Drug and target affinity prediction method based on graph neural network topology enhancement

Abstract

According to the method for predicting the affinity of the drug and the target based on graph neural network topology enhancement, the physicochemical properties of protein, especially the structure and the dynamics of a pocket region, can be better described, and key modes and characteristics in DTA prediction can be better captured, so that the prediction accuracy is improved. Comprising the following steps: (1) expressing a drug and a target as a structure of a graph, constructing a protein graph by using an amino acid sequence of protein and a contact graph, and linearly inputting a standard character string into a molecular graph of the drug based on simplified molecules; (2) extracting representations of drugs and targets, integrating protein pocket information into a topological structure, creating virtual nodes, incorporating the virtual nodes into a residue graph, and extracting representations by using different graph neural network modules after obtaining an enhanced residue graph and a molecular graph; and (3) connecting the representations and inputting the representations into a multi-layer perceptron to predict affinity.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedicine and image processing, and in particular to a method for predicting drug-target affinity based on graph neural network topology enhancement, which is mainly used for protein representation, drug screening, discovery and design. Background Art

[0002] The application areas of drug-target affinity prediction include drug discovery and development and disease treatment. In practical applications, neural networks are often used to predict the affinity between drugs and targets. Common deep learning-based methods, such as convolutional neural networks, are widely used in drug-target affinity prediction. These models usually convert three-dimensional compounds into one-dimensional sequences to extract basic characteristic information of binding affinity. However, one-dimensional sequences lack sufficient structural information, which limits the prediction accuracy.

[0003] Some recent methods use graph structures to represent compounds and proteins. In compounds, drug atoms are considered as nodes, and chemical bonds are edges. In proteins, residues are considered as nodes, and the contact graph defines the topology, which is a matrix with distances between residues. Compared with one-dimensional sequence input, graph-based models can capture more information from drugs and proteins, including atom types, bonds between atoms, and topology, thereby improving the accuracy of drug-target affinity prediction. These methods include representing compounds as graph structures, using atoms as nodes and chemical bonds as edges; representing proteins as graph structures, using residues as nodes, and contact graphs to define the topology. These methods can capture more structural information, thereby improving the accuracy of drug-target affinity prediction.

[0004] The structure of the graph can retain more structural information and increase the diversity of drug and protein features, which has promoted the development of various graph neural network-based methods to capture these features. For example, recent studies have enhanced the feature representation of graph neural networks through self-attention and pooling mechanisms of graph neural networks, as well as various pooling strategies. There are also methods that fuse drug and residue representations through attention mechanisms. Or some methods consider the impact of water molecules on the prediction of drug-target affinity, use ligand non-binding and ligand-binding graphs, and simulate the atomic graph between the drug and the target protein. Or a single undirected graph is used to represent the drug-target complex, incorporating covalent bonds in the protein.

[0005] However, despite their progress, these methods still fail to fully represent the physicochemical properties of proteins, especially the protein pocket structures. They also lack appropriate ways to transfer information after changing the structure of the graph. Therefore, accurately representing the topological structure close to the drug-target binding pocket is crucial to improve DTA (Drug Target Affinity) predictions. Summary of the invention

[0006] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a prediction method for drug-target affinity based on graph neural network topology enhancement, which can better describe the physicochemical properties of proteins, especially the structure and dynamics of pocket regions, and can better capture the key patterns and features in DTA prediction, thereby improving the accuracy of prediction.

[0007] The technical solution of the present invention is: the method for predicting drug-target affinity based on graph neural network topology enhancement comprises the following steps:

[0008] (1) Drugs and targets are represented as graph structures. The protein graph is constructed using the amino acid sequence and contact graph of the protein. The molecular graph of the drug is based on a simplified molecular linear input specification string.

[0009] (2) Extract representations of drugs and targets, integrate protein pocket information into topological structures, and create virtual nodes to incorporate them into the residue graph. After obtaining the enhanced residue graph and molecular graph, different graph neural network modules are used to extract representations.

[0010] (3) These representations are concatenated and input into a multilayer perceptron to predict affinity.

[0011] By integrating additional topological information into the original graph, more contextual information is provided for the relationships between nodes and edges in the graph, allowing the model to better capture patterns and features in the graph structure. In addition, protein pocket information (as topological information) is integrated into the protein residue graph structure, and a new protein representation method and virtual nodes are used to extract protein topological information. This method can better describe the physicochemical properties of proteins, especially the structure and dynamics of pocket regions. In this way, key patterns and features in DTA predictions can be better captured, thereby improving the accuracy of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The architecture of the model of the present invention is shown.

[0013] Figure 2 is the topology of the residue graph. (a) There is no connection between the virtual node and the nearby residue nodes. (b) Information is exchanged between the residue nodes and the virtual nodes.

[0014] Figure 3 It is a multi-topology graph neural network module, which consists of three graph convolutional layers using different topologies.

[0015] Figure 4The figure shows the performance of this method compared with other methods on the Davis dataset, where RMSE, MSE, Pearson, Spearman and CI represent various model evaluation indicators. The arrows after the indicators represent that the larger or smaller the value, the higher the corresponding evaluation of the model.

[0016] Figure 5 The performance of this method is compared with other methods on the KIBA dataset.

[0017] Figure 6 The performance of various message passing methods between virtual nodes and residual nodes is evaluated using the mean and standard deviation values ​​of six evaluation metrics.

[0018] Figure 7 Ablation experiments are performed on the Davis dataset to verify the effectiveness of this method.

[0019] Figure 8 It is a flow chart of the method for predicting drug-target affinity based on graph neural network topology enhancement of the present invention. DETAILED DESCRIPTION

[0020] like Figure 1 , Figure 8 As shown, this method for predicting drug-target affinity based on graph neural network topology enhancement includes the following steps:

[0021] (1) Drugs and targets are represented as graph structures. The protein graph is constructed using the amino acid sequence and contact graph of the protein. The molecular graph of the drug is based on a simplified molecular linear input specification string.

[0022] (2) Extract representations of drugs and targets, integrate protein pocket information into topological structures, and create virtual nodes to incorporate them into the residue graph. After obtaining the enhanced residue graph and molecular graph, different graph neural network modules are used to extract representations.

[0023] (3) These representations are concatenated and input into a multilayer perceptron to predict affinity.

[0024] By integrating additional topological information into the original graph, more contextual information is provided for the relationships between nodes and edges in the graph, allowing the model to better capture patterns and features in the graph structure. In addition, protein pocket information (as topological information) is integrated into the protein residue graph structure, and a new protein representation method and virtual nodes are used to extract protein topological information. This method can better describe the physicochemical properties of proteins, especially the structure and dynamics of pocket regions. In this way, key patterns and features in DTA predictions can be better captured, thereby improving the accuracy of predictions.

[0025] Preferably, in step (1), for a drug, the simplified molecular linear input specification string of the drug is converted into a molecular graph, wherein atoms are nodes and chemical bonds between atoms are edges; considering the self-action of atoms in the message transmission process, a self-loop is added to each atom, and the node features in the molecular graph are composed of five attributes: atomic symbol, atomic degree, number of hydrogen atoms bound to the atom, number of hydrogen atoms implicitly bound to the atom, and whether the atom is aromatic. These attributes are combined into a multidimensional feature vector, and the molecular graph is represented as:

[0026] G d =(V d ,A d )

[0027] Among them G d represents a bidirectional graph, V d Represents the features in the graph, A d is an adjacency matrix with self-loops;

[0028] For proteins, they are represented as residue graphs. In proteins, residues are considered nodes, and edges are determined by the contact graph of the protein. Node features include position-specific scoring matrices, one-hot encoding of residue symbols, and multiple residue attributes. In order to construct the residue graph, a contact matrix with a dimension of (L, L) is obtained by prediction, where L represents the length of the amino acid sequence. This matrix shows the probability of contact between residues, ranging from 0 to 1. Higher values ​​indicate greater contact probability. A threshold of 0.5 is applied to the contact graph matrix to form the adjacency matrix of the protein graph. The residue graph is represented as:

[0029] G p =(V p ,A p )

[0030] A p is an adjacency matrix with self-loops, V p Residue characteristics.

[0031] Preferably, in step (1), in order to enhance the expression of the residue graph, the pocket information of each protein is obtained from the protein pocket database and integrated into the residue graph. The pocket information of each protein is defined as {pt k |k∈l}, where l represents the number of protein pockets, represents the amino acid position in the kth pocket, and to facilitate communication between nodes associated with these pockets, a virtual node is created for each pocket in the residue graph. For each protein residue graph, an additional feature matrix V is created for the virtual nodes. vn = {v m+1 , …, v m+l}, where l represents the number of protein pockets and m represents the number of nodes. These virtual node vectors are initialized to zero.

[0032] Preferably, in step (1), in order to enhance protein representation, each protein pocket is represented as a virtual node, and this virtual node is connected to all adjacent residue nodes to capture local interactions within the pocket region.

[0033] Preferably, in step (1), information is transferred from the residue node to the virtual node, and the original residue graph is represented as:

[0034] G p =(V p ,A p )

[0035] Where V p =v 1 , ..., v m and A p =[a ij ] m×m , m is the number of nodes, and the structure with virtual nodes is represented as a pocket graph G pt = {V pt , A pt} and G′ pt = {V pt , A′ pt}, where V pt = {V p , V vn}={v 1 , …, v m , v m+1 , …, v m+1} is a node feature formed by connecting the residue feature and the virtual node, G {pt} and G′ {pt} The topological structures of pt =[b ij ] (l+m)×(l+m) and A′ pt =[b′ ij ] (l+m)×(l+m) ; G pt and G′ pt They are all different topological structures composed of residue graphs and virtual nodes. pt and A′ pt The calculation formula is as follows:

[0036]

[0037] Where m is the number of residue nodes, l is the number of protein pockets, pt jRepresents the amino acid position in the jth protein pocket. i and j are integer types, and their values ​​range between 1 and m+l.

[0038] Preferably, in step (2), in order to obtain drug representation, a graph convolutional network is used to extract features from the pocket graph and the molecule graph. For a graph G=(V, A), where V represents the feature matrix and A is the adjacency matrix of the graph, the message passing process of the graph convolutional network is determined by the following formula:

[0039]

[0040] H 0 =V

[0041]

[0042] Where σ represents a nonlinear activation function; corresponds to an adjacency matrix with self-loops; express The graph convolutional network aggregates information from neighboring nodes, combines this information with the node’s own features, and passes the combined information through a learnable weight matrix and then a nonlinear activation function.

[0043] Preferably, in step (2), in order to extract the graph representation, all node features are passed through a pooling layer to compress them into a unified graph representation, and three GCN layers are used to extract the drug representation:

[0044] V′ d =GCN(V d ,A d )

[0045]

[0046]

[0047]

[0048] Among them, Figure 3 As shown in the figure, GCN consists of three graph convolutional layers; Pooling is a global average pooling layer; Dropout represents a dropout layer with a dropout rate of 0.2; W and b represent weights and bias terms respectively; L represents the number of network layers; Relu activation function is used, and a dropout layer is added to prevent overfitting; given an input graph G d =(V d , A d ), the drug representation V is derived drug .

[0049] Preferably, in step (2), in order to extract protein representation, a multi-topology graph neural network block (such as Figure 2 This block uses different message passing strategies between its layers; to prevent over-smoothing, G′ is applied in the third layer. pt The topology of the pocket graph G, where messages are passed from residue nodes to virtual nodes; message passing near protein pockets, using unidirectional connections for message passing; pt and G′ pt It is input into the multi-topology graph neural network block, and the final protein representation is obtained through global average pooling; the obtained drug and protein representations are concatenated, and the final affinity value is predicted through a multi-layer perceptron:

[0050]

[0051] Among them, the MLP consists of two fully connected layers, each followed by a dropout and ReLU as a nonlinear activation function; represents the predicted affinity value.

[0052] like Figures 4 to 7 As shown, in order to verify the effectiveness of the proposed scheme, this model is trained on the Davis and KIBA datasets, and the model is evaluated on the test set to verify the effectiveness of the drug-target affinity method.

[0053] First, the amino acid sequences of proteins in the Davis dataset and the simplified molecular linear input specification string of molecules are extracted. Then the pocket information of proteins in the Davis and KIBA datasets is extracted, and protein and molecular graph representations are constructed.

[0054] The datasets are then divided into training set, validation set, and test set on Davis and KIBA, respectively. The model is then trained on the training set, and the model hyperparameters are adjusted on the validation set through k-fold cross validation. Finally, the training set and validation set are used as training sets, affinity is predicted on the test set, and the model is evaluated on the Davis and KIBA datasets, respectively. The effectiveness is verified through ablation experiments on the Davis dataset, and the impact of different message passing methods on affinity prediction is verified.

[0055] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting drug-target affinity based on graph neural network topology enhancement, characterized by: The method comprises the following steps: (1) Drugs and targets are represented as graph structures. The protein graph is constructed using the amino acid sequence and contact graph of the protein. The molecular graph of the drug is based on a simplified molecular linear input specification string. (2) Extract representations of drugs and targets, integrate protein pocket information into topological structures, and create virtual nodes to incorporate them into the residue graph. After obtaining the enhanced residue graph and molecular graph, different graph neural network modules are used to extract representations. (3) These representations are concatenated and input into a multilayer perceptron to predict affinity.

2. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 1, characterized in that: In the step (1), for a drug, the simplified molecular linear input specification string of the drug is converted into a molecular graph, in which atoms are nodes and chemical bonds between atoms are edges; considering the self-action of atoms in the message transmission process, a self-loop is added to each atom, and the node features in the molecular graph are composed of five attributes: atomic symbol, atomic degree, number of hydrogen atoms bound to the atom, number of hydrogen atoms implicitly bound to the atom, and whether the atom is aromatic. These attributes are combined into a multidimensional feature vector, and the molecular graph is represented as: G d =(V d ,A d ) Among them G d represents a bidirectional graph, V d Represents the features in the graph, A d is an adjacency matrix with self-loops; For proteins, they are represented as residue graphs. In proteins, residues are considered nodes, and edges are determined by the contact graph of the protein. Node features include position-specific scoring matrices, one-hot encoding of residue symbols, and multiple residue attributes. In order to construct the residue graph, a contact matrix with a dimension of (L, L) is obtained by prediction, where L represents the length of the amino acid sequence. This matrix shows the probability of contact between residues, ranging from 0 to 1. Higher values ​​indicate greater contact probability. A threshold of 0.5 is applied to the contact graph matrix to form the adjacency matrix of the protein graph. The residue graph is represented as: G p =(V p ,A p ) A p is an adjacency matrix with self-loops, V p Residue characteristics.

3. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 2, characterized in that: In the step (1), in order to enhance the expression of the residue graph, the pocket information of each protein is obtained from the protein pocket database and integrated into the residue graph. The pocket information of each protein is defined as {pt k |k∈l}, where l represents the number of protein pockets, represents the amino acid position in the kth pocket, and to facilitate communication between nodes associated with these pockets, a virtual node is created for each pocket in the residue graph. For each protein residue graph, an additional feature matrix V is created for the virtual nodes. vn = {v m+1 , …, v m+l }, where l represents the number of protein pockets and m represents the number of nodes, and these virtual node vectors are initialized to zero.

4. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 3, characterized in that: In step (1), in order to enhance protein representation, each protein pocket is represented as a virtual node, which is connected to all adjacent residue nodes to capture local interactions within the pocket region.

5. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 3, characterized in that: In step (1), information is transferred from the residue node to the virtual node, and the original residue graph is represented as: G p =(V p ,A p ) Where V p =v1,…,v m and A p =[a ij ] m×m , m is the number of nodes, and the structure with virtual nodes is represented as a pocket graph G pt =(V pt , A pt ) and G′ pt =(V pt , A′ pt ), where V pt = {V p , V vn }={v1,…,v m , v m+1 , …, v m+1 } is a node feature formed by connecting the residue feature and the virtual node, G {pt} and G′ {pt} The topological structures of pt =[b ij ] (l+m)×(l+m) and A′ pt =[b′ ij ] (l+m)×(l+m) ; G pt and G′ pt They are all different topological structures composed of residue graphs and virtual nodes. pt and A′ pt The calculation formula is as follows: Where m is the number of residue nodes, l is the number of protein pockets, pt j Represents the amino acid position in the jth protein pocket. i and j are integer types, and their values ​​range between 1 and m+l.

6. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 5, characterized in that: In step (2), in order to obtain drug representation, a graph convolutional network is used to extract features from the pocket graph and the molecule graph. For a graph G = (V, A), where V represents the feature matrix and A is the adjacency matrix of the graph, the message passing process of the graph convolutional network is determined by the following formula: Where σ represents a nonlinear activation function; corresponds to an adjacency matrix with self-loops; express The graph convolutional network aggregates information from neighboring nodes, combines this information with the node’s own features, and passes the combined information through a learnable weight matrix and then a nonlinear activation function.

7. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 6, characterized in that: In step (2), in order to extract the graph representation, all node features are passed through a pooling layer to compress them into a unified graph representation, and three GCN layers are used to extract the drug representation: V′ d =GCN(V d ,A d ) Among them, GCN consists of three graph convolutional layers; Pooling is a global average pooling layer; Dropout represents a dropout layer with a dropout rate of 0.2; W and b represent weights and bias terms respectively; L represents the number of network layers; Relu activation function is used, and a dropout layer is added to prevent overfitting; given an input graph G d =(V d , A d ), the drug representation V is derived drug .

8. The method for predicting drug-target affinity based on graph neural network topology enhancement according to claim 7, characterized in that: In step (2), in order to extract protein representation, a multi-topology graph neural network block is constructed, which uses different message passing strategies between its layers; in order to prevent over-smoothing, G′ is applied in the third layer pt The topology of the pocket graph G, where messages are passed from residue nodes to virtual nodes; message passing near protein pockets, using unidirectional connections for message passing; pt and G′ pt It is input into the multi-topology graph neural network block, and the final protein representation is obtained through global average pooling; the obtained drug and protein representations are concatenated, and the final affinity value is predicted through a multi-layer perceptron: Among them, the MLP consists of two fully connected layers, each followed by a dropout and ReLU as a nonlinear activation function; represents the predicted affinity value.

Images