Method for predicting protein-protein binding affinity based on graph convolutional network

Abstract

The invention provides a method for predicting protein-protein binding affinity based on a graph convolutional network, and relates to the field of bioinformatics. The basic principle is as follows: respectively converting a protein-protein compound and an interaction region into atomic-level graphs; intramolecular and intermolecular node features and edge features of a compound are learned through a model constructed on the basis of a graph convolutional network, graph structure features of protein are fully utilized, the protein-protein interaction relationship is deeply excavated, then the precision of predicting a protein-protein binding affinity value is improved, and the prediction precision of the protein-protein binding affinity value is improved. And the function of predicting the affinity of the multi-chain compound is realized by processing the data set.

Description

Technical Field

[0001] The invention belongs to the field of bioinformatics and relates to a method for predicting protein-protein binding affinity based on a graph convolutional network. Technical Background

[0002] The interaction between proteins plays a core role in the physiological processes of organisms, including immunity, metabolism, and signal transduction. Understanding protein-protein interactions is crucial for studying biological systems and drug design, and directly affects the development of protein and peptide drugs. A key feature of molecular interactions is binding affinity. Accurately obtaining protein-protein binding affinity is the key to understanding the biological functions caused by protein-protein interactions. However, traditional experimental methods require a lot of resources and time to determine affinity. At the same time, computational predictions such as molecular dynamics simulations and empirical energy functions have certain limitations in terms of computational requirements and accuracy.

[0003] The prediction of protein-protein binding affinity is a complex problem because the interaction between proteins involves multiple factors, such as structure, sequence, charge, etc. The interaction between these factors is very complex, and it is difficult to accurately establish a mathematical model to predict them; and the prediction of protein-protein binding affinity requires a large amount of data support, while the currently available data sets are relatively limited, which limits the application of deep learning methods in this field; in addition, the prediction of protein-protein binding affinity involves the interaction between multiple proteins, and most of the existing data sets only contain information on two protein chains, which limits the in-depth study and prediction capabilities of complex protein interactions.

[0004] Although there are quite a few deep learning methods to predict the dissociation constant (KD) or Gibbs free energy (ΔG) of protein-protein, protein-ligand and protein-nucleic acid complexes, a method that can accurately predict the characteristics of protein complexes of various properties has not yet been developed, especially in the field of protein-protein binding affinity prediction. This field remains underexplored and no stable and accurate prediction model has yet been constructed. Summary of the invention

[0005] In view of the problem that the existing protein-protein affinity prediction methods have low accuracy and are difficult to predict the affinity of multi-chain complexes, the present invention proposes a method for predicting protein-protein binding affinity based on graph convolutional networks, which involves the field of bioinformatics. The basic principle is as follows: Figure 1): The protein-protein complex and the interaction region are converted into atomic-level graphs respectively, and the intramolecular and intermolecular node features and edge features of the complex are learned through a model built based on a graph convolutional network. The graph structure characteristics of the protein are fully utilized to deeply explore the interaction relationship between proteins, thereby improving the accuracy of predicting the protein-protein binding affinity value, and realizing the function of predicting the affinity of multi-chain complexes by processing the data set.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] In one aspect, the present invention provides a method for predicting protein-protein binding affinity based on a graph convolutional network, the method comprising the following steps:

[0008] S1: Obtain protein-protein complex dataset;

[0009] S2: Extract interaction interfaces within a certain distance threshold in protein-protein complexes and generate protein-protein complex maps and interaction maps ( Figure 2 );

[0010] Specifically, the protein-protein complex graph uses atoms as nodes and the edges between two atoms; the interaction graph only has edges.

[0011] S3: Extract the initial node features and edge features of the graph;

[0012] S4: Update node features of protein-protein complex graphs;

[0013] S4.1: Use linear layers to project node features and edge features to specific dimensions, and then use activation functions to introduce nonlinear relationships, which allows the model to capture more complex, nonlinear relationships.

[0014] S4.2: Update the edge features again, and concatenate the node features connected by the edge with the edge features, so that the subsequent update of the node features takes into account the characteristics of the edge;

[0015] S4.3: Fuse the updated edge features with the node features to obtain new node features, and use the attention mechanism and gated recurrent unit (GRU) to capture the relationship between nodes and edges;

[0016] It should be understood that the traditional graph convolutional neural network (GCN) usually only performs weighted summation through the adjacency matrix, while the present invention uses the attention mechanism to weight the edge features, which means that the model can adjust the way information is propagated according to the "importance" of the edge, thereby capturing more meaningful edge information and improving the representation ability; and the GRU unit is used when fusing edge features and node features, which is not just a simple weighted summation or average, but uses GRU to dynamically combine the features of neighboring nodes and edge features, and then better capture the complex relationship between nodes and edges. Therefore, the combination of the attention mechanism and GRU to fuse node and edge features is a necessary condition for the present invention to improve the accuracy of predicting binding affinity between proteins.

[0017] S4.4: Continue to update the node features, aggregate the neighbor node features through the maximum pooling layer, and then splice them with the source node features.

[0018] It should be understood that the traditional GCN information aggregation method usually only aggregates the features of neighboring nodes through weighted average (or simple average), which may cause some important neighboring node features to be averaged, thereby losing some significant information, and in some cases, this average aggregation method cannot capture the local structural information of important nodes in the graph well; while the maximum pooling can select the node features with the largest values ​​from each feature channel, which means that this processing method can focus on the most significant or active features while ignoring unimportant information, thereby emphasizing key features. In graph data, the number of nodes may be large, and directly processing all nodes will result in excessive computational overhead and difficulty in extracting global features; maximum pooling can condense a large amount of local information into a small amount of global information, which not only reduces the amount of computation, but also highlights the main features. Therefore, the maximum pooling layer processing method is also a key factor affecting the accuracy of the prediction results.

[0019] S5: Update edge features of the interaction graph;

[0020] S6: The final prediction value of protein-protein binding affinity is output through the fully connected layer.

[0021] Furthermore, the activation function described in step S4.1 is h' v =LeakyRELU(W n h v +b n ), where h' v is the node feature after projection, W n represents the weight matrix of the linear transformation, h v represents the initial node feature, b n Represents the bias term.

[0022] The activation functions suitable for the hidden layer of the convolutional neural network mainly include Relu and Leaky Relu. The Leaky Relu function solves the neuron death problem of the Relu function and has the advantages of the Relu function, namely, high computational efficiency, fast network convergence, nonlinearity, etc. In addition, the Leaky Relu function has a small positive slope in the negative region, so it can perform back propagation even for negative input values. Therefore, the type of activation function is also a key factor affecting the accuracy of the prediction result, and only the Leaky Relu function is most suitable for the prediction method provided by the present invention.

[0023] Furthermore, the attention mechanism described in step S4.3 is where α uv represents the attention weight from node u to node v, e” uv represents the final feature after the edge (u, v) is updated, N(v) represents the set of neighbor nodes of node v, and e” vk represents the edge formed by node v and its neighbor node k.

[0024] Furthermore, the formula of the gated recurrent unit described in step S4.3 is in, represents the updated node features after fusing node features and edge features, h' v represents the node feature after projection, N(v) represents the set of neighbor nodes of node v, α uv represents the attention weight from node u to node v, e” uv Represents the final feature of the edge (u,v) after updating.

[0025] Furthermore, the maximum pooling layer formula described in step S4.4 is in, is the feature of the maximum pooled aggregated neighbor nodes, N(v) represents the set of neighbor nodes of node v, Indicates the updated node features after fusing node features and edge features.

[0026] Furthermore, the node feature concatenation formula in step S4.4 is: Among them, L represents the number of node update layers; Indicates the updated node features after fusing node features and edge features; Aggregate the features of neighbor nodes for max pooling.

[0027] Furthermore, the step S5 comprises the following steps:

[0028] S5.1: concatenate the features finally updated from each edge to the two nodes connected to it (i.e., the node features updated in step S4) to obtain new edge features;

[0029] S5.2: The edge features are further updated through a multi-layer perceptron (MLP), and the edge features are weighted and aggregated using edge weights to obtain the global features of the interaction graph; and layer normalization is applied in this module.

[0030] Furthermore, the formula used by the multilayer perceptron to update the edge features in step S5.2 is: where e' ij is the updated edge feature, It is the edge feature formed by the concatenation of the source node i feature, the target node j feature and the initial edge feature.

[0031] Furthermore, the step S2 is to pre-process the protein-protein complex data set, the purpose of which is to exclude the influence of irrelevant structures in the protein on the predicted binding affinity, and secondly, because there are many atoms in the protein, if the complete protein structure is used, the requirements on the computer are too high and the calculation time is prolonged. Step S2 specifically includes the following steps:

[0032] S2.1: Calculate the distance d between the Cα atoms of the residues on the interacting chains ij , extract All atoms in the range are considered as the protein-protein complex graph G; the d ij represents the distance between the Cα atoms of the i-th residue on one chain and the j-th residue on the other chain involved in the interaction;

[0033] The reason for keeping All atoms within the range are included, firstly because the range can fully cover the interaction area, and secondly because the atoms outside this range have relatively little effect on the interaction.

[0034] S2.2: Extract All atoms in the range are considered as the interaction graph G interacte ;

[0035] S2.3: Pass and distance threshold By comparison, the adjacency matrix A of the protein-protein complex graph and the interaction graph was calculated respectively. ij ;

[0036] When determining protein interaction regions, the commonly used distance threshold is usually arrive This threshold is used to determine whether the atoms or residues of two proteins are in contact, thereby inferring whether they interact with each other.

[0037] S2.4: In the protein-protein complex graph, atoms are used as nodes of the graph, and the adjacency matrix A ij = 1 as the edge of the graph; in the interaction graph, only the edges are included, not the nodes, and the adjacency matrix A ij = 1 as the edge of the graph.

[0038] The present invention uses other methods to preprocess the data set, but other preprocessing methods may affect the accurate extraction of the interaction region or introduce the influence of irrelevant atoms on the interaction prediction, interfere with the training of the model, and then reduce the accuracy of the prediction results. In short, only when the above preprocessing method is used in combination with the model provided by the present invention can the accuracy of the prediction results be guaranteed.

[0039] Furthermore, the node features include but are not limited to atom type, number of bonds, atomic charge, number of connected hydrogen atoms, chirality information, whether the atom is in an aromatic ring, and an AEV feature vector describing the atomic environment; edge features include but are not limited to chemical bond type, stereochemical information, whether it is a conjugated bond, the angle formed between atoms, and the distance between atoms.

[0040] In summary, the preprocessing of the protein-protein complex dataset and the updating of node and edge features ensure the prediction method of protein-protein binding affinity based on graph convolutional network proposed in the present invention ( Figure 3 )’s reliability and applicability to large-scale and complex scenarios.

[0041] Specifically, the specific steps of the method for predicting protein-protein binding affinity based on graph convolutional network provided by the present invention are as follows:

[0042] 1) The three-dimensional structure of the protein-protein complex was extracted using the toolkit RDKit and converted into a graph structure, with atoms as nodes and edges between two atoms.

[0043] 2) Compute adjacency matrices of protein-protein complex graphs and interaction graphs

[0044] 2.1) Calculate the Euclidean distance d between each atom pair ij , by the distance threshold In comparison, the adjacency matrix A of the protein-protein complex graph is calculated ij :

[0045]

[0046] 2.2) Calculate the Euclidean distance d between the two interacting chains ij , by the distance threshold In comparison, the adjacency matrix A of the interaction graph is calculated ij_interacte :

[0047]

[0048] 3) In the protein-protein complex graph, atoms are used as nodes of the graph, and the adjacency matrix A ij = 1 as the edge of the graph; in the interaction graph, only the edges are included, not the nodes, and the adjacency matrix A ij = 1 as the edge of the graph;

[0049] 4) Extract the node features and edge features of the graph through the RDkit toolkit to obtain the initial graph representation G(h,e), G interacte (e);

[0050] 4.1) Node features include atom type, number of bonds, atomic charge, number of connected hydrogen atoms, chirality information, whether the atom is in an aromatic ring, and the AEV feature vector describing the atomic environment;

[0051] 4.2) Edge features include chemical bond type, stereochemical information, whether it is a conjugated bond, the angle formed between atoms, and the distance between atoms;

[0052] 5) Updating and fusing node features and edge features of protein-protein complex graphs;

[0053] 5.1) For each node v, its initial features are projected through linear transformation and activation function to generate a new feature representation:

[0054] h' v =LeakyRELU(W n h v +b n )

[0055] Among them, W n represents the weight matrix of the linear transformation, b n represents the bias term, h' v Represents the node features after projection;

[0056] 5.2) Initial feature e of edge (u,v) uv The feature projection h' of the source node u u Perform splicing and then generate new edge features through linear transformation:

[0057]

[0058] in, represents the weight matrix of the linear transformation, represents the bias term;

[0059] 5.3) The projection feature h' of the target node v vThe first edge feature projection result e' uv Perform splicing and then generate the final edge features through linear transformation:

[0060]

[0061] in, represents the weight matrix of the linear transformation, represents the bias term;

[0062] 5.4) The model updates node features through message passing and attention mechanism. During the message passing process, the edge feature e of the neighbor node u is uv is used to update the features of the target node v;

[0063] 5.4.1) For each edge, a weight is assigned to the edge through a softmax normalized attention mechanism:

[0064]

[0065] Among them, N(v) represents the set of neighbor nodes of node v, α uv represents the attention weight from node u to node v, e” vk represents the edge formed by node v and its neighbor node k;

[0066] 5.4.2) Use GRU to fuse node features and edge features:

[0067]

[0068] 5.5) Update node features using neighbor node information;

[0069] 5.5.1) For each node v, use maximum pooling to aggregate the features of neighboring nodes:

[0070]

[0071] 5.5.2) The final feature of node v is the sum of its own feature and neighbor feature:

[0072]

[0073] Among them, L represents the number of nodes updated layer;

[0074] 5.6) After multiple layers of node update layers, node features will be continuously updated and accumulated at each layer:

[0075]

[0076] 6) Update the edge features in the interaction graph by combining the node features in the composite graph;

[0077] 6.1) Update and merge the edge features e of the input interaction graph and the node features and edge features of the protein-protein complex graph

[0078] 6.2) Calculate the feature e of each edge and the feature of the connected node to generate the edge message m:

[0079]

[0080] in, represents the characteristics of source node i, Represents the characteristics of the target node j;

[0081] 6.3) In the edge feature update, the current edge features and node features are combined:

[0082]

[0083] 6.4) Update edge features through multi-layer perceptron:

[0084]

[0085] 7) The edge features are processed through the linear layer and sigmoid activation function to calculate the weight of each edge:

[0086] w ij =σ(W·e' ij +b)

[0087] Among them, w ij represents the weight of edge (i, j), W represents the weight matrix of linear transformation, b represents the bias term, and σ represents the sigmoid activation function;

[0088] 8) Use edge weights to perform weighted aggregation on edge features to obtain the global feature h of the graph g :

[0089]

[0090] 9) Output the final prediction value through the fully connected layer;

[0091]

[0092] 10) During the training process, the error between the output value predicted by the model and the true value is calculated using the loss function:

[0093]

[0094] in, is the predicted value of the model, y i is the real target value;

[0095] 11) Use Adam to optimize and update the model parameters:

[0096]

[0097] Among them, g t represents the gradient of iteration step t; represents the gradient of parameter θ; f t (θ t -1) represents the objective function to be optimized, θ t -1 indicates the parameter value evaluated in the previous iteration.

[0098] On the other hand, the present invention provides a computer program product, comprising a processor and a memory, wherein the memory is used to store executable instructions of the processor; the processor is configured to perform the above-mentioned method for predicting protein-protein binding affinity based on a graph convolutional network by executing the executable instructions.

[0099] The beneficial effects of the present invention are mainly manifested in:

[0100] 1. For the atomic-level protein graph, the present invention not only integrates the features of adjacent edges during the node feature update process, but also generates node embedding through the node neighborhood sampling strategy, so as to better process large-scale graph data;

[0101] 2. By aggregating atomic information in local areas to capture key local structural features in proteins and capturing long-range dependencies through message passing mechanisms, this method can better model complex protein structures and their interactions, thereby improving the accuracy and scalability of predictions;

[0102] 3. When processing the same data set, the various evaluation indicators of the method provided by the present invention are superior to other protein-protein binding affinity prediction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] 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 work.

[0104] Figure 1 A flow chart of the method provided by the present invention;

[0105] Figure 2 Schematic diagram of protein-protein complex map and interaction map;

[0106] Figure 3 This is the architecture diagram of the graph convolutional network. DETAILED DESCRIPTION

[0107] The present invention is further elaborated in detail below in conjunction with the accompanying drawings and specific embodiments of the specification. The embodiments are only used to explain the present invention and are not used to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0108] Example 1

[0109] This example trains the model on the protein-protein complex dataset in PDBbind version 2020 and performs predictions on two public test sets. The specific steps are as follows:

[0110] 1) Obtain the protein-protein interaction affinity dataset from the public dataset PDBbind version 2020, which contains a total of 2852 complex structure files;

[0111] 1.1) For complexes with only two chains in the dataset, this example specifies these two chains as interacting components; for complexes containing multiple chains, manual structural inspection and labeling are performed to determine the chains involved in the interaction, and the remaining irrelevant chains are deleted;

[0112] 1.2) The negative logarithm (pKa) with base 10 was used as a unified measure of binding affinity, and the formula is as follows:

[0113] pKa=-log 10 K

[0114] Where K can represent any affinity value (Kd, Ki or IC50) in mol / L;

[0115] 2) The three-dimensional structure of the protein-protein complex was obtained from the PDB file of the protein-protein complex using the toolkit RDKit and converted into a graph structure with atoms as nodes; considering that there are many atoms in the protein, the PDB file of the complex was preprocessed before the protein was represented as a graph, and the distance d between the Cα atoms of the residues on the interacting chains was calculated. ij (represents the distance between the Cα atoms of the i-th residue on one chain and the j-th residue on the other chain involved in the interaction), extract All atoms in the range are extracted as the protein-protein complex graph G. All atoms in the range are considered as the interaction graph G interacte ;

[0116] 3) Calculate the adjacency matrix of protein-protein complex graphs and interaction graphs;

[0117] 3.1) Calculate the Euclidean distance d between each atom pair ij , by the distance threshold In comparison, the adjacency matrix A of the protein-protein complex graph is calculated ij :

[0118]

[0119] 3.2) Calculate the Euclidean distance d between the two interacting chains ij , by the distance threshold In comparison, the adjacency matrix A of the interaction graph is calculated ij_interacte :

[0120]

[0121] 4) In the protein-protein complex graph, atoms are used as nodes of the graph, and the adjacency matrix A ij = 1 as the edge of the graph; in the interaction graph, only the edges are included, not the nodes, and the adjacency matrix A ij = 1 as the edge of the graph;

[0122] 5) Use the RDkit tool to extract the node features and edge features of the graph and obtain the initial graph representation G(h,e), G interacte (E);

[0123] 5.1) Node features include atom type, number of bonds, atomic charge, number of connected hydrogen atoms, chirality information, whether the atom is in an aromatic ring, and AEV feature vector describing the atomic environment;

[0124] 5.2) Edge features include chemical bond type, stereochemical information, whether it is a conjugated bond, the angle formed between atoms, and the distance between atoms;

[0125] 6) Updating and fusing node features and edge features of protein-protein complex graphs;

[0126] 6.1) For each node v, its initial features are projected through linear transformation and activation function to generate a new feature representation:

[0127] h' v =LeakyRELU(W n h v +b n )

[0128] Among them, W n represents the weight matrix of the linear transformation, b nrepresents the bias term, h' v Represents the node features after projection;

[0129] 6.2) Initial feature e of edge (u,v) uv The feature projection h' of the source node u u Perform splicing and then generate new edge features through linear transformation:

[0130]

[0131] in, represents the weight matrix of the linear transformation, represents the bias term;

[0132] 6.3) The projection feature h' of the target node v v The first edge feature projection result e' uv Perform splicing and then generate the final edge features through linear transformation:

[0133]

[0134] in, represents the weight matrix of the linear transformation, represents the bias term;

[0135] 6.4) The model updates node features through message passing and attention mechanism. During the message passing process, the edge feature e of the neighbor node u is uv is used to update the features of the target node v;

[0136] 6.4.1) For each edge, a weight is assigned to the edge through a softmax-normalized attention mechanism:

[0137]

[0138] Among them, N(v) represents the set of neighbor nodes of node v, α uv represents the attention weight from node u to node v, e” vk represents the edge formed by node v and its neighbor node k;

[0139] 6.4.2) Use GRU to fuse node features and edge features:

[0140]

[0141] 6.5) Update node features using neighbor node information;

[0142] 6.5.1) For each node v, use maximum pooling to aggregate the features of neighboring nodes:

[0143]

[0144] 6.5.2) The final feature of node v is the sum of its own feature and neighbor feature:

[0145]

[0146] Among them, L nodes update the number of layers;

[0147] 6.6) After multiple update layers, node features will be continuously updated and accumulated at each layer:

[0148]

[0149] 7) Update edge features in the interaction graph by combining node features in the composite graph;

[0150] 7.1) Update and merge the edge features e of the input interaction graph and the node features and edge features of the protein-protein complex graph

[0151] 7.2) Calculate the feature e of each edge and the feature of the connected node to generate the edge message m:

[0152]

[0153] in, represents the characteristics of source node i, Represents the characteristics of the target node j;

[0154] 7.3) In the edge feature update, the current edge features and node features are combined:

[0155]

[0156] 7.4) Update edge features through multi-layer perceptron:

[0157]

[0158] 8) The edge features are processed through the linear layer and sigmoid activation function to calculate the weight of each edge:

[0159] w ij =σ(W·e' ij +b)

[0160] Among them, w ij represents the weight of edge (i, j), W represents the weight matrix of linear transformation, b represents the bias term, and σ represents the sigmoid activation function;

[0161] 9) Use edge weights to perform weighted aggregation on edge features to obtain the global feature h of the graph g :

[0162]

[0163] 10) Output the final prediction value through the fully connected layer:

[0164]

[0165] 11) During the training process, the error between the output value predicted by the model and the true value is calculated using the loss function:

[0166]

[0167] in, is the predicted value of the model, y i is the real target value;

[0168] 12) Use Adam to optimize and update the model parameters:

[0169]

[0170] Among them, g t represents the gradient of iteration step t; represents the gradient of parameter θ; f t (θ t -1) represents the objective function to be optimized, θ t -1 indicates the parameter value evaluated in the previous iteration.

[0171] 13) After each epoch, the performance of the model is evaluated using MAE (mean absolute error), RMSE (root mean square error), and Pearson correlation coefficient (R) on the training set and validation set:

[0172]

[0173]

[0174]

[0175] 14) The training data is input into the graph convolutional neural network model in batch form to train the entire model. The batch size is 32. When the performance of the model is no longer improved, the training is terminated, and then the model is used to predict the affinity of the protein-protein complex in the test set;

[0176] 15) This example uses the model provided by the present invention to make predictions on two test sets Test_1 (complex data sets containing multiple molecules) and Test_2 disclosed in the ProBAN method [1]; and compares the values ​​predicted by the ProBAN, PRODIGY, and PPI-Affinity models provided in the literature [1]. The specific results are shown in Table 1.

[0177] The ProBAN model is a method proposed in 2024 for predicting protein-protein complex binding affinity based on deep convolutional neural networks (CNN); the PRODIGY model is a method proposed in 2017 for predicting binding force by calculating the number of atomic contacts made at the protein-protein complex interface within a distance threshold range and based on the polar / non-polar / charged properties of the interacting amino acids; the PPI-Affinity model is a method proposed in 2022 for predicting binding affinity using a support vector machine (SVM) predictor.

[0178] Pearson correlation describes the degree of association between two variables. Here, it refers to the degree of association between the actual affinity value and the affinity value predicted by the model. The closer this value is to 1, the stronger the linear correlation between the predicted value and the actual value.

[0179] MAE (mean absolute error) represents the average value of the absolute error between the predicted value and the true value. When the predicted value is completely consistent with the true value, it is equal to 0, that is, a perfect model; conversely, the greater the error, the greater the value.

[0180] RMSE, root mean square error, represents the sample standard deviation of the difference between the predicted value and the true value (called residual). This indicator reflects the degree of sample dispersion. When doing nonlinear fitting, the smaller the RMSE, the better. When the predicted value and the true value are completely consistent, it is equal to 0, that is, a perfect model; conversely, the greater the error, the greater the value. These three parameters are indicators used to measure the ability of the prediction model.

[0181] The results show that the model provided by the present invention is superior to the existing methods in all indicators, especially MAE and RMSE, both of which are far lower than those of other methods. The model provided by the present invention is superior to other methods because the data preprocessing focuses the graph structure on the interaction area to avoid interference from irrelevant structures; at the same time, the node feature update method provided by the present invention has a stronger learning ability for the interaction features of the complex and a better prediction effect.

[0182] Table 1

[0183]

[0184]

[0185] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions described in the above embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0186] References

[0187] [1] Bogdanova EA, Novoseletsky VN. ProBAN: Neural network algorithm for predicting binding affinity in protein–protein complexes. Proteins. 2024; 92(9): 1127-1136. doi: 10.1002 / prot.26700.

Claims

1. A method for predicting protein-protein binding affinity based on graph convolutional networks, characterized in that: The method comprises the following steps: S1: Obtain protein-protein complex dataset; S2: Extract interaction interfaces within a certain distance threshold in protein-protein complexes and generate protein-protein complex maps and interaction maps; S3: Extract the initial node features and edge features of the graph; S4: Update the node features of the protein-protein complex graph, that is, project the node features and edge features into a specific dimension, and then use the activation function to introduce nonlinear relationships; The attention mechanism and gated recurrent unit are used to capture the relationship between nodes and edges. The neighbor node features are aggregated through the maximum pooling layer, and then the aggregated node features are spliced ​​to the node itself to complete the final update of the node features. S5: Update the edge features of the interaction graph and finally output the final prediction value of protein-protein binding affinity through a fully connected layer.

2. The method according to claim 1, characterized in that The activation function is h v '=LeakyRELU(W n h v +b n ), where h v ' represents the node features after projection, W n represents the weight matrix of the linear transformation, h v ' represents the initial node feature, b n Represents the bias term.

3. The method according to claim 1, characterized in that The attention mechanism is where α uv represents the attention weight from node u to node v, e' u ' v represents the final feature after the edge (u, v) is updated, N(v) represents the set of neighbor nodes of node v, e' v ' k represents the edge formed by node v and its neighbor node k.

4. The method according to claim 1, characterized in that The formula of the gated recurrent unit is in, represents the updated node features after fusing node features and edge features, h v ' represents the node feature after projection, N(v) represents the set of neighbor nodes of node v, α uv represents the attention weight from node u to node v, e' u ' v Represents the final feature of the edge (u,v) after updating.

5. The method according to claim 1, characterized in that The maximum pooling layer formula is in, is the feature of the maximum pooled aggregated neighbor nodes, N(v) represents the set of neighbor nodes of node v, Indicates the updated node features after fusing node features and edge features.

6. The method according to claim 1, characterized in that The node feature concatenation formula in step S4 is: Among them, L represents the number of node update layers; Indicates the updated node features after fusing node features and edge features; Aggregate the features of neighbor nodes for max pooling.

7. The method according to claim 1, characterized in that The S5 step comprises the following steps: S5.1: Concatenate the features finally updated from each edge to the two nodes connected to it to obtain a new edge feature; S5.2: The edge features are further updated through a multi-layer perceptron, and the edge features are weighted and aggregated using edge weights to obtain the global features of the interaction graph; And the application layer is normalized in this module.

8. The method according to claim 7, characterized in that The formula used by the multilayer perceptron to update the edge features in step S5.2 is: where e i ' j is the updated edge feature, It is the edge feature formed by the concatenation of the source node i feature, the target node j feature and the initial edge feature.

9. The method according to claim 1, characterized in that The S2 step comprises the following steps: S2.1: Calculate the distance d between the Cα atoms of the residues on the interacting chains ij , extract All atoms in the range are considered as the protein-protein complex graph G; the d ij represents the distance between the Cα atoms of the i-th residue on one chain and the j-th residue on the other chain involved in the interaction; S2.2: Extract All atoms in the range are considered as the interaction graph G interacte ; S2.3: Pass and distance threshold By comparison, the adjacency matrix A of the protein-protein complex graph and the interaction graph was calculated respectively. ij ; S2.4: In the protein-protein complex graph, atoms are used as nodes of the graph, and the adjacency matrix A ij = 1 as the edge of the graph; In the interaction graph, only edges are included, not nodes, with the adjacency matrix A ij = 1 as the edge of the graph.

10. A computer program product, characterized in that The product comprises a processor and a memory, wherein the memory is used to store executable instructions of the processor; the processor is configured to execute the method for predicting protein-protein binding affinity based on a graph convolutional network according to any one of claims 1 to 9 by executing the executable instructions.

Images