Multi-label protein interaction prediction method based on joint training

By combining graph convolution networks and one-dimensional convolution networks, the nominal features and sequence features of protein sequences are extracted, and the prediction model of multi-label protein interactions is constructed, which solves the problem of difficulty in effectively binding features in the existing technology, which significantly improves the Predictive accuracy of protein interaction categories.

CN119007819AInactive Publication Date: 2024-11-22NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411141189.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the multi-label type prediction of protein interaction, existing methods are difficult to effectively bind sequence characteristics and nominal characteristics, resulting in low prediction efficiency and insufficient accuracy.

Method used

Using a joint training method, combining graph convolutional network (GCN) and one-dimensional convolutional network (CNN), information is extracted from the nominal characteristics and sequence characteristics of protein sequences, and a prediction model for multi-label protein interaction is constructed. Full connection layer performs multi-label prediction.

Benefits of technology

The prediction accuracy of protein interaction categories has been significantly improved, and the Micro-F1 value has been increased by about 0.04, which is better than the existing technology methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007819A_ABST
    Figure CN119007819A_ABST
Patent Text Reader

Abstract

The invention discloses a joint training-based multi-label protein interaction prediction method, which comprises the following steps of: selecting interaction data from a database containing a plurality of PPI types as an original data source, preprocessing the data, constructing a PPI data set, performing feature extraction on a protein sequence, and performing feature extraction on the protein sequence; dividing the obtained features into nominal features and sequence features, respectively forming nominal feature vectors and sequence feature vectors, and then constructing a model based on a joint training structure; extracting a relation structure of the interaction network from the nominal feature vector by using a graph convolutional network; coding and extracting protein sequence information in the sequence features by using a one-dimensional convolutional neural network; and finally, performing multi-label prediction by using a full connection layer, and putting the model into use according to performance feedback. Compared with an existing method, the method can significantly improve the accuracy of protein interaction category prediction, and has a promotion effect on research in the fields of protein functions, drug development and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the fields of bioinformatics and computer science, and in particular relates to a method for predicting multi-label protein interactions based on joint training. Background Art

[0002] In the past few decades, with the continuous deepening of molecular biology research, the study of protein interactions (hereinafter referred to as PPIs) has attracted widespread attention from researchers. Although experimental methods such as yeast two-hybrid and tandem affinity purification can be used to explore PPIs, these methods are expensive, inefficient and prone to false positives. With the rapid growth of biological data, there is an urgent need for a computational method to predict the type and characteristics of PPIs to overcome the limitations of traditional experimental methods.

[0003] In recent years, deep learning has gradually become a hot research technology in the field of bioinformatics, and has demonstrated its powerful feature extraction and pattern recognition capabilities in many applications. Especially in processing sequence features (such as information based on amino acid sequences), deep learning can make effective predictions by learning the intrinsic patterns of protein sequences. However, most existing methods focus on the sequence features of sequence information, while ignoring nominal features such as the physicochemical properties of amino acids. Nominal features provide a direct description of the relevant properties of protein sequences, which in turn are directly related to the type of PPI.

[0004] Although existing methods have achieved some success in using sequence features for PPI prediction, how to effectively combine and utilize these two types of features (sequence features and nominal features) in the multi-label PPI type prediction problem has not been fully explored.

[0005] Based on this, the present invention proposes a multi-label PPI prediction method based on joint training, which processes the nominal features and sequence features in PPI data by comprehensively using graph convolutional network (GCN) and one-dimensional convolutional network (CNN). Summary of the invention

[0006] The purpose of the present invention is to solve the above-mentioned deficiencies of the prior art and provide a method for predicting multi-label protein interactions based on joint training, which uses deep learning and graph convolutional network technology to predict protein interactions.

[0007] To achieve the technical purpose, the present invention adopts the following technical solutions:

[0008] A method for predicting multi-label protein interactions based on joint training, comprising the following steps:

[0009] S1. Select interaction data from a database containing multiple PPI types as the original data source, preprocess the data, and construct a multi-label protein-protein interaction (PPI) dataset {(p i ,p j ,I ij )}, where p i ,p j Corresponding to the protein numbered i,j, I ij Corresponding interaction category;

[0010] S2. Extract features from the protein sequence, and divide the obtained features into nominal features and sequence features, forming a nominal feature vector and a sequence feature vector respectively;

[0011] S3. Consider the nominal feature vector and the sequence feature vector as descriptions of the protein from different perspectives, and build a model based on the joint training structure;

[0012] S4. Extract the relational structure of the interaction network from the nominal feature vector using a graph convolutional network (GCN) with spatial aggregation capabilities;

[0013] S5. Use one-dimensional convolutional neural network (1D CNN) to encode and extract protein sequence information from sequence features;

[0014] S6. Apply a fully connected layer for multi-label prediction. The fully connected layer uses a sigmoid activation function and binary cross entropy as the loss function to handle the output of multiple interaction types. The formula is as follows:

[0015]

[0016] Further, in step S1, the preprocessing includes:

[0017] The data was cleaned to remove the protein sequence parts with more than 512 amino acids and to fill the sequences with less than 512 amino acids with 0;

[0018] Then, a de-redundancy operation is performed to reduce the redundancy between protein sequences by executing a validated algorithm, including using a method to compare sequence similarity to determine and remove similar or identical protein sequences, avoiding duplication in multi-label classification problems, extracting interaction data from the STRING database, and removing protein sequences shorter than 50 amino acids and highly homogeneous proteins;

[0019] For interactions involving the same protein pair [p1, p2], if their interaction types are different, they are merged and the interaction types are integrated into one type vector In order to reduce the homogeneity and repetitiveness of the samples;

[0020] Then, a predefined threshold is used, a balanced sampling strategy is adopted, and multiple iterative perturbations are performed to construct a multi-label PPI dataset that is non-redundant, class-balanced, and diverse enough to ensure the diversity and fairness of the dataset.

[0021] Further, in step S2, the feature extraction generates a nominal feature vector of fixed dimension and a sequence feature vector of variable dimension based on a protein sequence of a certain length, wherein:

[0022] Nominal features are calculated directly based on the physicochemical properties of the protein sequence and are fixed in a lower dimension;

[0023] The sequence features are generated by encoding each amino acid in the protein sequence, where each amino acid is encoded as a 48-dimensional numerical vector; the encoding combines the following aspects:

[0024] i. Using the amino acid encoding obtained by the pre-trained Word2Vec model to capture and utilize the amino acid context information contained in the sequence;

[0025] ii. One-hot encoding based on known physicochemical properties of amino acids to capture the essential characteristics of amino acids;

[0026] iii. Position weight matrix encoding extracted based on the position-related information of amino acids in the protein sequence, adding position specificity to enhance the model's understanding of the sequence structure.

[0027] Further, in step S2, the feature extraction includes the following steps:

[0028] Apply multiple one-dimensional convolutional layers with different kernel sizes to the matrix X, where each kernel size is specifically adapted to capture the patterns and relationships between the corresponding number of amino acid residues;

[0029] For each convolutional layer of the kernel size, a maximum pooling layer is used to reduce the feature dimension, so as to extract key sequence information and reduce subsequent calculations;

[0030] A dropout layer with a loss ratio of 0.3 is introduced to reduce the risk of overfitting and optimize the generalization ability of the model;

[0031] The output feature maps of the three convolutional layers are concatenated into a unified feature vector after dimensionality reduction and inactivation, which serves as the input of the multi-label classification model.

[0032] Furthermore, in step S4, the method for extracting the relationship structure of the interaction network includes:

[0033] Construct an undirected graph G = (P, E) of multiple categories of PPIs, corresponding to the graph structure representation of the seven types of PPIs. In each undirected graph, the vertex set P represents proteins, and the edge set E represents one of the seven types of PPIs recorded in the STRING database.

[0034] A graph convolutional network (GCN) layer is applied to capture the inter-vertex relational structure from the undirected graph G. The GCN layer uses spatial neighborhood sampling technology to learn and infer node hierarchical structure features. The graph convolutional network (GCN) is used to encode the k-order neighborhood of vertex p, and the formula is used:

[0035]

[0036] The 1-hop to k-hop neighborhood of each protein p is obtained by iterative calculation;

[0037] For a given protein p, a set of proteins Neigh(p) that directly interact with p is selected, and uniform sampling is performed to extract a fixed number of neighborhood sets U(p), where the size and depth k of uniform sampling are both set to 2 to adapt to the complex PPI graph structure and reduce the computational complexity;

[0038] By formula:

[0039] Agg=σ(W·concat(V(p i ), V(p j ))+b)

[0040] The aggregation function AGG described in processes the data from the k-hop neighborhood to refine the feature set of each protein p and reflect their dynamic interaction properties and complex relationships between proteins;

[0041] According to each protein p and its k-order neighborhood set V(p) = {h 1 (p)...h k (p)}, a trainable aggregation function is used to aggregate the embedding representation of each k-order neighborhood of p to form a comprehensive feature of p;

[0042] During model training, information from different deep neighborhood hierarchies is integrated to capture the complex interaction networks and relationships between proteins.

[0043] Further, in step S5, the method for encoding and extracting the protein sequence information in the sequence feature is as follows:

[0044] Three one-dimensional convolutional layers with different core sizes are designed, each for sequence patterns of different lengths, to capture the local and global patterns of amino acid sequences and their contextual information;

[0045] The maximum pooling layer gradually performs spatial dimensionality reduction on the feature maps output by each convolutional layer to extract the most critical and representative features in the sequence;

[0046] Apply dropout ratio to reduce overfitting of the model, thereby improving the generalization performance of the model on new data;

[0047] The vector formed by concatenating the feature maps processed by all convolutional layers is used as the input of the multi-label prediction task. This step ensures that sequence information of different scales is fully considered by the model.

[0048] Feature fusion is performed based on the nominal features extracted and optimized by the graph convolution layer and the sequence features processed by the one-dimensional convolution layer. On this basis, the fully connected layer is used to perform the final multi-label prediction.

[0049] Beneficial effects of the present invention:

[0050] (1) The present invention provides a method for predicting multi-label protein interactions based on joint training. Compared with existing methods, this method comprehensively extracts information from nominal features and sequence features, and can achieve efficient prediction of protein interaction categories.

[0051] (2) The Micro-F1 values ​​of the method of the present invention under random partitioning, breadth-first partitioning, and random partitioning were 0.9019, 0.7157, and 0.7417, respectively, while the Micro-F1 values ​​of the best method in the prior art were 0.8613, 0.6874, and 0.7086, respectively. The method of the present invention significantly improved the accuracy of protein category prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of the prediction of multi-label protein interactions based on joint training of Example 1;

[0053] Figure 2 This is an architecture diagram of the model constructed based on the joint training structure in Example 1;

[0054] Figure 3 This is a schematic diagram of sampling neighborhood nodes in Example 1. DETAILED DESCRIPTION

[0055] To further illustrate the technical solution of the present application, the implementation of the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It should be noted that the following is only a preferred implementation of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

[0056] Example 1

[0057] like Figure 1 , a prediction method for multi-label protein interactions based on joint training, comprising the following steps:

[0058] S1. We selected interaction data from the STRING database version 10.5, which contains multiple PPI types, as the original data source. Each data instance format is the identifier of two proteins, the interaction type, and other interaction information. We preprocessed the data to construct a multi-label protein-protein interaction (PPI) dataset {(p i ,p j ,I ij )}, where p i ,p j Corresponding to the protein numbered i,j, I ij Corresponding interaction category;

[0059] In this step, the pretreatment specifically includes:

[0060] The data was cleaned to remove the protein sequence parts with more than 512 amino acids and to fill the sequences with less than 512 amino acids with 0;

[0061] Then, a de-redundancy operation is performed to reduce the redundancy between protein sequences by executing a validated algorithm, including using a method to compare sequence similarity to determine and remove similar or identical protein sequences, avoiding duplication in multi-label classification problems, extracting interaction data from the STRING database, and removing protein sequences shorter than 50 amino acids and highly homogeneous proteins;

[0062] For interactions involving the same protein pair [p1, p2], if their interaction types are different, they are merged and the interaction types are integrated into one type vector In order to reduce the homogeneity and repetitiveness of the samples;

[0063] Then, a predefined threshold is used, a balanced sampling strategy is adopted, and multiple iterative perturbations are performed to construct a multi-label PPI dataset that is non-redundant, class-balanced, and diverse enough to ensure the diversity and fairness of the dataset.

[0064] In the specific example, the constructed multi-label PPI datasets with no redundancy and balanced categories are SYS30K and SYS60K, where the number of samples in the datasets are 30,074 and 60,457 respectively, aiming to provide a more extensive and effective training and validation platform for subsequent model performance evaluation and comparative studies.

[0065] S2. Extract features from the protein sequence, and divide the obtained features into nominal features and sequence features, forming a nominal feature vector and a sequence feature vector respectively;

[0066] The feature extraction is to generate a nominal feature vector of fixed dimension and a sequence feature vector of variable dimension according to a protein sequence of a certain length, wherein:

[0067] The generation of nominal feature vectors directly calculates the nominal feature vector of the protein using the physicochemical properties of the protein sequence, such as amino acid composition, hydrophobicity or hydrophilicity, etc. The dimension of this feature vector is fixed and low to enhance the computational efficiency of the model.

[0068] In this experiment, the nominal features were directly calculated based on the physicochemical properties of the protein sequence and fixed at a lower dimension;

[0069] Sequence features are generated by encoding each amino acid in the protein sequence, where each amino acid is encoded as a 48-dimensional numerical vector. The encoding combines the following aspects:

[0070] i. Using the amino acid encoding obtained by the pre-trained Word2Vec model to capture and utilize the amino acid context information contained in the sequence;

[0071] ii. One-hot encoding based on known physicochemical properties of amino acids to capture the essential characteristics of amino acids;

[0072] iii. Position weight matrix encoding extracted based on the position-related information of amino acids in the protein sequence, adding position specificity to enhance the model's understanding of the sequence structure.

[0073] The feature extraction comprises the following steps:

[0074] 2.1 Apply multiple one-dimensional convolutional layers with different core sizes to process the matrix X, where each core size is specifically adapted to capture the patterns and relationships between the corresponding number of amino acid residues;

[0075] 2.2 For each convolutional layer of the core size, use the maximum pooling layer to reduce the feature dimension, so as to extract the key sequence information and reduce the subsequent calculation;

[0076] 2.3 A dropout layer with an inactivation ratio of 0.3 is introduced to reduce the risk of overfitting and optimize the generalization ability of the model;

[0077] 2.4 The output feature maps of the three convolutional layers are concatenated into a unified feature vector after dimensionality reduction and inactivation, which serves as the input of the multi-label classification model.

[0078] S3. Consider the nominal feature vector and the sequence feature vector as descriptions of the protein from different perspectives, and build a model based on the joint training structure;

[0079] Based on the joint training architecture, the models complement each other, maximize the interaction consistency, and improve the prediction accuracy. The specific architecture diagram of the model built based on the joint training structure is as follows: Figure 2 As shown in the figure, based on the constructed protein-protein interaction (PPI) graph and two sets of protein features, one-dimensional convolution and graph convolution networks are used to extract feature information in the graph structure and predict protein interactions;

[0080] S4. Extract the relational structure of the interaction network from the nominal feature vector using a graph convolutional network (GCN) with spatial aggregation capabilities;

[0081] The method for extracting the relationship structure of the interaction network comprises:

[0082] Construct an undirected graph G = (P, E) of multiple categories of PPIs, corresponding to the graph structure representation of the seven types of PPIs. In each undirected graph, the vertex set P represents proteins, and the edge set E represents one of the seven types of PPIs recorded in the STRING database.

[0083] A graph convolutional network (GCN) layer is applied to capture the inter-vertex relational structure from the undirected graph G. The GCN layer uses spatial neighborhood sampling technology to learn and infer node hierarchical structure features. The graph convolutional network (GCN) is used to encode the k-order neighborhood of vertex p, and the formula is used:

[0084]

[0085] The 1-hop to k-hop neighborhood of each protein p is obtained by iterative calculation;

[0086] For a given protein p, a set of proteins Neigh(p) that directly interact with p is selected, and uniform sampling is performed to extract a fixed number of neighborhood sets U(p), where the size and depth k of uniform sampling are both set to 2 to adapt to the complex PPI graph structure and reduce the computational complexity;

[0087] By formula:

[0088] Agg=σ(W·concat(V(p i ), V(p j ))+b)

[0089] The aggregation function AGG described in processes the data from the k-hop neighborhood to refine the feature set of each protein p and reflect their dynamic interaction properties and complex relationships between proteins;

[0090] According to each protein p and its k-order neighborhood set V(p) = {h 1 (p)...h k (p)}, a trainable aggregation function is used to aggregate the embedding representation of each k-order neighborhood of p to form a comprehensive feature of p;

[0091] During model training, information from different deep neighborhood hierarchies is integrated to capture the complex interaction networks and relationships between proteins. Figure 3 This is a schematic diagram of sampling neighborhood nodes, which shows the uniform sampling of a 2-hop neighborhood of a fixed size of 2 for protein p. The lines represent interactions between proteins, dark gray represents selected nodes, and light gray represents unselected nodes. This application adopts an advanced graph neural network architecture, and by introducing uniform sampling technology and neighborhood aggregation functions, it can effectively solve the computational complexity problems associated with traditional GNN and adjacency matrices, and improve the efficiency and accuracy of protein sequence feature representation. This method is particularly suitable for constructing a complete and redundant multi-label protein interaction dataset.

[0092] S5. Use a one-dimensional convolutional neural network (1DCNN) to encode and extract protein sequence information from sequence features; the specific operations are as follows:

[0093] 5.1 Design three one-dimensional convolutional layers with different core sizes, each designed for sequence patterns of different lengths to capture the local and global patterns of amino acid sequences and their context information;

[0094] 5.2 The maximum pooling layer gradually performs spatial dimensionality reduction on the feature maps output by each convolutional layer to extract the most critical and representative features in the sequence;

[0095] 5.3 Apply dropout ratio to reduce overfitting of the model, thereby improving the generalization performance of the model on new data;

[0096] 5.4 The vector formed by concatenating the feature maps processed by all convolutional layers is used as the input of the multi-label prediction task. This step ensures that sequence information of different scales is fully considered by the model.

[0097] In a specific embodiment, 1D convolutional layers of sizes 9, 11, and 13 are used to effectively process sequence feature vectors, including a maximum pooling layer that gradually reduces the dimension of the output feature map, and a dropout layer of 0.3 is used to reduce overfitting; through this multi-level, multi-scale design, the model ensures that both local and global patterns in the amino acid sequence are deeply analyzed and learned.

[0098] S6. Apply a fully connected layer for multi-label prediction. The fully connected layer uses a sigmoid activation function and binary cross entropy as the loss function to handle the output of multiple interaction types. The formula is as follows:

[0099]

[0100] On the datasets SHS27K, SYS30K and SYS60K with different partitioning schemes, we verified the impact of integrating nominal features and sequence features on the multi-label PPI prediction performance. We tested the corresponding components of CoMPPI through ablation experiments. The model using only sequence features and the model using only nominal features are called w / oNom and w / oOrd, respectively. From the experimental return results, in BFS mode, compared with SHS27K, SYS30K and SYS60K models, CoMPPI's Micro-F1 is 0.023, 0.026, and 0.025 higher than w / oOrd, and 0.1439, 0.1272, and 0.0764 higher than w / o Nom. Comprehensive return data shows that in all cases, the performance of the CoMPI architecture is significantly better than the three most advanced methods of AFTGAN, GNN-PPI and PIPR.

[0101] The combined experimental results of the present invention found that the accuracy of predicting protein interaction categories can be significantly improved, which has a driving effect on the research of protein function and drug development.

[0102] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0103] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0104] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0105] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0106] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A method for predicting multi-label protein interactions based on joint training, It is characterized in that The following steps are involved: S1. Select interaction data from a database containing multiple PPI types as the original data source, preprocess the data, and construct a PPI dataset {(p i ,p j ,I ij )}, where p i ,p j Corresponding to the protein numbered i,j, I ij Corresponding interaction category; S2. Extract features from the protein sequence, and divide the obtained features into nominal features and sequence features, forming a nominal feature vector and a sequence feature vector respectively; S3. Consider the nominal feature vector and the sequence feature vector as descriptions of the protein from different perspectives, and build a model based on the joint training structure; S4. Extracting the relational structure of the interaction network from the nominal feature vector using a graph convolutional network with spatial aggregation capabilities; S5. Use one-dimensional convolutional neural network to encode and extract protein sequence information from sequence features; S6. Apply a fully connected layer for multi-label prediction. The fully connected layer uses a sigmoid activation function and binary cross entropy as the loss function to handle the output of multiple interaction types. The formula is as follows:

2. The method for predicting multi-label protein interactions based on joint training as claimed in claim 1, It is characterized in that In step S1, the preprocessing includes: The data was cleaned to remove the protein sequence parts with more than 512 amino acids and to fill the sequences with less than 512 amino acids with 0; Then, a de-redundancy operation is performed to reduce the redundancy between protein sequences by executing a validated algorithm, including using a method to compare sequence similarity to determine and remove similar or identical protein sequences, avoiding duplication in multi-label classification problems, extracting interaction data from the STRING database, and removing protein sequences shorter than 50 amino acids and highly homogeneous proteins; For the same protein pairs [p 1 ,p 2 ], if their interaction types are different, they are merged and the interaction types are integrated into one type vector In order to reduce the homogeneity and repetitiveness of the samples; Then, a predefined threshold is used, a balanced sampling strategy is adopted, and multiple iterative perturbations are performed to construct a multi-label PPI dataset that is non-redundant, class-balanced, and diverse enough to ensure the diversity and fairness of the dataset.

3. The method for predicting multi-label protein interactions based on joint training as claimed in claim 1, It is characterized in that In step S2, the feature extraction is based on a protein sequence of a certain length to generate a fixed-dimensional nominal feature vector and a variable-dimensional sequence feature vector. in: Nominal features are calculated directly based on the physicochemical properties of the protein sequence and are fixed in a lower dimension; Sequence features are generated by encoding each amino acid in the protein sequence, where each amino acid is encoded as a 48-dimensional numerical vector.

4. The method for predicting multi-label protein interactions based on joint training as claimed in claim 1, It is characterized in that In step S2, the feature extraction includes the following steps: Apply multiple one-dimensional convolutional layers with different kernel sizes to the matrix X, where each kernel size is specifically adapted to capture the patterns and relationships between the corresponding number of amino acid residues; For each convolutional layer of the kernel size, a maximum pooling layer is used to reduce the feature dimension, so as to extract key sequence information and reduce subsequent calculations; A dropout layer with a loss ratio of 0.3 is introduced to reduce the risk of overfitting and optimize the generalization ability of the model; The output feature maps of the three convolutional layers are concatenated into a unified feature vector after dimensionality reduction and inactivation, which serves as the input of the multi-label classification model.

5. The method for predicting multi-label protein interactions based on joint training as claimed in claim 1, It is characterized in that In step S4, the method for extracting the relationship structure of the interaction network includes: Construct an undirected graph G = (P, E) of multiple categories of PPIs, corresponding to the graph structure representation of the seven types of PPIs. In each undirected graph, the vertex set P represents proteins, and the edge set E represents one of the seven types of PPIs recorded in the STRING database. The GCN layer is applied to capture the inter-vertex relational structure from the undirected graph G. The GCN layer uses the spatial neighborhood sampling technique to learn and infer node hierarchy features. GCN is used to encode the k-order neighborhood of vertex p and use the formula: The 1-hop to k-hop neighborhood of each protein p is obtained by iterative calculation; For a given protein p, a set of proteins Neigh(p) that directly interact with p is selected, and uniform sampling is performed to extract a fixed number of neighborhood sets U(p), where the size and depth k of uniform sampling are both set to 2 to adapt to the complex PPI graph structure and reduce the computational complexity; By formula: Agg=σ(W·concat(V(p i ),V(p j ))+b) The aggregation function AGG described in processes the data from the k-hop neighborhood to refine the feature set of each protein p and reflect their dynamic interaction properties and complex relationships between proteins; According to each protein p and its k-order neighborhood set V(p) = {h 1 (p)...h k (p)}, a trainable aggregation function is used to aggregate the embedding representation of each k-order neighborhood of p to form a comprehensive feature of p; During model training, information from different deep neighborhood hierarchies is integrated to capture the complex interaction networks and relationships between proteins.

6. The method for predicting multi-label protein interactions based on joint training as claimed in claim 1, It is characterized in that In step S5, the method for encoding and extracting the protein sequence information in the sequence feature is as follows: Three one-dimensional convolutional layers with different core sizes are designed, each for sequence patterns of different lengths, to capture the local and global patterns of amino acid sequences and their contextual information; The maximum pooling layer gradually performs spatial dimensionality reduction on the feature maps output by each convolutional layer to extract the most critical and representative features in the sequence; Apply dropout ratio to reduce overfitting of the model, thereby improving the generalization performance of the model on new data; The vector formed by concatenating the feature maps processed by all convolutional layers is used as the input of the multi-label prediction task to ensure that sequence information of different scales is fully considered by the model.

Citation Information

Cited By

  • Efficient protein stability prediction method for selective state space modeling

    CN120496640A

  • An efficient protein stability prediction method with selective state space modeling

    CN120496640B