A method, apparatus and medium for identifying protein interaction sites

Abstract

This invention discloses a method, apparatus, and medium for identifying protein-protein interaction sites. The method constructs a GHGPR-PPIS model, which includes five sequentially connected GraphHeat-GPR modules, a linear layer, a fully connected layer, and a softmax layer. Built upon a graph convolutional network, it employs a hot kernel and integrates generalized PageRank technology and edge self-attention feature processing blocks. This fully utilizes hidden information in the protein graph, significantly improving the performance of protein-protein interaction site prediction. Compared with other competing models, it not only reduces the number of layers and training parameters but also slightly improves testing performance, generalization ability, and practical application capability.

Description

Technical Field

[0001] This invention relates to the field of digital medical technology, and in particular to a method, apparatus and medium for identifying protein interaction sites. Background Technology

[0002] Proteins are crucial biological macromolecules that play a variety of roles in cells. They are the main driving forces of life processes and participate in almost all physiological functions. Proteins perform their biological functions through interactions with various molecules, including proteins, nucleic acids, peptides, and substrates. Accurate identification of protein-protein interaction sites (PPIS) at the molecular level is of great significance for annotating protein function, understanding protein-protein interaction networks, and studying the mechanisms of various diseases. Currently, methods for identifying PPIS can be divided into two categories: traditional laboratory wet experiments and computational methods. However, traditional laboratory wet experiments for PPIS identification are expensive and time-consuming. Therefore, seeking accurate computational methods to predict PPIS has become a prominent research hotspot in the field of protein function.

[0003] In recent years, numerous computational methods have been proposed for predicting PPIS. Among them, machine learning-based and deep learning-based methods have become widely used for PPIS identification. Machine learning-based methods utilize protein sequence and structural information, such as the original protein sequence, position-specific scoring matrix (PSSM), and definition of secondary structure of proteins (DSSP), as features. Machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), Naive Bayes, and eXtreme Gradient Boosting, are employed to predict PPIS. While machine learning-based methods demonstrate good predictive performance, the development and use of deep learning models have further improved the accuracy of PPIS identification.

[0004] Deep learning-based methods for predicting protein sequence identity information (PPIS) can be broadly categorized into three types: models based on Convolutional Neural Networks (CNNs), models based on Recurrent Neural Networks (RNNs), and models based on Graph Neural Networks (GNNs). CNN-based models capture local contextual features, such as PSSM and DSSP, from sequence and structural information for PPIS prediction. For example, DeepPPISP uses TextCNN to extract global features from protein sequences, concatenates these global and local features, and inputs them into fully connected (FC) layers for PPIS prediction. ProB-site utilizes 1D-CNN to generate three distinct feature sets from sequence information, then combines these feature sets and inputs them into FC layers for PPIS prediction. While CNN models can effectively extract local features from protein sequences and achieve good performance, they neglect long-range dependencies within the protein sequence.

[0005] To address the lack of long-range dependency features in protein sequences, RNNs with memory capabilities have been introduced into PPIS prediction research. For example, DELPHI combines CNNs and RNNs to extract local and long-range dependency features for PPIS prediction. DLPred utilizes a deep learning architecture based on Simplified Long Short-Term Memory (SLSTM) to improve the prediction of imbalanced protein interaction sites. Although CNN- and RNN-based models primarily utilize features of protein primary sequence and secondary structure for PPIS prediction, their performance has reached a bottleneck. Therefore, models utilizing protein tertiary structure features to identify PPIS are quietly emerging. However, due to the lack of definite tertiary structures for many proteins, research on PPIS identification using GNN-based deep learning models remains in its early stages. Until recently, the advent of Alphafold, a highly accurate protein structure prediction tool, alleviated this limitation by providing reliable protein structures.

[0006] In recent years, GNN-based models have gained widespread attention and achieved good results in extracting high-level structural features. For example, GraphPPIS transforms the PPIS prediction problem into a graph node classification task and employs Graph Convolutional Networks (GCNs) to distinguish between interacting and non-interacting amino acid residues. GraphPPIS outperforms five competing sequence-based methods in terms of performance. Another example is AGAT-PPIS, a GNN-based method that incorporates protein structural features and has achieved state-of-the-art performance in PPIS prediction to date. AGAT-PPIS adds edge features to the Graph Attention Network (GAT) to provide additional structural information and enhances translation and rotation invariance. It utilizes eight AGAT layers to mine node embeddings and uses a Multi-Layer Perceptron (MLP) to predict PPIS. Although GraphPPIS and AGAT-PPIS significantly improve PPIS prediction accuracy by leveraging the protein structural features of GNNs, GNN-based methods are still in their early stages. Furthermore, current GNN models, such as GCN, GAT, and many others, face two significant challenges. First, GNN-based methods aim to enhance graph-based semi-supervised learning by defining graph convolutions as weighted averages of neighboring nodes and then learning the graph's convolution kernels. This strategy leverages the inherent smoothness of graphs to improve performance. However, a major obstacle in GNN-based methods lies in determining the optimal neighborhood of each node in the graph. Second, as the depth of GNN models increases, features tend to become overly smooth, leading to performance degradation. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method, device, and medium for identifying protein interaction sites. It also proposes a graph network model called GHGPR-PPIS, which is based on a graph convolutional network with a graph heat kernel. Furthermore, it integrates Generalized PageRank (GPR) technology and edge self-attention feature processing blocks to fully utilize the hidden information in the protein graph to accurately identify protein-protein interaction sites.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for identifying protein interaction sites includes the following steps:

[0010] Obtain the protein-protein interaction dataset, divide the dataset into training and testing sets, and perform preprocessing.

[0011] Obtain the sequence feature representation, structural feature representation, and edge feature representation of protein chains in the training and testing sets, and use the sequence feature representation and structural feature representation of protein chains as node feature information of protein chains;

[0012] The training set is input into a pre-defined GHGPR-PPIS model, which includes five sequentially connected GraphHeat-GPR modules, a linear layer, a fully connected layer, and a softmax layer. The GraphHeat-GPR modules are used to obtain low-frequency signals between nodes of the protein chain through hot kernel convolution and update node features to obtain first node features; they are used to feed the first node features and edge features into an edge self-attention module to update the attention coefficients between nodes, obtaining updated second node features; they are used to linearly combine the second node features with the initial node features and feed them into a normalization layer as input to the next GraphHeat-GPR module; the linear layer is used to weightedly combine the outputs of each GraphHeat-GPR module to obtain the final result, which is then fed into the fully connected layer; the fully connected layer and the softmax layer are used to predict the results.

[0013] The network parameters are updated by comparing the true labels with the predicted results and minimizing the cross-entropy loss function until the maximum number of iterations is reached, at which point the training ends and the trained GHGPR-PPIS model is obtained.

[0014] Input the test set into the trained GHGPR-PPIS model to obtain protein-protein interaction sites.

[0015] Preferably, the preprocessing includes removing protein chains from the dataset that do not correspond to the corresponding structures, thus creating a new dataset.

[0016] Preferably, the sequence features of the protein data are obtained from the PSSM, the Hidden Markov Model matrix, and the atomic features of each amino acid residue. The PSSM and the Hidden Markov Model matrix are generated using PSI-BLAST and HHblits tools, respectively, to produce N*20 dimensional sequence feature matrices, which are then normalized, where N represents the length of the protein chain. The atomic features of the amino acid residues include atomic mass, B factor, whether it is a side chain atom, charge, the number of hydrogen atoms attached to its bonds, whether it is part of a ring, and the van der Waals radius of the atom.

[0017] Preferably, the structural feature representation of the protein data consists of a 14-dimensional matrix DSSP and pseudo-position embedding features of amino acid residues. The 14-dimensional matrix DSSP includes 9 dimensions representing the secondary structural states of the protein chain, 4 dimensions which are transformed into the twist angles PHI and PSI of the peptide backbone using sine and cosine functions, and 1 dimension which is converted into relative solvent accessibility by converting the solvent-accessible surface area. The pseudo-position embedding features of amino acid residues are the relative position information of each amino acid residue relative to a reference amino acid residue.

[0018] Preferably, the edge feature representation of the protein chain includes the Euclidean distance between two nodes and the cosine of the angle between the vector formed by the two nodes and the first amino acid residue of the protein chain.

[0019] Preferably, the step of using hot kernel convolution processing to obtain low-frequency signals between nodes of the protein chain and updating node features specifically includes the following steps:

[0020] The definition of the thermonucleus is as follows:

[0021]

[0022] In the formula, The eigenvalues ​​u of the normalized Laplacian matrix L are described. i Smoothness, L = ID -1 / 2 AD -1 / 2 And I represents the identity matrix, s is the scaling parameter, A is the protein chain diagram, i.e., the adjacency matrix, and D is the diagonal matrix of A.

[0023] Using heat kernels to analyze the node features of the input protein chain Processing is performed to obtain new node features. As shown in formula (2):

[0024]

[0025] In the formula, s is the scaling parameter (s≥0). and It is a learnable parameter matrix.

[0026] Preferably, the step of feeding the updated node features and edge features into the edge self-attention module to update the attention coefficients between nodes and obtain the updated node features specifically includes the following steps:

[0027] Introducing the edge self-attention coefficient g ij The updated node features and edge features are fed into the edge self-attention module to obtain the self-attention coefficients δ between nodes. ij As shown below:

[0028]

[0029]

[0030] β ij =σ(e ij +g ij (5)

[0031]

[0032] In the formula, i and j represent nodes i and j in the protein chain diagram, respectively, and b is the bias. Representing edge features, β ij and e ij W represents the weight of node j relative to node i and the attention coefficient between nodes, respectively. m W α and W v It is a learnable parameter matrix, N i Represents the set of neighboring nodes of node i;

[0033] Based on the self-attention coefficients, the updated node features are obtained. As shown below:

[0034]

[0035] Among them, W ε and W ρ is a learnable parameter matrix.

[0036] Preferably, the output results of each GraphHeat-GPR module are weighted and combined, specifically including the following steps:

[0037] The GPR weights are calculated as follows:

[0038] γ n =α(1-α) n ,γ N =(1-α) N (8)

[0039] Where, γ n Let α represent the GPR weights for the nth propagation step, α be a learnable parameter, and N be the final number of steps.

[0040] The outputs of each GraphHeat-GPR module are weighted and combined to obtain the final output feature Z, as shown below:

[0041]

[0042] In the formula, H (n) This is the network output of step n.

[0043] Based on the above, the present invention also discloses a protein interaction site recognition device, including a memory for storing a computer program; and a processor for executing the computer program to implement any of the methods described above.

[0044] Based on the foregoing, the present invention also discloses a computer-readable medium storing a computer program, which, when executed by a processor, implements any of the methods described above.

[0045] Based on the above technical solution, the beneficial effects of the present invention are as follows: The present invention provides a method for identifying protein interaction sites, which involves acquiring a protein-protein interaction dataset, dividing the dataset into a training set and a test set, and preprocessing the dataset; acquiring the sequence feature representation, structural feature representation, and edge feature representation of protein chains in the training set and the test set, and using the sequence feature representation and structural feature representation of protein chains as node feature information of protein chains; inputting the training set into a preset GHGPR-PPIS model, wherein the GHGPR-PPIS model includes five sequentially connected GraphHeat-GPR modules, a linear layer, a fully connected layer, and a softmax layer, wherein the GraphHeat-GPR module is used to obtain low-frequency signals between nodes of the protein chain through hot kernel convolution processing and update the node features to obtain the first The first node feature and edge feature are fed into the edge self-attention module to update the attention coefficients between nodes and obtain the updated second node feature. The second node feature is linearly combined with the initial node feature and fed into the normalization layer as input to the next GraphHeat-GPR module. The linear layer is used to weightedly combine the outputs of each GraphHeat-GPR module to obtain the final result, which is then fed into the fully connected layer. The fully connected layer and softmax are used to predict the result. The network parameters are updated by comparing the true label with the predicted result and minimizing the cross-entropy loss function until the maximum number of iterations is reached, at which point training ends, and the trained GHGPR-PPIS model is obtained. The test set is input into the trained GHGPR-PPIS model to obtain protein-protein interaction sites. This invention, by using a hot kernel on the basis of a graph convolutional network and integrating generalized PageRank technology and edge self-attention feature processing blocks, fully utilizes the hidden information in the protein graph, significantly improving the performance of protein-protein interaction site prediction. Compared with other competing models, it not only reduces the number of layers and training parameters but also slightly improves test performance, generalization ability, and practical application capability. Attached Figure Description

[0046] Figure 1 This is a general flowchart of a protein interaction site identification method in one embodiment;

[0047] Figure 2 This is a diagram of the GHGPR-PPIS model architecture in a protein interaction site identification method in one embodiment.

[0048] Figure 3 This is a schematic diagram of node updating in a protein interaction site identification method in one embodiment;

[0049] Figure 4 This is a schematic diagram of the self-attention coefficient update process in a protein interaction site identification method in one embodiment. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0051] like Figure 1 , 2 As shown, this embodiment provides a method for identifying protein-protein interaction sites, proposing the GHGPR-PPIS graph network model. This model is based on a graph convolutional network with a graph heat kernel, and also integrates Generalized PageRank (GPR) technology and edge self-attention feature processing blocks. It fully utilizes the hidden information in the protein graph to accurately identify protein-protein interaction sites, specifically including the following steps:

[0052] Step 1: Perform parameter tuning and performance testing on the proposed model using the same dataset as AGAT-PPIS. The AGAT-PPIS dataset is derived from the GraphPPIS dataset. The GraphPPIS dataset (Table 1) includes the following subsets: the Train_335 training dataset, and three test datasets named Test_315, Test_60, and Ubtest_31.

[0053] Table 1 Statistical analysis of the GraphPPIS dataset

[0054]

[0055] The AGAT-PPIS dataset (Table 2) was obtained by removing protein chains from the GraphPPIS dataset that did not correspond to the corresponding structures. It consists of four subsets: the Train_335-1 training set (one chain removed from Train_335) and three test datasets: Test_315-28 (28 chains removed from Test_315), Test_60-0 (same as Test_60), and Ubtest_31-6 (6 chains removed from Ubtest_31). The relatively small ratio of positive samples (interacting sites) to negative samples (non-interacting sites) in the training and test datasets represents an imbalance in the number of positive and negative samples. Throughout the work, Train_335-1 and Test_60-0 were used as benchmark datasets for model training and testing, while the remaining two test sets were used to evaluate the generalization performance of the models.

[0056] Table 2 Statistical analysis of the AGAT-PPIS dataset

[0057]

[0058] Step 2: Sequence feature representations of the protein data were obtained from PSSM, Hidden Markov Model (HMM) matrices, and atomic features (AF) of each amino acid residue. PSSM and HMM were generated using PSI-BLAST and HHblits tools, respectively, producing N*20 dimensional sequence feature matrices, where N represents the length of the protein chain. These matrices were then normalized. The atomic feature of each amino acid residue consisted of seven dimensions: atomic mass, B factor, whether it was a side chain atom, charge, number of hydrogen atoms attached to its bond, whether it was part of a ring, and the van der Waals radius of the atom.

[0059] Step 3: The structural feature representation of the protein data consists of a 14-dimensional matrix DSSP and pseudo-position embedding features of amino acid residues. The first nine dimensions represent the secondary structure states of the protein chain (H: α-helix, G: 310 helix, I: π-helix, E: extended β-chain, B: separated β-chain, T: rotation, S: bending, C: others, and unknown protein secondary structures). The next four dimensions are obtained by transforming the peptide backbone to the torsion angles PHI and PSI using sine and cosine functions. The last dimension is generated by converting the solvent-accessible surface area to relative solvent accessibility. Finally, the pseudo-position embedding feature (PPEF) of amino acid residues represents the relative position information of each amino acid residue relative to a reference amino acid residue.

[0060] For the edge feature representation of the protein chain, two features were selected: (1) the Euclidean distance between two nodes; and (2) the cosine of the angle between the vector formed by the two nodes and the first amino acid residue of the protein chain.

[0061] Step 4: Construct the protein chain diagram. First, calculate the coordinates of the side-chain centroids (SCs) of all amino acid residues from the protein's tertiary structure file (e.g., .PDB format).

[0062] Subsequently, the Euclidean distances between all pairs of SCs were calculated based on the amino acid residues on the protein chain. This step allows us to generate a distance matrix for the protein chain that captures the pairwise distances between all amino acid residues. To convert the distance matrix into an adjacency matrix, we set a distance threshold to... If the distance value is less than or equal to a threshold, it indicates that the two residues are very close. In this case, an edge is established between them in the graph representation of the protein chain. The corresponding position in the adjacency matrix is ​​set to 1. Conversely, if the distance value exceeds the threshold, the corresponding position is assigned a value of 0, indicating that there is no edge. Finally, the adjacency matrix is ​​used as the graph representation of the protein chain.

[0063] Step 5: Sequence feature representation and structural feature representation of the protein chain are collectively referred to as the node feature information of the protein chain. The node feature information and edge feature representation of the protein are respectively passed through a fully connected layer to obtain the embedded representations of the protein node features and edge features. The obtained embedded representations are used as the initial input to the network for training.

[0064] Step 6: In existing GNN-based PPIS prediction methods, graph convolution is usually defined as a weighted average of neighboring nodes, and performance is improved by learning graph convolution kernels to utilize smoothness. However, previous studies have lacked an effective graph convolution method to capture the inherent smoothness in the network structure. This application adopts a GraphHeat model that uses a hot kernel to enhance the low-frequency filter. This model gives greater importance to the low-frequency filter and significantly reduces the impact of high-frequency signal changes on the graph. By utilizing the smoothness of labels or features applied to nodes by the graph structure, GraphHeat has achieved significant success. The hot kernel is defined in Equation (1) as follows:

[0065]

[0066] In the formula, The eigenvalues ​​u of the normalized Laplacian matrix L are described. i The smoothness. Where L = ID -1 / 2 AD -1 / 2And I represents the identity matrix, s is the scaling parameter, A is the protein chain diagram, i.e., the adjacency matrix, and D is the diagonal matrix of A.

[0067] Use heat kernels to analyze the input protein node features Processing is performed to obtain new node features. As shown in formula (2):

[0068]

[0069] In the formula, s is the scaling parameter (s≥0), which is set to 2 here. and It is a learnable parameter matrix.

[0070] Step 7: After obtaining the node features processed by hot kernel convolution, the edge features and node features are used as input to the ESGRET block to further enhance the node features. The ESGRET block introduces the importance of edges and considers the different effects of neighboring nodes with different dependencies on the target node built on the Edge Aggregated Graph Attention Networks and Transfer Learning (EGRET). Figure 3 , 4 As shown, the specific process is that the ESGRET block introduces an edge self-attention coefficient g within the EGRET block. ij (As shown in Equation 3), it is integrated into the ESGRET block (as shown in Equation 5). Subsequently, the self-attention coefficient δ is calculated according to Equation 6. ij This completes the update of node features as shown in Formula 7.

[0071]

[0072]

[0073] β ij =σ(e ij +g ij (5)

[0074]

[0075]

[0076] In the formula, i and j represent nodes i and j in the protein chain diagram, respectively. b is the bias. Represents edge features. β ij and e ij Let W represent the weight of node j relative to node i and the attention coefficient between the nodes, respectively. m Wa W v W ε and W ρ It is a learnable parameter matrix, N i Let β represent the set of neighboring nodes of node i. ij The meaning and β ik same.

[0077] Table 3 shows the improved performance of the model on the Test_60 test set after adding edge self-attention coefficients. Table 3 verifies the importance of adding a self-attention mechanism to EGRET for performance improvement. Therefore, it is undeniable that our proposed ESGRET block helps improve the model's performance.

[0078] Table 3 Performance comparison between the improved edge module and the initial edge processing module

[0079]

[0080] Step 8: Linearly combine the updated protein node representations with the original protein node features and feed them into the normalization layer as input for the next layer. Table 4 shows the improvement in model performance after adding the original protein node features to the normalization layer. After updating the node features in each layer of the model, subsequent steps involve normalizing these updated features to ensure the reliability of the model results. Before normalizing the updated node features, the model performance is enhanced by merging them with the original node features. Then, the merged features are normalized using the ContraNorm normalization function. Experiments show that directly using ContraNorm to normalize the updated node features results in a significant reduction in all four key performance metrics of the model. Specifically, ACC, F1, MCC, and AUPRC decrease from initial values ​​of 0.843, 0.545, 0.454, and 0.552 to updated values ​​of 0.814, 0.534, 0.439, and 0.513, respectively. However, these metrics are significantly improved after introducing the original node features into the normalization process. Notably, the ACC, F1, MCC, and AUPRC metrics showed commendable increases of 0.046 (5.65%), 0.049 (9.18%), 0.062 (14.12%), and 0.083 (16.18%), respectively. These improvements were even more significant compared to metrics without ContraNorm application. Specifically, these metrics showed enhancements of 0.017 (2.02%), 0.038 (6.97%), 0.047 (10.35%), and 0.044 (7.97%). Based on these compelling findings, there is no doubt that the fusion of original node features with updated node features plays a crucial role in maximizing the integrity of key node information. This fusion not only protects against the loss of important details but also significantly improves the overall performance of the model.

[0081] Table 4 shows that the normalization of original node features helps improve model performance.

[0082]

[0083] Step 9: The difference between graph convolution and general convolution is that when multiple layers of graph convolution are stacked, it tends to homogenize node features. In other words, deep graph convolution often produces similar node features, which may lead to a decrease in model performance as the network becomes deeper. On the other hand, shallow graph convolution cannot effectively propagate node information. To address these challenges, this application introduces the GPR-GNN architecture, which supports N-Step propagation based on graph topology, thereby producing the propagation result H at each step. (1) →H (N)The GPR-GNN structure transforms the original node features into embedded features H through a linear layer transformation. (0) As input, GPR represents a linear combination of the propagation results at each step, and the weights of this linear combination are called GPR weights. The weights are learned by binding parameters, and the weight calculation process for each step is shown in Equation 8.

[0084] γ n =α(1-α) n ,γ N =(1-α) N (8)

[0085] Where, γ n Let α represent the GPR weights for the nth propagation step, α be a learnable parameter, and N be the final number of steps.

[0086] Ultimately, the output characteristic of GPR is a weighted combination of the propagation results at each step, where H (n) Z is the network output at step n, and Z is the final network output, as shown in Formula 9.

[0087]

[0088] Based on Equations 8 and 9, it is clear that if the result of the nth propagation has a small impact on model training, the corresponding GPR weights will decrease during gradient update. As the weights decrease, the impact of this step on model training diminishes, effectively avoiding the problem of excessive model smoothing.

[0089] While the experiments based on the 5-layer GraphHeat-GPR architecture yielded commendable performance, it is important to verify that model depth is a key factor influencing its overall performance. Considering this, and given the prevalence of 3- to 4-layer GNN models, a thorough evaluation of GraphHeat-GPR architectures spanning 2 to 6 layers was conducted. The results (Table 5) illustrate the model's performance at different depths. Notably, the ACC value obtained with a 5-layer architecture is slightly lower than that obtained using a 6-layer configuration (a negligible difference of 0.03). However, the model peaks on the three key evaluation metrics—F1, MCC, and AUPRC—on imbalanced datasets, achieving best values ​​of 0.583, 0.501, and 0.596, respectively. Finally, based on these comprehensive evaluations, the 5-layer GraphHeat-GPR architecture was chosen to construct the GHGPR-PPIS model, specifically designed for predicting protein-protein interactions.

[0090] Table 5. Impact of the 2-6 layer GraphHeat-GPR architecture on model performance

[0091]

[0092] Step 10: Compare the true labels with the predicted results and minimize the cross-entropy loss function. The calculation process of the loss function is shown in Equation 10.

[0093]

[0094] In the formula, n represents the number of samples, a i and b i These represent the actual label and the predicted label, respectively.

[0095] The GHGPR-PPIS model utilizes graph-heated kernel networks and GPR technology. By leveraging the node information aggregation capabilities of these technologies, it prevents extracted node features from becoming overly smoothed as the model depth increases, which could lead to performance degradation. Furthermore, it incorporates a self-attention mechanism to enhance edge feature processing. Interpretability analysis demonstrates that GHGPR-PPIS can effectively extract key information affecting protein-protein interaction site prediction from protein sequence and structural data.

[0096] As shown in Table 6, compared with current sequence-based or structure-based models, the GHGPR-PPIS model significantly improves the prediction performance of protein-protein interaction sites. Among them, the three key indicators used to evaluate the performance of the model based on the unbalanced dataset, F1, MCC and AUPRC, are 0.583, 0.501 and 0.596, respectively, all of which are higher than the existing models.

[0097] Table 6 compares the performance of the models on the Test_60 test set.

[0098]

[0099] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0100] One embodiment also provides a protein interaction site identification device, including a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the protein interaction site identification method as described in the above embodiments.

[0101] In one embodiment, a computer-readable medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the protein interaction site identification method as described in the above embodiments.

[0102] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0105] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A method for identifying protein-protein interaction sites, characterized in that, The method comprises the following steps: obtaining a protein-protein interaction dataset, dividing the dataset into a training set and a test set and preprocessing the dataset; obtaining sequence feature representations, structure feature representations and edge feature representations of protein chains in the training set and the test set, and taking the sequence feature representations and the structure feature representations of the protein chains as node feature information of the protein chains; inputting the training set into a preset GHGPR-PPIS model, wherein the GHGPR-PPIS model comprises five GraphHeat-GPR modules connected in sequence, a linear layer, a fully connected layer and a softmax, the GraphHeat-GPR module is configured to obtain low-frequency signals between nodes of the protein chains by adopting heat kernel convolution processing and update node features to obtain first node features, send the first node features and edge features to an edge self-attention module to update attention coefficients between the nodes to obtain second node features after updating, and perform linear combination on the second node features and initial node features and send the linear combination result to a normalization layer as input of a next GraphHeat-GPR module, the linear layer is configured to perform weighted combination on output results of each GraphHeat-GPR module to obtain a final result and send the final result to the fully connected layer, and the fully connected layer and the softmax are configured to predict a result; comparing a real label with the predicted result to update network parameters to minimize a cross-entropy loss function until a maximum number of iterations is reached, and the training is completed to obtain a trained GHGPR-PPIS model; inputting the test set into the trained GHGPR-PPIS model to obtain protein-protein interaction sites, wherein the heat kernel convolution processing is configured to obtain low-frequency signals between nodes of the protein chains and update node features, and the heat kernel convolution processing comprises the following steps: The definition of the hot core is as follows: (1) wherein A normalized Laplacian matrix is described Eigenvalues of the smoothness, and denotes the identity matrix, s is a scaling parameter, A is a protein chain graph, i.e. an adjacency matrix, and D is the diagonal matrix of A. Using hot kernel on node features of input protein chain Performing processing to obtain new node features As shown in equation (2): (2) wherein is a scaling parameter , and are learnable parameter matrices.

2. The method for identifying protein interaction sites according to claim 1, characterized in that, The preprocessing comprises removing protein chains in the dataset that do not correspond to corresponding structures to process the dataset into a new dataset.

3. The method for identifying protein interaction sites according to claim 1, characterized in that, The sequence features of the protein chain are obtained from a PSSM, a hidden Markov model matrix, and atomic features of each amino acid residue, wherein the PSSM and the hidden Markov model matrix are generated using PSI-BLAST and HHblits tools, respectively, resulting in a sequence feature matrix of the protein chain and normalizing it, wherein N represents the length of the protein chain; and the atomic features of the amino acid residue include atomic mass, B-factor, whether it is a side chain atom, charge, number of hydrogen atoms bonded to it, whether it is part of a ring, and van der Waals radius of the atom.

4. The method for identifying protein interaction sites according to claim 1, characterized in that, The structure feature representation of the protein chain is composed of a 14-dimensional matrix DSSP and pseudo-position embedding features of amino acid residues, the 14-dimensional matrix DSSP comprises nine dimensions representing secondary structure states of the protein chain, four dimensions being twist angles PHI and PSI of a peptide skeleton by using sine and cosine function transformation, and one dimension being relative solvent accessibility by converting solvent accessible surface area; and the pseudo-position embedding features of the amino acid residues are relative position information of each amino acid residue relative to a reference amino acid residue.

5. The method for identifying protein interaction sites according to claim 1, characterized in that, The edge feature representation of the protein chain comprises a Euclidean distance between two nodes and a cosine value of an included angle between a vector formed by the two nodes and a first amino acid residue of the protein chain.

6. The method for identifying protein interaction sites according to claim 1, characterized in that, The updated node features and the edge features are sent to the edge self-attention module to update attention coefficients between the nodes to obtain updated node features, and the sending comprises the following steps: Introducing edge self-attention coefficients The updated node features and edge features are sent to an edge self-attention module to obtain self-attention coefficients between nodes As follows: (3) (4) (5) (6) In the formula, and These represent nodes in the protein chain diagram. and nodes , It's a bias. Representing edge features, and Representing nodes respectively Relative to node The weights and attention coefficients between nodes, , and It is a learnable parameter matrix. Represents a node The set of adjacent nodes; Based on the self-attention coefficients, obtain the updated node features As follows: (7) wherein, and are learnable parameter matrices.

7. The method for identifying protein interaction sites according to claim 1, characterized in that, The output results of each GraphHeat-GPR module are weighted combined, and the weighted combination comprises the following steps: The GPR weight is calculated as follows: (8) wherein, GPR weights representing the n-th propagation step, are learnable parameters, is the final number of steps, The output results of each layer of the GraphHeat-GPR module are combined by weighting to obtain a final output feature As shown below: (9) wherein is the network output of the nth step.

8. A protein interaction site recognition apparatus, characterized by, The computer readable medium stores a computer program, and the computer program is executed by the processor to implement the method in any one of claims 1 to 7.

9. A computer readable medium characterized by The computer readable medium stores a computer program, and the computer program is executed by the processor to implement the method in any one of claims 1 to 7.

Images