Protein-protein interaction site prediction method and system based on multi-scale structure cooperation

Abstract

This invention discloses a method and system for predicting protein-protein interaction sites based on multi-scale structural coordination, belonging to the field of bioinformatics. The method first constructs initial residue features and a protein graph; then, it extracts local geometric features of residues using an isovariant graph neural network as the backbone; a structure-aware self-attention module is connected in parallel to capture topological associations between residues and global protein information; edge-based clustering and pooling information enhancement modules are inserted at intervals in the backbone, and common features of multi-residue clusters are extracted through clustering, pooling, graph convolution, and unpooling; finally, classification is performed using a multilayer perceptron. This invention coordinates multi-scale information from local residue structure, multi-residue associations, and the overall protein structure, significantly improving the prediction accuracy of protein-protein interaction sites.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting protein interaction sites based on multi-scale structural synergy. Background Technology

[0002] Protein-protein interaction sites (PPIS) are key regions on a protein's surface that enable specific binding to other proteins. Accurate identification of PPIS is crucial for understanding protein function, elucidating molecular mechanisms of diseases, and accelerating drug discovery and design. For example, in structure-based drug design, PPIS identification helps determine the binding pockets of small molecule drugs, providing a structural basis for molecular docking. In disease mechanism research, mutations in key residues often lead to the disruption or abnormal reconstruction of PPIS, thereby triggering dysregulation of cell signaling pathways and becoming a potential pathogenic factor for various genetic diseases and cancers. Therefore, developing efficient and accurate PPIS prediction methods has become a research hotspot in the field of bioinformatics.

[0003] With the rapid development of structural biology techniques, high-precision experimental methods such as X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (Cryo-EM) have been widely used to resolve the three-dimensional structure of proteins, providing a reliable data foundation for the identification of protein sequence and structure data (PPIS). However, these methods generally suffer from high equipment costs, long resolution cycles, and low throughput, making it difficult to meet the ever-increasing demand for processing protein sequence and structure data. Against this backdrop, computational PPIS prediction techniques, due to their high efficiency, low cost, and strong scalability, are gradually becoming an important research direction to replace or supplement experimental methods.

[0004] Existing PPIS calculation methods are mainly divided into two categories: those based on sequence information and those based on structure information. Early methods relied heavily on amino acid sequence features, such as position-specific scoring matrices and hidden Markov models, combined with machine learning models (such as support vector machines and random forests) for prediction. However, these methods struggle to fully capture the long-range interactions between residues in the three-dimensional space of proteins. With the rise of Graph Neural Networks (GNNs), more and more researchers are applying them to protein structure modeling, treating residues as nodes in a graph and constructing edges based on spatial proximity relationships, thereby enabling information propagation and feature learning on the graph.

[0005] Nevertheless, existing graph neural network-based methods still have the following shortcomings: First, most methods rely solely on increasing the number of network layers to expand the receptive field of residue nodes in order to indirectly capture long-range inter-residue interactions and overall protein structural information, but this approach is difficult to achieve effective coordination of multi-scale protein structural information; second, existing methods often neglect the common structural features contained in residue clusters (local structural units formed by multiple residues), while these cluster structures play a key role in the formation of protein interaction interfaces; finally, existing models have limited ability to model the global topological structure of proteins and are unable to fully integrate local residue geometry, inter-residue associations, and overall scaffold information.

[0006] In summary, there is an urgent need for a PPIS prediction method that can coordinate multi-scale structural information of proteins, enhance the ability to perceive the structure of residue clusters, and integrate global topological features, so as to further improve prediction accuracy and robustness. Summary of the Invention

[0007] To address this issue, this invention provides a method and system for predicting protein interaction sites based on multi-scale structural coordination. This method solves the problem in existing technologies where it is difficult to coordinate multi-scale information such as local residue geometry, multi-residue association information, and overall protein topology, leading to insufficient accuracy in predicting protein interaction sites.

[0008] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting protein interaction sites based on multi-scale structural coordination, the method comprising the following steps: Step S1: Obtain the sequence information and three-dimensional structure information of the target protein, extract the initial features of each residue, and construct the adjacency matrix of the protein map based on the positional relationship of the residues in three-dimensional space. Step S2: Input the initial features and adjacency matrix into a pre-constructed multi-scale structural cooperative model. The multi-scale structural cooperative model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters, enhance inter-cluster information on a coarsened graph, and then backpool back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture topological associations between residues and global protein structural information. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused and then processed by a feedforward neural network to obtain high-order residue features. Step S3: Input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

[0009] Preferably, in step S1, the initial features extracted from each residue include sequence features and structural features; The sequence features include: position-specific scoring matrix and hidden Markov model features; The structural features include: defining the protein secondary structure matrix, residue atom features, and residue pseudo-position features; The definition of the protein secondary structure matrix is ​​achieved by encoding each amino acid as a 14-dimensional vector. The 14-dimensional vector includes: 9-dimensional one-hot secondary structure features, 4-dimensional sine and cosine values ​​of peptide backbone torsion angle, and 1-dimensional relative solubility of amino acids obtained by normalizing the maximum solubility surface area. The residue atom features are obtained in the following way: For all atoms in the residue, seven attributes of each atom are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electron charge, number of connected hydrogen atoms, whether it is in a ring, and van der Waals radius of the atom. The average value of each attribute on all atoms is calculated, and the seven average values ​​are used as the seven-dimensional residue atom features. The residue pseudo-position features are obtained by calculating the centroid coordinates of the side chain of each residue as the pseudo-position of the residue, taking the first residue in the protein sequence as a reference, calculating the distance between the pseudo-position of each residue and the reference residue, and using this distance as a 1-dimensional residue pseudo-position feature.

[0010] Preferably, in step S1, the specific method for constructing the adjacency matrix is ​​as follows: calculate the centroid coordinates of the side chain of each residue according to the protein PDB file, calculate the Euclidean distance between all residue pairs to obtain a distance map, set a distance threshold, set elements with a distance greater than the threshold to 0, and set elements with a distance less than or equal to the threshold to 1, thereby obtaining the adjacency matrix of the protein map.

[0011] Preferably, in step S2, the backbone of the equivariant graph neural network is formed by the residual connection of multiple equivariant graph convolutional layers; the update process of the equivariant graph convolutional layers for node features and node coordinates is as follows: ; ; ; ; in, , Represents a node , Input features, , Represents a node , coordinates , , This represents a multilayer perceptron used for edge, coordinate, and node operations; Represents a node , Message passing between them Represent edge features; The updated values ​​for the node coordinates. Represents the normalization factor. Represents the total number of nodes in the protein graph; It is a node The result after aggregation These are the node features output by the isomorphic graph neural network.

[0012] Preferably, in step S2, the clustering pooling information enhancement module sequentially executes the clustering pooling step, the graph convolution step, and the unpooling step; The clustering pooling step includes: For any two nodes in the protein graph , and its feature vector Input linear layer And add a bias term After activation by tanh function Calculate edge importance score Set threshold Filter out all that meet the requirements The edges form a subset of edges to be merged. ; The set of edges to be merged Based on this, a connected component detection algorithm is used to assign each node to a cluster. , forming cluster sequences The union of all clusters equals the original node set. Isolated nodes that have not merged with any other nodes form their own independent clusters; this yields the cluster allocation matrix. , The original number of nodes. For the number of clusters, Represents a node Belongs to cluster ; Constructing the weight matrix The rules for its element values ​​are as follows: If the edge ,but ; like And nodes If not merged, then ; otherwise ; Using the cluster allocation matrix Weight matrix Original node feature matrix and the original adjacency matrix Thicken the graph using the following formula: , , ; in This is the coarsened cluster node feature matrix. For the hidden layer dimension, Represents the original number of edges between clusters. Matrix truncation yields the adjacency matrix of the coarsened graph. and based on Reconstructing the edge set of the coarsened graph This completes the clustering and pooling step, yielding a coarsened image and its node features. The graph convolution step includes: The coarsened image is then input into a graph convolutional network for information propagation. The algorithm learns the structural relationships between clusters, enhances cluster node features, and outputs the enhanced cluster node features. The enhanced cluster node feature matrix is ​​the output of the graph convolutional layer. The cluster node feature matrix is ​​input to the graph convolutional layer. For graph convolutional network operations; The unpooling step includes: The enhanced cluster node features are depooled back into the original residue node features using tensor broadcasting to obtain enhanced features with the same number of nodes as the original.

[0013] Preferably, the structure-aware self-attention module incorporates the topological information contained in the multi-step random walk matrix into the self-attention calculation to capture the topological associations between residues and the global structural information of the protein, specifically including: Based on the adjacency matrix of the protein graph Calculate the degree matrix The transition matrix of the first-step random walk is obtained. ;Will Step 1 to The step power matrices are stacked along the third dimension to form a tensor. ,in , The number of nodes; For the tensor Each position corresponding dimensional vector Attention bias is calculated through linear mapping. ,in For learnable weight vectors, As the bias term, we obtain the attention bias matrix. ; For any two nodes eigenvectors Each is processed by a learnable weight matrix. Calculate the dot product after linear transformation and divide by the scaling factor. , The hidden layer dimension, plus the attention bias. Finally, the attention coefficients are obtained by normalization using the softmax function: .

[0014] Preferably, in the multi-scale structural collaborative model, the structure-aware self-attention module is connected in parallel with the backbone of the equivariant graph neural network, and the outputs of the two are added and fused before being input into the feedforward neural network; the clustering pooling information enhancement module is inserted once every two equivariant graph convolutional layers to enhance the backbone network's perception of the common features of multi-residue clusters.

[0015] Preferably, the method further includes a model training step: using a binary cross-entropy loss function as the optimization objective to update the parameters of the multi-scale structural collaborative model; and in the prediction evaluation after the model training is completed, using at least one of the following indicators to evaluate the prediction performance: accuracy, recall, precision, F1 score, area under the receiver operating characteristic curve, area under the precision-recall curve, and Matthews correlation coefficient.

[0016] This invention also provides a protein interaction site prediction system based on multi-scale structural coordination. This system is used to implement the aforementioned protein interaction site prediction method based on multi-scale structural coordination, and specifically includes: The feature extraction and graph construction module is used to obtain the sequence information and three-dimensional structural information of the target protein, extract the initial features of each residue, and construct the adjacency matrix of the protein graph based on the positional relationship of the residues in three-dimensional space. A multi-scale structural cooperative model module is used to input the initial features and adjacency matrix into a pre-constructed multi-scale structural cooperative model. The multi-scale structural cooperative model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters, enhance inter-cluster information on a coarsened graph, and then backpool back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture topological associations between residues and global protein structural information. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused and then processed by a feedforward neural network to obtain high-order residue features. The classification module is used to input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

[0017] This invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions to cause a computer device to execute the above-described method for predicting protein interaction sites based on multi-scale structural coordination.

[0018] As can be seen from the above technical solutions, this invention application has the following beneficial effects: First, multi-scale structural synergy significantly improves prediction accuracy: This invention extracts local geometric features of residues through the backbone of an isovariant graph neural network, extracts common features of multi-residue clusters through an intermittently inserted clustering pooling information enhancement module, and captures topological associations between residues and overall protein backbone information through a parallel access structure-aware self-attention module. This achieves multi-scale structural information synergy from local residues, residue clusters to the global protein, effectively solving the problem that existing methods are difficult to integrate multi-scale information and significantly improving the prediction accuracy of protein interaction sites.

[0019] Second, it enhances the ability to perceive the structure of residue clusters: This invention introduces an edge-based clustering pooling information enhancement module, which filters the subset of edges to be merged by calculating the importance score of the edges, uses the connected component detection algorithm to cluster spatially adjacent and similar residues into clusters, and performs graph convolution on the coarsened graph to enhance the information interaction between clusters before backpooling back to the original residues. This enables the model to effectively perceive the functional cluster structure formed by multiple residues, making up for the shortcomings of existing methods that ignore the common features of residue clusters.

[0020] Third, fully capture the global topological structure of the protein: This invention adopts a structure-aware self-attention module, which incorporates the topological information contained in the multi-step random walk matrix as an attention bias into the self-attention calculation. This enables the model to directly capture the long-range dependencies between residues and the overall topological structure of the protein without relying on network depth, thereby enhancing the model's efficiency in utilizing global structural information and further improving the robustness of interaction site prediction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Referring to the drawings will make the features and advantages of the present invention clearer. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a protein interaction site prediction method based on multi-scale structural coordination provided by the present invention; Figure 2 This is a schematic diagram of the overall framework of the multi-scale structural collaborative model in this invention; Figure 3 This is a schematic diagram of the initial residue characterization and protein map establishment in this invention; Figure 4 This is a schematic diagram of the clustering pooling process in the clustering pooling information enhancement module of this invention; Figure 5 This is a schematic diagram of the calculation process of the structure-aware self-attention module in this invention; Figure 6 This is a performance comparison chart of the method of this invention with other methods on the Test_60 dataset; Figure 7 This is a block diagram of a protein interaction site prediction system based on multi-scale structural coordination provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: To address the problem of insufficient accuracy in predicting protein-protein interaction sites due to the difficulty in coordinating multi-scale information such as local residue geometry, multi-residue association information, and overall protein topology in existing technologies. For example... Figure 1 As shown, this invention proposes a method for predicting protein interaction sites based on multi-scale structural coordination. This method includes the following steps: Step S1: Obtain the sequence information and three-dimensional structure information of the target protein (e.g., from a PDB file), extract the initial features of each residue, and construct the adjacency matrix of the protein map based on the positional relationship of the residues in three-dimensional space.

[0024] Specifically, such as Figure 3 As shown, the initial features extracted for each residue include sequence features and residue structural features. Wherein: Sequence features include position-specific scoring matrix (PSSM) and hidden Markov model (HMM) features.

[0025] Residue structure features: These include defining the protein secondary structure matrix (DSSP), residue atom features, and residue pseudo-position embedding. The definition of the protein secondary structure matrix is ​​achieved by encoding each amino acid as a 14-dimensional vector. The 14-dimensional vector includes: 9-dimensional one-hot secondary structure features, 4-dimensional sine and cosine values ​​of peptide backbone torsion angle, and 1-dimensional relative solubility of amino acids obtained by normalizing the maximum solubility surface area. The residue atom features are obtained in the following way: For all atoms in the residue, seven attributes of each atom are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electron charge, number of connected hydrogen atoms, whether it is in a ring, and van der Waals radius of the atom. The average value of each attribute on all atoms is calculated, and the seven average values ​​are used as the seven-dimensional residue atom features. The residue pseudo-position features are obtained by calculating the centroid coordinates of the side chain of each residue as the pseudo-position of the residue, taking the first residue in the protein sequence as a reference, calculating the distance between the pseudo-position of each residue and the reference residue, and using this distance as a 1-dimensional residue pseudo-position feature.

[0026] The specific method for constructing the adjacency matrix is ​​as follows: Calculate the centroid coordinates of the side chains of each residue based on the protein PDB file, calculate the Euclidean distance between all residue pairs, and obtain the protein distance map; set a distance threshold (in this embodiment, it is set to...). The adjacency matrix of the protein graph is obtained by setting elements with a distance greater than a threshold to 0 and elements with a distance less than or equal to the threshold to 1. .

[0027] Step S2: Input the initial features and adjacency matrix into a pre-constructed multi-scale structural cooperative model. The multi-scale structural cooperative model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters and enhance inter-cluster information on a coarsened graph before back-pooling back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture the topological associations between residues and the global structural information of the protein. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused in parallel and then processed by a feedforward neural network to obtain high-order residue features.

[0028] Specifically, the overall architecture of the multi-scale structural collaborative model is as follows: Figure 2 As shown, it includes an equivariant graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module.

[0029] (1) Backbone of Equivariant Graph Neural Network The backbone of the equivariant graph neural network is composed of residual connections between multiple equivariant graph convolutional layers. In this embodiment, the backbone contains 6 equivariant graph convolutional layers, with a hidden layer dimension of [missing information]. Set it to 64.

[0030] The update process for node features and node coordinates in each isomorphic graph convolutional layer is as follows: compute nodes with neighboring nodes Message vectors between: ; Update node coordinates (ensuring equivariance): ; Aggregate neighbor messages: ; Update node characteristics: ; in, , Represents a node , Input features, , Represents a node , coordinates , , This represents a multilayer perceptron used for edge, coordinate, and node operations; Represents a node , Message passing between them Representing edge features (in this embodiment, distance or learnable edge embedding can be selected); The updated values ​​for the node coordinates. Represents the normalization factor. Represents the total number of nodes in the protein graph; It is a node The result after aggregation These are the node features output by the isomorphic graph neural network.

[0031] (2) Clustering Pooling Information Enhancement Module This module sequentially executes clustering pooling, graph convolution, and depooling steps, and is inserted once every two equal-variable graph convolutional layers (i.e., after layers 2, 4, and 6) to enhance the backbone network's perception of common features of multi-residue clusters.

[0032] Clustering pooling steps, such as Figure 3 The following is stated: 1. For any two nodes in the protein graph , and its feature vector Input linear layer And add a bias term After activation by tanh function Calculate edge importance scores: ; Set threshold (In this embodiment) ), filter out all that meet the requirements The edges form a subset of edges to be merged. : .

[0033] 2. with Based on this, a connected component detection algorithm is used to assign each node to a cluster. , forming cluster sequences The union of all clusters equals the original node set. The isolated nodes that are not merged each form an independent cluster. This yields the cluster allocation matrix. , The original number of nodes. For the number of clusters, Represents a node Belongs to cluster .

[0034] 3. Construct the weight matrix The rules for its element values ​​are as follows: If the edge ,but ; like And nodes If not merged, then ; otherwise .

[0035] 4. Graph coarsening: Using the cluster allocation matrix... Weight matrix Original node feature matrix and the original adjacency matrix Thicken the graph using the following formula: , , ; in This is the coarsened cluster node feature matrix. For the hidden layer dimension, Represents the original number of edges between clusters. Matrix truncation yields the adjacency matrix of the coarsened graph. and based on Reconstructing the edge set of the coarsened graph This completes the clustering and pooling steps, yielding a coarsened image and its node features.

[0036] Graph convolution steps: The coarsened image is fed into a graph convolutional network (GCN) for information propagation: ; The system learns the structural relationships between clusters, enhances cluster node features, and outputs the enhanced cluster node features. The enhanced cluster node feature matrix is ​​the output of the graph convolutional layer. The cluster node feature matrix is ​​input to the graph convolutional layer. This refers to operations on graph convolutional networks.

[0037] Unpooling steps: The enhanced cluster node features are depooled back into the original residue node features using tensor broadcasting, resulting in enhanced features with the same number of nodes as the original. These enhanced features are then residually concatenated or spliced ​​with the output of the isomorphic graph convolutional layer to continue propagation in the subsequent backbone network.

[0038] (3) Structure-aware self-attention module This module is connected in parallel to the backbone of the isomorphic graph neural network to capture the topological associations between residues and global protein structural information. For example... Figure 5 As shown, the calculation process is as follows: Based on the adjacency matrix of the protein graph Calculate the degree matrix The transition matrix of the first-step random walk is obtained. .

[0039] Will Step 1 to The step power matrices are stacked along the third dimension to form a tensor. ,in , This represents the number of nodes. In this embodiment... .

[0040] tensor Each position corresponding dimensional vector The attention bias is calculated using a linear mapping: ; Obtain the attention bias matrix .

[0041] For any two nodes eigenvectors Each is processed by a learnable weight matrix. Calculate the dot product after linear transformation and divide by the scaling factor. , For the hidden layer dimension, plus attention bias Finally, the attention coefficients are obtained by normalization using the softmax function: .

[0042] This attention coefficient is used for subsequent node feature weighted aggregation, enabling the model to perceive the global topological structure between residues based on random walks.

[0043] Fusion method: The structure-aware self-attention module and the backbone of the equivariant graph neural network are connected in parallel. The outputs of the two are added and fused, and then input into the feedforward neural network (FFN) to further learn the fused features and obtain high-order residue features.

[0044] Step S3: Input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

[0045] Specifically, the higher-order residue features are input into a multilayer perceptron (MLP) to perform binary classification prediction on whether each residue is a protein interaction site, and output the probability that each residue belongs to the positive class (interaction site).

[0046] Furthermore, this embodiment also includes a model training step. A binary classification cross-entropy loss function is used as the optimization objective to update the parameters of the multi-scale structural collaborative model. Training employs five-fold cross-validation: the training data is randomly divided into five folds, four folds are used to train the model, one fold is used to evaluate performance, this process is repeated five times, and the average performance is calculated to adjust the hyperparameters.

[0047] The specific hyperparameter settings are as follows: Distance threshold for constructing protein maps: ; The backbone of the isomorphic graph neural network consists of 6 layers, with 64 hidden layers. Clustering pooling information enhancement module: Inserted once every two isomorphic graph convolutional layers, with edge scoring threshold. ; Structure-aware self-attention module: random walk steps Attention count = 4, Dropout rate = 0.3; Optimizer: Adam, initial learning rate = 0.01; Training epochs: 30 epochs; Learning rate scheduler: ReduceLROnPlateau, when AUPRC stops increasing for 6 consecutive epochs, the learning rate is adjusted to 0.6 times the current rate.

[0048] In the prediction evaluation after model training, the following metrics are used to evaluate model performance: Accuracy (ACC), Recall, Precision, F1 score (F1), Area Under the Receiver Operating Characteristic (AUROC), Area Under the Precision-Recall (AUPRC), and Matthews Correlation Coefficient (MCC). The formulas for each metric are as follows: ; ; ; ; .

[0049] in, A true positive indicates a correctly predicted interaction site; A true negative result indicates a correctly predicted non-interacting site; A false positive indicates that the interaction site was predicted incorrectly. A false negative indicates an incorrect prediction of a non-interacting site.

[0050] This embodiment was validated on the publicly available benchmark test set Test_60 (containing 60 protein chains). Experimental results are as follows: Figure 6 As shown, the method of this invention (MSC-PPIS) significantly outperforms existing methods in multiple metrics such as ACC, Recall, Precision, F1, AUROC, AUPRC, and MCC, demonstrating the effectiveness of the multi-scale structural synergy mechanism.

[0051] Example 2: like Figure 7 As shown, this invention provides a protein interaction site prediction system based on multi-scale structural coordination. This system is used to implement the protein interaction site prediction method based on multi-scale structural coordination described in Embodiment 1 above, and specifically includes: The feature extraction and graph construction module 100 is used to obtain the sequence information and three-dimensional structural information of the target protein, extract the initial features of each residue, and construct the adjacency matrix of the protein graph based on the positional relationship of the residues in three-dimensional space. A multi-scale structural coordination model module 200 is used to input the initial features and adjacency matrix into a pre-constructed multi-scale structural coordination model. The multi-scale structural coordination model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters, enhance inter-cluster information on a coarsened graph, and then backpool back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture topological associations between residues and global protein structural information. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused and then processed by a feedforward neural network to obtain high-order residue features. The classification module 300 is used to input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

[0052] This embodiment provides a protein interaction site prediction system based on multi-scale structural coordination, used to implement the aforementioned protein interaction site prediction method based on multi-scale structural coordination. Therefore, the specific implementation of the protein interaction site prediction system based on multi-scale structural coordination can be found in the previous embodiment section of the protein interaction site prediction method based on multi-scale structural coordination. For example, the feature extraction and graph construction module 100, the multi-scale structural coordination model module 200, and the classification module 300 are used to implement steps S1, S2, and S3 in the aforementioned protein interaction site prediction method based on multi-scale structural coordination, respectively. Therefore, the specific implementation can be referred to the description of the corresponding embodiments. To avoid redundancy, it will not be repeated here.

[0053] Example 3: This invention provides a computer storage medium storing a computer software product, which includes several instructions to cause a computer device to execute the above-described method for predicting protein interaction sites based on multi-scale structural coordination.

[0054] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting protein-protein interaction sites based on multi-scale structural coordination, characterized in that, Includes the following steps: Step S1: Obtain the sequence information and three-dimensional structure information of the target protein, extract the initial features of each residue, and construct the adjacency matrix of the protein map based on the positional relationship of the residues in three-dimensional space. Step S2: Input the initial features and adjacency matrix into a pre-constructed multi-scale structural cooperative model. The multi-scale structural cooperative model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters, enhance inter-cluster information on a coarsened graph, and then backpool back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture topological associations between residues and global protein structural information. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused and then processed by a feedforward neural network to obtain high-order residue features. Step S3: Input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

2. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, In step S1, the initial features extracted for each residue include sequence features and structural features; The sequence features include: position-specific scoring matrix and hidden Markov model features; The structural features include: defining the protein secondary structure matrix, residue atom features, and residue pseudo-position features; The definition of the protein secondary structure matrix is ​​achieved by encoding each amino acid as a 14-dimensional vector. The 14-dimensional vector includes: 9-dimensional one-hot secondary structure features, 4-dimensional sine and cosine values ​​of peptide backbone torsion angle, and 1-dimensional relative solubility of amino acids obtained by normalizing the maximum solubility surface area. The residue atom features are obtained in the following way: For all atoms in the residue, seven attributes of each atom are extracted: atomic mass, B factor, whether it is a side chain atom of the residue, electron charge, number of connected hydrogen atoms, whether it is in a ring, and van der Waals radius of the atom. The average value of each attribute on all atoms is calculated, and the seven average values ​​are used as the seven-dimensional residue atom features. The residue pseudo-position features are obtained by calculating the centroid coordinates of the side chain of each residue as the pseudo-position of the residue, taking the first residue in the protein sequence as a reference, calculating the distance between the pseudo-position of each residue and the reference residue, and using this distance as a 1-dimensional residue pseudo-position feature.

3. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, In step S1, the specific method for constructing the adjacency matrix is ​​as follows: calculate the centroid coordinates of the side chain of each residue according to the protein PDB file, calculate the Euclidean distance between all residue pairs to obtain a distance map, set a distance threshold, set elements with a distance greater than the threshold to 0, and set elements with a distance less than or equal to the threshold to 1, thereby obtaining the adjacency matrix of the protein map.

4. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, In step S2, the backbone of the equivariant graph neural network is formed by the residual connections of multiple equivariant graph convolutional layers; the update process of the equivariant graph convolutional layers for node features and node coordinates is as follows: ； ； ； ； in, , Represents a node , Input features, , Represents a node , coordinates , , This represents a multilayer perceptron used for edge, coordinate, and node operations; Represents a node , Message passing between them Represent edge features; The updated values ​​for the node coordinates. Represents the normalization factor. Represents the total number of nodes in the protein graph; It is a node The result after aggregation These are the node features output by the isomorphic graph neural network.

5. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, In step S2, the clustering pooling information enhancement module sequentially executes the clustering pooling step, the graph convolution step, and the unpooling step; The clustering pooling step includes: For any two nodes in the protein graph , and its feature vector Input linear layer And add a bias term After activation by tanh function Calculate edge importance score Set threshold Filter out all that meet the requirements The edges form a subset of edges to be merged. ; The set of edges to be merged Based on this, a connected component detection algorithm is used to assign each node to a cluster. , forming cluster sequences The union of all clusters equals the original node set. Isolated nodes that have not merged with any other nodes form their own independent clusters; this yields the cluster allocation matrix. , The original number of nodes. For the number of clusters, Represents a node Belongs to cluster ; Constructing the weight matrix The rules for its element values ​​are as follows: If the edge ,but ; like And nodes If not merged, then ; otherwise ; Using the cluster allocation matrix Weight matrix Original node feature matrix and the original adjacency matrix Thicken the graph using the following formula: ， ， ； in This is the coarsened cluster node feature matrix. For the hidden layer dimension, Represents the original number of edges between clusters. Matrix truncation yields the adjacency matrix of the coarsened graph. and based on Reconstructing the edge set of the coarsened graph This completes the clustering and pooling step, yielding a coarsened image and its node features. The graph convolution step includes: The coarsened image is then input into a graph convolutional network for information propagation. The algorithm learns the structural relationships between clusters, enhances cluster node features, and outputs the enhanced cluster node features. The enhanced cluster node feature matrix is ​​the output of the graph convolutional layer. The cluster node feature matrix is ​​input to the graph convolutional layer. For graph convolutional network operations; The unpooling step includes: The enhanced cluster node features are depooled back into the original residue node features using tensor broadcasting to obtain enhanced features with the same number of nodes as the original.

6. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, The structure-aware self-attention module incorporates the topological information contained in the multi-step random walk matrix into the self-attention calculation to capture the topological associations between residues and the global structural information of the protein, specifically including: Based on the adjacency matrix of the protein map Calculate the degree matrix The transition matrix of the first-step random walk is obtained. ;Will Step 1 to The step power matrices are stacked along the third dimension to form a tensor. ,in , The number of nodes; For the tensor Each position corresponding dimensional vector Attention bias is calculated through linear mapping. ,in For learnable weight vectors, As the bias term, we obtain the attention bias matrix. ; For any two nodes eigenvectors Each is processed by a learnable weight matrix. Calculate the dot product after linear transformation and divide by the scaling factor. , The hidden layer dimension, plus the attention bias. Finally, the attention coefficients are obtained by normalization using the softmax function: 。 7. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, In the multi-scale structural collaborative model, the structure-aware self-attention module is connected in parallel with the backbone of the equivariant graph neural network, and the outputs of the two are added and fused before being input into the feedforward neural network; the clustering pooling information enhancement module is inserted once every two equivariant graph convolutional layers to enhance the backbone network's perception of the common features of multi-residue clusters.

8. The method for predicting protein interaction sites based on multi-scale structural coordination according to claim 1, characterized in that, The method further includes a model training step: using a binary cross-entropy loss function as the optimization objective to update the parameters of the multi-scale structural collaborative model; in the prediction evaluation after the model training is completed, at least one of the following indicators is used to evaluate the prediction performance: accuracy, recall, precision, F1 score, area under the receiver operating characteristic curve, area under the precision-recall curve, and Matthews correlation coefficient.

9. A protein-protein interaction site prediction system based on multi-scale structural coordination, characterized in that, The system is used to implement the protein interaction site prediction method based on multi-scale structural coordination as described in any one of claims 1 to 8, specifically including: The feature extraction and graph construction module is used to obtain the sequence information and three-dimensional structural information of the target protein, extract the initial features of each residue, and construct the adjacency matrix of the protein graph based on the positional relationship of the residues in three-dimensional space. A multi-scale structural cooperative model module is used to input the initial features and adjacency matrix into a pre-constructed multi-scale structural cooperative model. The multi-scale structural cooperative model includes an isomorphic graph neural network backbone, a clustering pooling information enhancement module, and a structure-aware self-attention module. The isomorphic graph neural network backbone extracts local geometric features of residues layer by layer. The clustering pooling information enhancement module is intermittently inserted into the backbone network to cluster multiple residues into clusters, enhance inter-cluster information on a coarsened graph, and then backpool back to the original residues. The structure-aware self-attention module is connected in parallel with the isomorphic graph neural network backbone to capture topological associations between residues and global protein structural information. The outputs of the isomorphic graph neural network backbone and the structure-aware self-attention module are fused and then processed by a feedforward neural network to obtain high-order residue features. The classification module is used to input the higher-order residue features into a multilayer perceptron and perform binary classification prediction on whether each residue is a protein-protein interaction site.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, which includes several instructions for causing a computer device to execute the protein interaction site prediction method based on multi-scale structural coordination as described in any one of claims 1 to 8.

Images