A decoupled graph network algorithm for predicting protein-protein interaction relationships

Abstract

The present invention provides a decoupled graph network algorithm for predicting protein interactions, which relates to the technical field of protein interaction prediction and includes the following steps: constructing a protein interaction graph, calculating the conditional probability of another interaction existing simultaneously when any one protein interaction exists to obtain a conditional probability matrix, performing initial word embedding on protein nodes to obtain protein vector representations, passing the protein vector representations through N multi-layer perceptrons to obtain protein vector representations of different protein vectors in various categories, using a graph neural network to update the protein vector representations of specific categories, performing dot product operations on the feature representations of the N categories of protein pairs to obtain protein correlations and constraining them with a conditional probability matrix, thereby performing classification predictions through edge category information. The present invention decouples the category information of proteins for classification predictions and constrains them with conditional probabilities, thereby improving the prediction effect.

Description

Technical Field

[0001] The present invention relates to the technical field of protein interaction relationship prediction, and in particular to a decoupled graph network algorithm for protein interaction relationship prediction. Background Art

[0002] Protein-protein interactions (PPIs) relationship prediction plays a key role, and its wide applications include but are not limited to industrial production, health care, disease diagnosis and treatment, drug design, etc. These fields all involve protein macromolecules and their properties, and both 0.1 prediction and multi-label prediction of these properties are crucial. In addition to direct physical binding, PPIs have many other indirect ways of cooperation and mutual regulation, or jointly promote specific biological functions. It can be said that the study of PPIs and their interaction types is crucial to understanding cell biological processes in normal and disease states, which in turn helps to identify therapeutic targets and design new drugs.

[0003] The study of polymorphic protein-protein interactions provides a systematic understanding of biological processes and reveals the basis of disease mechanisms. Currently, there are still some difficulties in multi-label prediction of protein interaction relationships. When tested on new datasets, the performance of existing methods will drop significantly.

[0004] At present, the paper "Learning Unknown from Correlations: Graph Neural Network for Inter-novel-protein Interaction Prediction" discloses a method for predicting protein interaction relationships, designs a new evaluation framework, including the starting point of random data partitioning of proteins, namely BFS and DFS, and proposes a GNN-PPI method based on GNN to simulate the correlation between PPIs. This method can better predict the interaction between new proteins. Then, the method of this paper only regards PPIs as a graph, uses GNN to aggregate the information of neighboring nodes, blends all protein category information together, and then updates the protein features, and directly uses it for classification prediction. It is unable to identify each specific interaction relationship separately, and does not decouple the fused label information well, so it does not achieve better results. Summary of the invention

[0005] In order to solve the problems raised in the background technology and further improve the performance of protein interaction relationship prediction, the present application provides a decoupled graph network algorithm for protein interaction relationship prediction, comprising the following steps:

[0006] construct protein interaction maps;

[0007] Traverse the data set and calculate the conditional probability of the existence of another interaction when any protein interaction exists, forming an N*N conditional probability matrix, where N is the number of protein interaction relationship types;

[0008] Perform initial word embedding on protein nodes to obtain protein vector representation;

[0009] The protein vector representation is passed through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the N category information in the protein vector representation;

[0010] Using graph neural networks to update the protein vector representations of specific categories;

[0011] The edge index is used to extract the N categories of protein vector representations of the two updated protein nodes at both ends of the edge. The dot product operation is performed on the two proteins to obtain the protein correlation relationship and constrained by the conditional probability matrix, so as to make classification predictions based on the category information of the edge.

[0012] Specifically, the initial word embedding of the protein node to obtain the protein vector representation includes:

[0013] The protein nodes are word-embedded and encoded to obtain a first protein vector representation; the first protein vector representation is encoded through an independent encoding module to obtain a second protein vector representation, and the second protein vector representation is used as the final protein vector representation obtained in this step.

[0014] Preferably, the independent encoding module includes a first convolutional layer, a normalization layer, a first pooling layer, an LSTM layer, a second pooling layer and a fully connected layer.

[0015] Specifically, the use of graph neural networks to update the protein vector representation of a specific category is to perform convolution operations on the feature vector representations of N different categories through GIN respectively, aggregate the information of neighboring protein nodes according to the protein interaction graph, and then update the feature representations of N categories of different protein nodes.

[0016] Specifically, the updating of the feature representations of N categories of different protein nodes is as follows: inputting the feature representations into a first linear layer, and then passing through a Relu layer, a Dropout layer, and a second linear layer respectively.

[0017] Specifically, the method extracts N categories of protein vector representations of two updated protein nodes at both ends of the edge through edge indexes, performs dot product operations on the protein vector representations of the two proteins to obtain protein correlations and constrains them with conditional probability matrices, so as to make classification predictions through the category information of the edge. Specifically, the method comprises: extracting N categories of feature representations of two updated protein nodes at both ends of the edge through edge indexes; performing dot product operations on the feature representations of the N categories of the two proteins and then performing linear transformation to obtain N fused features; splicing the N fused features and inputting them into the FC layer for feature compression, and multiplying the result after dimensionality reduction with the N*N conditional probability matrix to obtain a prediction result.

[0018] Preferably, the loss function of the algorithm corresponding to the model is cross entropy loss, and the formula is as follows:

[0019]

[0020] Among them, X train represents the training set, Represents protein pair x ij The prediction results of the interaction relationship.

[0021] Preferably, during the training process and the testing process, the algorithm divides the data set by any one of random division, depth-first division or breadth-first division.

[0022] At the same time, the present invention also discloses a decoupling graph network system for predicting protein interaction relationships, comprising:

[0023] The protein interaction graph construction module is used to construct a graph of protein pairs in the data set, where the nodes of the graph are proteins and the edges of the graph are the interaction relationships between protein pairs;

[0024] The conditional probability matrix calculation module is used to traverse the data set to obtain the conditional probability matrix. The conditional probability matrix is ​​the conditional probability that another interaction exists when any one protein interaction exists; wherein N is the number of interaction relationship types of proteins;

[0025] The word embedding module is used to perform initial word embedding on protein nodes to obtain protein vector representation;

[0026] The decoupling module is used to pass the protein vector representation through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the N category information in the protein vector representation;

[0027] A protein vector representation update module, which is used to update the protein vector representation of a specific category through a graph neural network;

[0028] The prediction module is used to extract the N categories of protein vector representations of the two updated protein nodes at both ends of the edge through the edge index, perform dot product operations on the two proteins on the N categories of protein vector representations to obtain the protein correlation relationship and constrain it with a conditional probability matrix, so as to make classification predictions based on the category information of the edge.

[0029] Furthermore, the present invention also discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the decoupled graph network algorithm for predicting protein interaction relationships as described above when executing the computer program.

[0030] Furthermore, the present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the decoupled graph network algorithm for predicting protein interaction relationships as described above is implemented.

[0031] After adopting the above scheme, the beneficial effects of the present invention are:

[0032] (1) The method of the present invention can identify each specific category of interaction separately, decouple the fused information, achieve unique category representation of protein information through decoupling, enhance the correlation between categories, and then use the prior information of the N*N conditional probability matrix to constrain the protein relationship. Therefore, the effect of the method of the present invention is better than that of the prior art.

[0033] (2) Since there is sequence correlation between protein sequences, the present invention adds an LSTM module to the encoding module, and extracts the relationship between protein sequences through the serialization representation of LSTM, which can improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 It is a flow chart of an algorithm in a specific embodiment of the present invention;

[0036] Figure 2 It is a framework diagram provided in a specific embodiment of the present invention;

[0037] Figure 3It is a schematic diagram of an application scenario of the method or system of a specific embodiment of the present invention;

[0038] Figure 4 It is a schematic diagram of the structure of a computer device provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0039] The principles and features of the present invention are described below. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0040] See also Figure 1 and Figure 2 , Figure 1 It is a key step flow chart of the present invention, Figure 2 The present invention will be described more clearly and completely below through specific embodiments.

[0041] The present invention specifically implements a decoupled graph network algorithm for predicting protein interactions, comprising the following steps:

[0042] S100. Construct a protein interaction graph; specifically, use a protein interaction dataset to construct a protein interaction graph, in which paired interacting protein data are used to construct a protein interaction graph, in which proteins serve as nodes and protein interaction relationships serve as edges. In this specific implementation, the datasets used are SHS27K and SHS148K, both of which are public datasets. The types of protein interaction relationships include 7 categories, namely reaction, binding, post-translational modifications (ptmod), activation, inhibition, catalysis, and expression, which represent reaction, binding, post-translational modification, activation, inhibition, catalysis, and expression, which include direct and indirect effects. Use any one dataset to carry out the experiment of the present invention, and the results of the two datasets can be seen in the experimental results below. Now assume that the protein sequence is:

[0043] P={p0,p1,...,p i ,...,p m}

[0044] Among them, p i represents a protein node;

[0045] The protein-protein interaction (PPI) sets are:

[0046] X={x ij ={p i ,pj}|i≠j,p i ∈P,p j ∈P}

[0047] Among them, x ij Represents protein node p i and protein node p j The protein pairs that make up;

[0048] The protein relationship diagram is as follows:

[0049] G=(P,X)

[0050] The protein interaction relationship is encoded by One-Hot as follows:

[0051] Y={y ij =(y ij1 ,y ij2 ,...,y ijN )|y ijt ={0,1},t∈[1,N],i∈M,j∈M}

[0052] Among them, y ij Represents protein pair x ij interaction relationship, N is the number of protein interaction relationship categories, in this specific implementation, N is 7, and M is the total number of protein pairs.

[0053] For the training process and the testing process, this specific implementation method divides the protein-protein relationship data set by DFS (Depth-First Search), BFS (Breadth-First Search) or Random method to obtain a training set and a test set for model training and testing. More specifically, DFS is a depth-first traversal, and BFS is a breadth-first traversal. The specific data division steps can refer to the pseudo code given in the paper mentioned in the background technology and the public detailed code, which can be clearly obtained by those skilled in the art. In this specific implementation, three methods are used to divide the data set for each data set to obtain experimental results.

[0054] In order to facilitate subsequent operations, the storage method of the protein interaction graph (PPI Graph) is updated to the edge index (edge_index) method.

[0055] S200. Traverse the data set and calculate the conditional probability that another interaction exists when any protein interaction exists, forming an N*N conditional probability matrix, where N is the number of types of protein interaction relationships. The data set here refers to the training set. It can be understood that Figure 2The size of the conditional probability matrix in is 3*3 for illustration only. For this specific implementation, the size of the conditional probability matrix is ​​7*7.

[0056] S300. Perform initial word embedding on protein nodes to obtain protein vector representation; the protein nodes here can be obtained from the graph or from the original data set, specifically, including the following steps:

[0057] S310. Perform word embedding encoding on the protein node to obtain a first protein vector representation. The specific encoding process is prior art and will not be further expanded here.

[0058] S320. Encode the first protein vector representation through an independent encoding module to obtain a second protein vector representation, and use the second protein vector representation as the protein vector representation finally obtained in this step. The independent encoding module includes a first convolutional layer (Conv1D layer), a normalization layer (BN layer), a first pooling layer (first Pooling layer), an LSTM layer, a second pooling layer (second Pooling layer) and a fully connected layer (FC layer). Specifically, the first protein vector representation is first convolved through the first convolutional layer, and then pooled through the first pooling layer. The obtained vector representation is passed through the LSTM layer to learn the correlation between protein sequences, and then passed through the second pooling layer and the fully connected layer to obtain the updated vector representation, i.e., the second protein vector representation. Figure 2 The feature encoding process in is not a complete process, but only highlights the LSTM layer. In this embodiment, the input channel of the first convolutional layer operation is 13, the output channel is 1, the convolution kernel size is 3, the bias is 0, and a BN layer is provided after the first convolutional layer; the first pooling layer operation uses the maximum pooling operation (MaxPool1d), the pooling size is 3, and the step size is 3; the input channel and output channel of the LSTM layer are both 1, and numlayers is set to 1; the second pooling layer is an average pooling layer with a pooling size of 3 and a step size of 3; the fully connected layer input dimension is 2000 / 3, and the output dimension is 256.

[0059] S400. The protein vector representation is passed through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the 7 category information in the protein vector representation; it should be understood that Figure 2 The three decoupling modules drawn in the figure are only for illustration. Specifically, the second protein vector representation obtained in the above steps is input into N multi-layer perceptrons to obtain N different types of feature representations, which are decoupled category information; in this specific implementation, N is 7, because there are 7 types of protein interactions in the dataset used, and the input dimension and output dimension of the 7 multi-layer perceptrons in this specific implementation are both 256.

[0060] S500. Using a graph neural network to update a protein vector representation of a specific category; comprising the following steps:

[0061] S510. Convolution operations are performed on N different categories of feature vector representations through GIN respectively, and information of neighboring protein nodes is aggregated according to the protein interaction graph, and then the feature representations of N categories of different protein nodes are updated; the graph isomorphic network is represented by GIN, wherein the sum of neighboring node features is used as an aggregation function, and multilayer perceptrons (MLPs) are used for updating, which can be expressed by the following formula:

[0062]

[0063] Where q is the number of iterations, is the feature of node p at the qth iteration, MLP represents multi-layer perceptron, Ne(p) represents all the neighborhoods of protein node P in the protein interaction graph, and ∈ represents a learnable parameter. In this specific implementation, the multi-layer perceptron MLPs in GIN use two linear layers, each of which is followed by a Relu layer, and there is a BN operation after the second linear operation and the Relu operation. It should be understood that the GIN graph neural network is only an optional method, and the present invention can also use the GNN graph neural network.

[0064] The updating of the feature representations of N categories of different protein nodes specifically includes: inputting the feature representation into the first linear layer, and then passing through the Relu layer, the Dropout layer and the second linear layer respectively, the output dimension of the first linear layer is 512, the Dropout layer is used to prevent overfitting, and after passing through the second linear layer, the output dimension is 512, and the Dropout layer is only used in the training process.

[0065] S600. Extract the N categories of protein vector representations of the two updated protein nodes at both ends of the edge through the edge index (edg_index), perform dot product operations on the two proteins and the N categories of protein vector representations to obtain the protein correlation relationship and constrain it with the conditional probability matrix, so as to make classification predictions through the category information of the edge. Specifically including:

[0066] S610. Extracting feature representations of N categories of two updated protein nodes corresponding to both ends through edge indexes, wherein the edge indexes are used to guide protein representations of each category after mapping;

[0067] S620. Perform dot product operations on the feature representations of the N categories of the two proteins and then perform linear transformation to obtain N fused features, each with a dimension of 64.

[0068] S630. Concatenate the N fused features and input them into the FC layer for feature compression. Multiply the dimension reduction result with the N*N conditional probability matrix of the data set obtained above to obtain the prediction result. The dimension after concatenation is 1*448, the dimension after compression is 1*7, and the dimension of the conditional probability matrix is ​​7*7.

[0069] It should be noted that the above process Figure 2 Only the key parts are drawn in the figure, and not all steps are reflected in the figure.

[0070] The model of the present invention is an end-to-end model, and the loss function is a cross entropy loss function:

[0071]

[0072] Among them, X train represents the training set, Represents protein pair x ij The prediction results of the interaction relationship.

[0073] After adopting the above-mentioned specific implementation content, this specific implementation chooses to test the PPIs in the test set, which accounts for 20%, and uses BFS, DFS and Random methods to segment the data. The BFS or DFS segmentation algorithm has completely different results for different root nodes. The root node degree threshold in BFS and DFS is set to 5. In order to eliminate the impact of the randomness of data partitioning on the performance of the PPI method, this specific implementation also repeated the experimental results under 3 different random seeds, and the experimental statistical results are the average values. In this specific implementation, protein features based on amino acid sequences are used, refer to (Multifaceted protein–protein interaction prediction based on siameseresidual rcnn).

[0074] Micro-F1 is used to evaluate the prediction performance of multi-label PPI. The specific results and the comparison results with the existing methods are shown in Table 1. In Table 1, the PIPR method is the method used in the paper "Multifaceted protein-protein interaction prediction based on Siamese residual RCNN", and the GNN_PPI method is the method "Learning Unknown from Correlations: Graph Neural Network for Inter-novel-protein Interaction Prediction" mentioned in the background technology. It can be seen from the table that for the SHS27K dataset and the SHS148K dataset, no matter which dataset division method is used, the method of the present invention is superior to the two existing methods, and for the BFS dataset division method, the improvement effect is the most obvious.

[0075] See also Figure 3 , Figure 3 This is a schematic diagram of an application scenario of a decoupling graph network system for predicting protein interrelationships provided in an embodiment of the present application, which may include a terminal and a server. The area denial method for low-speed drones provided in the present application may be implemented through a terminal or a server.

[0076] Table 1 Comparison of micro-F1 results

[0077]

[0078] like Figure 3 As shown, the terminal and the server are connected via a network, such as a wired or wireless network connection. The terminal may include but is not limited to portable terminals such as mobile phones and tablets installed with various network platform applications, as well as fixed terminals such as computers, query machines, and advertising machines. The server provides users with various business services, including service push servers, user recommendation servers, etc.

[0079] The terminal may include the following modules:

[0080] The protein interaction graph construction module is used to construct a graph of protein pairs in the data set, where the nodes of the graph are proteins and the edges of the graph are the interaction relationships between protein pairs;

[0081] The conditional probability matrix calculation module is used to traverse the data set to obtain the conditional probability matrix. The conditional probability matrix is ​​the conditional probability that another interaction exists when any one protein interaction exists; wherein N is the number of interaction relationship types of proteins;

[0082] The word embedding module is used to perform initial word embedding on protein nodes to obtain protein vector representation;

[0083] The decoupling module is used to pass the protein vector representation through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the N category information in the protein vector representation;

[0084] A protein vector representation update module, which is used to update the protein vector representation of a specific category through a graph neural network;

[0085] The prediction module is used to extract the N categories of protein vector representations of the two updated protein nodes at both ends of the edge through the edge index, perform dot product operations on the two proteins on the N categories of protein vector representations to obtain the protein correlation relationship and constrain it with a conditional probability matrix, so as to make classification predictions based on the category information of the edge.

[0086] In some embodiments, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store relevant data of the image acquisition device. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above-mentioned decoupling graph network algorithm for predicting protein interaction relationships is implemented.

[0087] In some embodiments, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input system connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a decoupling graph network algorithm for predicting protein interaction relationships is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input system of the computer device can be a touch layer covered on the display screen, or a button, a trackball or a touch pad set on the computer device housing, or an external keyboard, touch pad or mouse, etc.

[0088] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0089] In some embodiments, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0090] In some embodiments, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above-mentioned method embodiments when executed by a processor.

[0091] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0092] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A decoupled graph network algorithm for predicting protein interaction relationships, characterized in that: The steps include: construct protein interaction maps; Traverse the data set and calculate the conditional probability of the existence of another interaction when any protein interaction exists, forming an N*N conditional probability matrix, where N is the number of protein interaction relationship types; Perform initial word embedding on protein nodes to obtain protein vector representation; The protein vector representation is passed through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the N category information in the protein vector representation; Using a graph neural network to update a specific category of protein vector representations; the updating of a specific category of protein vector representations using a graph neural network is specifically to perform convolution operations on N different categories of feature vector representations through GIN, aggregate information of neighboring protein nodes according to a protein interaction graph, and then update feature representations of N categories of different protein nodes; The N categories of protein vector representations of the two updated protein nodes at both ends of the edge are extracted through the edge index. The dot product operation is performed on the two proteins to obtain the protein correlation relationship and constrained by the conditional probability matrix, so as to make classification predictions based on the category information of the edge. The specific steps are as follows: Extract the feature representations of N categories of the two updated protein nodes corresponding to both ends through the edge index; Perform dot product operations on the feature representations of the N categories of the two proteins and then perform linear transformation to obtain N fused features; The N fused features are concatenated and input into the FC layer for feature compression. The result after dimensionality reduction is multiplied by the N*N conditional probability matrix to obtain the prediction result.

2. A decoupled graph network algorithm for predicting protein interaction relationships according to claim 1, characterized in that: The initial word embedding of the protein node to obtain the protein vector representation includes: The protein nodes are word-embedded and encoded to obtain a first protein vector representation; the first protein vector representation is encoded through an independent encoding module to obtain a second protein vector representation, and the second protein vector representation is used as the final protein vector representation obtained in this step.

3. A decoupled graph network algorithm for predicting protein interaction relationships according to claim 2, characterized in that: The independent encoding module includes a first convolutional layer, a normalization layer, a first pooling layer, an LSTM layer, a second pooling layer and a fully connected layer.

4. According to the decoupled graph network algorithm for predicting protein interaction relationships according to claim 1, the feature representation of N categories of different protein nodes is updated specifically by inputting the feature representation into the first linear layer, and then passing through the Relu layer, the Dropout layer and the second linear layer respectively.

5. A decoupled graph network algorithm for predicting protein interaction relationships according to claim 1, characterized in that: The loss function of the model corresponding to the algorithm is cross entropy loss, and the formula is as follows: Among them, X train represents the training set, Represents protein pair x ij The prediction results of the interaction relationship.

6. A decoupled graph network algorithm for predicting protein interaction relationships according to claim 1, characterized in that: During the training and testing process, the algorithm divides the data set by random division, depth-first division or breadth-first division.

7. A decoupled graph network system for predicting protein interaction relationships, characterized in that: include: The protein interaction graph construction module is used to construct a graph of protein pairs in the data set, where the nodes of the graph are proteins and the edges of the graph are the interaction relationships between protein pairs; The conditional probability matrix calculation module is used to traverse the data set to obtain the conditional probability matrix. The conditional probability matrix is ​​the conditional probability that another interaction exists when any one protein interaction exists; wherein N is the number of interaction relationship types of proteins; The word embedding module is used to perform initial word embedding on protein nodes to obtain protein vector representation; The decoupling module is used to pass the protein vector representation through N multi-layer perceptrons to obtain the feature vector representation of different protein vectors in each category, thereby decoupling the N category information in the protein vector representation; A protein vector representation update module, which is used to update the protein vector representation of a specific category through a graph neural network; The prediction module is used to extract the N categories of protein vector representations of the two updated protein nodes at both ends of the edge through the edge index, perform dot product operations on the two proteins on the N categories of protein vector representations to obtain the protein correlation relationship and constrain it with a conditional probability matrix, so as to make classification predictions based on the category information of the edge.

8. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the decoupled graph network algorithm for predicting protein interaction relationships as described in any one of claims 1 to 6 is implemented.

Images