the features of a node. or This represents the result of node u based on the (k-1)th iteration of the graph convolutional network; represents the set of neighbors of node v; UPDATE represents the node feature update function; or This represents the result of node v through k-1 iterations of a graph convolutional network; or ω represents the feature representation of node v in the k-th iteration; MLP represents a multilayer perceptron neural network; ω is a learnable parameter or a constant.

[0043] 3) The original node feature representations are fused using the following formula. First perturbation node feature representation The fused node feature representation is obtained from the feature representation of the second perturbation node.

[0044] Z' = f Fusion ([Z0,Z1,Z2]);

[0045] In the formula: f Fusion Indicates the fusion function;

[0046] 4) Obtain the fused edge feature representation by fusing the node feature representation Z'.

[0047] e ij =z′ i ⊙z′ j e ij ∈E;

[0048] In the formula: ⊙ represents the Hadamard product; and Let i and j represent the feature representations of node i and node j, respectively.

[0049] Preferably, in step S204, the self-learning label relationship graph module obtains the label name embedding representation through the pre-trained model BERT;

[0050] X L =BERT(L NAME );

[0051] In the formula: Indicates the label name; The word vector representing the tag name, i.e., the tag name embedding representation;

[0052] X is represented by embedding the label name. L and learnable parameter matrix Construct a label relationship graph G L =(A L ,X L );

[0053] Label relationship graph G L =(A L ,X LIn a convolutional neural network with an input graph, the output label node feature representation

[0054] In the formula: Initialize Z L (0) =X L ; Degree matrix; W (l-1) Denotes the learnable parameter matrix; σ represents the sigmoid activation function; A L It is initialized as an identity matrix.

[0055] Preferably, in step S205, the edge feature representation of the protein relationship graph is calculated using the following formula;

[0056]

[0057] In the formula: E represents the edge feature representation of the protein-protein relationship graph; E represents the fusion edge feature representation. Tag node feature representation.

[0058] Preferably, in step S206, the multi-label loss calculation module, in the self-supervised learning task, uses the original node feature representation... and the first perturbation node feature representation Calculate the first disturbance loss function Based on the original node feature representation Second perturbation node feature representation Calculate the second disturbance loss function First perturbation loss function Second perturbation loss function This is the self-supervised learning loss function;

[0059]

[0060]

[0061]

[0062]

[0063] In the formula: (z 1,i ,z 0,i ) represents a positive sample pair; z 0,i ∈Z0, z 1,i ∈Z1;θ(z 0,i ,z 1,i ) indicates the calculation of z 0,i and z 1,i The cosine similarity; τ represents the temperature parameter; N represents the set of all nodes.

[0064] Preferably, in step S206, the multi-label loss calculation module calculates the following supervised learning loss function in the supervised learning task using the edge feature representation of the protein relationship graph;

[0065]

[0066]

[0067] p ij =Softmax(e ij );

[0068]

[0069] In the formula: The value represents the supervised learning loss; t represents the number of label categories. p represents the training set of the edge set; ij This represents the probability distribution of the relationship between proteins i and j; The predicted relationship represents the connection between proteins i and j; c represents the specific category of the label. This represents the true labels of proteins i and j in category c; This represents the prediction results of proteins i and j in category c; argmax represents taking the largest index in the set.

[0070] Preferably, in step S206, the training loss function is expressed by the following formula;

[0071]

[0072] In the formula: Indicates training loss; Indicates the loss in supervised learning; Indicates the loss from the first disturbance; λ represents the second disturbance loss; λ1 and λ2 represent the set hyperparameters.

[0073] The tag-guided protein interaction prediction method of this invention has the following beneficial effects:

[0074] The prediction model of this invention generates multi-scale graph representations through graph data augmentation, learns multi-scale protein feature representations by strengthening the feature representations of neighboring nodes in the graph neural network, and eliminates the differences in protein feature representations at different scales by introducing contrastive learning to further improve protein characterization capabilities. In other words, by fully exploring the original dataset through graph data augmentation, graph neural network processing, and contrastive learning, the generalization ability of protein feature representations can be improved, thereby enhancing the prediction effect of protein-protein interactions and enabling better analysis of the cellular biological processes of the protein's parent organism in normal and disease states.

[0075] Meanwhile, the prediction model of this invention introduces label information and learns the relationship between labels by constructing a self-learning label relationship graph to obtain label feature representation. Then, the label feature representation guides the learning of protein interaction relationships. That is, by learning the relationship between labels, it can fully explore the interrelationship information generated by multiple interaction relationships between proteins, which can improve the classification accuracy of the prediction model and further improve the prediction effect of protein interaction relationships, so as to better analyze the cell biological processes of the protein's host in normal and disease states. Attached Figure Description

[0076] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0077] Figure 1 This is a block diagram of a tag-guided protein interaction prediction method.

[0078] Figure 2 This is a network architecture diagram of the prediction model (LGMG-PPI);

[0079] Figure 3 This is a schematic diagram for verifying the validity of the SL-LRG topology;

[0080] Figure 4 This is a schematic diagram for validating the features of SL-LRG nodes. Detailed Implementation

[0081] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0082] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the figures, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, the terms "horizontal," "vertical," etc., do not indicate that the component is required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0083] The following detailed explanation illustrates the specific implementation methods:

[0084] Example:

[0085] This embodiment discloses a tag-guided method for predicting protein interaction relationships.

[0086] like Figure 1As shown, the tag-guided protein interaction prediction method includes the following steps:

[0087] S1: Obtain a test protein to be predicted;

[0088] S2: Input a protein to be tested into the trained prediction model and output the corresponding prediction relationship;

[0089] The prediction model first performs graph data augmentation on the protein to be tested to obtain a multi-scale graph representation; then, the multi-scale graph representation is input into a graph neural network to obtain a multi-scale protein feature representation, and contrastive learning is introduced to eliminate the differences in protein feature representations at different scales; then, a self-learning label relationship graph is constructed and the relationships between labels are learned to obtain label feature representations; finally, the label feature representations are used to correct the protein feature representations and guide the prediction of protein interaction relationships, thereby outputting the corresponding predicted relationships.

[0090] In this embodiment, as Figure 2 As shown, the prediction model includes a protein feature encoder module for extracting local and global features of proteins, a multi-scale graph data augmentation (MS-GDA) module for data augmentation and obtaining multi-scale protein feature representations, a self-learning label relation graph (SL-LRG) module for learning the relationships between labels, and a multi-label loss calculation module for performing self-supervised learning tasks and supervised learning tasks.

[0091] S3: The predicted relationship of a protein to be tested is used as the prediction result of its protein interaction relationship. Then, based on the prediction result of the protein interaction relationship, the cell biological processes of the subject to which the protein to be tested belongs in normal and disease states are analyzed.

[0092] In this embodiment, the cellular biological processes of the target subject (human or animal) under normal and disease states can be analyzed based on the protein-protein interaction prediction results of the target protein. This allows for the identification of therapeutic targets and the design of new drugs based on the analyzed data. The analysis of cellular biological processes, the identification of therapeutic targets, and the design of new drugs all utilize existing mature technologies. This invention improves the effectiveness of cellular biological process analysis, therapeutic target identification, and new drug design by obtaining better protein-protein interaction prediction results.

[0093] The prediction model of this invention generates multi-scale graph representations through graph data augmentation, learns multi-scale protein feature representations by strengthening the feature representations of neighboring nodes in the graph neural network, and eliminates the differences in protein feature representations at different scales by introducing contrastive learning to further improve protein characterization capabilities. In other words, by fully exploring the original dataset through graph data augmentation, graph neural network processing, and contrastive learning, the generalization ability of protein feature representations can be improved, thereby enhancing the prediction effect of protein-protein interactions and enabling better analysis of the cellular biological processes of the protein's parent organism in normal and disease states.

[0094] Meanwhile, the prediction model of this invention introduces label information and learns the relationship between labels by constructing a self-learning label relationship graph to obtain label feature representation. Then, the label feature representation guides the learning of protein interaction relationships. That is, by learning the relationship between labels, it can fully explore the interrelationship information generated by multiple interaction relationships between proteins, which can improve the classification accuracy of the prediction model and further improve the prediction effect of protein interaction relationships, so as to better analyze the cell biological processes of the protein's host in normal and disease states.

[0095] In the specific implementation process, the prediction model is trained through the following steps:

[0096] S201: Obtain a pair of proteins for training and input them into the prediction model;

[0097] S202: Extract local and global features of proteins through the protein feature encoder module to obtain protein feature representations with local and global information;

[0098] S203: Construct an original graph of protein interactions based on protein feature representations; perturb the original graph using a multi-scale graph neural network module to obtain a corresponding perturbed graph; then input the original graph and the perturbed graph into the graph neural network, outputting the original node feature representation and the perturbed node feature representation, i.e., the multi-scale protein feature representation; then fuse the original node feature representation and the perturbed node feature representation through contrastive learning to obtain the fused node feature representation; finally, calculate the fused edge feature representation using the fused node feature representation.

[0099] S204: Obtain the tag name embedding representation through the self-learning tag relationship graph module and construct the tag relationship graph; then input the tag relationship graph into the graph convolutional neural network and output the tag node feature representation;

[0100] S205: Correct the fusion edge feature representation by using the label node feature representation to obtain the protein relationship graph edge feature representation;

[0101] S206: The multi-label loss calculation module performs self-supervised learning through the original node feature representation and the perturbed node feature representation to obtain the self-supervised learning loss function; then it performs supervised learning through the edge feature representation of the protein relationship graph to obtain the supervised learning loss function; finally, it calculates the training loss function based on the self-supervised learning loss function and the supervised learning loss function, and then optimizes and updates the parameters of the prediction model through the training loss function.

[0102] S207: Repeat steps S201 to S206 until the prediction model converges.

[0103] This invention generates multi-scale graph representations through graph data augmentation during predictive model training. It learns multi-scale protein feature representations by strengthening the feature representations of neighboring nodes in a graph neural network. Contrastive learning is introduced to eliminate differences in protein feature representations at different scales and further improve protein characterization capabilities. Simultaneously, label information is introduced, and a self-learning label relationship graph is constructed to learn the relationships between labels, resulting in label feature representations. These label feature representations then guide the learning of protein-protein interaction relationships. In other words, graph data augmentation, graph neural network processing, and contrastive learning fully explore the original data, and by learning the relationships between labels, fully explore the interrelationships generated by various protein interactions. This improves the generalization ability of protein feature representations and the classification accuracy of the predictive model, thereby further enhancing the predictive performance of protein interaction relationships.

[0104] It should be noted that the prediction model of this invention can be regarded as a label-guided multi-scale graph neural network (LGMG-PPI) prediction model for protein interaction relationships.

[0105] Proteins are composed of sequences of amino acids; there are 20 common amino acids. (Amino acid glossary definition follows.) protein collection in definition Let I be the set of PPIs (protein interactions), where I represents whether there is a relationship between two proteins. If I(x) ij ) = 1, indicating that protein p i and protein p j There is an interaction relationship between them; if I(x) ij ) = 0, indicating that protein p i and protein p jThere is no interaction between them, or it means that no interaction has been found between them in current research. Based on the above definition, a PPI graph is constructed, with proteins as nodes and PPIs as edges.

[0106] Protein-protein interaction relationships only indicate whether an interaction exists between proteins; however, multiple interaction relationships may exist between proteins. The objective of this invention is to predict these multiple interaction relationships, which is a multi-label classification task. This invention defines the label set for PPIs as follows: Where t represents the existence of t kinds of interaction relationships.

[0107] In practice, the protein feature encoder module includes a local feature encoder and a global feature encoder.

[0108] The local feature encoder, consisting of a convolutional neural network (CNN) and a global max pooling (GMP) layer, extracts the input protein using the following formula. Local feature representation h in i ;

[0109] h i =f GMP (f CNN (p i ;θ CNN ));

[0110]

[0111] In the formula: f CNN Indicates the convolution operation; f GMP This represents the max pooling layer operation; Represents a collection of proteins; This refers to a defined amino acid vocabulary; a j This refers to amino acids in the amino acid vocabulary; θ CNN Represents the training parameters for the convolution operation;

[0112] The global feature encoder consists of a bidirectional gated recurrent unit (BiGRU) and a global average pooling (GAP) layer, which extracts the local feature representation h of the input using the following formula. i The global feature representation in the model yields a protein feature representation x that contains both local and global information. i ∈X;

[0113] x i =f GAP (f BiGRU (hi ;θ BiGRU ));

[0114] In the formula: f BiGRU Indicates a bidirectional gated loop operation; f GAP Represents the global average pooling layer operation; θ BiGRU represents the training parameters for the bidirectional gated loop operation; X represents the protein feature representation obtained based on the protein feature encoder module.

[0115] This invention extracts local and global features of proteins through feature encoding, obtaining protein feature representations with both local and global information, thereby improving the ability to characterize proteins.

[0116] In practice, the Multi-Scale Graph Data Augmentation (MS-GDA) module mainly includes two types of graph data augmentation functions.

[0117] Define the original graph G = (X, A), and the node features. (Using the protein feature representation obtained from the protein feature encoder module as node features) and adjacency matrix

[0118] 1) The multi-scale graph neural network module first applies random graph data augmentation functions from two different perspectives. and The edges and node features of the original graph G=(X,A) are perturbed respectively to obtain the first perturbed graph G1=(X,A1) and the second perturbed graph G2=(X2,A);

[0119] Perturb the edges of the original graph G = (X, A) and randomly delete the edges that disrupt the topology of the original graph.

[0120] The node features of the original graph G=(X,A) are perturbed by randomly setting some columns of the node features of the original graph to 0.

[0121]

[0122]

[0123]

[0124] ∈1~Bernoulli(N,1-δ1);

[0125]

[0126]

[0127] In the formula: ∈1 represents the N Bernoulli distribution result obtained based on the hyperparameter δ1∈(0,1); This indicates that the graph data augmentation function is applied based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; Let represent the set of edges in the original graph; Bernoulli represents the Bernoulli distribution; δ1∈(0,1) is a hyperparameter representing the ratio of edges removed. X1 represents the uniform distribution result obtained based on the hyperparameter δ2∈(0,1); X2 represents the graph data augmentation function based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; X represents the node features of the original graph; Unifrom represents a uniform distribution; δ2∈(0,1) is a hyperparameter representing the ratio of node features set to 0;

[0128] 2) Input the original graph G = (X,A), the first perturbation graph G1 = (X,A1), and the second perturbation graph G2 = (X2,A) into a graph convolutional network (GIN is used in this embodiment) and output the original node feature representation. First perturbation node feature representation Second perturbation node feature representation

[0129] GNN is one of the most effective graph representation learning methods currently available. Its main idea is to aggregate the features of neighboring nodes and then update the feature representation of its own node. Through k iterations of aggregation and update, the node representation aggregates the representations of its k-hop neighboring nodes.

[0130] A graph neural network with k iterations is represented as:

[0131]

[0132]

[0133] In the formula: AGG represents the representation of node v after aggregating the features of its neighboring nodes; AGG represents the function that aggregates the features of a node. or This represents the result of node u based on the (k-1)th iteration of the graph convolutional network; represents the set of neighbors of node v; UPDATE represents the node feature update function; or This represents the result of node v through k-1 iterations of a graph convolutional network; or ω represents the feature representation of node v in the k-th iteration; MLP represents a multilayer perceptron neural network; ω is a learnable parameter or a constant.

[0134] 3) The original node feature representations are fused using the following formula. First perturbation node feature representation The fused node feature representation is obtained from the feature representation of the second perturbation node.

[0135] Z' = f Fusion ([Z0,Z1,Z2]);

[0136] In the formula: f Fusion Indicates the fusion function;

[0137] 4) Obtain the fused edge feature representation by fusing the node feature representation Z'.

[0138] e ij =z′ i ⊙z′ j e ij ∈E;

[0139] In the formula: ⊙ represents the Hadamard product; and Let i and j represent the feature representations of node i and node j, respectively.

[0140] In specific implementation, this invention adopts a self-learning approach to obtain the relationship representation between labels and construct a self-learning label relationship graph (SL-LRG).

[0141] First, set a learnable parameter. T represents the number of label categories. AL is initialized as the identity matrix, serving as the initial topology of the label relationship graph.

[0142] Then, the label name embedding representation is obtained through the pre-trained model BERT;

[0143] X L =BERT(L NAME );

[0144] In the formula: Indicates the label name; The word vector representing the tag name, i.e., the tag name embedding representation;

[0145] X is represented by embedding the label name. L and learnable parameter matrix Construct a label relationship graph G L =(A L,X L );

[0146] Label relationship graph G L =(A L ,X L The input is fed into a Graph Convolutional Network (GCN), which outputs the feature representation of the labeled nodes.

[0147]

[0148] In the formula: Initialize Z L (0) =X L ; Degree matrix; W (l-1) Denotes the learnable parameter matrix; σ represents the sigmoid activation function; A L It is initialized as an identity matrix. During model training, parameters A are updated via gradient backpropagation. L This allows the system to learn the implicit label relationships within the data, thereby achieving the goal of self-learning label relationship graphs.

[0149] In the specific implementation process, the edge feature representation of the protein relationship graph is calculated using the following formula;

[0150]

[0151] In the formula: E represents the edge feature representation of the protein-protein relationship graph; E represents the fusion edge feature representation. Tag node feature representation.

[0152] In practice, data often contains noise, which can prevent the model from accurately representing the original data distribution and severely impact its learning performance. To address this issue, this invention introduces a self-supervised learning task into the prediction model. The purpose of this task is to enhance the accuracy of the primary learning task and improve the model's performance by adding an auxiliary task.

[0153] In self-supervised learning tasks, the multi-label loss calculation module uses the original node feature representation. and the first perturbation node feature representation Calculate the first disturbance loss function Based on the original node feature representation Second perturbation node feature representation Calculate the second disturbance loss function First perturbation loss function Second perturbation loss function This is the self-supervised learning loss function;

[0154]

[0155]

[0156]

[0157]

[0158] In the formula: (z 1,i ,z 0,i ) represents a positive sample pair; z 0,i ∈Z0, z 1,i ∈Z1;θ(z 0,i ,z 1,i ) indicates the calculation of z 0,i and z 1,i The cosine similarity; τ represents the temperature parameter, which controls the model's ability to distinguish negative samples. A smaller value indicates that the model pays more attention to difficult negative samples; N represents the set of all nodes.

[0159] In the specific implementation process, the multi-label loss calculation module calculates the following supervised learning loss function in the supervised learning task through the edge feature representation of the protein relationship graph;

[0160]

[0161]

[0162]

[0163]

[0164] In the formula: The value represents the supervised learning loss; t represents the number of label categories. p represents the training set of the edge set; ij This represents the probability distribution of the relationship between proteins i and j; The predicted relationship represents the connection between proteins i and j; c represents the specific category of the label. This represents the true labels of proteins i and j in category c; This shows the prediction results of proteins i and j in category c; argmax represents taking the largest index in the set.

[0165] In practice, the training loss function is represented by the following formula;

[0166]

[0167] In the formula: Indicates training loss; Indicates the loss in supervised learning; Indicates the loss from the first disturbance; λ represents the second disturbance loss; λ1 and λ2 represent the set hyperparameters.

[0168] To better illustrate the advantages of the technical solution of the present invention, the following experiments are also disclosed in this embodiment.

[0169] 1. Dataset

[0170] This experiment uses the same dataset setup as previous work (published in LV GF, HU ZQ, BI YG et al. Learning Unknown from Correlations: Graph Neural Network for Inter-novel-protein Interaction Prediction), and uses PPI data from the STRING database (published in SZKLARCZYK D, GABLE AL, LYON D, et al. STRING v11: protein–protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets) to evaluate the model.

[0171] The STRING database collects, scores, and integrates most publicly available PPIs data, and establishes a comprehensive and objective PPIs network. Furthermore, Chen et al. (published in CHEN M, JU CJT, ZHOU G, et al. Multifaceted protein–protein interaction prediction based on Siamese residual RCNN) extracted two subsets from STRING, named SHS27k and SHS148k. Specific information for the three datasets is shown in Table 1. The original dataset is a protein network graph, where nodes represent proteins and edges represent interactions between proteins. Secondly, since proteins are composed of amino acid sequences, the average length of the amino acid sequences constituting the proteins in each dataset was statistically analyzed, as shown in Table 1.

[0172] Table 1 Dataset Statistics

[0173]

[0174] 2. Experimental setup and evaluation indicators

[0175] In this experiment, 20% of the data was randomly selected from the dataset as the test set. Furthermore, to eliminate the impact of randomness in data partitioning on the performance of the PPI method, the experimental results were repeated under three different random seeds. This experiment used protein features based on amino acid sequences, referring to the amino acid embedding method used by Chen et al. (published in CHEN M, JU CJT, ZHOU G, et al. Multifaceted protein–protein interaction prediction based on Siamese residual RCNN) to represent each amino acid. The model used the Adam algorithm to update all trainable parameters. This experiment followed the experimental setup of previous work (published in LV GF, HU ZQ, BI YG et al. Learning Unknown from Correlations: Graph Neural Network for Inter-novel-protein Interaction Prediction), using micro-F1 as the evaluation metric.

[0176] 3. Benchmark Method

[0177] 3.1 Machine Learning Benchmark Methods

[0178] This experiment selected three representative machine learning (ML) algorithms as benchmark methods: Support Vector Machine (SVM) (disclosed in GUO Y, YU L, WEN Z, et al. Using support vector machine combined with auto covariance to predict protein-protein interactions from protein sequences), Logistic Regression (LR) (disclosed in SILBERBERG Y, KUPIEC M, SHARAN RA method for predicting protein-protein interaction types), and Random Forest (RF) (disclosed in WONGL, YOU ZH, LI S, et al. Detection of protein-protein interactions from aminoacid sequences using a rotation forest model with a novel PR-LPQ descriptor).

[0179] 3.2 Deep Learning Benchmark Methods

[0180] This experiment selected four deep learning (DL) algorithms for the PPI prediction task: DPPI (disclosed in HASHEMIFARS, NEYSHABUR B, KHAN AA, et al. Predicting protein–protein interactions through sequence-based deep learning), DNN-PPI (disclosed in HASHEMIFARS, NEYSHABUR B, KHAN AA, et al. Predicting protein–protein interactions through sequence-based deep learning), PIPR (disclosed in CHEN M, JU CJT, ZHOU G, et al. Multifaceted protein–protein interaction prediction based on Siameseresidual RCNN), and GNN-PPI (disclosed in LV GF, HU ZQ, BI YG et al. Learning Unknown from Correlations: Graph Neural Network for Inter-novel-protein Interaction Prediction).

[0181] 4. Comparative Experiment

[0182] Table 2 shows the performance of different calculation methods on different datasets. The results are in the form of micro-F1 mean ± standard deviation under three different random seeds, where LGMG-PPI is the model method proposed in this experiment.

[0183] Table 2 Comparative Study of Experimental Results

[0184]

[0185] The following results were obtained through observation and analysis:

[0186] 1) Deep learning algorithms generally outperform machine learning algorithms, indicating that deep learning-based techniques are superior in encapsulating various types of information about protein pairs (such as amino acid composition and co-occurrence) and automatically extracting robust information suitable for the learning objective. Secondly, the performance of various methods increases with the size of the dataset. This is because the increased data volume allows the model to learn more thoroughly and has stronger generalization ability.

[0187] 2) Compared to the best baseline method GNN-PPI, the model method proposed in this experiment (LGMG-PPI) demonstrates better prediction performance across all data types and exhibits greater stability. Specifically, the micro-F1 score is improved by 2.01% on the SHS27k dataset, 0.94% on the SHS148k dataset, and 0.93% on the STRING dataset. Given the already reliable nature of the best baseline method, the fact that the proposed model method can further improve upon it demonstrates its superiority.

[0188] 5. Ablation test

[0189] To further analyze the role of each module in the model, experiments were conducted by removing different modules to verify their effectiveness. Therefore, the following ablation experiments were set up in this experiment:

[0190] (1) This indicates the removal of elements from the multi-scale graph neural network module. This type of data augmentation, i.e., data augmentation methods that do not use perturbation graph edges;

[0191] (2) This indicates the removal of elements from the multi-scale graph neural network module. This type of data augmentation refers to data augmentation methods that do not use perturbation graph node features;

[0192] (3) w / o MS-GDA: This means that the multi-scale graph neural network module is completely removed, that is, no graph data augmentation strategy is used;

[0193] (4) w / o SL-LRG: This means removing the label relationship graph module, i.e., not using label information for guided learning.

[0194] Table 3 Ablation Experiment

[0195]

[0196] The experimental results are shown in Table 3. The results show that the data augmentation method for perturbed graph node features is slightly better than the data augmentation method for perturbed graph edges, and both graph data augmentation methods are beneficial to the model. This indicates that graph data augmentation methods can enhance the model's generalization ability by perturbing the original graph data. Furthermore, removing the label relationship graph module reduces the model's performance on all datasets. This suggests that introducing the label relationship graph module allows the model to learn the implicit relationships between labels, thereby obtaining the hidden states of the labels and guiding the final prediction results. Overall, each sub-module of the model proposed in this experiment is beneficial to the entire model.

[0197] 6. Experiment on the effectiveness of self-learning label relationship graph

[0198] 6.1 Topology Validity Experiment

[0199] Adaptive label graphs learn label features by introducing a self-learning topology. To verify the effectiveness of the topology, the topology that does not use label graphs is replaced with a Multi-Layer Perceptron (MLP) instead of a Generic Network (GCN). Specifically, the formula is... Replace with Z L =f MLP (X L ).

[0200] Experimental results are as follows Figure 3 As shown in the figure, the experimental results demonstrate that the topological structure incorporating labels yields significantly better results. This indicates that there are certain relationships between labels in the PPI prediction task, and the implicit relationships between labels can be effectively learned through a self-learning label relationship graph, further proving the effectiveness of the proposed method.

[0201] 6.2 Node Feature Effectiveness Experiment

[0202] The initial representation of the node features in the self-learned label relationship graph is the word embedding representation. This experiment uses the pre-trained model BERT (disclosed in Devlin J, Chang MW, Lee K, et al. Bert: Pre-training of deep bidirectional transformers for language understanding) to obtain the word embedding representation. In this section, the performance of the model under different word embedding representations will be evaluated. Specifically, the model performance under BERT and One-Hot embedding representations will be compared through experiments.

[0203] Experimental results are as follows Figure 4 As shown in the figure, the multi-label recognition accuracy is not significantly affected when different word embeddings are used as input to the GCN. This indicates that the improved performance achieved by the model does not entirely come from the semantic information derived from the word embeddings. Furthermore, using powerful word embeddings can lead to better performance. One possible reason is that word embeddings learned from large text corpora retain certain semantic information, and these word embeddings have certain relationships in the embedding space. The model can utilize these implicit relationships to further improve its predictive ability.

[0204] 7. Summary

[0205] This invention proposes a label-guided multi-scale graph neural network method for predicting protein interactions. Graph data augmentation is used to obtain graph representations at multiple scales, which are then input into the graph neural network to obtain protein feature representations at various scales. Contrastive learning is introduced to further enhance protein representation capabilities. Furthermore, a self-learning label relationship graph is constructed to learn the relationships between labels, thereby obtaining label information representations to guide the final protein interaction prediction. Experimental results on three publicly available datasets demonstrate the effectiveness of this model in predicting protein interactions, outperforming the best baseline methods.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A tag-guided method for predicting protein-protein interactions, characterized in that, Includes the following steps: S1: Obtain a test protein to be predicted; S2: Input a protein to be tested into the trained prediction model and output the corresponding prediction relationship; The prediction model first performs graph data augmentation on the protein to be tested to obtain a multi-scale graph representation; then, the multi-scale graph representation is input into a graph neural network to obtain a multi-scale protein feature representation, and contrastive learning is introduced to eliminate the differences in protein feature representations at different scales; then, a self-learning label relationship graph is constructed and the relationships between labels are learned to obtain label feature representations; finally, the label feature representations are used to correct the protein feature representations and guide the prediction of protein interaction relationships, thereby outputting the corresponding predicted relationships. The prediction model includes a protein feature encoder module for extracting local and global features of proteins, a multi-scale graph neural network module for data augmentation, graph neural network processing and contrastive learning, a self-learning label relationship graph module for learning the relationship between labels, and a multi-label loss calculation module for self-supervised learning tasks and supervised learning tasks. Train the prediction model using the following steps: S201: Obtain a pair of proteins for training and input them into the prediction model; S202: Extract local and global features of proteins through the protein feature encoder module to obtain protein feature representations with local and global information; S203: Construct an original graph of protein interactions based on protein feature representations; perturb the original graph using a multi-scale graph neural network module to obtain a corresponding perturbed graph; then input the original graph and the perturbed graph into the graph neural network, outputting the original node feature representation and the perturbed node feature representation, i.e., the multi-scale protein feature representation; then fuse the original node feature representation and the perturbed node feature representation through contrastive learning to obtain the fused node feature representation; finally, calculate the fused edge feature representation using the fused node feature representation. S204: Obtain the tag name embedding representation through the self-learning tag relationship graph module and construct the tag relationship graph; then input the tag relationship graph into the graph convolutional neural network and output the tag node feature representation; S205: Correct the fusion edge feature representation by using the label node feature representation to obtain the protein relationship graph edge feature representation; S206: The multi-label loss calculation module performs self-supervised learning through the original node feature representation and the perturbed node feature representation to obtain the self-supervised learning loss function; then it performs supervised learning through the edge feature representation of the protein relationship graph to obtain the supervised learning loss function; finally, it calculates the training loss function based on the self-supervised learning loss function and the supervised learning loss function, and then optimizes and updates the parameters of the prediction model through the training loss function. S207: Repeat steps S201 to S206 until the prediction model converges; S3: The predicted relationship of a protein to be tested is used as the prediction result of its protein interaction relationship. Then, based on the prediction result of the protein interaction relationship, the cell biological processes of the subject to which the protein to be tested belongs in normal and disease states are analyzed.

2. The tag-guided protein interaction prediction method as described in claim 1, characterized in that: In step S202, the protein feature encoder module includes a local feature encoder and a global feature encoder; The local feature encoder, consisting of a convolutional neural network and a max-pooling layer, extracts the input protein using the following formula. Local feature representation h in i ; h i =f GMP (f CNN (p i ;θ CNN )); In the formula: f CNN Indicates the convolution operation; f GMP This represents the max pooling layer operation; A represents a set of proteins; A represents a defined amino acid vocabulary; a j This refers to amino acids in the amino acid vocabulary; θ CNN Represents the training parameters for the convolution operation; The global feature encoder consists of a bidirectional gated recurrent unit and a global average pooling layer, which extracts the local feature representation h of the input using the following formula. i The global feature representation in the model yields a protein feature representation x that contains both local and global information. i ∈X; x i =f GAP (f BiGRU (h i ;θ BiGRU )); In the formula: f BiGRU Indicates a bidirectional gated loop operation; f GAP Represents the global average pooling layer operation; θ BiGRU represents the training parameters for the bidirectional gated loop operation; X represents the protein feature representation obtained based on the protein feature encoder module.

3. The tag-guided protein interaction prediction method as described in claim 1, characterized in that: In step S203, the original graph G = (X, A) is defined, and the node features are... and adjacency matrix 1) The multi-scale graph neural network module first applies random graph data augmentation functions from two different perspectives. and The edges and node features of the original graph G=(X,A) are perturbed respectively to obtain the first perturbed graph G1=(X,A1) and the second perturbed graph G2=(X2,A); ∈1~Bernoulli(N,1-δ1); In the formula: ∈1 represents the N Bernoulli distribution result obtained based on the hyperparameter δ1∈(0,1); This indicates that the graph data augmentation function is applied based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; Let represent the set of edges in the original graph; Bernoulli represents the Bernoulli distribution; δ1∈(0,1) is a hyperparameter representing the ratio of edges removed. X1 represents the uniform distribution result obtained based on the hyperparameter δ2∈(0,1); X2 represents the graph data augmentation function based on the first perturbation graph G1=(X,A1). The obtained protein feature representation; X represents the node features of the original graph; Unifrom represents a uniform distribution; δ2∈(0,1) is a hyperparameter representing the ratio of node features set to 0; 2) Input the original graph G = (X,A), the first perturbation graph G1 = (X,A1), and the second perturbation graph G2 = (X2,A) into the graph neural network, and output the original node feature representations. First perturbation node feature representation Second perturbation node feature representation A graph neural network with k iterations is represented as: In the formula: AGG represents the representation of node v after aggregating the features of its neighboring nodes; AGG represents the function that aggregates the features of a node. or This represents the result of node u based on the (k-1)th iteration of the graph convolutional network; represents the set of neighbors of node v; UPDATE represents the node feature update function; or This represents the result of node v through k-1 iterations of a graph convolutional network; or ω represents the feature representation of node v in the k-th iteration; MLP represents a multilayer perceptron neural network; ω is a learnable parameter or a constant. 3) The original node feature representations are fused using the following formula. First perturbation node feature representation The fused node feature representation is obtained from the feature representation of the second perturbation node. Z'=f Fusion ([Z0,Z1,Z2]); In the formula: f Fusion Indicates the fusion function; 4) Obtain the fused edge feature representation by fusing the node feature representation Z'. yes ij =z′ i ⊙z′ j , yes ij ∈E; In the formula: ⊙ represents the Hadamard product; and Let i and j represent the feature representations of node i and node j, respectively.

4. The tag-guided protein interaction prediction method as described in claim 3, characterized in that: In step S204, the self-learning label relationship graph module obtains the label name embedding representation through the pre-trained model BERT; X L =BERT(L NAME ); In the formula: Indicates the label name; The word vector representing the tag name, i.e., the tag name embedding representation; X is represented by embedding the label name. L and learnable parameter matrix Construct a label relationship graph G L =(A L ,X L ); Label relationship graph G L =(A L ,X L In a convolutional neural network with an input graph, the output label node feature representation In the formula: Initialize Z L (0) =X L ; Degree matrix; W (l-1) Denotes the learnable parameter matrix; σ represents the sigmoid activation function; A L It is initialized as an identity matrix.

5. The tag-guided protein interaction prediction method as described in claim 4, characterized in that: In step S205, the edge feature representation of the protein relationship graph is calculated using the following formula; In the formula: E represents the edge feature representation of the protein-protein relationship graph; E represents the fusion edge feature representation. Tag node feature representation.

6. The tag-guided protein interaction prediction method as described in claim 5, characterized in that: In step S206, the multi-label loss calculation module, in the self-supervised learning task, uses the original node feature representation... and the first perturbation node feature representation Calculate the first disturbance loss function Based on the original node feature representation Second perturbation node feature representation Calculate the second disturbance loss function First perturbation loss function Second perturbation loss function This is the self-supervised learning loss function; In the formula: (z 1,i ,z 0,i ) represents a positive sample pair; z 0,i ∈Z0, z 1,i ∈Z1;θ(z 0,i ,z 1,i ) indicates the calculation of z 0,i and z 1,i The cosine similarity; τ represents the temperature parameter; N represents the set of all nodes.

7. The tag-guided protein interaction prediction method as described in claim 6, characterized in that: In step S206, the multi-label loss calculation module calculates the following supervised learning loss function in the supervised learning task using the edge feature representation of the protein relationship graph; p ij =Softmax(e ij ); In the formula: Indicates the loss in supervised learning; t represents the number of label categories; p represents the training set of the edge set; ij This represents the probability distribution of the relationship between proteins i and j; The predicted relationship represents the connection between proteins i and j; c represents the specific category of the label. This represents the true labels of proteins i and j in category c; This represents the prediction results of proteins i and j in category c; argmax represents taking the largest index in the set.

8. The tag-guided protein interaction prediction method as described in claim 7, characterized in that: In step S206, the training loss function is expressed by the following formula; In the formula: Indicates training loss; Indicates the loss in supervised learning; Indicates the loss from the first disturbance; λ represents the second disturbance loss; λ1 and λ2 represent the set hyperparameters.

Images