Protein interaction prediction method and system based on dynamic graph attention

Abstract

The invention belongs to the technical field of protein interaction prediction, and discloses a protein interaction prediction method and system based on dynamic graph attention, and the method comprises the following steps: 1, constructing a protein total graph; step 2, graph node feature enhancement: adjusting feature representation of residue nodes in the protein graph by using a gating mechanism, GRU time sequence updating and an additive attention mechanism, and performing adaptive enhancement on residue node features; step 3, predicting protein-protein interaction: aggregating information from adjacent nodes by adopting a dynamic graph attention mechanism, and capturing local and global interaction between residues; and splicing the output features of the two protein maps, and generating a score for predicting protein-protein interaction after passing through a full connection layer. According to the method, more accurate and stable protein interaction prediction is realized through dynamic feature fusion of a double-protein map structure.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of protein interaction prediction, and particularly relates to a protein interaction prediction method and system based on dynamic graph attention. BACKGROUND

[0002] Protein-protein interaction (PPI) is the basis of many key life activities in living organisms, and its research has important significance for revealing the molecular mechanisms of life activities, promoting disease diagnosis and treatment, and new drug research and development. Predicting PPI, that is, judging whether two target proteins will interact through a computational method, not only helps to deepen the understanding of biological activities, but also provides support for protein engineering research.

[0003] With the progress of computers and artificial intelligence technology, PPI prediction has made rapid development, mainly divided into three categories:

[0004] 1. Network-based method. This method uses protein network information to predict interactions, which can consider the relationship between proteins. For example, many studies have introduced the collective initiative of the International Network of Medical Union to evaluate the performance of 26 representative network-based methods; another study has proposed a hierarchical graph architecture that integrates 3D structure graph learning and sequence-structure fusion through a gating attention mechanism. However, this method has the disadvantage of often dealing with incomplete protein network information, and may have limitations for newly discovered proteins.

[0005] 2. Structure-based method. This method uses three-dimensional structure information of proteins to predict interactions, which has the advantage of capturing the real structure of proteins and has high prediction accuracy. For example, a study combined AlphaFold2 and optimized multiple sequence alignment to predict heterodimer protein complexes, and another study developed a motif matching algorithm for identifying proteins with sequence or structural similar motifs to a given query motif. However, this method usually requires the known structure of the target protein, and the prediction process is complex and time-consuming. Due to high cost, it is difficult to implement large-scale screening.

[0006] 3. Sequence-based method. Using protein sequence information to predict interactions has the characteristics of simplicity, low cost and fast calculation, and is widely used in large-scale PPI prediction and screening. There are many existing sequence-based methods, such as using LSTM model combined with word2vec, double module architecture combined with multi-scale representation learning and graph-based relationship modeling, support vector machine model, Res2vec residue representation method, deep residual recurrent convolutional neural network, and masked parallel convolutional neural network. However, many existing sequence-based methods still have shortcomings, such as the need to improve the prediction accuracy, the strict input length limit, the difficulty in capturing the spatial structure of the protein and the interaction between residues, and the information loss caused by truncating the sequence when dealing with long proteins, which limits the application of prediction.

[0007] The sequence-based PPI prediction method is widely used because of the easy acquisition of sequence data, but there are fundamental limitations:

[0008] First, the input length limit leads to information loss. The existing model architecture strictly limits the input sequence length (such as some methods only support ≤1500 residues), and long sequences need to be truncated, resulting in the loss of key residue information and directly reducing the prediction accuracy.

[0009] Second, the long-range dependence and residue interaction capture is insufficient. The long-range dependence of residues in protein sequences (such as the spatial action of remote residues) and subtle interactions (such as local action in alpha-helix and beta-fold) are key to function, but existing methods are difficult to effectively model these patterns, resulting in incomplete reflection of protein functional association and low prediction sensitivity.

[0010] Third, the evolution and biochemical feature extraction is insufficient. The evolutionary conservation and physical and chemical characteristics (such as hydrophobicity and charge) contained in protein sequences are important basis for interaction, but existing methods are insufficient in screening and optimizing these features, resulting in low feature utilization and limited prediction accuracy. SUMMARY

[0011] In view of the shortcomings of the prior art, the present application provides a protein-protein interaction prediction method and system based on dynamic graph attention, which can realize PPI prediction based on protein sequences only. First, an ESM-2 model is used to construct a graph structure at the residue level for each protein (using a contact graph as the adjacency matrix of the graph); then, the protein full graph construction module is used to process protein sequences of any length, solving the problem of sequence length limitation, and then the graph node feature enhancement module is used to optimize the node features in the graph; then, the graphs of the two proteins are input into the dynamic graph attention mechanism for feature extraction; finally, the extracted features are spliced and input into the full connection layer, and the protein-protein interaction prediction result is output.

[0012] The technical problems solved by the present application are as follows:

[0013] (1) Overcome the input length limit of sequence-based methods. Design a protein full graph construction module (referred to as CMP module) to realize the contact map prediction of protein sequences of any length by segmenting long sequences and fusing the overlapping regions, break the length limit of protein sequences of existing methods, and avoid information loss caused by sequence truncation.

[0014] (2) Optimize the extraction of evolutionary and biochemical characteristics in protein sequences. Design a graph node feature enhancement module (referred to as SSSM module), combine the gating mechanism and Gated Recurrent Unit (GRU) unit, and adaptively and dynamically optimize the node features in the protein graph, enhance the ability to capture key biochemical characteristics of residues (such as hydrophobicity and long-range dependence), and solve the problem that existing methods cannot effectively extract evolutionary and biochemical information in sequences.

[0015] (3) Design a dual-protein independent modeling and feature fusion framework, which does not rely on protein structure or network information, effectively fuse the features extracted by graph convolution and dynamic attention of two proteins and input them into the full connection layer, realize accurate PPI prediction based on sequence data only, and get rid of the dependence on protein structure or domain annotation, and expand the scope of application of the method (such as unknown or newly discovered proteins).

[0016] (4) Improve the sensitivity to subtle interactions between residues. Through dynamic feature fusion of dual-protein graph structure, more accurate and stable PPI prediction is realized, and a reliable tool is provided for large-scale proteomics research.

[0017] In order to solve the above technical problems, the technical scheme adopted by the present application is:

[0018] The protein interaction prediction method based on dynamic graph attention includes the following steps:

[0019] Step 1, protein full graph construction:

[0020] A contact map is constructed from two protein sequences using a pre-trained language model, which is a contact probability matrix used to capture the pairwise interaction probability between residues;

[0021] Threshold processing is applied to the contact probability matrix to generate the adjacency matrix of the protein graph;

[0022] A node feature matrix containing 33 physicochemical features is constructed for each residue node;

[0023] The adjacency matrix and the node feature matrix are processed by a graph neural network to construct a protein graph, where residues are nodes of the protein graph and interactions between residues are edges;

[0024] Step 2, graph node feature enhancement:

[0025] The feature representation of the residue node in the protein graph is adjusted using a gating mechanism, GRU temporal update, and additive attention mechanism to adaptively enhance the residue node feature.

[0026] Step 3, protein-protein interaction prediction:

[0027] A dynamic graph attention mechanism is used to aggregate information from adjacent nodes to capture local and global interactions between residues. The output features of the two protein graphs are spliced, and after passing through a fully connected layer, a score for predicting protein-protein interactions is generated.

[0028] Further, in the step of constructing a protein graph, the pre-trained language model is an ESM-2 model, which processes protein sequences through 33 layers of Transformer to extract the context representation of each residue from the input protein sequence, generates a hidden state embedding for each residue, and calculates the output contact probability matrix based on dot product and cross-attention weight.

[0029] Further, in the step of constructing a protein graph, a threshold of 0.5 is applied to the contact probability matrix to determine the presence or absence of interaction, generating an interaction matrix C. The interaction matrix C is used as an adjacency matrix to construct a protein graph; where each element C ij of the interaction matrix C is defined as follows:

[0030]

[0031] Further, in the step of constructing a protein graph, a segmentation and assembly strategy is used to process long sequence proteins, specifically including: dividing a protein sequence with a length of N, which exceeds a set threshold, into overlapping sub-sequences, where the threshold N cut is 3000; predicting the contact graph for each sub-sequence, taking the average of the contact probabilities in the overlapping region to generate a complete contact graph, and realizing the construction of a protein sequence amino acid level graph of any length.

[0032] Further, after constructing the protein graph, in the step of enhancing the graph node feature, the node feature representation in the graph structure data is enhanced through a graph node feature enhancement module. The graph node feature enhancement module includes: an adaptive gating mechanism for feature selection, a state transfer temporal update mechanism based on GRU unit, and a context-aware feature optimization module based on additive attention mechanism. The specific operations of the graph node feature enhancement module include:

[0033] The dimension of the input feature matrix is expanded to a high-dimensional space through a learnable projection.

[0034] The timing feature update is implemented using a gating mechanism and a GRU unit;

[0035] The attention weight is calculated by an additive attention mechanism and an enhanced node feature representation is generated.

[0036] Further, in the gating mechanism, the gate vector g t is calculated as follows:

[0037] g t =σ(W g ·[x t ;h t-1 ]+b g ) (2);

[0038] where x t represents the current node feature, h t-1 represents the previous hidden state, W g and b g are learnable parameters, and σ is a sigmoid activation function.

[0039] In the state passing timing update mechanism based on the GRU unit, the candidate hidden state h' t is calculated by time-dynamic gating integration as follows:

[0040] h' t =g t ⊙GRU(x t ;h t-1 ) (3);

[0041] where ⊙ represents element-wise multiplication, and h' t represents the updated hidden state.

[0042] In the context-aware feature optimization module based on the additive attention mechanism, the attention score a t and the attention weight α t are calculated as follows:

[0043] a t =tanh(W a [h' t ;h' t ]+b a ) (4);

[0044]

[0045] where W a and W α are weight matrices, tanh is the hyperbolic tangent activation function, [h' t ; h' t ] is the vector concatenation operation, and b ah' is the bias term. t The updated hidden state uses softmax as the normalization function, and the final hidden state integrates these attention weights through feature scaling.

[0046] h t =α t ⊙h' t (6);

[0047] Then, collect the hidden states [h1, h2, ..., h] for all time steps. N Finally, output the hidden state x at the last time step. ssm =h N , where N is the protein sequence length.

[0048] Furthermore, the dynamic attention mechanism employs GATv2, where the central residue i is the query node and the neighboring residue j is the bond node. The features of the query node and the bond node are decoupled through a weight matrix transformation, allowing the attention score to adapt to the unique features of each query. Based on the attention score, the neighbor information of the residue node in the protein graph is aggregated to update the node features. At the same time, electrostatic complementarity priors are introduced to enhance the protein functional relevance. The updated node features are represented as follows:

[0049]

[0050] in This represents the embedding representation of residue i in the (l+1)th layer, i.e., the updated node feature vector. This indicates a feature concatenation operation, which concatenates the outputs of two attention heads to expand the feature dimension. There are two parallel attention heads that compute independently before concatenation. k = 1, 2 represents the k-th attention head. Represents the set of neighbors of residue i. Summing all residues j in , This represents the attention weight of residue j to residue i in the k-th attention head, and its specific calculation is given by Equation 8. This represents the weight matrix of the embedding features of the transformed neighbor residue j in the k-th attention head. This represents the embedding representation of residue j in layer l, i.e., the node feature vector before the update; The initial input is x ssm ;

[0051] Among them, attention weight Used to characterize the electrostatic interaction between residues, the sign of the weight reflects electrostatic attraction or repulsion, and the numerical value represents the strength of the interaction;

[0052]

[0053] in, represents the unnormalized attention score of residue j to residue i in the kth attention head, represents the unnormalized attention score of residue m to residue i in the kth attention head, represents the absolute value of

[0054]

[0055] where exp(·) represents the exponential function, a T represents the transpose of the attention vector a, LeakyReLU(·) is an activation function, represents the weight matrix of the embedding feature of the transformation center residue i in the kth attention head, h i the embedding of residue i in the lth layer, h j the embedding of residue j in the lth layer, represents the sum of all residues in the neighbor set of residue i, m represents the serial number of the neighbor residue, h m represents the embedding feature of the neighbor residue m, λ is the prior weight, E(i,j) is the electrostatic complementarity quantification, the attention weight is dynamically adjusted by the electrostatic complementarity quantification E(i,j), the weight sign is determined by E(i,j) and can be positive or negative;

[0056]

[0057] where charge(i), charge(j) represent the charge values of residues i and j, respectively, pI i , pI j represent the isoelectric points of residues i and j, respectively, max pI and min pI represent the maximum and minimum isoelectric points of all residues.

[0058] Further, the loss function is a composite loss, and the formula is as follows:

[0059] L = L b + λ1·L e + λ2·L a (15);

[0060] where L b is the binary cross-entropy loss, which ensures the consistency of the model output and the true interaction label, L e is the electrostatic property constraint loss, which strengthens the physical meaning of the signed attention, L a is the attention sparsity constraint, which focuses on key interaction residues to reduce the interference of irrelevant residues, and λ1 and λ2 are balance coefficients.

[0061]

[0062] where M represents the number of samples, y i represents the real protein-protein interaction, represents the predicted protein-protein interaction, and a ij The signed attention weight is shown in formula 8, E(i,j) is the electrostatic complementarity quantitative value shown in formula 10, and sign() is a sign function.

[0063] The application also provides a protein-protein interaction prediction system based on dynamic graph attention, which is used to implement the protein-protein interaction prediction method based on dynamic graph attention as described above, and the system comprises a protein full graph construction module, a graph node feature enhancement module and a protein-protein interaction prediction module.

[0064] The protein full graph construction module uses a pre-trained language model to construct a contact graph from two protein sequences, applies threshold processing to a contact probability matrix, and generates an adjacency matrix of the protein graph; a node feature matrix containing 33 physicochemical features is constructed for each residue node; the adjacency matrix and the node feature matrix are processed through a graph neural network to construct the protein graph.

[0065] The graph node feature enhancement module adjusts the feature representation of the residue nodes in the protein graph by using a gating mechanism, GRU time sequence updating and additive attention mechanism, and adaptively enhances the residue node features.

[0066] The protein-protein interaction prediction module aggregates information from adjacent nodes by using a dynamic graph attention mechanism, captures local and global interactions between residues, splices the output features of two protein graphs, and generates a score of the predicted protein-protein interaction through a fully connected layer.

[0067] Compared with the prior art, the application has the following advantages:

[0068] (1) Overcome the input length limitation of sequence-based methods. The protein full graph construction module is designed to realize contact graph prediction for protein sequences of any length by segmenting long sequences and fusing overlapping regions, breaking the length limitation of existing methods on protein sequences, and avoiding information loss caused by sequence truncation.

[0069] (2) Optimize the extraction of evolutionary and biochemical features in protein sequences. The graph node feature enhancement module is designed to adaptively and dynamically optimize the node features in the protein graph by combining a gating mechanism and a gated recurrent unit, enhance the ability to capture key biochemical features of residues (such as hydrophobicity and long-range dependence), and solve the problem that existing methods cannot effectively extract evolutionary and biochemical information in sequences.

[0070] (3) Design a double-protein independent modeling and feature fusion framework, without relying on protein structure or network information, effectively fuse the features extracted by graph convolution and dynamic attention of the two proteins and input them into the full connection layer, realize accurate PPI prediction based on sequence data only, get rid of the dependence on protein structure or domain annotation, and expand the scope of application of the method (such as unknown or newly discovered proteins).

[0071] (4) Improve the sensitivity to subtle interactions between residues. Through dynamic feature fusion of double-protein graph structure, more accurate and stable PPI prediction is realized, and a reliable tool is provided for large-scale proteomics research. BRIEF DESCRIPTION OF DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0073] Figure 1 The schematic diagram of the framework of the present application;

[0074] Figure 2 The schematic diagram of the protein full graph construction module of the present application;

[0075] Figure 3 The schematic diagram of the graph node feature enhancement module of the present application;

[0076] Figure 4 The GATV-2 model architecture diagram in the embodiment of the present application;

[0077] Figure 5 The area under the curve (AUC) comparison diagram of the present application and the existing method;

[0078] Figure 6 The accuracy rate comparison diagram of the interaction class of the present application and the existing method;

[0079] Figure 7 The accuracy rate comparison diagram of the non-interaction class of the present application and the existing method;

[0080] Figure 8 The performance on the BioGRID yeast dataset in the embodiment;

[0081] Figure 9 The performance on the BioGRID Homo sapiens dataset in the embodiment;

[0082] Figure 10 The performance on the yeast core dataset in the embodiment;

[0083] Figure 11 This example shows the performance across all multi-species datasets.

[0084] Figure 12 The performance of the multi-species dataset in the example is shown when the identity of the dataset is less than or equal to 0.4.

[0085] Figure 13 The performance of the multi-species dataset in the example is shown on a homology of less than or equal to 0.25.

[0086] Figure 14 The performance of the multi-species dataset in the example is shown when the identity of the dataset is less than or equal to 0.1.

[0087] Figure 15 The performance of the multi-species dataset in the example is shown when the identity of the dataset is less than or equal to 0.01.

[0088] Figure 16 A visual network diagram for predicting human-virus protein interactions. Detailed Implementation

[0089] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0090] Example 1

[0091] Combination Figure 1 As shown, this embodiment discloses a protein-protein interaction prediction method based on dynamic graph attention, which aims to predict the activity of protein-protein interaction (PPI) by processing the sequences of two proteins as input. The basic workflow includes protein full graph construction, graph node feature enhancement module, and final protein-protein interaction prediction.

[0092] This invention designs a protein full-map construction module and a graph node feature enhancement module to achieve graph construction at the amino acid level for protein sequences of arbitrary length and enhance the feature representation of amino acids in graph nodes. The protein full-map construction module employs a segmentation and assembly strategy, processing overlapping subsequences to achieve graph construction at the amino acid level for protein sequences of arbitrary length. The graph node feature enhancement module utilizes gating and attention mechanisms to adjust the feature representation of amino acid nodes (residue nodes) in the protein graph.

[0093] First, a contact map (a probability matrix of pairwise interactions between protein residues) is constructed from the protein sequence using a pre-trained language model (ESM-2 model is used in this embodiment). Based on this contact map, a protein map containing spatial and interaction features is then constructed.

[0094] To handle protein sequences of arbitrary length, a protein full-map construction module was designed.Figure 2 ), for generating contact maps between protein residues, thus constructing protein graphs. After constructing protein graphs, the present application employs a graph node feature enhancement module ( Figure 3 ) to amplify relevant physicochemical properties, such as hydrophobicity and charge, while suppressing noise features. This module models local interactions between amino acid residues, with a particular focus on short-range interactions found in structural motifs such as alpha helices and beta sheets.

[0095] Subsequently, the present application employs a dynamic graph attention mechanism that aggregates information from neighboring nodes, effectively capturing local and global structural patterns within protein graphs. Finally, the output features from both protein graphs are concatenated, and after passing through a fully connected layer, the model generates a predicted interaction score. This end-to-end training approach enables efficient learning from protein sequence input to the final prediction output.

[0096] As a preferred embodiment, the method of the present embodiment, in particular, comprises the following steps:

[0097] Step 1, protein graph construction:

[0098] To predict protein-protein interactions, the first step is to extract structural information from protein sequences, which can be represented by contact maps. A pre-trained language model is used to construct contact maps from two protein sequences, which is a contact probability matrix that captures the probability of pairwise interactions between residues as an adjacency matrix for constructing protein graphs. Here, residues are nodes, and their interactions form edges. In the present embodiment, the pre-trained model "esm2_t33_650M_UR50D" of the ESM-2 model is used to extract the context representation of amino acids, which is the basis for contact map prediction. This model takes the original protein sequence as input, processes the protein sequence through 33 layers of Transformer, extracts the context representation of each residue from the input protein sequence, generates hidden state embeddings for each residue, and uses the pairwise relationships between these embedding vectors (including dot product and cross-attention weights) to calculate the output contact probability matrix.

[0099] Then, a thresholding process is applied to the contact probability matrix to generate the adjacency matrix of the protein graph. In the present embodiment, a threshold of 0.5 is applied to the contact probability matrix to determine the presence or absence of interactions, generating an interaction matrix C of dimension (N x N), which serves as an adjacency matrix for constructing a protein graph; where each element C ij is defined as follows:

[0100]

[0101] This method can adaptively process sequences of any length, because the self-attention mechanism of ESM-2 can dynamically capture long-distance dependencies without explicit length restrictions. In constructing the protein graph, the interaction matrix C serves as the adjacency matrix of the graph, and the protein graph is constructed.

[0102] Then, a node feature matrix containing 33 physicochemical features is constructed for each residue node. The adjacency matrix and the node feature matrix are processed by a graph neural network to construct the protein graph, where the residues are the nodes of the protein graph and the interactions between residues are the edges.

[0103] Specifically, when using a graph neural network to extract protein features, in addition to the generation method of the edge, it is also crucial to determine the features of each node in the graph. Since the protein graph is constructed with residues as nodes, the selection of features should reflect the differences between different residues. Residues are composed of multiple parts, including common carboxyl, amino and central carbon atoms. The main source of differences between different residues lies in the unique side chain (R group) structure. These differences encompass properties such as polarity, charge, aromaticity, etc. A total of 33 features were selected to describe the residue nodes, as shown in Table 1. It should be noted that although existing technologies also extract bioinformatics features of residues, the present application focuses more on the feature representation of residues themselves, and the bioinformatics features used are not the same. The number of features used in the present application is smaller, and the computational complexity is much smaller. For a protein graph with a sequence length of N, the node feature matrix can be represented by a matrix with a shape of (N, 33).

[0104] Table 1 Protein node features

[0105]

[0106] In addition, as a preferred embodiment, in the construction of the full protein graph, a segmentation and assembly strategy is used to process ultra-long sequence proteins (protein length exceeding 3000) to save memory consumption.

[0107] By inputting the adjacency matrix and node feature matrix of the protein into the graph neural network, the corresponding feature vector representation of the protein can be obtained. However, for the running device, insufficient memory can hinder the prediction of the contact map and cause a memory overflow. It is known that residues that are closer in sequence are more likely to contact each other. In the contact map matrix of the protein, the interaction probability in the diagonal region is larger. Therefore, if the device memory is insufficient, the contact map of a large protein can be segmented and assembled. Assuming that the length of the target large protein is N, a fragment length and a step size can be set, so that several overlapping subsequences are obtained. Then the contact map of each subsequence can be predicted, and after assembly, the complete contact map of the protein is formed. For the overlapping region in the contact map, the average value can be used to obtain the interaction probability. The contact map processing process of the large protein is as shown in Figure 2 .

[0108] Specifically comprising: segmenting a protein sequence with a length of N and exceeding a set threshold into overlapping subsequences, wherein the threshold N cut is 3000, and the step size is set to 500; predicting the contact map of each subsequence respectively, and taking the average value of the contact probability of the overlapping region to generate a complete contact map, thereby realizing the graph construction of the amino acid level of the protein sequence of any length.

[0109] As shown in Figure 2 , the steps of constructing a protein graph from a protein sequence are as follows:

[0110] (1) Contact map generation

[0111] Input the protein sequence, and use a pre-trained model to predict the contact probability matrix between residues (i.e., which residues can interact in space).

[0112] Output a symmetric matrix (N x N, N is the number of residues), and the matrix element value is 1 (contact) or 0 (no contact).

[0113] (2) Edge generation

[0114] Convert the positions with a value of 1 in the contact map to edges of the graph (e.g., if residues i and j are in contact, an edge is built), which represents the interaction between residues.

[0115] (3) Connection processing and peptide bond processing

[0116] Connect adjacent residues: add edges between adjacent residues in the graph (corresponding to peptide bonds), ensure that adjacent residues (i and i+1) are forced to connect, and ensure the continuity of the protein chain.

[0117] Combine the contact map edges: the edge set of the final graph = the peptide bond edge + the predicted / calculated contact edge.

[0118] (4) Generate graph

[0119] Node: Each residue is a node, with features (e.g. amino acid type, charge, hydrophobicity, etc., see Table 1) assigned.

[0120] Edge: Contains two types, structural edge: non-covalent interactions based on contact map. Sequence edge: peptide bond connection between adjacent residues.

[0121] Output: Graph data structure (adjacency matrix + node feature matrix), which can be directly input into a graph neural network (GNN).

[0122] Final protein graph: Node = residue, edge = spatial contact + peptide bond, feature = physicochemical property.

[0123] Step 2, graph node feature enhancement:

[0124] After building a protein graph using the protein full graph construction module, the present application designs a novel neural network architecture to enhance the node feature representation in the graph structure data through a graph node feature enhancement module.

[0125] The proposed graph node feature enhancement module adopts a framework composed of three parts:

[0126] 1. Adaptive gating mechanism for feature selection;

[0127] 2. State transfer time series update mechanism based on GRU unit;

[0128] 3. Context-aware feature optimization module based on additive attention mechanism.

[0129] The feature representation of the residue node in the protein graph is adjusted using the gating mechanism, GRU time series update, and additive attention mechanism, and the residue node feature is adaptively enhanced.

[0130] The specific operations of the graph node feature enhancement module include:

[0131] The dimension of the input feature matrix is expanded to a high-dimensional space through a learnable projection;

[0132] Temporal feature update is achieved using a gating mechanism and a GRU unit;

[0133] Attention weights are calculated and enhanced node feature representations are generated through an additive attention mechanism.

[0134] Given an input feature matrix where N is the number of nodes and D is the feature dimension (where D = 33, as shown in Table 1), the graph node feature enhancement module ( Figure 3 ) first expands the input to a higher-dimensional space (66 in this embodiment) through a learnable projection, and then applies hierarchical variable protein transformation to capture multi-scale structure patterns.

[0135] In combination Figure 3 The gating mechanism regulates the information flow through parameterized state updates. The gate vector g t is computed as:

[0136] g t = σ(W g · [x t ; h t-1 ] + b g ) (2).

[0137] where x t denotes the current node feature, h t-1 represents the previous hidden state, W g and b g are learnable parameters, and σ is the sigmoid activation function. This adaptive gating mechanism enables context-dependent feature modulation, enhancing task-relevant signals while suppressing noise components.

[0138] In the state passing time update mechanism based on GRU units, the candidate hidden state h' t is computed through time-dynamic gated integration:

[0139] h' t = g t ⊙ GRU(x t ; h t-1 ) (3).

[0140] where ⊙ denotes element-wise multiplication, and h' t represents the updated hidden state; the Gated Recurrent Unit (GRU unit) can capture sequential patterns in data, making it suitable for tasks involving dynamic or sequential node features.

[0141] In the context-aware feature optimization module based on additive attention mechanism, the attention score a t and the attention weight a t are computed as follows:

[0142] a t = tanh(W a [ h' t ; h' t ] + b a ) (4).

[0143]

[0144] where W a and W α are weight matrices, tanh is the hyperbolic tangent activation function, and [h' t ; h't ] is a vector concatenation operation, b a is a bias term, h' t is the updated hidden state, softmax is a normalization function, and the final hidden state integrates these attention weights through feature scaling:

[0145] h t = a t ⊙h' t (6) ;

[0146] Then, the hidden states of all time steps [h1, h2, …, h N ] are collected, and the final output is the hidden state of the last time step x ssm = h N , where N is the length of the protein sequence.

[0147] Step 3, Protein-protein interaction prediction:

[0148] The dynamic graph attention mechanism is used to aggregate information from adjacent nodes, capturing local and global interactions between residues. The output features of the two protein graphs are concatenated, and after passing through a fully connected layer, the score of the predicted protein-protein interaction is generated.

[0149] (1) Dynamic graph attention

[0150] In graph neural networks, the classic graph attention network (GAT) relies on a static attention mechanism. Its core limitation is that for all query nodes, the attention ranking of neighbor nodes (keys) is fixed globally. For example, in a protein-protein interaction network, if GAT considers protein i as a query, the attention ranking of its interaction partners (e.g., protein j -> protein k) remains unchanged even if protein k becomes the query object. This rigid, query-independent ranking method cannot capture the context-specific interactions inherent in protein-protein interactions. Different proteins often prefer different interaction partners based on their functional states, but the static attention mechanism cannot model this dynamic preference.

[0151] To address this limitation in protein-protein interaction analysis, the invention uses a multi-head graph attention network v2 (a dynamic attention mechanism (GATv2)), with the central residue i as the query node and the neighbor residue j as the key node. The features of the query node and the key node are decoupled through a weight matrix transformation, allowing the attention score to adapt to the unique features of each query.

[0152] Based on the attention score aggregation of neighbor information of residue nodes in the protein graph, the node features are updated. At the same time, electrostatic complementarity prior is introduced to enhance the relevance of protein function. The updated node features are represented as follows:

[0153]

[0154] in This represents the embedding representation of residue i in the (l+1)th layer, i.e., the updated node feature vector. This indicates a feature concatenation operation, which concatenates the outputs of two attention heads to expand the feature dimension. There are two parallel attention heads that compute independently before concatenation. k = 1, 2 represents the k-th attention head. Represents the set of neighbors of residue i. Summing all residues j in , This represents the attention weight of residue j to residue i in the k-th attention head, and its specific calculation is given by Equation 8. This represents the weight matrix of the embedding features of the transformed neighbor residue j in the k-th attention head. This represents the embedding representation of residue j in layer l, i.e., the node feature vector before the update; The initial input is x ssm .

[0155] Among them, attention weight Used to characterize electrostatic interactions between residues, the sign of the weight reflects electrostatic attraction or repulsion, and the numerical value indicates the strength of the interaction; signed attention weights are used to interpret electrostatic complementarity, enhancing their ability to predict functional relationships between proteins based on their electrostatic properties.

[0156]

[0157] in, This represents the unnormalized attention score of residue j relative to residue i in the k-th attention head. This represents the unnormalized attention score of residue m to residue i in the k-th attention head. express The absolute value;

[0158]

[0159] Where exp(·) represents the exponential function, a T This represents the transpose of the attention vector a, and LeakyReLU(·) is the activation function. h represents the weight matrix h of the embedding features of the transformation center residue i in the k-th attention head. i The embedding of residue i in layer l, h j Embedding of residue j in layer l, Represents the set of neighbors of residue i. The summation of all residues, where m represents the index of the neighboring residue, h m Let λ represent the embedding feature of the neighboring residue m, λ be the prior weight, and E(i,j) be the electrostatic complementarity quantization. The attention weight is dynamically adjusted through the electrostatic complementarity quantization E(i,j). The sign of the weight is determined by E(i,j) and can be positive or negative.

[0160]

[0161] Where charge(i) and charge(j) represent the charge values ​​of residues i and j, respectively, and pI i pI j Let i and j represent the isoelectric points of residues i and j, respectively, and maxpI and minpI represent the maximum and minimum isoelectric points of all residues.

[0162] By processing the features of residues with different roles in the diagram (central residue vs. neighboring residues) separately, the information from both is ultimately fused through an attention mechanism to capture the correlation between residues.

[0163] (2) PPI forecast

[0164] This invention uses a three-layer GATV2. The first layer employs two multi-head attention mechanisms, among which... The output is a 132-dimensional vector. The second and third layers use single-head attention, where... This model design captures local contacts through head-to-head attention divergence, enabling it to identify subtle interactions at the atomic level. It captures local contacts through divergent head-to-head attention differences, thereby identifying minute interactions at the atomic level; simultaneously, it integrates a multi-hop messaging mechanism to effectively map allosteric transport paths, thus tracing complex connection networks between distant residues.

[0165] Furthermore, the model utilizes signed attention weights to interpret electrostatic complementarity, enhancing its ability to predict functional relationships between proteins based on their electrostatic properties. These features collectively enable a comprehensive understanding of protein activity, improving the accuracy of predictions for interactions and activities. Interactions are predicted after global average pooling as follows:

[0166]

[0167] Where, Φ GAP This is a global average pooling function used to convert matrix-form features (such as the node embedding matrix of a protein map) into a fixed-dimensional vector. It achieves this by averaging the features across all nodes. represents the node embedding matrix of the first protein after being processed by the third layer graph attention network (GATv2), f1 represents the feature vector of the first protein after being processed by global average pooling, represents the node embedding matrix of the second protein after being processed by the third layer graph attention network, f2 represents the feature vector of the second protein after being processed by global average pooling, [f1||f2] represents a feature splicing operation, W f represents a weight matrix for feature fusion, b f represents a bias term in the feature fusion process, ReLU is a linear rectification activation function, f c represents the joint feature vector obtained after feature fusion and activation, represents the weight matrix of the output layer, b o represents the bias term of the output layer, and sigma is a Sigmoid activation function that maps the output to the interval [0, 1], representing the prediction probability of the interaction between the two proteins, represents the final interaction prediction result (a probability value between 0 and 1, and the closer to 1, the higher the interaction possibility). The architecture of the entire neural network is shown in Figure 4 .

[0168] Figure 4 In the formula, the left part is a dynamic graph attention mechanism, 3-layer GATv2Conv+ReLU: GATv2Conv can dynamically learn the importance weight between nodes. Dynamic graph attention (GATv2): better than the original GAT, can more flexibly calculate attention. ReLU activation function is used after each layer to introduce nonlinearity. Dropout: to prevent overfitting. The right side is a graph node feature enhancement module. After obtaining the dynamic attention graph of the protein, feature fusion is performed, and further non-linear transformation and regularization are performed through ReLU->Dropout, full connection, ReLU->Dropout, final feature extraction, and Sigmoid output of a probability value between 0 and 1, which is used for PPI prediction.

[0169] Loss calculation:

[0170] The loss function of the application is a composite loss, and the formula is as follows:

[0171] L=L b +λ1·L e +λ2·L a (15);

[0172] wherein L b is a binary cross-entropy loss, which ensures the consistency of the model output and the true interaction label, L e is an electrostatic property constraint loss, which strengthens the physical meaning of signed attention, and L aFor attention sparsity constraint, focusing on key interaction residues reduces the interference of irrelevant residues, λ1, λ2 are balance coefficients;

[0173]

[0174] Wherein, M represents the number of samples, y i represents the real protein interaction, represents the predicted protein interaction, α ij The signed attention weight formula 8 is shown, and E(i,j) is the electrostatic complementarity quantitative value formula 10, and sign() is a symbol function.

[0175] Embodiment 2

[0176] The embodiment discloses a protein-protein interaction prediction system based on dynamic graph attention, which comprises a protein full graph construction module, a graph node feature enhancement module and a protein-protein interaction prediction module.

[0177] The protein full graph construction module is used to realize the graph construction of the amino acid level of the protein sequence of any length. The protein full graph construction module uses a pre-trained language model to construct a contact graph from two protein sequences, applies threshold processing to a contact probability matrix, generates an adjacency matrix of the protein graph, constructs a node feature matrix containing 33 physicochemical features for each residue node, and processes the adjacency matrix and the node feature matrix through a graph neural network to construct the protein graph.

[0178] The graph node feature enhancement module adjusts the feature representation of the residue nodes in the protein graph by using a gating mechanism, GRU time sequence updating and additive attention mechanism, and adaptively enhances the residue node features.

[0179] The protein-protein interaction prediction module adopts a dynamic graph attention mechanism to aggregate information from adjacent nodes, capture local and global interactions between residues, splice the output features of two protein graphs, pass through a fully connected layer, and generate a score of the predicted protein-protein interaction.

[0180] The structure and function implementation of each module can be referred to the detailed steps of the protein-protein interaction prediction method based on dynamic graph attention described in Embodiment 1, which will not be repeated here.

[0181] Embodiment 3

[0182] As another embodiment of the present application, a computer readable storage medium is also provided, which stores a computer program, and the computer program is executed by a processor to implement the protein-protein interaction prediction method based on dynamic graph attention as described in Embodiment 1.

[0183] The computer readable storage medium can be a non-transitory computer readable storage medium. For example, and without limitation, this can include a portable computer diskette, a hard disk, random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

[0184] Experiments:

[0185] To verify the effectiveness of the invention, the invention was tested on four datasets.

[0186] (1) BioGRID dataset. BioGRID is a publicly accessible database that compiles experimentally verified protein-protein and genetic interaction data for a variety of organisms, thereby supporting biological research through comprehensive and reliable interaction networks. The invention used the multiple-validated human and yeast physical interaction datasets from BioGRID as benchmarks for training and evaluation. The Saccharomyces cerevisiae (yeast) dataset contains 13,462 positive and 13,462 negative items derived from 17,015 proteins, while the Homo sapiens dataset contains 31,164 positive and 31,164 negative items obtained from 38,869 proteins.

[0187] (2) Saccharomyces cerevisiae core dataset. This dataset is derived from the Saccharomyces cerevisiae core dataset, containing 2,438 proteins, forming 5,271 positive and 5,266 negative items.

[0188] (3) Multi-species dataset. This dataset combines data from Caenorhabditis elegans, Escherichia coli, and Drosophila melanogaster to create a multi-species dataset. In this dataset, during the dataset construction phase, different sequence identity thresholds (40%, 25%, 10%, 1%, and the unfiltered full dataset) were used to pre-filter proteins. The invention performed 5-fold cross-validation on these pre-filtered subsets, without any additional sequence identity filtering during cross-validation. In the final dataset, the different filtered datasets contain 10,747, 12,641, 19,458, 25,916, and 32,959 positive items, and 8,065, 9,819, 15,827, 22,012, and 32,959 negative items, respectively.

[0189] (4) Virus-Human Interaction Dataset: The Virus-Human Interaction Dataset contains 8,929 positive and 8,929 negative entries from 12,652 proteins, which is used as an independent test set in the experiment to test the effectiveness of the present application.

[0190] Performance of the present application (DPEG) on BioGRID:

[0191] Table 2: 5-fold cross-validation performance of protein-protein interaction (PPI) prediction on BioGRID Saccharomyces Cerevisiae Dataset

[0192] Methods Accuracy Precision Sensitivity Specificity MCC F1 -score DeepFE-PPR 85.24±0.52 85.49±1.41 84.99±2.77 85.49±2.11 70.57±1.06 85.19±0.79 PIPR 95.76±0.25 94.61±0.53 97.06±0.41 94.47±0.55 91.56±0.48 95.82±0.24 DeepDuo 97.06±0.28 98.06±0.51 96.02±0.35 98.10±0.50 94.14±0.57 97.02±0.30 DeepTrio 97.55±0.38 98.95±0.20 96.12±0.74 98.98±0.21 95.15±0.74 97.52±0.40 DPEG (the present invention) 98.04±0.32 97.47±0.66 98.64±0.19 97.44±0.69 96.09±0.63 98.05±0.31

[0193] Table 3: 5-fold cross-validation performance of protein-protein interaction (PPI) prediction on BioGRID Homo sapiens Dataset

[0194] Methods Accuracy Precision Sensitivity Specificity MCC F1 -score DeepFE-PPI 87.66±0.57 89.42±1.05 85.47±2.27 89.85±1.4 75.44±1.09 87.37±0.78 PIPR 97.6±0.08 97.57±0.35 97.63±0.44 97.56±0.36 95.2±0.15 97.6±0.1 DeepDuo 98.04±0.05 98.83±0.28 97.23±0.28 98.85±0.27 96.09±0.1 98.02±0.05 DeepTrio 98.12±0.12 99.0±0.17 97.23±0.28 99.01±0.17 96.26±0.23 98.11±0.13 DPEG (the present invention) 98.64±0.13 98.28±0.31 99.02±0.18 98.26±0.32 97.29±0.25 98.65±0.13

[0195] Performance of the present application on Saccharomyces Cerevisiae Core Dataset:

[0196] Table 4: 5-fold cross-validation performance of protein-protein interaction (PPI) prediction on Saccharomyces Cerevisiae Core Dataset

[0197] Methods Accuracy Precision Sensitivity Specificity MCC F1 -score SVM-AC 87.35±1.38 87.82±4.84 87.3±5.23 87.41±6.33 87.34±1.33 75.09±2.51 SVM-MCD 91.36±0.4 91.94±0.69 90.67±0.77 NA 91.3±0.73 84.21±0.59 DeepFE-PPI 94.78±0.61 96.45±0.87 92.99±0.66 NA NA 89.62±1.23 PIPR 92.16±0.55 96.57±1.22 87.46±1.46 96.83±1.27 91.78±0.59 84.71±1.1 DeepDuo 92.26±0.44 94.17±0.65 90.11±0.56 94.42±0.56 92.09±0.53 84.6±0.89 DeepTrio 92.57±0.63 96.33±0.88 88.53±1.19 96.62±0.83 92.26±0.65 85.43±1.22 DPEG 92.53±0.32 94.19±0.50 90.66±1.14 94.32±0.62 85.13±0.63 92.31±0.42

[0198] Performance of the present application on Multi-Species Dataset:

[0199] Table 5: 5-fold cross-validation performance of protein-protein interaction (PPI) prediction on Multi-Species Dataset

[0200]

[0201] Cross-experiment performance of the present application on all mixed datasets:

[0202] Table 6: 5-fold cross-experiment performance on all mixed datasets

[0203]

[0204] The cross-experiment results on all mixed datasets are similar to the results of single dataset experiments, which proves the generalization of the present method.

[0205] The virus-human interaction dataset is used as an independent test set, and the comparison of each method is performed on this dataset. The experimental results are as follows Figure 5 - Figure 7 .

[0206] Figure 8 - Figure 15For ablation experiments (to test the effect of each module), where Figure 8 For performance on the BioGRID Yeast dataset in the examples; Figure 9 For performance on the BioGRID Homo sapiens dataset in the examples; Figure 10 For performance on the Yeast Core dataset in the examples; Figure 11 For performance on all the multi-species datasets in the examples; Figure 12 For performance on the multi-species datasets with identity <= 0.4 in the examples; Figure 13 For performance on the multi-species datasets with identity <= 0.25 in the examples; Figure 14 For performance on the multi-species datasets with identity <= 0.1 in the examples; Figure 15 For performance on the multi-species datasets with identity <= 0.01 in the examples.

[0207] To quantify the contribution of the designed graph node feature enhancement module (SSSM module) and dynamic graph attention (GATv2) of the present application, four model variants were systematically evaluated on multiple datasets. DPEG-DG (only dynamic graph attention), DPEG-SG (graph node feature enhancement module + static graph attention), DPEG-NG (only static graph attention), and DPEG (graph node feature enhancement module + dynamic graph attention) of the present application. According to the results of the ablation experiment, it is verified that the complete DPEG model designed by the present application is always superior to all simplified versions, which indicates that there is a synergistic effect between the core components.

[0208] Dynamic graph attention: DPEG vs. DPEG-SG, which can improve the MCC value of each dataset. For example, on the BioGRID Saccharomyces cerevisiae dataset, DPEG (MCC = 96.09) is superior to DPEG-SG (MCC = 95.25). In the low sequence identity scenario (such as multi-species sequence identity <0.01), the MCC value of DPEG is improved by 1.84% (95.43 vs. 93.59). This highlights the stronger modeling capability of GATV2 for adaptive residue level interactions in proteins with a long evolutionary distance when it works in synergy with the node feature enhancement module.

[0209] Node feature enhancement module: DPEG vs. DPEG-DG, which further improves performance, especially in the sensitivity (0.74% improvement) and specificity (0.38% improvement) of human and yeast BioGRID datasets, indicating its role in capturing conserved biochemical motifs. It is worth noting that when combined with static GAT, the node feature enhancement module provides key stability - under all test conditions, the MCC value of DPEG-SG (node feature enhancement module + static vs. static only) is 0.16-0.30% higher than DPEG-NG.

[0210] As Figure 16 For the visualization of the Human-virus protein interaction prediction network, the application constructs a protein interaction prediction network to visualize the potential interaction relationship between human proteins and virus proteins, providing researchers with intuitive visualization effects to explore the molecular mechanisms and potential biological functions.

[0211] Of course, the above description is not a limitation of the application, and the application is not limited to the above examples. Within the scope of the application, changes, modifications, additions or substitutions made by those skilled in the art should be within the scope of the application.

Claims

1. A protein-protein interaction prediction method based on dynamic graph attention, characterized in that, The method comprises the following steps: Step 1, protein graph construction: A pre-trained language model is used to construct a contact graph from two protein sequences, which is a contact probability matrix for capturing the probability of pairwise interaction between residues; Threshold processing is applied to the contact probability matrix to generate an adjacency matrix of the protein graph; A node feature matrix containing 33 physicochemical characteristics is constructed for each residue node; An adjacency matrix and a node feature matrix are processed by a graph neural network to construct a protein graph, in which residues are nodes of the protein graph and interactions between residues are edges; Step 2, graph node feature enhancement: The feature representation of the residue node in the protein graph is adjusted by using a gating mechanism, GRU time series update and additive attention mechanism to adaptively enhance the residue node feature; Step 3, protein-protein interaction prediction: A dynamic graph attention mechanism is used to aggregate information from adjacent nodes to capture local and global interactions between residues; the output features of the two protein graphs are spliced, and then a full connection layer is used to generate a score for predicting protein-protein interaction.

2. The dynamic graph attention based protein-protein interaction prediction method of claim 1, wherein, In the step of protein graph construction, the pre-trained language model is an ESM-2 model, which processes protein sequences through 33 layers of Transformer to extract the context representation of each residue from the input protein sequence, generates a hidden state embedding for each residue, and calculates an output contact probability matrix based on dot product and cross-attention weight. 3.The dynamic graph attention-based protein-protein interaction prediction method of claim 1, wherein, In the step of constructing the whole protein map, a threshold of 0.5 is applied to the contact probability matrix to determine the existence of interaction, and an interaction matrix C is generated, which is used as an adjacency matrix to construct the protein map; wherein each element C of the interaction matrix C is defined as follows: C = 1 if the contact probability is greater than or equal to 0.5, otherwise C = 0. ij ​ 4. The dynamic graph attention based protein-protein interaction prediction method of claim 1, wherein, In the protein full map construction, a segmentation and assembly strategy is used to process long sequence proteins, specifically including: protein sequences with a length of N and exceeding a set threshold are segmented into overlapping subsequences, where the threshold N is 3000 cut ; the contact map of each subsequence is predicted respectively, the contact probability of the overlapping region is averaged, a complete contact map is generated, and the amino acid level map construction of protein sequences of any length is realized.

5. The dynamic graph attention based protein-protein interaction prediction method of claim 1, wherein, After constructing the protein graph, in the step of graph node feature enhancement, the node feature representation in the graph structure data is enhanced by a graph node feature enhancement module; the graph The graph node feature enhancement module includes an adaptive gating mechanism for feature selection, a state transfer time series update mechanism based on GRU unit, and a context-aware feature optimization module based on additive attention mechanism; the specific operation of the graph node feature enhancement module includes: The dimension of the input feature matrix is expanded to a high-dimensional space through a learnable projection; A gating mechanism and a GRU unit are used to realize time series feature update; An additive attention mechanism is used to calculate attention weights and generate enhanced node feature representation.

6. The dynamic graph attention based protein-protein interaction prediction method of claim 5, wherein, In the gating mechanism, the gate vector g t is calculated as follows: g t = σ(W g · [x t ; h t-1 ] + b g ) (2); where x t represents the current node feature, h t-1 represents the previous hidden state, W g and b g are learnable parameters, and σ is a sigmoid activation function; In the state passing time update mechanism based on the GRU unit, the candidate hidden state h ' t Computed by a time-dynamic gated integration: h' t = g t ⊙ GRU(x t ; h t-1 ) (3) where denotes element-wise multiplication, h t denotes the updated hidden state; In the context-aware feature optimization module based on the additive attention mechanism, the attention score a t and the attention weight a t is calculated as follows: a t = tanh(W a [h' t ; h' t + b a ) (4); where W a and W α are weight matrices, tanh is the hyperbolic tangent activation function, [h' t ; h' t ] is a vector concatenation operation, b a is a bias term, h' t is the updated hidden state, and softmax is a normalization function. The final hidden state integrates these attention weights through a feature scaling: h t = a t ⊙ h' t (6); Then, collect the hidden states of all time steps [h1, h2, …, h N ], and finally output the hidden state of the last time step x ssm = h N , where N is the length of the protein sequence.

7. The dynamic graph attention based protein-protein interaction prediction method of claim 1, wherein, The dynamic attention mechanism uses GATv2, with the center residue i as the query node and the neighbor residue j as the key node. The features of the query node and the key node are decoupled through weight matrix transformation, so that the attention score is adapted to the unique features of each query. The neighbor information of the residue node in the protein graph is aggregated based on the attention score to update the node feature. At the same time, the electrostatic complementarity prior is introduced to enhance the protein function correlation, and the updated node feature representation is as follows: in This represents the embedding representation of residue i in the (l+1)th layer, i.e., the updated node feature vector. This indicates a feature concatenation operation, which concatenates the outputs of two attention heads to expand the feature dimension. There are two parallel attention heads that compute independently before concatenation. k = 1, 2 represents the k-th attention head. Represents the set of neighbors of residue i. Summing all residues j in , This represents the attention weight of residue j to residue i in the k-th attention head, and its specific calculation is given by Equation 8. This represents the weight matrix of the embedding features of the transformed neighbor residue j in the k-th attention head. This represents the embedding representation of residue j in layer l, i.e., the node feature vector before the update; The initial input is x ssm ; where the attention weight For representing the electrostatic interaction between residues, the sign of the weight reflects electrostatic attraction or repulsion, and the magnitude indicates the strength of the interaction; wherein, represents the unnormalized attention score of residue j to residue i in the kth attention head, represents the unnormalized attention score of residue m to residue i in the kth attention head, represents the absolute value of . where exp(·) denotes the exponential function, denotes the transpose of attention vector a, LeakyReLU(·) is an activation function, denotes the weight matrix of the embedding feature of the transformation center residue i in the kth attention head, h i the embedding of residue i in the lth layer, h j the embedding of residue j in the lth layer, denotes the summation of all residues in the neighbor set of residue i denotes the neighbor residue number m, h m denotes the embedding feature of the neighbor residue m, λ is a prior weight, E(i,j) is a static complementarity quantification, the attention weight is dynamically adjusted through the static complementarity quantification E(i,j), the weight sign is determined by E(i,j) and can be positive or negative; where charge(i), charge(j) represent the charge values of residues i, j, respectively, pI i , pI j represent the isoelectric points of residues i, j, respectively, max pI, min pI represent the maximum and minimum values of isoelectric points among all residues.

8. The dynamic graph attention based protein-protein interaction prediction method of claim 1, wherein, The loss function is a composite loss, and the formula is as follows: L = L b + λ1 · L e + λ2 · L a (15); where L b is the binary cross-entropy loss, which ensures the consistency between the model output and the real interaction label, L e is the electrostatic property constraint loss, which strengthens the physical meaning of signed attention, L a is the attention sparsity constraint, which focuses on key interaction residues to reduce the interference of irrelevant residues, and λ1, λ2 are balance coefficients. where M represents the number of samples, y i represents the true protein-protein interaction, represents the predicted protein-protein interaction, a ij The signed attention weight is shown in Equation 8, E(i,j) is the electrostatic complementarity quantification value shown in Equation 10, and sign() is a sign function.

9. A protein-protein interaction prediction system based on dynamic graph attention, characterized in that, The system is used to implement the dynamic graph attention-based protein interaction prediction method according to any one of claims 1-8, and the system comprises a protein graph construction module, a graph node feature enhancement module and a protein-protein interaction prediction module. The protein full-map construction module uses a pre-trained language model to construct a contact map from two protein sequences, applies threshold processing to the contact probability matrix to generate an adjacency matrix of the protein map, and constructs a node feature matrix containing 33 physicochemical features for each residue node. A protein graph is constructed by processing the adjacency matrix and node feature matrix using a graph neural network. The graph node feature enhancement module uses gating mechanism, GRU temporal update and additive attention mechanism to adjust the feature representation of residue nodes in protein graph and adaptively enhance the residue node features. The protein-protein interaction prediction module uses a dynamic graph attention mechanism to aggregate information from neighboring nodes and capture local and global interactions between residues. The output features of the two protein graphs are spliced ​​together and passed through a fully connected layer to generate a score predicting protein-protein interactions.

Images