Protein interaction site prediction method based on graph neural network

Abstract

The invention belongs to the technical field of biological information, and particularly relates to a protein interaction site prediction method based on a graph neural network. The invention provides a novel protein-protein interaction site prediction model GTE-PPIS, and according to the prediction model, an EGNN module and a Graph transfer module are used for jointly extracting features. And finally, the node features are fed to a multi-layer perceptron to obtain a prediction result, and residual connection is introduced, so that the problem of excessive smoothness is relieved, and the model can more fully learn node representation. The result shows that on a benchmark test set, the GTE-PPIS is superior to other models in various evaluation indexes, and the effectiveness of the GTE-PPIS model is fully proved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and in particular relates to a method for predicting protein interaction sites based on graph neural networks. Background Art

[0002] As one of the most common molecules in cells, proteins are crucial for regulating various biological processes in organisms. Proteins rarely act alone, but often interact with other proteins by forming chemical bonds. Protein-protein interactions are an important mode of interaction between biological molecules in cells and are essential for the normal physiological functions of organisms. Protein-protein interactions (PPIs) play an important role in many biological processes such as signal transduction, transport, and metabolism. The amino acid residues that bind during the interaction are protein-protein interaction sites. Predicting PPI sites helps to construct protein-protein interaction networks, predict protein functions, reveal the molecular mechanisms of diseases and the development of new treatments.

[0003] There are two main methods for predicting protein-protein interaction sites that are currently widely used: one is a sequence-based method and the other is a structure-based method. Sequence-based methods only extract feature descriptors from primary protein sequences. The descriptors commonly used in this field can be roughly divided into four categories: sequence information, evolutionary information, physicochemical properties, and predicted structural features. Many sequence-based protein-protein interaction site prediction methods are based on traditional machine learning methods to process the above sequence features. Commonly used machine learning techniques include support vector machines, random forests or other tree-based methods, shallow neural networks, and simple regression algorithms. In response to the problem of unbalanced data sets in protein interaction prediction, Wang et al. proposed methods such as the downsampling method based on the nearest neighbor rule (ENNS), and used support vector machines to establish a predictor to infer interface residues. Sanchez-Garcia et al. proposed the BIPSPI method, which used xgboost to predict protein-protein interaction sites. Qiu et al. introduced the concept of natural language processing and identified protein binding sites through shallow neural networks. Murakami et al. used Bayes classifier (NBC) and kernel density estimation (KDE) are used to predict residues in protein sequences that bind to other proteins. Recently, deep learning has been used in this field and has provided good prediction results. For example, methods such as CNN and LSTM are used as effective classifiers for PPI sites. Wang et al. combined CNN with feature selection rotation forest (FSRF) and proposed a new method for PPI site prediction based on protein sequence, namely CNN-FSRF. Tsukiyama et al. first used word2vec to develop an LSTM model to predict PPI sites between humans and viruses using amino acid sequences alone, named LSTM-PHV. In addition, neural network-based architectures have been implemented to process local and global sequence contexts. For example, Kang et al. proposed a PPI site prediction hybrid network HN-PPISP that integrates CNN, MLP-Mixer and LSTM models, using a complex network architecture to extract local and global features of proteins to achieve better prediction results.

[0004] Compared with sequence-based methods, structure-based methods can obtain the three-dimensional structure of proteins, thereby extracting very informative geometric and structural features, and can often infer PPI sites more accurately. The GraphPPIS model proposed by Yuan et al. is the first method to use deep graph convolutional networks to predict PPI sites. Zhou et al. proposed an enhanced version of the graph attention network AGAT, which added edge features and provided more structural information. Zhong et al. used the GCN module and the GraphSage module to extract global and local features respectively, and jointly predict PPI sites. At present, the use of graph neural networks to predict PPI sites is in its infancy and has great research potential.

[0005] Since experimental-based protein-protein interaction (PPI) site prediction methods such as two-hybrid experiments and affinity purification are time-consuming and labor-intensive, it is necessary to develop high-precision computational methods as auxiliary tools for routine experiments to predict protein-protein interaction (PPI) sites. Summary of the Invention

[0006] The present invention proposes a protein interaction site prediction method based on graph neural network, which solves the problem of time-consuming and labor-intensive protein-protein interaction site prediction, improves the prediction ability of protein-protein interaction sites, and makes the prediction more accurate.

[0007] The technical solution of the present invention is achieved as follows:

[0008] A graph neural network-based protein interaction site prediction model, called GTE-PPIS, consists of the following parts:

[0009] Graph construction: proteins are represented as graph data structures using nodes and edges. Nodes represent amino acid residues, and edges represent the positional relationships between amino acids. The protein graph data structure will serve as subsequent input.

[0010] The feature extraction module includes the EGNN module and the Graph Transformer module. These two modules are parallel and extract features from the protein graph data structure simultaneously without interfering with each other. The Graph Transformer module focuses on the global structure and long-range dependencies of the graph, while the EGNN module captures the fine 3D geometry and local characteristics. The results of the two modules are fused.

[0011] The classification module performs final processing on the results of the fusion of the two modules through a multi-layer perceptron and outputs each node, that is, whether the amino acid residue is a protein interaction site.

[0012] Through the above technical solution, the GTE-PPIS model innovatively uses the EGNN module (equivariant graph neural network) and the Graph transformer module (graph transformer (GT)). The two modules work together to extract protein features, improving the model's ability to extract protein information.

[0013] A method for predicting protein interaction sites based on graph neural networks, comprising the following steps:

[0014] S1. Extract node features and edge features from proteins and represent proteins as graph structures, where nodes represent amino acid residues and edges represent positional relationships between amino acids.

[0015] S2. Input the protein graph structure into the GTE-PPIS model. The GTE-PPIS model extracts protein features through the Graphtransformer module and the EGNN module respectively, and then fuses the output results of the two modules.

[0016] S3. The multi-layer perceptron outputs the fusion of the Graph transformer module and the EGNN module as the final result, that is, whether each amino acid residue in the protein is an interaction site.

[0017] Using the above technical solution, step S2 utilizes a graph transformer module with a self-attention mechanism to focus on the global structure and long-range dependencies of the graph. Simultaneously, it is combined with an EGNN module to capture detailed three-dimensional geometry and local characteristics. This dual module collaboratively extracts protein features, resulting in enhanced feature extraction capabilities compared to existing technologies. Step S3 fuses the protein features extracted by the two modules, using a multilayer perceptron to output the final prediction: whether each amino acid residue is a PPI site.

[0018] Optionally, in step S2, the equivariant graph convolutional layer achieves equivariance by replacing the standard message passing mechanism with equivariant message passing and introducing a coordinate update mechanism in the graph neural network;

[0019]

[0020] in, and represents the feature embedding of nodes i and j at layer l, Represents coordinate x i and x j The square of the distance between ij represents edge features, and are multilayer perceptrons for edge and coordinate operations, Indicates the information transmission between nodes, Represent the coordinates of node i in layers l and l+1 respectively, represents the message after node i is aggregated at layer l, C is a normalization constant, C = 1 / (N v ―1), where N v is the number of nodes, Represents the displacement vector between node i and node j in layer l.

[0021] Through this technical solution, the EGNN module explicitly models the coordinates of nodes in the protein's three-dimensional structure, ensuring the model's equivariance to geometric transformations such as rotation and translation. This is difficult to achieve in traditional graph neural networks, as they typically ignore coordinate information or rely only on relative distances.

[0022] Optionally, in step S3, the node feature of the i-th node and edge features between nodes (i, j) First, two linear layers are initialized to d dimensions. and

[0023]

[0024]

[0025] in, and The node and edge embeddings are linearly transformed and then input into the GT layer, where they are stacked multiple times to generate the final embeddings of the nodes and edges. is the i-th node feature after initialization, is the edge feature between nodes i and j after initialization, and Respectively represent the feature extraction of node features and edge features by weight matrix, is the learnable bias vector;

[0026] The l-th layer graph transformer (the l-th layer GT layer) updates the node and edge embeddings through the message passing mechanism and the improved multi-head self-attention mechanism;

[0027]

[0028] in, represents the feature embedding of node i at layer l, and are the learnable parameters of the linear layer, ∈R 2d×d Indicates that its shape is 2k×d, k∈1,...,H represents the number of attention heads, d k is the dimension of each head, j∈N(i) represents the neighboring nodes of node i, Indicates that the information on the edge connecting node i and adjacent node j is aggregated by summing, Softmax j∈N(i) Represents the SoftMax operation on the adjacent node j, Norm represents layer normalization, Concat represents the connection operation, SiLU represents a class of activation function, represents the normalized attention weight of node j to node i in the lth layer and the kth attention head, represents the key vector of node j, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the final generated node features and edge features, The query vector of node i, The key vector of node j, represents the characteristics of edge (i,j), Represents the edge features between (i, j) in the lth layer, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the key vector of node j.

[0029] Through the above technical solution, the Graph Transformer module captures the interactions between long-range amino acids in proteins, such as the associations between different structural domains, through the self-attention mechanism, overcoming the "over-smoothing" problem caused by the limitation of the number of message passing layers in traditional graph neural networks.

[0030] Optionally, in step S4, the probability of protein interaction site prediction is defined as follows:

[0031] Y = Softmax(WH+b);

[0032] Where H is the final node representation of GTE-PPIS, W and b are the weight and bias of the MLP output layer, respectively, and Y∈R n×2 , where n represents the number of residues and Y has two columns, representing the probability of protein interaction sites and the probability of non-protein interaction sites, respectively.

[0033] Through the above technical solution, the above formula can achieve binary classification probability output, map the node-level (amino acid) features extracted by the model to the probability distribution of the PPI site, and obtain the result of whether it is a PPI site.

[0034] Optionally, in step S1, the node features include PSSM (20 dimensions), HMM (20 dimensions), DSSP (14 dimensions), and AF (7 dimensions), and these four features have a total of 61 dimensions. The edge features include Euclidean distance and angle cosine value, and these two features have a total of 2 dimensions.

[0035] Through the above technical solution, in step S1, four node features (covering evolutionary information, structural data, and atomic properties) and two edge features (Euclidean distance and cosine value) are used to comprehensively represent proteins, allowing subsequent neural networks to learn with sufficient protein information. Compared with existing technologies, the representation is more comprehensive.

[0036] Optionally, in step S1, the protein interaction site prediction problem is regarded as a graph node classification problem, and each protein is represented as an undirected graph G = (V, E). Represents N v Node features of amino acid residues.

[0037] Through the above technical solution, node features and undirected graphs are necessary inputs for graph neural networks.

[0038] After adopting the above technical solution, the beneficial effects of the present invention are:

[0039] The present invention designs a novel structure-based PPI site prediction model GTE-PPIS. The present invention deploys four different features to construct the initial node representation, covering evolutionary information, structural data, and atomic properties. A graph transformer module with a self-attention mechanism is used to focus on the global structure and long-range dependencies of the graph. At the same time, combined with the EGNN module, it captures fine three-dimensional geometric structures and local characteristics. In addition, residual connections are introduced to alleviate the over-smoothing problem, enabling the model to more fully learn node representations. Experimental results show that the performance of the model of the present invention is better than other advanced models. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 1 is a diagram of the overall architecture of GTE-PPIS in the embodiment;

[0042] Figure 2 are the AUROC (A) and AUPRC (B) values ​​of the model at each fold in the 5-fold cross validation of the validation dataset and Test_60 in the embodiment;

[0043] Figure 3 This is an example of GTE-PPIS and three other models predicting PPI sites in the embodiment (PDB ID: 6LKI, Chain A). DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] The embodiments of the present application disclose a method for predicting protein interaction sites based on graph neural networks.

[0046] Example

[0047] according to Figures 1 to 3 As shown, a protein interaction site prediction method based on graph neural network includes the following contents.

[0048] 1. Materials and Methods

[0049] 1.1 Dataset

[0050] The dataset used in this experiment is the dataset used by the AGAT-PPIS Institute, which consists of a training dataset (Train_335) and a test dataset (Test_60, Test_315, and UBtest_31). The above-mentioned Train_335 and Test_60 are from the widely used public datasets Dset_186, Dset_72, and Dset_164, respectively. After 6 steps of filtering, redundant proteins with high sequence similarity or overlap are removed using BLASTClust. Test_315 is another test dataset proposed by Yuan et al., with the aim of further verifying the generalization ability of the model. UBtest_31 contains 31 unbound protein structures, corresponding to the 31 proteins with known monomeric structures in Test_60 (Btest_31), which are used to evaluate the robustness of the model and the impact of conformational changes on model performance. In previous experiments, it was found that a few protein sequences in these data sets were inconsistent with the sequences in the corresponding protein PDB files on the PDB website. Therefore, these abnormal proteins were excluded from this experiment, and the training dataset (Train_334) and the test dataset (Test_60, Test_315-28, and UBtest_25) were finally obtained. The specific statistical details of the above datasets are shown in Table 1.

[0051] Table 1 Dataset information

[0052]

[0053] 1.2 Protein expression

[0054] The present invention regards the PPI site prediction problem as a graph node classification problem and represents each protein as an undirected graph G = (V, E). Represents N v Node features of amino acid residues. Represents the edge features between pairs of amino acids. Based on the sequence and structural features, the protein feature matrix is ​​obtained and the protein edge feature matrix where d v Indicates the characteristic dimension of each node (d v =61), d e Indicates the characteristic dimension of each edge (d e =2), n represents the length of the protein chain. The features used are described below.

[0055] PSSM,PSSM is generated by using PSI-BLAST to perform multiple sequence alignment on the sequence, reflecting the frequency and correlation of different amino acids at each position. When using PSI-BLAST, the e value is set to 0.001 and the number of iterations is set to 3. Then the values ​​of the PSSM matrix are normalized to a score between 0 and 1 using formula (1), where x represents the input feature value and x min and x max are the minimum and maximum values ​​of a column of the input feature matrix respectively.

[0056]

[0057] HMM. The HMM profile was generated by aligning the HHblits v3.0.3 algorithm with the UniClust30 database using default parameters and then normalized.

[0058] To obtain residue secondary structure information, the DSSP algorithm is used to calculate protein sequences. First, a 9-dimensional one-hot encoded feature vector is extracted from the secondary structure state. Second, a sine-cosine transform is performed on the two torsion angles PHI and PSI of the polypeptide backbone to obtain a 4-dimensional feature vector. Third, the solvent-accessible surface area is converted to relative solvent accessibility, resulting in a 1-dimensional feature vector. Ultimately, the DSSP dimension is 14.

[0059] AF, the present invention adds the features of each atom (excluding hydrogen atoms) that constitutes the residue to the node feature matrix. Seven atomic features are extracted from the protein PDB file: whether it is located in a ring, the van der Waals radius of the atom, the B factor, whether it is a side chain atom of the residue, the number of hydrogen atoms bonded to it, the electron charge, and the atomic mass. Since the number of atoms constituting each residue is inconsistent, the present invention takes the average of these seven atomic features as the final atomic feature, as shown in formula (2), where Represents the feature i of residue atom j, N a represents the number of residue atoms, {x i} i＝1,...,7 Represents the atomic feature i of the residue. After processing, the seven-dimensional atomic features of each residue are obtained.

[0060]

[0061] For protein sequences, the present invention calculates the Euclidean distance between each residue pair and normalizes it. In order to fix the relative position relationship between nodes, the present invention also calculates and Angle θ ijThe cosine value of is shown in formula (3), where · represents the dot product of the vector. The relative position relationship of the residues can be represented by the distance information and the cosine value information, which increases the translation and rotation invariance of the model.

[0062]

[0063] 1.3 GTE-PPIS Architecture

[0064] Figure 1 Figure 1 shows the overall architecture of the GTE-PPIS model. (A) shows a schematic overview of GTE-PPIS. (B) shows the equivariant graph neural network. (C) shows the architecture of the first multi-head attention layer in the graph transformer module (GT). Figure 1 In the figure, (A) shows the GTE-PPIS model, which includes graph construction, feature extraction, and classification. In the feature extraction part, a dual-module structure is used to extract features. The EGNN module (equivariant graph neural network) uses the features of the edges as conditional inputs to adjust the information fusion mechanism in the message passing process, thereby enhancing the representation ability of graph data. (B) shows the equivariant graph neural network. The Graph transformer module combines the advantages of the graph neural network and the transformer model. It can not only process the local structure in the graph data, but also effectively capture global dependencies through the self-attention mechanism. (C) shows the architecture of the first multi-head attention layer in the graph transformer module. Then, the features extracted by the two modules are fused to obtain the final node representation. Finally, a multi-layer perceptron is used to predict whether the amino acid residue is a PPI site.

[0065] 1.3.1EGNN Module

[0066] EGNN is composed of a series of stacked equivariant graph convolutional layers (EGCLs). It uses edge information and the coordinates and node embeddings of the previous layer to perform a series of transformations on the input to update the coordinates and node embeddings. EGCL achieves equivariance by replacing the standard message passing mechanism with equivariant message passing and introducing a coordinate update mechanism in graph neural networks.

[0067]

[0068] in, and represents the feature embedding of nodes i and j at layer l, Represents coordinate x i and x j The square of the distance between ij Represents edge features. and Multilayer Perceptrons for edge and coordinate operations, respectively. Represents the information transfer between nodes. The coordinate update of node i is obtained by weighted summation of the differences in the previous layer’s coordinate embeddings, and a normalization factor C = 1 / (N v ―1) Normalize, where N v is the number of nodes in the graph. The weight of the weighted sum is generated by a coordinate-operated MLP (Multi-layer Perceptron) applied to equivariant message passing on each edge (i, j)

[0069] 1.3.2 Graph transformer module

[0070] The Graph transformer module can effectively capture the global dependencies in the overall protein structure through the self-attention mechanism, especially the interactions between different domains and long-range amino acids, thereby helping the model understand which regions may be interaction sites. and edge features between nodes (i, j) First, two linear layers are initialized to d dimensions. and

[0071] in, and The node and edge embeddings are linearly transformed and then input into the GT layer, where they are stacked multiple times to generate the final embeddings of the nodes and edges. is the i-th node feature after initialization, is the edge feature between nodes i and j after initialization, and Respectively represent the feature extraction of node features and edge features by weight matrix, is the learnable bias vector.

[0072] The lth layer of graph transformer (the lth layer of GT, the graph transformer module includes 4 GT layers in total) updates the node and edge embeddings through the message passing mechanism and the improved multi-head self-attention (MHA) mechanism.

[0073]

[0074] in, represents the feature embedding of node i at layer l, and are the learnable parameters of the linear layer, ∈R 2d×dIndicates that its shape is 2d×d, k∈1,…,H represents the number of attention heads, d k is the dimension of each head, j∈N(i) represents the neighboring nodes of node i, Indicates that the information on the edge connecting node i and adjacent node j is aggregated by summing, Softmax j∈N(i) Represents the SoftMax operation on the adjacent node j, Norm represents layer normalization, Concat represents the connection operation, SiLU represents a class of activation function, represents the normalized attention weight of node j to node i in the lth layer and the kth attention head, represents the key vector of node j, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the final generated node features and edge features, The query vector of node i, The key vector of node j, represents the characteristics of edge (i,j), Represents the edge features between (i, j) in the lth layer, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the key vector of node j.

[0075] 1.3.3 Multilayer Perceptron

[0076] The present invention uses a multi-layer perceptron to convert the final node representation into the probability of PPI site prediction. The multi-layer perceptron (MLP) is defined as follows.

[0077] Y=Softmax(WH+b) (14)

[0078] Where H is the final node representation of GTE-PPIS, W and b are the weight and bias of the MLP output layer respectively. n×2 , where n represents the number of residues and Y has two columns, representing the probability of PPI sites and the probability of non-PPI sites, respectively.

[0079] 2. Results

[0080] 2.1 Implementation Details

[0081] The present invention uses five-fold cross validation to evaluate the performance of the GTE-PPIS model on the Train_334 dataset. In each fold, 20% of the training data is used as an independent validation set to select the best performing model. Finally, these five models (referring to the models obtained in each fold of the five-fold cross validation) are evaluated on the test set respectively, and their average performance is used as the final result. After experimental analysis, the EGC layer of the EGNN module is set to 10 layers, and the hidden layer embedding size is set to 256. The Graph transformer module uses 4 GT layers, the number of attention heads is 4, and layer normalization is used to stabilize the training process. Dropout (random deactivation rate) is set to 0.1, and the batch size is set to 1. The present invention adopts a learning rate decay strategy, with an initial learning rate of 0.001 and a minimum learning rate of 10 ―6 , trained for 50 epochs using the Adam optimizer.

[0082] 2.2 Evaluation Metrics

[0083] The present invention uses seven statistical indicators, including accuracy, precision, recall, F1-score (F1), area under the precision-recall curve (AUPRC), area under the receiver operating characteristic curve (AUROC) and Matthews correlation coefficient (MCC), to evaluate the model performance.

[0084]

[0085]

[0086] Among them, TP and TN represent the number of correctly predicted positive and negative samples, and FP and FN represent the number of incorrectly predicted positive and negative samples. Accuracy refers to the proportion of correctly identified samples among all samples. Precision represents the probability that among all samples predicted to be positive, they are actually positive. Recall describes the proportion of correctly identified samples among all positive samples. F1 is the harmonic mean of precision and recall. MCC takes TP, TN, FP and FN into consideration and is a more balanced evaluation metric. Even in unbalanced datasets, MCC can still be used as an effective evaluation metric. AUROC and AUPRC are evaluation metrics that do not depend on specific thresholds and can provide a measure of the overall performance of the method.

[0087] 2.3 Performance comparison with other models

[0088] We first compared the performance of GTE-PPIS with nine other state-of-the-art PPI site prediction models on the independent test set Test_60, including ScanNet, DELPHI, HN-PPISP, DeepPPISP, MaSIF-site, EDLMPPI, GraphPPIS, RGCNPPIS, and AGAT-PPIS. As shown in Table 2, GTE-PPIS achieved superior performance across all metrics. GTE-PPIS achieved 1.3%, 2.2%, 3.3%, and 6.4% higher recall, F1-score, MCC, and AUPRC, respectively, than the next best method.

[0089] To further evaluate the model's generalization and robustness, we compared GTE-PPIS with other predictors on the test sets Test_315-28 and UBtest_25. The results are shown in Table 3. As can be seen, GTE-PPIS demonstrated superior MCC and AUPRC on the Test_315-28 and UBtest_25 datasets. On UBtest_25, its AUPRC was slightly lower than that of the predictive model RGCNPPIS, but it demonstrated superior MCC. Overall, GTE-PPIS is currently the model with the best predictive ability.

[0090] Table 2 Performance comparison with existing models on Test_60

[0091]

[0092] Table 3 Performance comparison with other models on Test_315-28 and UBtest_25

[0093]

[0094] 2.4 Model Analysis

[0095] In order to prove the necessity of each module of GTE-PPIS, the present invention conducted module ablation experiments on the validation set and the independent test set Test_60. The present invention created four variants of GTE-PPIS: NoEGNN (excluding the EGNN module and focusing only on the GT module), NoGT (excluding the GT module and focusing only on the EGNN module), NoResidual (excluding residual connections), and NoPrevious (excluding node representations of the previous layer). The present invention uses two indicators AUROC and AUPRC for evaluation. It can be seen from Table 4 that whenever one architecture of the model is eliminated, the predictive ability of the model decreases on the validation dataset and Test_60, which shows the necessity of each module in GTE-PPIS. It can also be seen that the NoEGNN variant of the model has AUROC (-4.4%) and AUPRC (-15.3%) on the validation dataset, and AUROC (-4.1%) and AUPRC (-17.2%) on Test_60. The NoGT variant of the model has AUROC (-8.6%) and AUPRC (-0.2%) on the validation dataset, and AUROC (-8.8%) and AUPRC (-2.0%) on Test_60. The NoResidual variant of the model has AUROC (-2.2%) and AUPRC (-8.4%) on the validation dataset, and AUROC (-12.9%) and AUPRC (-36.5%) on Test_60. The NoPrevious variant of the model has AUROC (-1.8%) and AUPRC (-6.6%) on the validation dataset, and AUROC (-1.8%) and AUPRC (-7.4%) on Test_60. As can be seen from the results, the NoResidual variant of the model has a certain decline in effect on the validation dataset, and its overall performance on Test_60 is the worst, and its predictive ability is greatly affected. This shows that residual connections play a vital role in GTE-PPIS. In contrast, we can also see that pruning node representations from earlier layers has the least impact on the model’s predictive power.

[0096] Table 4 AUROC and AUPRC of GTE-PPIS and its variants on 5-fold cross validation and independent test set Test_60

[0097]

[0098] 2.5 Feature Ablation Analysis

[0099] In the experiment, the present invention used four features of protein: PSSM, HMM, DSSP and AF. Similar to the model analysis part, the present invention performed feature ablation analysis on the validation set and the independent test set Test_60, and the evaluation indicators were AUROC and AUPRC. Each time a feature was removed, the impact of different features on the prediction performance was evaluated. It can be seen from Table 5 that when any feature is removed, the prediction ability of the trained model for the validation dataset and Test_60 will be reduced, which shows that each feature plays a certain role and there is no feature redundancy. In particular, when AF is removed, the AUROC value on the validation dataset is 0.808, and the AUPRC value is 0.481, and the AUROC value on Test_60 is 0.812, and the AUPRC value is 0.482, achieving the lowest value of the model prediction effect. It can be seen that AF has a huge impact on the model learning ability. In contrast, after removing the HMM feature, the AUROC value on the validation dataset is 0.860, the AUPRC value is 0.583, and the AUROC value on Test_60 is 0.869, the AUPRC value is 0.590. It can be seen that the HMM feature has the least impact on the model performance and is less important than other features. The present invention calculates the AUROC and AUPRC values ​​of the model at each fold in the 5-fold cross validation of the validation dataset and Test_60, and draws a box plot, as shown in Figure 2. Figure 2 As shown. Figure 2 As can be seen more intuitively in the figure, for the AUROC and AUPRC indicators, when AF is removed, the model has the lowest mean and outlier values ​​on both the validation dataset and Test_60. When DSSP is removed, the model's AUROC and AUPRC indicators fluctuate the most on both the validation and test sets, indicating that the absence of DSSP amplifies the randomness of model performance.

[0100] Table 5 AUROC and AUPRC of feature ablation on 5-fold cross validation and independent test set Test_60

[0101]

[0102] 2.6 Case Study

[0103] To more intuitively compare the prediction performance of the models, we used a protein from Staphylococcus aureus (PDB ID: 6LKI, Chain A) to visualize the results. This protein contains 294 amino acids and includes 20 sites. Table 6 shows the statistical details of this protein example. GraphPPIS, RGCNPPIS, and AGAT-PPIS correctly predicted 18, 18, and 17 sites, respectively, while GTE-PPIS correctly predicted 19 sites. GTE-PPIS demonstrated the best performance by correctly predicting the most sites. Figure 3The results predicted by GTE-PPIS, GraphPPIS, RGCNPPIS, and AGAT-PPIS are shown and compared with the true values. In the visualization results, blue indicates correctly predicted sites and yellow indicates incorrectly predicted non-sites. Figure 3 From the yellow part, it can be clearly seen that GTE-PPIS has the least false positive sites predicted compared with the other three methods, which is closer to the actual situation.

[0104] Table 6 Statistical details of protein examples (PDB ID: 6LKI, Chain A)

[0105]

[0106] In summary, the present invention proposes a new PPI site prediction model, GTE-PPIS, which uses the EGNN module and the graph transformer module, which extract structural features from spatial neighborhoods. Experiments on an independent test set confirm that the GTE-PPIS model outperforms currently popular methods. Further module and feature ablation studies and case studies confirm that GTE-PPIS is an effective tool for predicting PPI sites. In addition, the structure of GTE-PPIS can also be used to predict interaction sites between proteins and other ligands, such as peptides and drugs, thereby enhancing the versatility of GTE-PPIS.

[0107] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the technical solutions of the present invention should be included in the protection scope of the present invention.

Claims

1. A protein interaction site prediction model based on graph neural network, characterized in that: The name of the model is GTE-PPIS. The GTE-PPIS model consists of the following parts: Graph construction: proteins are represented as graph data structures using nodes and edges. Nodes represent amino acid residues, and edges represent the positional relationships between amino acids. The protein graph data structure will serve as subsequent input. The feature extraction module includes the EGNN module and the Graph Transformer module. These two modules are parallel and extract features from the protein graph data structure simultaneously without interfering with each other. The Graph Transformer module focuses on the global structure and long-range dependencies of the graph, while the EGNN module captures the fine 3D geometry and local characteristics. The results of the two modules are fused. The classification module performs final processing on the results of the fusion of the two modules through a multi-layer perceptron and outputs each node, that is, whether the amino acid residue is a protein interaction site.

2. A protein interaction site prediction method based on graph neural network, characterized in that: Prediction is performed using a protein interaction site prediction model based on a graph neural network as claimed in claim 1, specifically comprising the following steps: S1. Extract node features and edge features from proteins and represent proteins as graph structures, where nodes represent amino acid residues and edges represent positional relationships between amino acids. S2. Input the protein graph structure into the GTE-PPIS model. The GTE-PPIS model extracts protein features through the Graph transformer module and the EGNN module respectively, and then fuses the output results of the two modules. S3. The multi-layer perceptron outputs the fusion of the Graph transformer module and the EGNN module as the final result, that is, whether each amino acid residue in the protein is an interaction site.

3. The method for predicting protein interaction sites based on graph neural network according to claim 1, characterized in that: In step S2, the equivariant graph convolutional layer achieves equivariance by replacing the standard message passing mechanism with equivariant message passing and introducing a coordinate update mechanism in the graph neural network; in, and represents the feature embedding of nodes i and j at layer l, Represents coordinate x i and x j The square of the distance between ij represents edge features, and are multilayer perceptrons for edge and coordinate operations, Indicates the information transmission between nodes, Represent the coordinates of node i in layers l and l+1 respectively, represents the message after node i is aggregated at layer l, C is a normalization constant, C = 1 / (N v ―1), where N v is the number of nodes, Represents the displacement vector between node i and node j in layer l.

4. The method for predicting protein interaction sites based on graph neural network according to claim 1, characterized in that: In step S3, the node feature of the i-th node and edge features between nodes (i, j) First, two linear layers are initialized to d dimensions. and in, and The node and edge embeddings are linearly transformed and then input into the GT layer, where they are stacked multiple times to generate the final embeddings of the nodes and edges. is the i-th node feature after initialization, is the edge feature between nodes i and j after initialization, and Respectively represent the feature extraction of node features and edge features by weight matrix, is the learnable bias vector; The l-th layer graph transformer, i.e., the l-th layer GT layer, updates the node and edge embeddings through a message passing mechanism and an improved multi-head self-attention mechanism; in, represents the feature embedding of node i at layer l, and are the learnable parameters of the linear layer, ∈R 2V×V Indicates that its shape is 2d×V, k∈1,...,H represents the number of attention heads, d k is the dimension of each head, j∈N(i) represents the neighboring nodes of node i, Indicates that the information on the edge connecting node i and adjacent node j is aggregated by summing, Softmax j∈N(i) Represents the SoftMax operation on the adjacent node j, Norm represents layer normalization, Concat represents the connection operation, SiLU represents a class of activation function, represents the normalized attention weight of node j to node i in the lth layer and the kth attention head, represents the key vector of node j, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the final generated node features and edge features, The query vector of node i, The key vector of node j, represents the characteristics of edge (i,j), Represents the edge features between (i, j) in the lth layer, represents the preliminary node features of the l+1 layer without residual and activation, represents the preliminary edge features of the l+1th layer without residual and activation, Represents the key vector of node j.

5. The method for predicting protein interaction sites based on graph neural network according to claim 1, characterized in that: In step S4, the probability of protein interaction site prediction is defined as follows: Y = Softmax(WH+b); Where H is the final node representation of GTE-PPIS, W and b are the weight and bias of the MLP output layer, respectively, and Y∈R n×2 , where n represents the number of residues and Y has two columns, representing the probability of protein interaction sites and the probability of non-protein interaction sites, respectively.

6. The method for predicting protein interaction sites based on graph neural network according to claim 1, characterized in that: In step S1, the node features include PSSM, HMM, DSSP, and AF, which have a total of 61 dimensions. The edge features include Euclidean distance and angle cosine value, which have a total of 2 dimensions.

7. The method for predicting protein interaction sites based on graph neural network according to claim 1, characterized in that: In step S1, the protein interaction site prediction problem is regarded as a graph node classification problem, and each protein is represented as an undirected graph G = (V, E). Represents N v Node features of amino acid residues.

Images