Protein site prediction method based on hierarchical graph and geometric vector perceptron
Through the method based on hierarchical graph and geometric vector perceptron, the problem of insufficient multi-scale information and symmetry processing in protein site prediction is solved, and more accurate and efficient protein site prediction is achieved.
Patent Information
- Application Number
- CN202411194410.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-29
AI Technical Summary
The existing graph neural network methods fail to fully consider the multi-scale molecular information and the symmetry of protein molecules in three-dimensional space in protein site prediction, resulting in poor prediction results.
Using a method based on hierarchical graphs and geometric vector perceptrons, a graph structure of atomic hierarchy and residue hierarchy is constructed, and aggregation and update of node features is combined with geometric vector perceptrons to process multi-scale information and spatial denaturation of proteins.
It significantly improves the accuracy and efficiency of protein site prediction, and can learn more comprehensively the different scale patterns and spatial structural characteristics of proteins, improving the prediction effect.
Smart Images

Figure CN118692564B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of bioinformatics, and particularly relates to a protein site prediction method based on a hierarchical graph and a geometric vector perceptron. Background Art
[0002] In the study of protein action, identifying binding sites in protein-protein interactions is a key issue, which is crucial to revealing the functional mechanism of proteins and their application in medical research. In recent years, many researchers have focused on deep learning-based methods to achieve rapid prediction of protein sites, among which graph neural networks (GNNs) have shown excellent performance. However, current graph neural network methods have shortcomings in site prediction, especially the insufficient modeling of information at different scales of protein molecules, and the incomplete treatment of the natural symmetry of protein molecules in three-dimensional space.
[0003] Protein-protein interactions are an important area of biological research. Such interactions play a key role in processes such as cell signaling, metabolic regulation, and cell cycle. A deeper understanding of protein-protein interactions helps to resolve the complex molecular networks within cells, reveal disease mechanisms, and promote the development of new drugs. One of the related research areas is the identification of protein-protein interaction sites. These sites, i.e., the interface residues between interacting proteins, not only determine how proteins bind to their interacting partners, but also directly affect the function and regulatory mechanism of proteins. Through in-depth research on these sites, the interaction mechanism between proteins can be revealed, providing key information for disease treatment and new drug design. However, traditional experimental methods such as X-ray crystallography and two-hybrid screening have the problems of long time and high cost in identifying these sites. Therefore, the development of fast and efficient computational methods for the identification of protein-protein interactions (PPIS) has become a research focus.
[0004] At present, graph-based protein site prediction methods usually fail to fully consider the modeling of multi-scale molecular information when processing two-dimensional protein graphs, especially the full utilization of atomic information. In addition, research on molecular representation learning has shown that considering the spatial equivariance of molecules plays an important role in improving prediction results. However, existing graph neural network methods often ignore the inherent symmetry of protein molecules in three-dimensional space in protein site prediction, such as the impact of variability such as rotation, reflection and translation on graph representation learning. This neglect may limit the performance of the model in processing complex spatial structures. Therefore, combining multi-scale hierarchical information and spatial equivariance processing will help improve the accuracy and effect of protein site prediction. Summary of the invention
[0005] The present invention proposes a protein site prediction method based on a hierarchical graph and a geometric vector perceptron to solve the above problems.
[0006] The technical solution of the present invention is achieved in this way:
[0007] The protein site prediction method based on hierarchical graph and geometric vector perceptron is characterized by comprising the following steps:
[0008] S1. Extracting protein node features and edge features from protein sequence and protein structure information. These features include residue types, atomic coordinates, chemical bond information, etc., to comprehensively describe the structure and function of proteins.
[0009] S2. Based on the extracted node features and edge features, a hierarchical graph representation of the protein is constructed as the input of the RAGVP-PPIS model. The hierarchical graph representation includes graph structures at the atomic level and the residue level, which can represent different scale information of the protein.
[0010] S3, build geometric vector perceptrons focusing on the atomic level and the residue level respectively to learn the feature information of the atomic scale and the residue scale;
[0011] S4, hierarchical graph representation based on the RAGVP-PPIS model, inputs the node features and edge features of the protein graph into the geometric vector perceptron of the corresponding scale to perform aggregate updates of node embeddings;
[0012] S5. The RAGVP-PPIS model performs feature dimensionality reduction processing through the prediction layer and outputs the final prediction result for accurate prediction of protein sites. The function of the prediction layer is to reduce the dimensionality of the previously learned features. The previous features may be vectors of dozens of dimensions, while the final output prediction result has only two dimensions, representing the probability that the predicted site is or is not a binding site. Therefore, the prediction layer is required for dimensionality reduction. The prediction layer in the present invention is also stacked by GVP.
[0013] Optionally, the hierarchical graph construction of the protein in step S2 is obtained by modeling the graph structure at the atomic level and the residue level;
[0014] In the modeling of atomic level graphs, each atom is regarded as a node in the graph. The characteristics of the node usually include the type of atom, position coordinates, and chemical environment. This information is represented by a feature vector as follows:
[0015] ;
[0016] in, From atoms The extracted feature vectors, the definition of the edge is based on the chemical bonds between atoms, usually connecting pairs of atoms connected by chemical bonds. The characteristics of the edge include distance and bond type, expressed as:
[0017] ;
[0018] in, It's an atom and The distance between
[0019] The resulting atomic level diagram Including node sets and edge sets, its graph structure can be expressed as:
[0020] ;
[0021] In the modeling process of the residue hierarchy graph, each residue is regarded as a node in the graph. The characteristics of the node include residue type, position, and contact information with adjacent residues. These characteristics are represented by feature vectors as follows:
[0022] ;
[0023] in, From the residue The extracted feature vectors, the definition of the edge is based on the contact relationship between residues, usually connecting residue pairs that are close to each other in space or have certain interactions. The feature of the edge is expressed as:
[0024] ;
[0025] The graph structure of the residue hierarchy graph obtained can be expressed as:
[0026] ;
[0027] This hierarchical graph structure can simultaneously process the features of proteins at the atomic level and the residue level, providing comprehensive input data for subsequent graph neural network models. Through the integration of this multi-scale information, the model can better capture the complex structural features of proteins, thereby improving the accuracy and effectiveness of predictions.
[0028] Optionally, in step S3, for the graph structures at the atom level and the residue level constructed in S2, respectively, a graph node feature learning model is constructed based on a geometric vector perceptron. The geometric vector perceptron (GVP) is a simple module for learning vector-valued functions and scalar-valued functions on geometric vectors and scalars. Given a tuple , which includes scalar features s and vector features V. Here, the previously obtained features are divided. The vector features include the previously mentioned location information, and the scalar features are the remaining non-location information. Calculate the new features , and its update formula is as follows:
[0029] ;
[0030] ;
[0031] ;
[0032] ;
[0033] ;
[0034] ;
[0035] ;
[0036] ;
[0037] GVP first uses two linear transformation matrices , , for vector features Perform feature transformation. Then represent the intermediate vector and Take the L2 norm respectively and get and , With the initial scalar representation Splicing , and then through the linear transformation and offset get , the final output vector represents Through a nonlinear function Transformed Compared with the previous vector representation The final output scalar is obtained by multiplying element by element. A nonlinear function of Transformed, in this process, only scalar multiplication, linear combination and L2 norm transformation are performed on the vector value input, so the rotation and reflection operations of GVP in three-dimensional Euclidean space are equivariant and invariant for vectors and scalars respectively;
[0038] in, , , represents the linear transformation matrix, represents the offset matrix, , , , represents vector features, , , , , , represents a scalar feature, and refers to a nonlinear function;
[0039] At the same time, in order to use GVP on the graph structure, GVP-GNN was built, and its update formula is as follows:
[0040] ;
[0041] ;
[0042] First, the messages in the message passing mechanism are calculated, where gvp represents the GVP update process mentioned above, which is responsible for updating the characteristics of node i and the information from neighbor node j. Node i receives the message between node j. By analyzing the node features of node j and edge features between nodes Then, the neighbor information on the graph is passed through messages, and the messages passed to all neighbor nodes of node i are Average and then compare it with the original feature of node i Add; finally get the new features of node i in the l+1th layer , LayerNorm is used for normalization in the aggregation process, and Dropout is added in the middle to prevent overfitting; i and j both represent nodes, represents the edge feature between node i and node j in layer l, represents the node features of node i in layer l, represents the node features of node j in layer l, represents the node features of node i in the l+1th layer, Represents the message sent by node j to node i.
[0043] Optionally, in step S4, feature learning based on geometric vector perceptron features is performed on the protein hierarchy graph, mainly involving aggregating and updating node embeddings on the atom hierarchy graph and the residue hierarchy graph to capture the structural information of the protein and optimize the prediction performance;
[0044] First, at the atomic level Uplink feature update, the update method is as follows:
[0045] ;
[0046] Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features;
[0047] Similarly, in the residue hierarchy diagram The update method on is as follows:
[0048] ;
[0049] Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features.
[0050] Optionally, the RAGVP-PPIS model includes three parts: input, feature extraction, and output. The RAGVP-PPIS model is built using pytorch. The values of important hyperparameters of the RAGVP-PPIS model are as follows:
[0051] The cutoff distance of edge representation in the protein residue hierarchy graph is set to 14Å, and the edge representation in the atomic hierarchy graph is determined by chemical bonds; the number of GVP-GNN layers in the residue hierarchy graph is 5, and the number of GVP-GNN layers in the atomic hierarchy graph is 2; the feature embedding dimension updated by the geometric vector perceptron is consistent with the initial input feature dimension; the output dimensions of the prediction layer network are 10 and 2 respectively; the learning rate for RAGVP-PPIS model training is 0.001, and the number of training rounds is 50; the cross entropy loss function is used for gradient update in RAGVP-PPIS model training; and relu is selected as the activation function.
[0052] Optionally, the node features in step S2 include residue node features and atomic node features, wherein the residue node feature representation includes PSSM, HMM and DSSP, and the atomic node feature representation includes atom type and chemical features.
[0053] After adopting the above technical solution, the beneficial effects of the present invention are:
[0054] In this invention, a protein site prediction method based on hierarchical graph and geometric vector perceptron is proposed, wherein the RAGVP-PPIS model significantly improves the accuracy and efficiency of protein site prediction by comprehensively utilizing multi-scale molecular hierarchical graph information and geometric vector perceptron graph neural network (GVP-GNN). The RAGVP-PPIS model pays special attention to the hierarchical graph structure of proteins, and enhances the understanding and prediction ability of protein structure through information at two scales, atomic hierarchical graph and residue hierarchical graph.
[0055] The core advantage of the RAGVP-PPIS model lies in its effective integration of multi-scale information. In the model, feature learning is performed separately on the atomic level graph and the residue level graph, and node features are aggregated and updated through GVP-GNN. This method can more comprehensively learn the different scale patterns of proteins and significantly enhance the learning ability of graph neural networks for protein structural features. In addition, in order to fully consider the three-dimensional spatial symmetry of protein molecules, the RAGVP-PPIS model applies GVP-GNN in the feature learning process. GVP-GNN introduces a coordinate update mechanism when aggregating node features to ensure that the feature learning process meets the symmetry requirements such as rotation, reflection and translation in three-dimensional space. This mechanism enables the model to effectively capture the spatial equivariance of protein molecules in actual space and improve the accuracy of prediction.
[0056] In terms of performance evaluation, the RAGVP-PPIS model outperforms the existing model (SOTA) on the benchmark dataset. The RAGVP-PPIS model achieves good results in various evaluation indicators such as accuracy, precision, recall, F1 score, Matthews correlation coefficient, area under the receiver operating characteristic curve, and area under the precision-recall curve. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0058] Figure 1 This is the overall architecture diagram of the RAGVP-PPIS model;
[0059] Figure 2 Schematic diagram of GVP-GNN feature update in RAGVP-PPIS model; DETAILED DESCRIPTION
[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0061] The embodiment of the present application discloses a protein site prediction method based on a hierarchical graph and a geometric vector perceptron.
[0062] Example: A protein site prediction method based on hierarchical graph and geometric vector perceptron, the specific contents are as follows:
[0063] 1. Dataset
[0064] The benchmark dataset used in the present invention is the same as the dataset used by AGAT-PPIS, including a training set (Train_335-1) and a test set (Test_60, Test_315-28, Ubtest_31-6).
[0065] The AGAT-PPIS dataset is obtained by fine-tuning the GraphPPIS dataset, and proteins with sequences inconsistent with those on the PDB website are removed from the GraphPPIS dataset. The detailed information of the GraphPPIS dataset and the AGAT-PPIS data are shown in Tables 1 and 2.
[0066] Table 1 Statistical details of the GraphPPIS dataset
[0067]
[0068] Table 2 Statistical details of the AGAT-PPIS dataset
[0069]
[0070] 2. Protein expression
[0071] The present invention uses an undirected graph (G, V) to describe a protein residue hierarchy graph and an atom hierarchy graph.
[0072] In the residue hierarchy graph, nodes are used to represent amino acids, and the edges between nodes are used to represent the connection between amino acids. In simple terms, it is considered that when the distance between two amino acids is less than the set cutoff distance, the two are connected. Based on the sequence and structure information of the protein, the node feature matrix and adjacency matrix corresponding to the residue hierarchy graph are obtained.
[0073] The nodes in the atomic hierarchy graph are used to represent the atoms in the amino acids, and the edges between the nodes are used to represent the chemical bonds between the atoms. When there is a chemical bond between two atoms, they are connected. Based on the atomic type and chemical feature sequence of the atoms, the node feature matrix and adjacency matrix corresponding to the atomic hierarchy graph are obtained. It is worth noting that the features obtained from the atomic hierarchy graph will be used as part of the node features of the residue hierarchy graph.
[0074] The hierarchical graph construction of proteins is obtained by modeling the graph structure at the atomic level and the residue level;
[0075] In the modeling of atomic level graphs, each atom is regarded as a node in the graph. The characteristics of the node usually include the type of atom, position coordinates, and chemical environment. This information is represented by a feature vector as follows:
[0076] ;
[0077] in, From atoms The extracted feature vectors, the definition of the edge is based on the chemical bonds between atoms, usually connecting pairs of atoms connected by chemical bonds. The characteristics of the edge include distance and bond type, expressed as:
[0078] ;
[0079] in, It's an atom and The distance between
[0080] The resulting atomic level diagram Including node sets and edge sets, its graph structure can be expressed as:
[0081] ;
[0082] In the modeling process of the residue hierarchy graph, each residue is regarded as a node in the graph. The characteristics of the node include residue type, position, and contact information with adjacent residues. These characteristics are represented by feature vectors as follows:
[0083] ;
[0084] in, From the residue The extracted feature vectors, the definition of the edge is based on the contact relationship between residues, usually connecting residue pairs that are close to each other in space or have certain interactions. The feature of the edge is expressed as:
[0085] ;
[0086] The graph structure of the residue hierarchy graph obtained can be expressed as:
[0087] ;
[0088] This hierarchical graph structure can simultaneously process the features of proteins at the atomic level and the residue level, providing comprehensive input data for subsequent graph neural network models. Through the integration of this multi-scale information, the model can better capture the complex structural features of proteins, thereby improving the accuracy and effectiveness of predictions.
[0089] 3. Node feature representation
[0090] The residue node features are obtained by processing the protein sequence and protein structure information, and the atomic node features are obtained based on the atom type and chemical features.
[0091] Residue node features include position-specific scoring matrix (PSSM) and hidden Markov model (HMM) matrix. The elements in the PSSM matrix represent the frequency or probability distribution of amino acids at the corresponding positions, reflecting the evolutionary information of the protein. HMM gives scores to each amino acid according to its frequency of occurrence at each position, which indicates the transition probability, insertion and deletion status. Based on the protein structure information, the protein secondary structure definition (DSSP) matrix is used to obtain the DSSP matrix containing the secondary information of the protein by processing the protein 3D structure through the DSSP program.
[0092] Atomic node features include atom types, which are statistics of atom types contained in proteins and obtained through one-hot encoding. Chemical features include information such as chemical charge, atomic radius, hybridization state, and chemical bond type. All these features together describe the chemical environment of atoms and their role in proteins.
[0093] 4. Edge Features
[0094] In the present invention, the edge features of the residue hierarchy graph include the distance relationship between amino acid nodes, and the edge features of the atom hierarchy graph include the chemical bond relationship between atom nodes. The residue hierarchy graph obtains the amino acid position information in the protein PDB file, and calculates the Euclidean distance between all amino acids based on this. Then, according to the cutoff distance, it is judged whether the Euclidean distance between two amino acids satisfies the relationship. If the Euclidean distance is less than the cutoff value, it is considered that there is an edge between the two amino acid nodes, otherwise it is considered that there is no edge. The edge features in the atom hierarchy graph are determined based on whether the atoms are connected by chemical bonds.
[0095] 5. Overall architecture of the RAGVP-PPIS model
[0096] The RAGVP-PPIS model proposed in this paper presents an innovative architecture design. Figure 1As shown in the figure. The model mainly consists of an input module, a feature extraction module and an output module, and aims to improve the prediction performance of protein-protein interaction sites through hierarchical graphs and geometric vector perceptron graph neural networks (GVP-GNN). In the input module, the RAGVP-PPIS model first extracts the features of atom and residue nodes based on the sequence and structural information of the protein. These node features include atom type, chemical charge, atomic radius, etc., as well as the geometric and chemical properties of the residues. In addition, edge features are extracted to describe the relationship between nodes. These features are used as the input of the model by constructing graph representations of atom hierarchical graphs and residue hierarchical graphs. In the feature extraction module, the model conducts in-depth feature learning of atom hierarchical graphs and residue hierarchical graphs through geometric vector perceptron graph neural networks (GVP-GNN). The design of GVP-GNN enables the model to effectively handle symmetry problems in protein molecules, including spatial transformations such as rotation and translation. Specifically, GVP-GNN replaces multi-layer linear perceptrons with geometric vector perceptrons when aggregating node features, and processes the input vector features and scalar features separately to ensure that the features are equivariant and invariant to spatial transformations during the learning process. This mechanism enables the model to effectively capture the spatial equivariance of protein molecules in real space, improving the accuracy of prediction. In the output module, the node features after feature update are aggregated to the final prediction layer. The integrated features are processed by linear layer for dimensionality reduction and the final prediction results are output.
[0097] In general, the RAGVP-PPIS model, through the combination of hierarchical graphs and GVP-GNN, can achieve remarkable results in multi-scale feature learning and spatial symmetry processing, and improve the accuracy and efficiency of protein-protein interaction site prediction.
[0098] 6. Geometric Vector Perceptron
[0099] The graph node feature learning model is built based on the geometric vector perceptron. The geometric vector perceptron (GVP) is a simple module for learning vector-valued functions and scalar-valued functions on geometric vectors and scalars. Given a tuple , which includes scalar features s and vector features V. Here, the previously obtained features are divided. The vector features include the previously mentioned location information, and the scalar features are the remaining non-location information. Calculate the new features , and its update formula is as follows:
[0100] ;
[0101] ;
[0102] ;
[0103] ;
[0104] ;
[0105] ;
[0106] ;
[0107] ;
[0108] GVP first uses two linear transformation matrices , , for vector features Perform feature transformation. Then represent the intermediate vector and Take the L2 norm respectively and get and . With the initial scalar representation Splicing , and then through the linear transformation and offset get The final output vector represents Through a nonlinear function Transformed Compared with the previous vector representation The final output scalar is obtained by multiplying element by element. A nonlinear function of In this process, only scalar multiplication, linear combination and L2 norm transformation are performed on the vector value input, so the rotation and reflection operations of GVP in three-dimensional Euclidean space are equivariant and invariant for vectors and scalars respectively. , , represents the linear transformation matrix, represents the offset matrix, , , , represents vector features, , , , , , represents a scalar feature, and Refers to a nonlinear function.
[0109] At the same time, in order to use GVP on the graph structure, GVP-GNN was built. The schematic diagram of GVP-GNN feature update is as follows Figure 2 As shown, the update formula of GVP-GNN is as follows:
[0110] ;
[0111] ;
[0112] First, the messages in the message passing mechanism are calculated, where gvp represents the GVP update process mentioned above, which is responsible for updating the characteristics of node i and the information from neighbor node j. Node i receives the message between node j. By analyzing the node features of node j and edge features between nodes Then, the neighbor information on the graph is passed through messages, and the messages passed to all neighbor nodes of node i are Average and then compare it with the original feature of node i Add; finally get the new features of node i in the l+1th layer , LayerNorm is used for normalization in the aggregation process, and Dropout is added in the middle to prevent overfitting; i and j both represent nodes, represents the edge feature between node i and node j in layer l, represents the node features of node i in layer l, represents the node features of node j in layer l, represents the node features of node i in the l+1th layer, Represents the message sent by node j to node i.
[0113] Feature learning based on geometric vector perceptron features is performed on protein hierarchical graphs, which mainly involves aggregating and updating node embeddings on atomic hierarchical graphs and residue hierarchical graphs to capture protein structural information and optimize prediction performance;
[0114] First, at the atomic level Uplink feature update, the update method is as follows:
[0115] ;
[0116] Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features;
[0117] Similarly, in the residue hierarchy diagram The update method on is as follows:
[0118] ;
[0119] Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features.
[0120] 7. Experimental setup of the RAGVP-PPIS model
[0121] The RAGVP-PPIS model is built using pytorch. The values of important hyperparameters of the RAGVP-PPIS model are as follows:
[0122] The cutoff distance of the edge representation in the protein residue hierarchy graph is set to 14Å, and the edge representation in the atomic hierarchy graph is determined by chemical bonds; the number of GVP-GNN layers in the residue hierarchy graph is 5, and the number of GVP-GNN layers in the atomic hierarchy graph is 2. The geometric vector perceptron updates the feature embedding dimension consistent with the initial input feature dimension; the output dimensions of the prediction layer network are 10 and 2 respectively; the learning rate of the RAGVP-PPIS model training is 0.001, and the training rounds are 50; the cross entropy loss function is used for gradient update in the RAGVP-PPIS model training; the activation function is selected as relu.
[0123] 8. Performance comparison of RAGVP-PPIS model with other prediction models
[0124] Table 3 Performance comparison of RAGVP-PPIS model and other models on Test_60
[0125]
[0126] The performance comparison experiment of the RAGVP-PPIS model and other protein-protein interaction (PPI) site prediction models was conducted on the independent test set Test_60. The comparison indicators include: accuracy, precision, recall, F1 score, Matthews correlation coefficient, area under the receiver operating characteristic curve, and area under the precision-recall curve. As shown in Table 3, the RAGVP-PPIS model outperforms other models in all seven evaluation indicators, showing significant improvement. Specifically, compared with sequence-based methods (PSIVER, ProNA2020, SCRIBER, DLPred), the RAGVP-PPIS model significantly improves the prediction effect by introducing structural information. These sequence-based models rely only on the sequence data of proteins and cannot fully utilize the spatial structural characteristics of proteins, so their performance is relatively low. The advantage of the RAGVP-PPIS model is that it integrates structural information, thereby achieving better results in indicators such as accuracy, precision, recall, F1 score, Matthews correlation coefficient, area under the receiver operating characteristic curve, and area under the precision-recall curve. When compared with graph-based methods, such as GraphPPIS, RGN, DeepProSite, AGAT-PPIS, and MEG-PPIS, the RAGVP-PPIS model also showed superior performance. These graph-based methods usually take structural information into account, but the RAGVP-PPIS model further improves the feature aggregation and update strategies during the graph learning process, resulting in varying degrees of performance improvement. In particular, compared with DeepPPISP, SPPIDER, and MaSIF-site, which only consider structural information but do not use graph learning, the RAGVP-PPIS model also shows significant advantages in all evaluation indicators. Compared with MEG-PPIS, a model with relatively good performance, RAGVP-PPIS shows significant improvements in all indicators. The comprehensive performance of the RAGVP-PPIS model surpasses that of MEG-PPIS, showing its excellent effect in the site prediction task of protein-protein interactions.
[0127] 9. Feature Ablation Experiment
[0128] Table 4 Performance comparison of RAGVP-PPIS models with different residue feature groups on the independent test set test_60
[0129]
[0130] Table 5 Performance comparison of RAGVP-PPIS models with different atomic feature groups on the independent test set test_60
[0131]
[0132] In this paper, we explored the impact of different features on the performance of the RAGVP-PPIS model through systematic feature ablation experiments. These experiments aim to compare the effects of residue-level graph features with those of atom-level graph features to guide the final feature selection and model optimization. Specifically, we evaluated three features in the residue-level graph: PSSM, HMM, and DSSP, and two features in the atom-level graph: atom type features and chemical features. Through ablation experiments, we analyzed the impact of these features on the model prediction performance when used alone and in combination.
[0133] Experimental results show that PSSM features are highly effective in providing sequence background information and amino acid conservation. HMM features also perform well in modeling sequence dependencies and predicting amino acid hidden states, especially when dealing with complex sequence patterns. DSSP features provide important secondary structure information for the model, helping to improve the understanding of structural features. However, when these residue hierarchy features are used alone, the performance of the model on certain evaluation indicators is different, indicating that each feature contributes differently to the model effect in different aspects. Among the atomic hierarchy features, the atom type features significantly improve the model's ability to predict the local chemical environment by refining atomic information and structural details. Chemical features enhance the understanding of molecular structure and function by providing detailed chemical properties such as charge, radius, and chemical bond type. These atomic hierarchy features show significant advantages in capturing atomic interactions and chemical background.
[0134] Taking into account the performance of different features, the experimental results show that combining the PSSM, HMM and DSSP features of the residue hierarchy diagram with the atom type features and chemical features of the atom hierarchy diagram can significantly improve the prediction accuracy and comprehensiveness of the model. This comprehensive feature selection strategy not only optimizes the performance of the RAGVP-PPIS model, but also provides an important reference for future feature selection and model design.
[0135] Table 6. Performance comparison of RAGVP-PPIS models with different hierarchical graphs on Test_60
[0136]
[0137] 10. Model architecture analysis experiment
[0138] The impact of different layers of graphs on our model performance
[0139] In the present invention, the effects of residue hierarchy diagram and atom hierarchy diagram on the performance of RAGVP-PPIS model were systematically compared through experiments.
[0140] The experimental results show that the residue hierarchy graph features have obvious advantages in model prediction. The PSSM, HMM and DSSP features in the residue hierarchy graph can effectively capture the conservation, sequence dependency and secondary structure information of protein sequences, which are important for identifying key interaction sites. Specifically, the PSSM feature is outstanding in revealing the sequence context and conservation of amino acids, while the HMM feature shows strong ability in processing complex sequence patterns and amino acid states. The DSSP feature provides key secondary structure information and enhances the model's understanding of protein folding structure.
[0141] The introduction of atomic hierarchy features also had a significant impact on model performance. The atom type feature enhances the ability to capture the details of the protein's internal structure by refining the type and position of each atom. The chemical feature further improves the understanding of the protein's chemical environment by providing information on the charge, radius, and chemical bonds of atoms. These features show obvious advantages in capturing local chemical properties and interactions between atoms.
[0142] A comprehensive comparison of the effects of the two hierarchical graphs shows that when the two are used together, the model shows significant improvements in key evaluation indicators such as accuracy, precision and recall. This shows that the residue hierarchical graph and the atomic hierarchical graph features each focus on features in different scales, and the comprehensive use of the features of these two levels can significantly enhance the overall performance of the model.
[0143] In summary, this paper proposes an advanced PPIS prediction model RAGVP-PPIS, which integrates the multi-scale information of the residue level graph and the atomic level graph, and effectively captures the complex symmetry and structural characteristics of protein molecules in three-dimensional space by introducing the geometric vector perceptron graph neural network (GVP-GNN). When predicting protein-protein interaction sites, the RAGVP-PPIS model not only utilizes rich sequence and structural information, but also considers detailed chemical environment characteristics, thereby significantly improving the prediction accuracy and comprehensive performance. Experimental results show that compared with existing mainstream methods, the RAGVP-PPIS model shows better performance in multiple key evaluation indicators, especially in terms of accuracy, precision, recall rate and F1 score.
[0144] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A protein site prediction method based on hierarchical graph and geometric vector perceptron, characterized in that: The following steps are involved: S1. Extracting protein node features and edge features from protein sequence and protein structure information, including residue types, atomic coordinates, and chemical bond information; S2. Based on the extracted node features and edge features, a hierarchical graph representation of the protein is constructed as the input of the RAGVP-PPIS model. The hierarchical graph representation includes graph structures at the atomic level and the residue level, which are used to represent information of proteins at different scales. The RAGVP-PPIS model includes three parts: input, feature extraction, and output. The RAGVP-PPIS model is built using pytorch. S3. Build GVP-GNN focusing on the atomic level and residue level respectively to learn the feature information at the atomic scale and residue scale; S4. Based on the hierarchical graph representation of the RAGVP-PPIS model, the node features and edge features of the protein graph are input into the GVP-GNN of the corresponding scale to perform aggregate updates on the node embeddings; S5, RAGVP-PPIS model performs feature dimensionality reduction through the prediction layer and outputs the final prediction result; The construction of the protein hierarchical graph in step S2 is obtained by modeling the graph structure at the atomic level and the residue level; In the modeling of atomic level graphs, each atom is regarded as a node in the graph. The characteristics of the node include the type of atom, position coordinates, and chemical environment. This information is represented by a feature vector as follows: ; in, From atoms The extracted feature vectors, the definition of the edge is based on the chemical bonds between atoms, connecting the atom pairs connected by chemical bonds. The characteristics of the edge include distance and bond type, expressed as: ; in, It is an atom and The distance between The resulting atomic level diagram Including node sets and edge sets, its graph structure can be expressed as: ; In the modeling process of the residue hierarchy graph, each residue is regarded as a node in the graph. The characteristics of the node include the residue type, position, and contact information with adjacent residues. These characteristics are represented by feature vectors as follows: ; in, From the residue The extracted feature vectors, the definition of the edge is based on the contact relationship between the residues, connecting the residue pairs that are close to each other in space or have interactions. The edge feature is expressed as: ; The graph structure of the residue hierarchy graph obtained can be expressed as: ; In step S3, for the graph structures at the atomic level and residue level constructed in S2, a graph node feature learning model is constructed based on GVP. Given a tuple , which includes scalar features s and vector features V. The vector features include location information, and the scalar features are the remaining non-location information. Calculate the new features , and its update formula is as follows: ; ; ; ; ; ; ; ; in, , , represents the linear transformation matrix, represents the offset matrix, , , , represents vector features, , , , , , represents a scalar feature, and refers to a nonlinear function; At the same time, in order to use GVP on the graph structure, GVP-GNN was built, and its update formula is as follows: ; ; Among them, gvp represents the update process, which is normalized by LayerNorm, and Dropout is added to prevent overfitting. The information of neighbor nodes is integrated through a weighted average mechanism. i and j both represent nodes. represents the edge feature between node i and node j in layer l, represents the node features of node i in layer l, represents the node features of node j in layer l, represents the node features of node i in the l+1th layer, Represents the message sent by node j to node i; In step S4, feature learning based on geometric vector perceptron features is performed on the protein hierarchy graph, which mainly involves aggregating and updating node embeddings on the atom hierarchy graph and the residue hierarchy graph to capture the structural information of the protein and optimize the prediction performance; First, at the atomic level The feature update is performed on the , and the update method is as follows: ; Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features; Similarly, in the residue hierarchy diagram The update method on is as follows: ; Among them, Layer and The atomic hierarchy graph node features of the layers are represented as and ; Represents the updated features.
2. The protein site prediction method based on hierarchical graph and geometric vector perceptron according to claim 1, characterized in that: The values of important hyperparameters of the RAGVP-PPIS model are as follows: The cutoff distance of edge representation in the protein residue hierarchy graph is set to 14Å, and the edge representation in the atomic hierarchy graph is determined by chemical bonds; the number of GVP-GNN layers in the residue hierarchy graph is 5, and the number of GVP-GNN layers in the atomic hierarchy graph is 2. The feature embedding dimension updated by the geometric vector perceptron is consistent with the initial input feature dimension; the output dimensions of the prediction layer network are 10 and 2, respectively; the learning rate for RAGVP-PPIS model training is 0.001, and the number of training rounds is 50; the cross entropy loss function is used for gradient update in RAGVP-PPIS model training; and relu is selected as the activation function.
3. The protein site prediction method based on hierarchical graph and geometric vector perceptron according to claim 1, characterized in that: The node features in step S2 include residue node features and atomic node features, wherein the residue node feature representation includes PSSM, HMM and DSSP, and the atomic node feature representation includes atom type and chemical features.
