A method and system for protein site prediction using dynamic feature perception and hierarchical isomorphic graph networks

Abstract

This invention relates to the field of protein structure analysis technology, specifically to a method and system for protein site prediction using dynamic feature perception and hierarchical isotropic graph networks. The method includes acquiring the amino acid sequence and three-dimensional structural information of the protein to be predicted, extracting residue-level multi-source features and constructing a residue space map; adaptively weighting the multi-source features based on context relevance through a dynamic feature importance perception mechanism to generate a context-aware residue representation; inputting the residue representation and three-dimensional coordinates into a hierarchical isotropic graph neural network module, and performing multi-level feature extraction using graph attention pooling and gating fusion mechanisms; and outputting the probability value of each residue belonging to a protein-protein interaction site based on the fused residue representation. This invention, while maintaining spatial isotropy, can accurately learn multi-scale spatial features, significantly improving the prediction accuracy of protein-protein interaction sites, especially showing outstanding advantages in the identification of isolated sites and continuous regions.

Description

Technical Field

[0001] This invention relates to the field of protein structure analysis technology, specifically to a method and system for predicting protein sites using dynamic feature perception and hierarchical isomorphic graph networks. Background Technology

[0002] Protein-protein interactions (PPIs) are fundamental to the execution of most intracellular life activities, and the identification of their interaction sites is of vital importance for understanding disease mechanisms, drug target discovery, and protein function annotation.

[0003] Currently, while experimental methods such as X-ray crystallography and nuclear magnetic resonance can accurately determine interaction sites, these methods are typically expensive and time-consuming. With the development of deep learning technology, computational prediction-based methods have gradually become a research hotspot. Existing computational methods have evolved from sequence-based models to using graph neural networks (GNNs) to model protein structures.

[0004] However, existing technologies still have the following drawbacks: 1. Most methods use a static assembly approach for multiple feature sets (such as PSSM, HMM, etc.), which makes it difficult to adaptively distinguish the importance of different feature sets based on the contextual relationship of the protein sequence.

[0005] 2. Insufficient perception of the complex layered three-dimensional spatial geometry of proteins, and the tendency for oversmoothing to occur during propagation in deep networks, resulting in insufficient characterization ability of local key sites.

[0006] 3. It has weak ability to identify isolated interaction sites or continuous regions, and lacks an effective interaction mechanism for cross-scale information. Summary of the Invention

[0007] The purpose of this invention is to propose a protein site prediction method and system based on dynamic feature perception and hierarchical isomorphic graph networks, in order to solve the problems of weak static fusion capability of multi-source features, insufficient three-dimensional geometric structure modeling, and insufficient cross-scale information interaction in the existing technology.

[0008] According to a first aspect of the embodiments of this disclosure, a method for protein site prediction using dynamic feature perception and hierarchical isomorphic graph networks is provided, comprising the following steps: Obtain the amino acid sequence and three-dimensional structure information of the protein to be predicted, extract residue-level multi-source features, and construct a residue space map; The multi-source features are subjected to context-dependent adaptive weighting through a dynamic feature importance-aware mechanism to obtain a context-aware residue representation. The context-aware residue representation and three-dimensional coordinates are input into a hierarchical E(n) variable graph neural network module, and multi-level feature extraction is performed by combining graph attention pooling mechanism and gating fusion mechanism to learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues. Based on the residue representation after multi-level gating fusion, a probability value is generated for each residue belonging to a protein-protein interaction site.

[0009] In one embodiment, the residue-level multi-source features include at least the position-specific scoring matrix feature PSSM, the hidden Markov model feature HMM, the protein secondary structure definition feature DSSP, the atomic physicochemical feature AF, and the spatial position feature PF.

[0010] In one embodiment, the specific method for constructing the graph structure is as follows: each residue in the protein is used as a graph node, and the Euclidean distance between the Cα atoms of the residues is used as the basis for edge connection. When the Euclidean distance between any two Cα atoms of residues is less than or equal to a preset value, an undirected edge is established between the corresponding residue nodes, and the three-dimensional spatial coordinates of each residue are used as the geometric input features of the hierarchical E(n) equivariant graph neural network module.

[0011] In one embodiment, the method for obtaining the residue representation specifically includes: Perform mean pooling on each feature group to extract the global context; The importance weights of each feature group are learned through a multilayer perceptron containing hidden layers. : Where σ represents the Sigmoid activation function, and Represents the learnable weight matrix. and Indicates the bias term. Indicates the current feature group input; By combining the scaling factor β with the feature weights for smooth adjustment, we obtain the weighted feature h′: The weighted multi-source features are fused using the learnable residual fusion parameter γ. With original features H By performing fusion, a context-aware residue representation is obtained. : Where γ is a learnable residual fusion parameter.

[0012] In one embodiment, the hierarchical E(n) equivariant graph neural network module consists of multiple E(n) equivariant graph convolutional layers; the graph attention pooling mechanism uses a multi-head self-attention approach to obtain node importance scores while maintaining the original edge relationships between the selected nodes.

[0013] In one embodiment, the gating fusion mechanism includes intra-layer gating fusion and multi-level gating fusion; the intra-layer gating fusion is used to fuse the previous state representation with the current isovariant geometric update representation; the multi-level gating fusion is used to fuse global average information and local peak information in subgraph representations at different levels.

[0014] In one embodiment, in the intra-layer gating fusion and multi-level gating fusion, the gating coefficient... for: Where σ represents the Sigmoid activation function, Represents the learnable weight matrix. Indicates the bias term. This indicates that two features are concatenated; a fused representation is obtained based on the gating coefficient. : In the intra-layer gated fusion, This represents the state of the previous layer. This represents the geometric update representation obtained from the current equivariant graph convolutional layer; in the multi-level gated fusion, the set of subgraph representations from different levels {H} is aggregated. s}, and will Set as global average information mean(H) s ),Will Set to local peak information max(H) s Then, the above-mentioned gating fusion method is used for joint fusion.

[0015] According to a second aspect of the present disclosure, a protein site prediction system based on dynamic feature perception and hierarchical isomorphic graph networks is provided, comprising: The data preparation and graph construction module obtains the amino acid sequence information and three-dimensional structure information of the protein to be predicted, extracts residue-level multi-source features, and constructs a residue space map. The dynamic feature perception module performs context-dependent adaptive weighting on the multi-source features through a dynamic feature importance perception mechanism to obtain a context-aware residue representation. The hierarchical feature extraction module inputs the context-aware residue representation and three-dimensional coordinates into the hierarchical E(n) variable graph neural network module, and performs multi-level feature extraction by combining graph attention pooling mechanism and gating fusion mechanism to learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues. The site prediction output module generates a probability value for each residue belonging to a protein-protein interaction site based on the residue representation after multi-level gating fusion.

[0016] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the protein site prediction method of dynamic feature perception and hierarchical isomorphic graph network.

[0017] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the protein site prediction method of the dynamic feature perception and hierarchical isomorphic graph network.

[0018] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention introduces a dynamic feature-aware mechanism, overcoming the limitations of static integration caused by the simple splicing of multi-source features (such as PSSM, HMM, etc.) in traditional methods. Through context-dependent adaptive weighting, the importance of different feature groups can be dynamically adjusted according to the specific environment of the protein sequence. This mechanism can more accurately capture the key attributes of residues and significantly improve the discriminative power of feature representation.

[0019] 2. This invention constructs a hierarchical E(n) equivariant graph neural network framework, which can effectively extract hierarchical three-dimensional spatial geometric information of protein structures. Unlike traditional graph neural networks that only focus on local neighborhoods, this invention learns multi-scale spatial features through a hierarchical architecture, preserving both local fine structural information and capturing global topological constraints. This deep modeling of complex geometric constraints provides a more comprehensive understanding of protein three-dimensional structures.

[0020] 3. The E(n) equivariant graph neural network used in this invention ensures strict equivariance of the model during translation and rotation operations in 3D space. This means that regardless of how the protein is positioned in space, the feature representation and prediction results extracted by the model remain consistent. This characteristic not only conforms to physical laws but also greatly enhances the model's robustness and generalization ability when processing data with different conformations.

[0021] 4. This invention provides more accurate and reliable prediction results for protein-protein interaction sites. It exhibits particularly outstanding performance in tasks involving the identification of isolated sites (dispersed active sites) and continuous regions (large-area binding interfaces). This solves the problem of easily missing isolated sites or having unclear boundaries in existing technologies.

[0022] 5. Through dynamic feature weighting and hierarchical structure learning, this invention not only performs well in predictive metrics, but also reveals the contribution of different feature groups to specific binding sites and key spatial structural patterns, thus providing a biologically explanatory basis for understanding the molecular mechanisms of protein interactions. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0024] Figure 1 This is a schematic diagram illustrating the principle of a protein site prediction method based on dynamic feature perception and hierarchical isomorphic graph networks. Detailed Implementation

[0025] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0029] Example 1: like Figure 1 As shown, this embodiment provides a protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks, including the following steps: S1. Obtain the amino acid sequence information and three-dimensional structure information of the protein to be predicted, extract residue-level multi-source features and construct a residue space map; The multi-source features include at least 20-dimensional position-specific scoring matrix features (PSSM), 20-dimensional hidden Markov model features (HMM), 14-dimensional protein secondary structure definition features (DSSP), 7-dimensional atomic physicochemical features (AF), and 2-dimensional spatial position features (PF).

[0030] Specifically, in this embodiment, the A chain of protein 2v9t is used as an example to obtain the primary amino acid sequence and three-dimensional structural coordinates of the protein chain. Subsequently, five sets of features are extracted for each residue in the protein chain: position-specific scoring matrix (PSSM), hidden Markov model (HMM), protein secondary structure definition (DSSP), atomic physicochemical features (AF), and spatial position features (PF). These features are then concatenated in residue order to form a 63-dimensional initial feature vector, which is used as input for subsequent models.

[0031] An undirected graph G(V, E) is constructed using each residue in the protein as a graph node and the spatial proximity between residues as graph edges. Edge connections are established based on the Euclidean distance between Cα atoms of the residues. When the distance is not greater than a preset value (such as 14 Å), graph edges are established between the corresponding residue nodes, and the three-dimensional coordinates of the residues are used as the geometric input of the graph neural network.

[0032] Specifically, for the A chain of 2v9t, each residue is considered as a graph node, and the node coordinates are the three-dimensional spatial coordinates of the corresponding Cα atom. The Euclidean distance between any two Cα atoms of residues is obtained. When the distance is less than or equal to 14 Å, an undirected edge is established between the corresponding nodes, thereby obtaining the residue-level spatial graph structure G(V, E) for protein-protein interaction site prediction.

[0033] S2. The multi-source features are subjected to context-dependent adaptive weighting through a dynamic feature importance-aware mechanism to obtain a context-aware residue representation; The obtained multi-source features are input into the dynamic embedding module, and context-dependent adaptive weighting is performed on different feature groups. Specifically, mean pooling is first performed on each feature group to extract the global context, then the importance weights of each feature group are learned through a multilayer perceptron (MLP) with hidden layers, and the feature weights are smoothly adjusted by combining a scaling factor. Finally, the weighted result is fused with the original features through learnable residual fusion parameters to obtain a context-aware residue representation.

[0034] Specifically, the five residue features corresponding to the A chain of 2v9t are input into the dynamic embedding module. For the current feature group input h, its importance weights are learned through a weight recalibration network. : Where σ represents the Sigmoid activation function, W1 and W2 represent learnable weight matrices, b1 and b2 represent bias terms, and h represents the current feature group input. Preferably, the multilayer perceptron includes a single hidden layer with a dimension of 32, and Dropout is added after the ReLU activation function, with a preferred Dropout rate of 0.1.

[0035] To avoid drastic fluctuations in feature weights that could lead to training instability, this invention introduces a scaling factor β to smoothly adjust the feature weights. For each set of feature inputs h, the importance weights are adjusted according to the aforementioned weight recalibration network. Obtain the corresponding weighted feature h': Subsequently, multiple weighted features h′ are combined to form an overall weighted feature H′, and H′ is fused with the original feature H using a learnable residual fusion parameter γ to obtain a context-aware residue representation. : Here, γ is a learnable residual fusion parameter, preferably initialized to 0.3. Through the above dynamic embedding, while preserving the original multi-source feature information, it is possible to adaptively highlight feature groups more relevant to protein-protein interaction site prediction, thereby improving the discriminative ability of residue representation.

[0036] S3. Input the context-aware residue representation and three-dimensional coordinates into the hierarchical E(n) variable graph neural network module, and combine graph attention pooling mechanism and gating fusion mechanism to perform multi-level feature extraction, and learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues; The hierarchical E(n) isovariant graph neural network module includes multiple E(n) isovariant graph convolutional layers. In each layer, node features and node coordinates are updated simultaneously, thereby learning the three-dimensional geometric constraints of proteins and the interaction relationships between residues while maintaining spatial isovariability.

[0037] Specifically, the context-aware residue representation Hout corresponding to the A chain of 2v9t is input together with the node coordinates into a hierarchical E(n) equivariant graph neural network module. Preferably, the module contains 6 layers of E(n) equivariant graph convolutional layers; in each layer, a message is generated based on the current features of the node, the squared Euclidean distance between nodes, and the edge features, and the residue node features and spatial coordinates are updated synchronously, thereby maintaining the equivariant modeling capability for geometric transformations such as translation and rotation in the model.

[0038] In the hierarchical E(n) variable graph neural network module, a graph attention pooling mechanism is introduced to score the importance of nodes and select key nodes according to a preset pooling ratio to construct a sub-layer hierarchical representation. Among them, multi-head self-attention is preferred to obtain the node importance score and maintain the original edge relationship between the selected nodes to form a hierarchical structure representation from the whole graph to the subgraph.

[0039] Specifically, a graph attention pooling mechanism is used to score the importance of each residue node on the 2v9t A-chain graph structure, and key nodes are selected based on the scoring results to form a sub-layer hierarchical representation. Preferably, the pooling ratio is set to [1, 0.5], that is, the first layer retains the full graph representation, and subsequent layers select key nodes proportionally; the scoring method preferably adopts a multi-head self-attention mechanism, and the top k nodes with the highest scores are retained, while maintaining the original adjacency relationship between the retained nodes, thereby realizing hierarchical modeling from the full graph to the subgraph.

[0040] In the hierarchical representation learning process, intra-layer gated fusion and multi-level gated fusion are set up. Intra-layer gated fusion is used to fuse the previous state representation with the current isovariant geometric update representation, while multi-level gated fusion is used to fuse global average information and local peak information in subgraph representations of different levels, thereby suppressing oversmoothing, reducing information loss and enhancing the expressive power of cross-scale structural patterns.

[0041] Specifically, in the hierarchical representation learning process of the A-chain in 2v9t, intra-layer gated fusion is first performed after each layer's isovariant update, adaptively fusing the previous layer's state representation H1 with the current geometric update representation H2. For any two feature representations H1 and H2 to be fused, the gating coefficients G are first obtained. fusion : Where σ represents the Sigmoid activation function, W3 represents the learnable weight matrix, b3 represents the bias term, and [H1 | | H2] represents the concatenation of the two feature representations. The fused representation H is obtained based on the gating coefficients. f Its expression is: In intra-layer gated fusion, H1 represents the state representation of the previous layer, and H2 represents the geometric update representation obtained by the current equivariant graph convolutional layer; S4. Based on the residue representation after multi-level gating fusion, generate the probability value of each residue belonging to a protein-protein interaction site.

[0042] In multi-level gating fusion, the set of subgraph representations at different levels {H} is first aggregated. s}, and set H1 as the global average information mean(H s Set H2 to the local peak information max(H s Then, the data is jointly fused using the gating fusion method described above. This gating fusion mechanism can adaptively allocate fusion weights according to the importance of information at different levels, thereby reducing information loss during hierarchical pooling and deep propagation, alleviating the oversmoothing problem, and enhancing the model's ability to express local key interaction sites and global structural patterns.

[0043] The obtained fused residues are used to represent the input and output prediction modules. The output prediction module is preferably a multilayer perceptron (MLP). The probability value of each residue belonging to a protein-protein interaction site is output by the sigmoid function, and the model training and parameter optimization are completed by the cross-entropy loss function.

[0044] Specifically, the A-chain residue representations of the fused 2v9t are input into the multilayer perceptron prediction module, and the Sigmoid activation function outputs the predicted probability that each residue belongs to a protein-protein interaction site. During the training phase, the cross-entropy loss function and the Adam optimizer are preferably used for parameter updates, with a learning rate preferably set to 5.00 × 10⁻⁶. -4The preferred training epochs are 50, and the preferred batch size is 1. The model predicts 15 true positives (TP), 72 true negatives (TN), 12 false positives (FP), and 8 false negatives (FN).

[0045] To verify the overall performance of the method of this invention, it was compared with AGAT-PPIS and MEG-PPIS on the Test 60 independent test set for protein-protein interaction site prediction. As shown in Table 1, the method of this invention outperforms AGAT-PPIS and MEG-PPIS in terms of accuracy (ACC), F1 score (F1), Matthews correlation coefficient (MCC), and area under the precision-recall curve (AUPRC). This indicates that the method of this invention has better overall performance in the task of predicting protein-protein interaction sites.

[0046] Table 1. Performance comparison of different models on the Test 60 dataset Example 2: This embodiment provides a protein site prediction system based on dynamic feature perception and hierarchical isomorphic graph networks, including: The data preparation and graph construction module obtains the amino acid sequence information and three-dimensional structure information of the protein to be predicted, extracts residue-level multi-source features, and constructs a residue space map. The dynamic feature perception module performs context-dependent adaptive weighting on the multi-source features through a dynamic feature importance perception mechanism to obtain a context-aware residue representation. The hierarchical feature extraction module inputs the context-aware residue representation and three-dimensional coordinates into the hierarchical E(n) variable graph neural network module, and performs multi-level feature extraction by combining graph attention pooling mechanism and gating fusion mechanism to learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues. The site prediction output module generates a probability value for each residue belonging to a protein-protein interaction site based on the residue representation after multi-level gating fusion.

[0047] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0048] Example 3: An electronic device is provided for running the aforementioned "Protein Site Prediction Method Based on Dynamic Feature Perception and Hierarchical Isovariant Graph Network". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S4 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain the amino acid sequence information and three-dimensional structure information of the protein to be predicted, extract residue-level multi-source features and construct a residue space map; S2. The multi-source features are subjected to context-dependent adaptive weighting through a dynamic feature importance-aware mechanism to obtain a context-aware residue representation; S3. Input the context-aware residue representation and three-dimensional coordinates into the hierarchical E(n) variable graph neural network module, and combine graph attention pooling mechanism and gating fusion mechanism to perform multi-level feature extraction, and learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues; S4. Based on the residue representation after multi-level gating fusion, generate the probability value of each residue belonging to a protein-protein interaction site.

[0049] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0050] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to perform the method steps S1 to S4 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0051] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0053] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks, characterized in that, Includes the following steps: Obtain the amino acid sequence and three-dimensional structure information of the protein to be predicted, extract residue-level multi-source features, and construct a residue space map; The multi-source features are subjected to context-dependent adaptive weighting through a dynamic feature importance-aware mechanism to obtain a context-aware residue representation. The context-aware residue representation and three-dimensional coordinates are input into a hierarchical E(n) variable graph neural network module, and multi-level feature extraction is performed by combining graph attention pooling mechanism and gating fusion mechanism to learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues. Based on the residue representation after multi-level gating fusion, a probability value is generated for each residue belonging to a protein-protein interaction site.

2. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, The residue-level multi-source features include at least the position-specific scoring matrix feature PSSM, the hidden Markov model feature HMM, the protein secondary structure definition feature DSSP, the atomic physicochemical feature AF, and the spatial position feature PF.

3. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, The specific method for constructing the graph structure is as follows: each residue in the protein is used as a graph node, and the Euclidean distance between the Cα atoms of the residues is used as the basis for edge connection. When the Euclidean distance between any two Cα atoms of residues is less than or equal to a preset value, an undirected edge is established between the corresponding residue nodes, and the three-dimensional spatial coordinates of each residue are used as the geometric input features of the hierarchical E(n) equivariant graph neural network module.

4. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, The methods for obtaining residue representations include: Perform mean pooling on each feature group to extract the global context; The importance weights of each feature group are learned through a multilayer perceptron containing hidden layers. : Where σ represents the Sigmoid activation function, and Represents the learnable weight matrix. and Indicates the bias term. Indicates the current feature group input; By combining the scaling factor β with the feature weights for smooth adjustment, we obtain the weighted feature h′: The weighted multi-source features are fused using the learnable residual fusion parameter γ. With original features H By performing fusion, a context-aware residue representation is obtained. : Where γ is a learnable residual fusion parameter.

5. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, The hierarchical E(n) equivariant graph neural network module consists of multiple E(n) equivariant graph convolutional layers; the graph attention pooling mechanism uses a multi-head self-attention approach to obtain node importance scores while maintaining the original edge relationships between the selected nodes.

6. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, The gating fusion mechanism includes intra-layer gating fusion and multi-level gating fusion; the intra-layer gating fusion is used to fuse the previous state representation with the current isovariant geometric update representation; the multi-level gating fusion is used to fuse global average information and local peak information in subgraph representations at different levels.

7. The protein site prediction method using dynamic feature perception and hierarchical isomorphic graph networks according to claim 1, characterized in that, In the intra-layer gated fusion and multi-level gated fusion, the gate coefficient is... for: Where σ represents the Sigmoid activation function, Represents the learnable weight matrix. Indicates the bias term. This indicates that two features are concatenated; a fused representation is obtained based on the gating coefficient. : In the intra-layer gated fusion, This represents the state of the previous layer. This represents the geometric update representation obtained from the current equivariant graph convolutional layer; in the multi-level gated fusion, the set of subgraph representations from different levels {H} is aggregated. s }, and will Set as global average information mean(H) s ),Will Set to local peak information max(H) s Then, the above-mentioned gating fusion method is used for joint fusion.

8. A protein site prediction system based on dynamic feature perception and hierarchical isomorphic graph networks, characterized in that, include: The data preparation and graph construction module obtains the amino acid sequence information and three-dimensional structure information of the protein to be predicted, extracts residue-level multi-source features, and constructs a residue space map. The dynamic feature perception module performs context-dependent adaptive weighting on the multi-source features through a dynamic feature importance perception mechanism to obtain a context-aware residue representation. The hierarchical feature extraction module inputs the context-aware residue representation and three-dimensional coordinates into the hierarchical E(n) variable graph neural network module, and performs multi-level feature extraction by combining graph attention pooling mechanism and gating fusion mechanism to learn the three-dimensional geometric structure constraints of the protein and the interaction relationship between residues. The site prediction output module generates a probability value for each residue belonging to a protein-protein interaction site based on the residue representation after multi-level gating fusion.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the protein site prediction method of dynamic feature perception and hierarchical isomorphic graph network as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the protein site prediction method of dynamic feature perception and hierarchical isomorphic graph network as described in any one of claims 1-7.

Images