Protein interaction prediction method based on hybrid graph attention neural network

Abstract

The application discloses a protein interaction prediction method based on a hybrid graph attention neural network, and belongs to the technical field of protein interaction prediction, and comprises the following steps: preprocessing of protein data; extracting and fusing protein features; constructing a TCN neural network; constructing a hybrid graph attention neural network; model training; and protein interaction prediction. The application takes the fusion of two kinds of features as protein feature expression, extracts protein global features by using a TCN network, and does not need to extract global features by using a convolutional neural network; samples GAT and GCN networks to extract protein local features, and does not need to separately extract local features by using a sliding window technology.

Description

Technical Field

[0001] The present invention relates to the technical field of protein interaction prediction, and in particular to a protein interaction prediction method based on a hybrid graph attention neural network. Background Art

[0002] Protein-protein interaction (PPI) is a very important research topic in biology. It refers to the functional connection between two or more protein molecules inside and outside the cell through specific interactions. PPI relationships are crucial for understanding cell functions, signal transduction, metabolic processes, etc. Interactions between proteins are the basis of intracellular biological processes, and they play a key role in a variety of biological activities, including gene expression regulation, cell cycle, immune response, signal transduction, etc. The study of protein interactions is crucial for a deep understanding of the role of proteins in various biological processes in organisms. This not only helps to promote disease treatment and drug development, but also promotes progress in the medical field and deepens the understanding of molecular mechanisms. However, current predictions of protein-protein interactions mainly rely on laboratory methods, which are usually time-consuming and labor-intensive, require a lot of resources, and can only be applied to specific proteins, and are difficult to be universally applied to all proteins.

[0003] Therefore, there is an urgent need for a highly accurate and reliable method to predict undiscovered protein-protein interactions. With the continuous advancement of neural network technology, deep learning and graph neural networks have gradually been applied to the prediction of protein-protein interactions. Neural network-based PPI prediction methods can be divided into two categories: one that relies solely on amino acid sequences, and the other that combines amino acid sequences with protein structural information. Methods that rely solely on amino acid sequences represent proteins by extracting sequence features, which ignores the protein's structural information and therefore affects the accuracy of predictions to a certain extent. In contrast, methods that combine amino acid sequences and protein structural features not only consider sequence information but also incorporate structural features, which can more comprehensively express protein characteristics. This method compensates for the shortcomings of relying solely on sequences and enriches feature expression, but the process of extracting protein structural features is very complex, increasing the computational and implementation difficulty of the model.

[0004] In summary, laboratory methods for predicting protein interactions are inefficient, time-consuming, and labor-intensive. Methods that rely solely on amino acid sequences ignore protein structural information. While methods that combine amino acid sequences and protein structural features can compensate for the lack of structural information, the extraction process of these structural features is very complex. Therefore, a protein interaction prediction method based on a hybrid graph attention neural network is proposed. Summary of the Invention

[0005] The technical problem addressed by this invention is how to extract key features that represent proteins and, based on these representative features, accurately and efficiently predict protein-protein interactions. Furthermore, this method uses automated machine learning to reduce laboratory research costs. A protein interaction prediction method based on a hybrid graph attention neural network is provided.

[0006] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0007] S1: Preprocessing of protein data

[0008] Obtain multiple benchmark data sets, filter the benchmark data sets, and then merge the benchmark data sets to obtain training sets and test sets;

[0009] S2: Extract and fuse protein features

[0010] The evolutionary characteristics of the protein sequence are obtained through the amino acid sequence of the protein, and the atomic characteristics and secondary structure characteristics of the protein are obtained according to the protein PDB file. Then, the positional embedding characteristics of the protein are calculated based on the position coordinates of the amino acid residues. Finally, the protein's evolutionary characteristics, secondary structure characteristics, atomic characteristics, and positional embedding characteristics are fused to obtain the protein's characteristic expression, which is used as the input data feature of the protein-protein interaction prediction model.

[0011] S3: Building a TCN Neural Network

[0012] The encoding part of the transformer is spliced ​​and fused with the AGCN network to obtain the TCN network;

[0013] S4: Building a Hybrid Graph Attention Neural Network

[0014] Construct a hybrid graph attention neural network, fuse the GCN network with the GAT network to extract local features of proteins, and use the TCN network to extract global features of proteins, to obtain a protein-protein interaction hybrid graph attention neural network;

[0015] S5: Model training

[0016] The protein-protein interaction mixed graph attention neural network is trained using the training set to obtain the trained weight parameters, which are then saved to obtain a protein-protein interaction prediction model.

[0017] S6: Protein-protein interaction prediction

[0018] The test set data is input into the protein-protein interaction prediction model for prediction to obtain the prediction results.

[0019] Furthermore, in step S1, the specific processing process is as follows: obtaining the amino acid sequence of the protein and the public interaction dataset, and performing six-step filtering processing on the dataset in sequence, wherein the six-step filtering processing is to exclude structures with more than 30% missing residues, remove repeated chains with the same UniProtKB / Swiss-Prot identifier, remove transmembrane proteins, remove oligomeric structures, remove proteins with buried surface accessibility and interface polarity below a set threshold, and remove redundant proteins.

[0020] Furthermore, in step S2, the specific processing process is as follows:

[0021] S21: PSSM and HMM features are calculated based on the amino acid sequence of the protein. The evolutionary characteristics of the protein are obtained by splicing the PSSM and HMM features. The protein structure and sequence information are obtained through the protein PDB file, and the secondary structure characteristics of the protein DSSP are calculated according to the DSSP algorithm;

[0022] S22: Based on the protein PDB files on the protein data bank website and the intrinsic properties of the residues, seven features of each atom constituting the residue are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electronic charge, number of hydrogen atoms bonded to it, whether it is cyclic, and van der Waals radius of the atom; the corresponding average value of the seven features of all atoms constituting the residue is taken as the seven atomic features of the residue;

[0023] S23: The coordinates of the residue side chain center of mass are used as the residue pseudo-positions, and the protein position embedding features are calculated using the Euclidean distance formula. The reference position for calculating the position embedding features is the first residue in each protein sequence.

[0024] S24: Finally, the four protein features of protein evolutionary features, secondary structure features, atomic features and position embedding features are fused and spliced ​​to obtain the 62-dimensional features of the protein.

[0025] Furthermore, in step S3, the encoding part of the transformer is connected in series with the AGCN network to form a TCN network.

[0026] Furthermore, in step S4, the GCN network and the GAT network are connected in parallel to obtain the embedding representation of each node respectively.

[0027] Furthermore, in step S4, the calculation formula of the GCN network is as follows:

[0028] H (l+1) =σ(((1-α)PH (l) +αH(0) )((1-β l )I n +β l W (l) ))

[0029] Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) Represents the node feature embedding before and after the layer runs, W (l) Represents the weight matrix, α and β l is a hyperparameter.

[0030] Furthermore, in step S4, the calculation formula of the GAT network is as follows:

[0031] H (l+1) =σ((1-β l )((1-α)H (l) +αH (0) )+(β l (H (l) ||H (0) )W (l) ))

[0032] Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) represents the node feature embedding before and after the layer is run, || represents the connection operation, W (l) Denotes linear learnable parameters, α and β l is a hyperparameter.

[0033] Furthermore, in step S4, in the protein-protein interaction mixed graph attention neural network, the feature expression of each protein is called a node, and the protein feature expression directly connected to the protein feature expression to be updated is called the neighborhood node of the protein.

[0034] Furthermore, in step S5, during the model training process, the loss function is a cross entropy loss function, and the formula is as follows:

[0035]

[0036] Where N is the total number of samples, y i is the true label of the i-th sample, p i is the predicted probability of the i-th sample.

[0037] Furthermore, in step S6, the prediction results of protein-protein interactions are calculated with the original labels to obtain performance parameters of the model.

[0038] Compared with the existing technology, the present invention has the following advantages: the protein interaction prediction method based on hybrid graph attention neural network fuses two features as protein feature expression, uses TCN network to extract global protein features, and does not need to use convolutional neural network to extract global features; samples GAT and GCN networks to extract local features of proteins, and does not need to use sliding window technology to extract local features separately; both GAT and GCN networks use residual structure and identity mapping, which can effectively alleviate the transition smoothing problem of deep GAT and GCN networks. Through this operation, deeper protein features can be extracted. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a schematic flow chart of a protein interaction prediction method based on a hybrid graph attention neural network in an embodiment of the present invention;

[0040] Figure 2 Schematic diagram of the TCN structure in the protein-protein interaction prediction model in an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of the GCN and GAT structures in the protein-protein interaction prediction model according to an embodiment of the present invention;

[0042] Figure 4 It is a schematic diagram of the overall structure of the protein-protein interaction prediction model in the example of the present invention. DETAILED DESCRIPTION

[0043] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0044] This embodiment provides a technical solution: a protein interaction prediction method based on a hybrid graph attention neural network, comprising

[0045] like Figure 1 As shown, this embodiment provides a technical solution: a protein interaction prediction method based on a hybrid graph attention neural network, comprising the following steps:

[0046] S1: Preprocessing of protein data.

[0047] In step S1, the following contents are specifically included:

[0048] S11: Protein data were refined through a six-step filtering process, including excluding structures with more than 30% missing residues, removing duplicate chains with the same UniProtKB / Swiss-Prot identifiers, removing transmembrane proteins, removing oligomeric structures (higher than dimers), removing proteins with buried surface accessibility and interface polarity below a certain threshold, and removing redundant proteins. The processed data and the distribution of interacting and non-interacting residues in the data are shown in Table 1.

[0049] Table 1 Statistics of three benchmark datasets and the training and test sets used in this example

[0050]

[0051]

[0052] S2: Extract multiple features of proteins and perform feature fusion;

[0053] In step S2, the following contents are specifically included:

[0054] S21: Obtain protein sequence and structural features through computational programs;

[0055] In step S21, the evolutionary characteristics of the protein are obtained through the amino acid sequence of the protein, including the position-specific score matrix (PSSM) and the hidden Markov model matrix (HMM). The PSSM is generated by running the alignment tool PSI-BLAST v2.10.1, which has an E value of 0.001 after three iterations. The HMM matrix is ​​generated by running the HHblits v3.0.3 algorithm using default parameters. DSSP is a program that calculates three types of structural characteristics of the protein. The structure and sequence information of the protein is obtained through the PDB file of the protein, and the secondary structure feature matrix (DSSP) of the protein is obtained according to the DSSP algorithm tool. Among them, the shapes of the PSSM and HMM matrices are both N v ×20, the DSSP matrix shape is N v ×14.

[0056] S22: Obtain the atomic features of the protein through the protein PDB file, and then calculate the positional embedding features of the protein through the position coordinates of the amino acid residues;

[0057] In step S22, seven features of each atom (excluding hydrogen atoms) constituting the residue are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electronic charge, the number of hydrogen atoms bonded to it, whether it is cyclic, and the van der Waals radius of the atom; since the number of atoms of different residues is different, the corresponding average value of the seven features of all atoms constituting the residue is taken as the seven atomic features of the residue, and its matrix shape is N v ×7; the position embedding feature of the residue contains the relative position information of each residue and the reference residue; in this embodiment, the coordinates of the center of mass of the residue side chain are used as the pseudo position of the residue, and the shape is N v ×3, the matrix shape is calculated by the Euclidean distance formula to be N v ×1 protein embedding features.

[0058] S22: Finally, the above four features are fused. First, PSSM and HMM are spliced ​​to form the evolutionary features of the protein, then the secondary structure features of the protein are spliced, and finally the atomic features and position embedding features of the protein are spliced ​​to obtain the 62-dimensional features of the protein, which are used as the input data features of the protein-protein interaction prediction model.

[0059] S3: Building the TCN Network

[0060] In step S3, the encoding part of the transformer is fused with the AGCN network to form a TCN network, which is used to extract the global features of proteins.

[0061] In step S3, the encoding part of the transformer is fused with the AGCN network. The fusion method is to obtain a TCN network by connecting the transformer encoding part and AGCN in series, which helps to improve the expressiveness and generalization ability of the model.

[0062] S4: Building a Hybrid Graph Attention Neural Network

[0063] In step S4, a protein-protein interaction hybrid graph attention neural network is constructed, using protein feature expressions as nodes and protein-protein interactions as edges. In this network, the GCN and GAT networks add residual structures and identity mappings, effectively alleviating the oversmoothing problem caused by deep graph neural networks. The two different models extract richer node features.

[0064] In this embodiment, in the protein-protein interaction hybrid graph attention neural network, the fine-grained relationship (local features) captured by the self-attention mechanism and the global structural features (global features) extracted by the TCN network are fused through the combination of the GAT network and the GCN network. This fusion not only enriches the feature space of the model, but also enhances the differences between nodes and improves the discriminability of classification and prediction tasks. At the same time, adding residuals to the deep graph neural network helps to alleviate the over-smoothing problem and enhance the discriminability of feature expression, and the residual connection facilitates the training of deep networks. The formula of the GCN network is as follows:

[0065] H (l+1) =σ(((1-α)PH (l) +αH (0) )((1-β l )I n +β l W (l) ))

[0066] Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) Represents the node feature embedding before and after the layer runs, W (l) Represents the weight matrix, α and β l is a hyperparameter.

[0067] The formula of the GAT network is as follows:

[0068] H (l+1) =σ((1-β l )((1-α)H (l) +αH (0) )+(β l (H (l) ||H (0) )W (l) ))

[0069] Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) represents the node feature embedding before and after the layer is run, || represents the connection operation, W (l) Denotes linear learnable parameters, α and β l is a hyperparameter.

[0070] It should be noted that the above two formulas use initial residual and identity mapping calculations to alleviate the transition smoothing problem of the eight-layer GAT and GCN networks.

[0071] It should be noted that residual connections add the output of the previous layer to the output of the current layer and use them as the final output of that layer, forming a "skip connection" to facilitate information flow and prevent gradient vanishing. Identity mapping means that there is no transformation or convolution operation, and the input is passed directly to the output.

[0072] S5: Model network training

[0073] The training set is used to train a hybrid graph attention neural network to obtain a protein-protein interaction prediction model;

[0074] Specifically, in step S5, the specific training scheme is as follows:

[0075] S51: The initial learning rate is set to 0.001, and the model is optimized using the cross entropy loss function and the Adam optimizer. Each training includes 50 epochs.

[0076] S52: In this experiment, 5-fold cross-validation training was performed on the training set.

[0077] S53: During training, to demonstrate the relative importance of each feature, we performed feature ablation experiments by using a single feature or excluding a feature from the final feature combination. We also evaluated model performance using 5-fold cross-validation (CV) and independent Test_60 for AUROC and AUPRC. Table 2 shows the impact of these features on model performance:

[0078] Table 2 Using individual features or removing individual features from the combined features

[0079]

[0080] S6: Protein-protein interaction prediction

[0081] The test dataset (Test_60) was fed into the protein-protein interaction prediction model to generate predictions. Table 4 shows that our method outperforms other methods. The first five methods are sequence-based, while the last four utilize protein structure information. As shown in Table 3, our model (ours) significantly outperforms other methods on the benchmark dataset Test_60 across seven metrics.

[0082] Table 3 Performance comparison with other methods on Test_60

[0083]

[0084]

[0085] In summary, the protein-protein interaction prediction method based on a hybrid graph attention neural network in the above embodiment first utilizes three publicly available and widely used benchmark datasets from previous research: Dset_186, Dset_72, and Dset_164, named after the number of proteins in the datasets. These datasets are refined through a six-step filtering process, including excluding structures with more than 30% missing residues, removing chains with the same UniprotKB / Swiss-Prot accessions, removing transmembrane proteins, removing oligomeric structures (higher than dimers), removing proteins with buried surface accessibility and interface polarity below a certain threshold, and removing redundant proteins. A hybrid graph attention neural network for protein-protein interactions is constructed, using protein expressions as nodes and protein-protein interactions as edges. The network model is then trained, using a cross-entropy loss function to accurately measure the binary classification of the data and predict the error. This also drives the model toward higher accuracy and confidence, thereby improving model performance. Finally, unknown proteins are input into the protein-protein interaction prediction model for prediction, obtaining prediction results. By fusing four protein features, a TCN, GAT, and GCN network were constructed to extract global and local protein features. The performance of different models was verified and compared on a test set. The experimental results demonstrate that these techniques can accelerate network convergence and enhance robustness. The proposed model achieves high prediction accuracy and is expected to be deployed on edge devices for production and application in the future.

[0086] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A protein interaction prediction method based on a hybrid graph attention neural network, characterized by: The following steps are involved: S1: Preprocessing of protein data Obtain multiple benchmark data sets, filter the benchmark data sets, and then merge the benchmark data sets to obtain training sets and test sets; S2: Extract and fuse protein features The evolutionary characteristics of the protein sequence are obtained through the amino acid sequence of the protein, and the atomic characteristics and secondary structure characteristics of the protein are obtained according to the protein PDB file. Then, the positional embedding characteristics of the protein are calculated based on the position coordinates of the amino acid residues. Finally, the protein's evolutionary characteristics, secondary structure characteristics, atomic characteristics, and positional embedding characteristics are fused to obtain the protein's characteristic expression, which is used as the input data feature of the protein-protein interaction prediction model. S3: Building a TCN Neural Network The encoding part of the transformer is spliced ​​and fused with the AGCN network to obtain the TCN network; S4: Building a Hybrid Graph Attention Neural Network Construct a hybrid graph attention neural network, fuse the GCN network with the GAT network to extract local features of proteins, and use the TCN network to extract global features of proteins, to obtain a protein-protein interaction hybrid graph attention neural network; S5: Model training The protein-protein interaction mixed graph attention neural network is trained using the training set to obtain the trained weight parameters, which are then saved to obtain a protein-protein interaction prediction model. S6: Protein-protein interaction prediction The test set data is input into the protein-protein interaction prediction model for prediction to obtain the prediction results.

2. The protein interaction prediction method based on hybrid graph attention neural network according to claim 1, characterized in that In step S1, the specific processing process is as follows: the amino acid sequence of the protein and the public interaction dataset are obtained, and the dataset is sequentially subjected to a six-step filtering process, wherein the six-step filtering process is to exclude structures with more than 30% missing residues, remove duplicate chains with the same UniProtKB / Swiss-Prot identifier, remove transmembrane proteins, remove oligomeric structures, remove proteins with buried surface accessibility and interface polarity below a set threshold, and remove redundant proteins.

3. The protein interaction prediction method based on hybrid graph attention neural network according to claim 1, characterized in that In step S2, the specific processing process is as follows: S21: PSSM and HMM features are calculated based on the amino acid sequence of the protein. The evolutionary characteristics of the protein are obtained by splicing the PSSM and HMM features. The protein structure and sequence information are obtained through the protein PDB file, and the secondary structure characteristics of the protein DSSP are calculated according to the DSSP algorithm; S22: Based on the protein PDB files on the protein data bank website and the intrinsic properties of the residues, seven features of each atom constituting the residue are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electronic charge, number of hydrogen atoms bonded to it, whether it is cyclic, and van der Waals radius of the atom; the corresponding average value of the seven features of all atoms constituting the residue is taken as the seven atomic features of the residue; S23: The coordinates of the residue side chain center of mass are used as the residue pseudo-positions, and the protein position embedding features are calculated using the Euclidean distance formula. The reference position for calculating the position embedding features is the first residue in each protein sequence. S24: Finally, the four protein features of protein evolutionary features, secondary structure features, atomic features and position embedding features are fused and spliced ​​to obtain the 62-dimensional features of the protein.

4. The protein interaction prediction method based on hybrid graph attention neural network according to claim 3, characterized in that In step S3, the encoding part of the transformer is connected in series with the AGCN network to form a TCN network.

5. The protein interaction prediction method based on hybrid graph attention neural network according to claim 4, characterized in that In step S4, the GCN network and the GAT network are connected in parallel to obtain the embedding representation of each node.

6. The protein interaction prediction method based on hybrid graph attention neural network according to claim 5, characterized in that In step S4, the calculation formula of the GCN network is as follows: H (l+1) =σ(((1-α)PH (l) +αH (0) )((1-b l )I n +b l W (l) )) Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) Represents the node feature embedding before and after the layer runs, W (l) Represents the weight matrix, α and β l is a hyperparameter.

7. The protein interaction prediction method based on hybrid graph attention neural network according to claim 5, characterized in that In step S4, the calculation formula of the GAT network is as follows: H (l+1) =σ((1-β l )((1-a)H (l) +αH (0) )+(β l (H (l) ||H (0) )W (l) )) Where σ represents the RELU activation function, P represents the normalized adjacency matrix in the equation, and I n represents the identity matrix, H (0) represents the input node feature matrix, H (l) and H (l+1) represents the node feature embedding before and after the layer is run, || represents the connection operation, W (l) Denotes linear learnable parameters, α and β l is a hyperparameter.

8. The protein interaction prediction method based on hybrid graph attention neural network according to claim 1, characterized in that In step S4, in the protein-protein interaction mixed graph attention neural network, the feature expression of each protein is called a node, and the protein feature expression directly connected to the protein feature expression to be updated is called the neighborhood node of the protein.

9. The protein interaction prediction method based on hybrid graph attention neural network according to claim 1, characterized in that In step S5, during the model training process, the loss function is the cross entropy loss function, and the formula is as follows: Where N is the total number of samples, y i is the true label of the i-th sample, p i is the predicted probability of the i-th sample.

10. The protein interaction prediction method based on hybrid graph attention neural network according to claim 1, characterized in that In step S6, the prediction results of protein-protein interactions are calculated with the original labels to obtain the performance parameters of the model.

Images