Drug target prediction model and method based on graph neural network

Abstract

The invention is applicable to the technical field of biological information, and provides a drug target prediction method based on a graph neural network, which comprises the following steps of: respectively taking a drug and a target node as a core, and automatically extracting local sub-graphs of the drug and the target node to capture local context information of the nodes in a complex biological network; taking different types of edges as feature information and fusing the feature information into a feature learning process, dynamically adjusting contribution degrees of the different types of edges to node representation through an attention mechanism, introducing residual connection and a normalization processing strategy, and fusing node features; and performing deep mining on the fused node features by using a two-layer graph convolutional network, and calculating a potential interaction prediction score between the drug and the target. The invention further provides a drug target prediction model based on the graph neural network. According to the method, independent feature extraction and training are carried out on predicted drugs and target spots, and the accuracy and richness of node representation are effectively improved while the flexibility of the model is kept.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and in particular relates to a drug target prediction model and method based on graph neural network. Background Art

[0002] In practical applications, drug target prediction technology has shown great potential. For example, in the development of anticancer drugs, multiple new cancer targets can be predicted through bioinformatics technology, and some compounds with potential therapeutic effects have been discovered. These compounds have been further verified by experiments and have shown good anticancer activity, providing new ideas and methods for the development of anticancer drugs. In addition, drug target prediction technology can also be used to predict the side effects and drug resistance of drugs, helping people optimize drug structures, reduce adverse drug reactions, and improve the safety and effectiveness of drugs. In the future, with the continuous advancement and improvement of technology, drug target interaction prediction technology will be more efficient and accurate, bringing new breakthroughs to the development of new drugs. On the one hand, with the continuous development of bioinformatics, computer science and artificial intelligence technology, drug target prediction methods will become more diversified and intelligent; on the other hand, with the continuous accumulation and sharing of biological data, people can use more experimental data to train and verify models and improve the reliability and accuracy of predictions.

[0003] DTI (drug-target interaction) prediction, as a key task to analyze the potential connection between drugs and targets in the body, is usually conceived as a binary classification problem, which aims to determine whether a given drug-target pair (DTP) has an interaction. However, in practical applications, labeled DTP data is extremely scarce and expensive to obtain. This bottleneck seriously restricts the efficiency and depth of the model's learning of DTP relationships. In order to break through this limitation, heterogeneous biological data, with its rich multi-source information and multi-dimensional perspectives, has opened up new ways to model the complex relationship between drugs and targets. These data not only cover the structural characteristics of drugs and the biological properties of targets, but also involve information on disease associations, drug side effects, and other aspects, providing strong support for a comprehensive understanding of DTI mechanisms.

[0004] Currently, the research on DTI prediction using biological heterogeneous data shows a diversified trend, mainly including three directions:

[0005] The first is the similarity-based method, which infers potential interactions by calculating the similarity between drugs and targets. The similarity-based method has a solid theoretical foundation in the field of drug-target interaction (DTI) prediction. Its core assumption is that in biology and pharmacology, drugs with similar functions or structures tend to bind to the same targets. Conversely, targets with similar properties may also attract drugs with the same or similar functions. This assumption has driven the proposal of a series of three-step strategies aimed at more accurately predicting the interaction between drugs and targets.

[0006] However, although these strategies have demonstrated their effectiveness in many aspects, they have also exposed some limitations. Specifically, these methods are unable to cope with complex data relationships. In real biological systems, drugs with very different chemical structures may bind to the same target due to specific molecular mechanisms. This non-intuitive binding mode is often ignored by similarity-based methods. In addition, the separation of feature learning from prediction tasks may also result in the learned representation not being fully suitable for subsequent prediction tasks, because the goal of feature learning may deviate from the optimal solution of the prediction task. This suboptimal solution limits the accuracy and generalization ability of DTI predictions, prompting researchers to continuously explore more integrated and intelligent prediction methods.

[0007] The second is a method based on knowledge graph (KG), which constructs a complex network containing drugs, targets and their relationships, and uses the structural information of the graph to assist prediction. Knowledge graph (KG)-based technologies are increasingly widely used in the biomedical field. They construct an information-rich biological heterogeneous graph by taking biological entities (such as genes, proteins, drugs, diseases, etc.) as nodes and the complex relationships between these entities as edges. In order to effectively mine knowledge and perform reasoning in this huge graph, researchers have introduced knowledge graph embedding models such as TransE, DistMult, and ComplEx, which aim to map the nodes and relationships in the graph into a low-dimensional vector space, thereby simplifying the processing of complex relationships and promoting computational feasibility. However, although these shallow embedding models have made significant progress in representation learning of biological heterogeneous graphs, they still have some shortcomings in modeling compositional relationships. Specifically, the relationships in biological heterogeneous graphs are often highly complex and diverse, including but not limited to the interactions between drugs and targets, the associations between genes and diseases, and the interaction networks between proteins. These relationships are not only huge in number, but also often present multi-level and multi-scale characteristics, making it difficult for simple embedding models to fully capture their intrinsic structure and semantic information.

[0008] The third is the method based on graph neural network (GNN), which can automatically learn the representation of nodes in the graph and capture the deep interaction pattern between drugs and targets, thereby achieving more accurate DTI prediction. When processing complex biological heterogeneous data based on the graph neural network (GNN) method, the existing technology integrates multi-source biological data into multiple interrelated drug and target networks. These networks not only cover drug-target interactions, but may also include drug-drug similarity, target-target similarity, drug-side effects, target-disease associations and other types of information. By adopting models based on graph convolutional networks (GCN) or graph attention networks (GAT), GNN can effectively extract deep representations of drugs and targets from these complex networks. These representations are rich in topological structures and semantic information between nodes, providing strong support for subsequent DTI (drug-target interaction) predictions.

[0009] IMCHGAN and SGCL-DTI are two representative models in this field. They introduce the concept of meta-paths to mine richer drug (or target) representations from heterogeneous graphs. A meta-path is a path pattern defined on a heterogeneous graph that can reveal complex relationships between different node types. However, these two models have certain limitations in utilizing meta-paths. They each extract meta-paths independently from each individual network, ignoring the cross-information and potential synergistic effects between networks. In addition, they fail to fully utilize contextual information when processing meta-paths, that is, the attribute information and structural information of the nodes on the path that the meta-path passes through, which is crucial for a deep understanding of the interaction mechanism between drugs and targets. In contrast, models such as MHGNN and MHTAN-DTI pay more attention to the contextual information of meta-paths when extracting feature representations. They not only aggregate features along the meta-paths, but also consider the characteristics of each node on the path and the interactions between nodes, so as to more comprehensively capture the complex relationship between drugs and targets. However, in the dual-channel feature extraction process of the MHGNN model, the selection of its input meta-path relies on manual design, which not only increases the subjectivity and complexity of the model, but also may limit the generalization ability of the model on different data sets. In addition, there is still room for further improvement in the algorithm efficiency and implementation details of MHGNN to meet the needs of large-scale biological heterogeneous graph processing. Summary of the invention

[0010] The purpose of the embodiment of the present invention is to provide a drug target prediction method based on graph neural network, aiming to solve the problems raised in the above background technology.

[0011] The embodiment of the present invention is implemented as follows: a drug target prediction method based on graph neural network, comprising the following steps:

[0012] Taking drug and target nodes as the core, local subgraphs are automatically extracted to capture the local contextual information of nodes in complex biological networks.

[0013] Different types of edges are considered as feature information and integrated into the feature learning process. The contribution of different types of edges to node representation is dynamically adjusted through the attention mechanism. Residual connection and normalization processing strategies are introduced to fuse node features.

[0014] A two-layer graph convolutional network is used to deeply mine the fused node features and calculate the potential interaction prediction score between drugs and targets.

[0015] Preferably, the step of automatically extracting local subgraphs with drug and target nodes as the core to capture local context information of nodes in complex biological networks is specifically as follows:

[0016] Construct a biological heterogeneous graph, which includes multiple types of nodes such as drugs, targets, diseases and side effects, and each type of node represents different feature information;

[0017] A dual-channel learning strategy is adopted to learn the features of the two nodes, drug and target. In each channel, the drug node (D) and the target node (T) are taken as the central nodes, and subgraphs D and T are extracted from the complete heterogeneous graph as the starting points of dual-channel representation learning.

[0018] Preferably, the step further includes: setting the initial representation of nodes such as drugs, targets, diseases and side effects to one-hot encoding form, using a linear transformation layer with a bias term, and applying it independently to each node type to map their features into a common, unified feature space.

[0019] Preferably, the step further includes: defining the heterogeneous graph as Where V is the set of nodes, E is the set of edges, each node v has a type φ(v), each edge e has a type Ψ(e), and the possible sets of node types and edge types are represented as and When |Tv|=|Te|=1, the graph degenerates into a normal homogeneous graph.

[0020] Preferably, the steps of considering different types of edges as feature information and integrating them into the feature learning process, dynamically adjusting the contribution of different types of edges to node representation through the attention mechanism, introducing residual connection and normalization processing strategies, and fusing node features are specifically as follows:

[0021] Learnable edge type feature embedding: In each layer of the graph attention network, each edge type Ψ(e)∈T(e) is assigned a specific dl-dimensional embedding feature, denoted as Taking both node embedding and edge type embedding into account, the attention score α is calculated as:

[0022]

[0023] Where Ψ(<i, j>) represents the type of edge with nodes i and j as endpoints, W r is a learnable matrix for type embedding transformation;

[0024] Residual connection: A node residual connection is added between the cross-layer node representations, and the node aggregation on the lth layer is represented as:

[0025]

[0026] in, is the attention weight about edge <i, j>, σ is an activation function (default is ELU[]);

[0027] When the dimension changes at layer l, a learnable linear transformation needs to be added:

[0028]

[0029] According to formula (2), K independent attention mechanisms are executed and their results are connected as the final representation. The formula is parsed as:

[0030]

[0031]

[0032]

[0033] In the above formula, || represents the concatenation operation. Represents the kth linear transformation The calculated attention scores are then averaged in the last layer (layer L), i.e.:

[0034]

[0035] L2 Normalization: Through L2 normalization, all features are scaled to the same magnitude:

[0036]

[0037] Among them, f i represents the output feature embedding of node i, The final expression from formula (7).

[0038] Preferably, the step of using a two-layer graph convolutional network to deeply mine the fused node features and calculate the potential interaction prediction score between the drug and the target is specifically:

[0039] Construct a DTP correlation graph to discover the potential relationships between DTPs, where each node is a DTP;

[0040] By concatenating the drug feature representation and the target feature representation, the comprehensive representation fdt of each node in the reconstructed DTP graph is obtained, that is, fdt = fd||ft, where d∈VD, t∈VT;

[0041] When processing a batch of DTPs, the representation z of each DTP calculated in the above way is combined into a representation matrix Z, in which the b-th row of Z is the eigenvector of the b-th DTP in the batch. The adjacency matrix A is constructed from the DTP representation and is expressed as:

[0042] A ij =softmax((ZZ T )) ij (9);

[0043] Among them, A ij is the element in the i-th row and j-th column of A, which represents the normalized similarity between the i-th drug and the j-th target in the same batch of data. It can be seen that the matrix A is a symmetric matrix. ij The larger the value, the stronger the association between the i-th drug and the j-th target DTP;

[0044] The DTP matrix Z is input into the two-layer GCN, and the final output y is used to predict DTI. The calculation process of y is:

[0045] y = A(ReLU(AZW1))W2 (10);

[0046] Where W1 and W2 represent the learnable parameters of the first and second layers of GCN respectively;

[0047] Binary cross entropy loss is used to evaluate the difference between the predicted DTI and the true DTI, and the formula is:

[0048]

[0049] Among them, B represents the number of DTPs in a batch training, y b represents the actual value of DTP. When D and P interact, y b =1, otherwise, y b =0,y′ b Represents the predicted value of DTP, which is normalized to y′b ∈[0, 1].

[0050] Another object of an embodiment of the present invention is to provide a drug target prediction model based on a graph neural network, which is used to implement the above-mentioned drug target prediction method based on a graph neural network, including:

[0051] The subgraph construction module is used to automatically extract local subgraphs with drug and target nodes as the core to capture the local context information of nodes in complex biological networks;

[0052] The attention fusion module is used to consider different types of edges as feature information and integrate them into the feature learning process. It dynamically adjusts the contribution of different types of edges to node representation through the attention mechanism, introduces residual connection and normalization processing strategies, and fuses node features.

[0053] The prediction analysis module is used to conduct in-depth mining of the fused node features and calculate the potential interaction prediction scores between drugs and targets.

[0054] The drug target prediction method based on graph neural network provided by the embodiment of the present invention abandons the traditional meta-path dependence and adopts a node aggregation strategy without meta-path to solve the problems of high sensitivity and poor flexibility in the node aggregation process of manually selected meta-path. By constructing a dual-channel subgraph, independent feature extraction and training are performed for the two types of predicted core nodes (drugs and targets), which effectively improves the accuracy and richness of node representation while maintaining the flexibility of the model.

[0055] An embodiment of the present invention provides a drug target prediction model based on a graph neural network, which removes the dependence on meta-paths and adopts dual-channel sub-graph learning and attention fusion mechanism to provide a more flexible, efficient and powerful solution for drug-target interaction prediction. It is of great significance in computer-assisted drug discovery and reuse. In particular, the system does not require manual setting of meta-paths, has higher flexibility and universality, can automatically adapt to different data sets and task requirements, reduces the complexity of human intervention and design, and can help many people who do not understand the relevant fields of the data sets used to get started faster. In addition, the accuracy of the embodiment of the present invention is better than that of common drug target prediction models. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A prediction flow chart of a drug target prediction module based on a graph neural network is provided for an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of drug / target feature representation learning provided by an embodiment of the present invention;

[0058] Figure 3 ROC curves of the model provided in the embodiment of the present invention and the MHGNN and LDMLP-DTI models on specific tasks:

[0059] Figure 4 PR curves of the model provided in the embodiment of the present invention and the MHGNN and LDMLP-DTI models on specific tasks. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0061] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0062] like Figure 1 As shown, a prediction flow chart of a drug target prediction module based on a graph neural network provided by an embodiment of the present invention includes:

[0063] Subgraph construction module: This module automatically extracts local subgraphs around drug and target nodes to capture local contextual information of nodes in complex biological networks.

[0064] Attention Fusion Module: This module considers different types of edges as feature information and integrates them into the feature learning process. It dynamically adjusts the contribution of different types of edges to node representation through the attention mechanism. At the same time, it introduces residual connection and normalization processing strategies to effectively alleviate the gradient vanishing problem in deep networks and improve the learning efficiency and stability of the model.

[0065] Prediction and analysis module: This module uses a two-layer graph convolutional network (GCN) to deeply mine the fused node features and calculate the potential interaction prediction score between drugs and targets. It not only simplifies the complexity of traditional path dependence, but also significantly improves the accuracy and robustness of the prediction.

[0066] Specifically, a drug target prediction method based on graph neural network includes the following steps:

[0067] (1) Construction and initialization of drug / target subgraph:

[0068] Based on heterogeneous biological data from a variety of public database resources, a biological heterogeneous graph is first constructed, which covers multiple types of nodes such as drugs, targets, diseases and side effects, and each type of node represents different feature information. In order to fully capture and retain the specific attributes of drugs and targets, a dual-channel learning strategy is used to learn the characteristics of these two nodes separately, so as to ensure that when processing complex biological information, the essential differences and internal connections between different entities can be accurately reflected. The difference is that this model gets rid of the inefficiency of meta-path selection in most existing models. In each channel, the drug node (D) and the target node (T) are respectively used as the central nodes, and subgraphs D and T are extracted from the complete heterogeneous graph as the starting point of dual-channel representation learning;

[0069] Specifically, taking Subgraph-D as an example, all edges with node D as the endpoint are extracted from the initial graph and reconstructed into a new heterogeneous graph. This new heterogeneous graph is used to further learn the feature representation of node D. Since there are no TS type edges in the dataset, there are only three types of nodes in the subgraph Subgraph-T, namely D, T, and I.

[0070] In order to effectively integrate and process different types of nodes such as drugs, targets, diseases, and side effects in a unified framework, the initial representation of these nodes is first set to one-hot encoding (e.g. Figure 2 As shown in the figure, given that different types of nodes may have different initial feature dimensions, a linear transformation layer with a bias term (also called a fully connected layer) is used independently for each node type to map their features into a common, unified feature space. The linear layer parameters in this mapping process will be continuously optimized with the subsequent training process of the network model to ensure that the interactions and associations between nodes can be effectively modeled and analyzed under a unified framework.

[0071] A heterogeneous graph is defined as Where V is the set of nodes, E is the set of edges, each node v has a type φ(v), each edge e has a type Ψ(e), and the possible sets of node types and edge types are represented as and When |Tv|=|Te|=1, the graph degenerates into a normal homogeneous graph;

[0072] (2) Drug / target feature representation learning, such as Figure 2 As shown:

[0073] Inspired by the advantages of the Graph Attention Network (GAT) model applied to complex heterogeneous graph neural networks, a simple and effective heterogeneous graph representation learning method is proposed, which uses the graph attention mechanism as the backbone and adds three contents applicable to the model of the embodiment of the present invention: learnable edge type feature embedding, residual connection and L2 normalization;

[0074] Learnable edge type feature embedding: GAT does not consider the diversity of nodes or edges and is not directly applicable to heterogeneous graphs (graphs containing multiple types of nodes and edges). In order to make up for this deficiency and enhance the adaptability of GAT to heterogeneous graphs, an extended method is proposed to integrate edge type information into the attention calculation process. Specifically, in each layer of the graph attention network, a specific dl-dimensional embedding feature is assigned to each edge type Ψ(e)∈T(e), expressed as These embedding vectors can capture the unique semantic or structural information contained in different edge types. Considering both node embedding and edge type embedding, the attention score α is calculated as:

[0075]

[0076] Among them, Ψ(<i,j> ) represents the type of edge with nodes i and j as endpoints, W r is a learnable matrix for type embedding transformation;

[0077] Residual connection: In order to alleviate the problems of gradient disappearance and gradient explosion, a node residual connection is added between cross-layer node representations, providing a path to bypass nonlinear transformations so that gradients can be directly passed to deeper layers. At the same time, through carefully designed pre-activation implementation, residual connections play a powerful role in GNNs.

[0078] The node aggregation on layer l is expressed as:

[0079]

[0080] in, It's about the edge<i,j> The attention weight, σ is an activation function (default is ELU[]);

[0081] When the dimension changes at layer l, a learnable linear transformation needs to be added:

[0082]

[0083] Similar to GAT, a multi-head attention mechanism is used to enhance the expressiveness of the model. Specifically, K independent attention mechanisms are executed according to formula (2), and their results are connected as the final representation. The formula is parsed as follows:

[0084]

[0085]

[0086]

[0087] In the above formula, || represents the concatenation operation. Represents the kth linear transformation The calculated attention score;

[0088] In the last layer (layer L), the obtained representations are averaged, that is:

[0089]

[0090] L2 normalization: Through L2 normalization, all features can be scaled to the same magnitude, making the model easier and faster to converge:

[0091]

[0092] Among them, f i represents the output feature embedding of node i, The final expression from formula (7);

[0093] (3) Prediction of drug-target interactions:

[0094] Since drugs with similar structures tend to interact with targets with similar structures, it can be inferred that DTPs with similar structural drugs or targets have stronger correlations than DTPs without similar structures. A DTP correlation graph is constructed to discover the potential relationship between DTPs, where each node is a DTP. By concatenating the drug feature representation and the target feature representation, the comprehensive representation fdt of each node in the reconstructed DTP graph is obtained, that is, fdt = fd||ft, where d∈VD, t∈VT;

[0095] When processing a batch of DTPs, the representation z of each DTP calculated in the above way is combined into a representation matrix Z, in which the b-th row of Z is the eigenvector of the b-th DTP in the batch. The adjacency matrix A is constructed from the DTP representation and is expressed as:

[0096] A ij =softmax((ZZ T )) ij (9);

[0097] Among them, A ijis the element in the i-th row and j-th column of A, which represents the normalized similarity between the i-th drug and the j-th target in the same batch of data. It can be seen that the matrix A is a symmetric matrix. ij The larger the value, the stronger the association between the i-th drug and the j-th target DTP;

[0098] The DTP matrix Z is input into the two-layer GCN, and the final output y is used to predict DTI. The calculation process of y is:

[0099] y = A(ReLU(AZW1))W2 (10);

[0100] Where W1 and W2 represent the learnable parameters of the first and second layers of GCN respectively;

[0101] The binary cross entropy (BCE) loss is used to evaluate the difference between the predicted DTI and the true DTI, and the formula is:

[0102]

[0103] Among them, B represents the number of DTPs in a batch training, y b represents the actual value of DTP. When D and P interact, y b =1, otherwise, y b =0,y′ b Represents the predicted value of DTP, which is normalized to y′ b ∈[0, 1].

[0104] The performance of the model proposed in the embodiment of the present invention in drug-target interaction (DTI) prediction is compared with the other six models in the prior art (NeoDTI, DTI-MGNN, SGLC-DTI, HAN and MGNN), as follows:

[0105] (1) Dataset:

[0106] The first dataset Hetero-A was collected through DTINet. This dataset contains four different types of biological data and six different types of relationships between them. The types and numbers of nodes are 708 drugs (D), 1512 targets (T), 5603 diseases (I), and 4192 side effects (S). The types and numbers of edges are 1923 D-T, 10036 D-D, 199214 D-I, 80164 D-S, 7363 T-T, and 1596745 TI.

[0107] Since insufficient positive samples will lead to false negative risks in data modeling and model training, which in turn leads to a large error between the predicted results and the actual results, a more complete data set is chosen to further train the model in order to optimize the accuracy of model prediction. A new data set Hetero-B is obtained through MHGNN, which is an update of data set A. The types of nodes and edges in data set B are the same as those in data set A, but the number has changed accordingly. The types and numbers of nodes are 2214 drugs (D), 1968 targets (T), 7025 diseases (I), and 3935 side effects (S). The types and numbers of edges are 8750 DT, 1091870 DD, 542970 DI, 104629 D-S, 456592 TT, and 2922064 T-I.

[0108] The model of the embodiment of the present invention was evaluated on two biologically heterogeneous datasets, Hetero-A and Hetero-B.

[0109] (2) Existing technology model:

[0110] NeoDTI first applies GCN to each independent relational network to capture the local and global features of each node (drug or target) and the complex interaction patterns between them. Then, using these learned representations as features, combined with known drug-target interaction data, an inductive matrix completion (IMC) model is trained to predict unknown drug-target interactions (DTI) through known representation features;

[0111] The DTI-MGNN model designs two parallel graph attention network branches, which focus on capturing the interactions between nodes in the topological graph and feature graph based on different strengths. This design allows the model to carefully distinguish and learn information flows from different data sources, thereby more accurately reflecting the complex relationships between nodes. At the same time, the model also introduces a graph convolutional network with a shared weight matrix to extract and integrate the common information between the two graphs, promoting the deep fusion of topological structure and semantic features;

[0112] IMCHGAN uses a two-level graph attention method to analyze the interaction between different nodes and meta-paths, and obtains the potential feature representations of drugs and targets from the DTI heterogeneous network. Finally, the score is predicted by the inductive matrix completion (IMC) model to obtain the interaction relationship between drugs and targets.

[0113] The core idea of ​​the SGCL-DTI model is to use the rich information in the graph structure data to optimize the feature representation of drugs and targets, and to strengthen the distinction between these representations through a contrastive learning mechanism. First, the graph convolutional network technology is used to extract the feature representation of drugs and targets from the neighborhood based on the meta-path. Then, with the drug-target pair as the core node, a topological graph and a semantic graph are constructed respectively. A contrastive learning strategy is used to further refine and optimize the feature representation of drugs and targets by comparing and distinguishing the information contained in the two graphs.

[0114] The HAN model is based on meta-paths and attention mechanisms, and adopts a hierarchical structure to process information within and between meta-paths to obtain feature representations of central nodes. Since the HAN model is based on node classification and clustering tasks, we designed two parallel HAN-based channels to learn the feature representations of drugs and targets respectively according to the task objectives. Finally, the outputs of the two channels are connected and the final prediction results are obtained through two linear layers;

[0115] MAGNN is an improved model based on HAN. This is because HAN only aggregates the information of two terminal nodes when processing information within a meta-path, while MAGNN aggregates all node representations in a meta-path instance. At the same time, MAGNN establishes a DTP correlation graph, assigns different weights to different DTP associations, and further optimizes the feature learning module and DTI prediction module.

[0116] (3) Comparison results:

[0117] It is measured by multiple evaluation indicators, including AUC, AUPR, Precision, Recall, F1 and MCC. Each model shows its performance data on different partitioned subsets of cross-validation, which are presented in the form of mean ± standard deviation. The results are shown in Table 1 and Table 2:

[0118] Table 1 Comparison results on the Hetero-A dataset

[0119]

[0120]

[0121] Table 2 Comparison results on the Hetero-B dataset

[0122]

[0123] According to Tables 1 and 2, it can be seen that the HAN and MHGNN models, especially the model of the embodiment of the present invention, show high performance in multiple evaluation indicators. In particular, the model of the embodiment of the present invention has reached a high level in indicators such as AUC, AUPR, Precision and F1, showing its superiority in the DTI prediction task. In addition, the model of the embodiment of the present invention also performs well in the MCC (Matthews Correlation Coefficient) indicator, which shows that the correlation between its prediction results and the actual situation is very high. In contrast, although the NeoDTI model also performs well in some indicators, it is relatively weak in comprehensive indicators such as F1 and MCC; and the performance of the SGLC-DTI model in multiple indicators is relatively low, which may indicate that the model has certain limitations in processing the current DTI prediction task. In general, the embodiment of the present invention has excellent performance and has potential application and research value.

[0124] The performance of the model proposed in the embodiment of the present invention is compared with that of the MHGNN and LDMLP-DTI models on a specific task by drawing their precision-recall curves:

[0125] Precision and recall are two key indicators for evaluating classifier performance. Precision reflects the proportion of samples predicted by the model to be positive and actually positive, while recall reflects the proportion of samples that are actually positive and correctly predicted by the model to be positive. Figure 3 , Figure 4 As shown in the figure, it can be seen that the curve of the model of the embodiment of the present invention is closest to the upper left corner of the coordinate axis (that is, the ideal point where the precision is 1.0 and the recall rate is 1.0), and its corresponding area (area=0.96) is the largest, which means that the model maintains a high recall rate while also maintaining a high precision, and has the best performance among the three models.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A drug target prediction method based on graph neural network, characterized in that: The following steps are involved: Taking drug and target nodes as the core, local subgraphs are automatically extracted to capture the local contextual information of nodes in complex biological networks. Different types of edges are considered as feature information and integrated into the feature learning process. The contribution of different types of edges to node representation is dynamically adjusted through the attention mechanism. Residual connection and normalization processing strategies are introduced to fuse node features. A two-layer graph convolutional network is used to deeply mine the fused node features and calculate the potential interaction prediction score between drugs and targets.

2. The drug target prediction method based on graph neural network according to claim 1, characterized in that: The steps of automatically extracting local subgraphs with drug and target nodes as the core to capture local context information of nodes in complex biological networks are as follows: Construct a biological heterogeneous graph, which includes multiple types of nodes such as drugs, targets, diseases and side effects, and each type of node represents different feature information; A dual-channel learning strategy is adopted to learn the features of the two nodes, drug and target. In each channel, the drug node (D) and the target node (T) are taken as the central nodes, and subgraphs D and T are extracted from the complete heterogeneous graph as the starting points of dual-channel representation learning.

3. The drug target prediction method based on graph neural network according to claim 2, characterized in that: The steps also include: setting the initial representation of nodes such as drugs, targets, diseases and side effects to one-hot encoding, using a linear transformation layer with a bias term, and applying it independently to each node type to map their features into a common, unified feature space.

4. The drug target prediction method based on graph neural network according to claim 2, characterized in that: The steps also include: defining the heterogeneous graph as Where V is the set of nodes, E is the set of edges, each node v has a type φ(v), each edge e has a type Ψ(e), and the possible sets of node types and edge types are represented as and When |Tv|=|Te|=1, the graph degenerates into a normal homogeneous graph.

5. The drug target prediction method based on graph neural network according to claim 4, characterized in that: The steps of treating different types of edges as feature information and integrating them into the feature learning process, dynamically adjusting the contribution of different types of edges to node representation through the attention mechanism, introducing residual connection and normalization processing strategies, and fusing node features are specifically as follows: Learnable edge type feature embedding: In each layer of the graph attention network, each edge type Ψ(e)∈T(e) is assigned a specific dl-dimensional embedding feature, denoted as Taking both node embedding and edge type embedding into account, the attention score α is calculated as: Among them, Ψ(<i,j> ) represents the type of edge with nodes i and j as endpoints, W r is a learnable matrix for type embedding transformation; Residual connection: A node residual connection is added between the cross-layer node representations, and the node aggregation on the lth layer is represented as: in, It's about the edge<i,j> The attention weight, σ is an activation function (default is ELU[]); When the dimension changes at layer l, a learnable linear transformation needs to be added: According to formula (2), K independent attention mechanisms are executed and their results are connected as the final representation. The formula is parsed as: In the above formula, || represents the concatenation operation. Represents the kth linear transformation The calculated attention scores are then averaged in the last layer (layer L), i.e.: L2 Normalization: Through L2 normalization, all features are scaled to the same magnitude: Among them, f i represents the output feature embedding of node i, The final expression from formula (7).

6. The drug target prediction method based on graph neural network according to claim 1, characterized in that: The steps of using a two-layer graph convolutional network to deeply mine the fused node features and calculate the potential interaction prediction score between the drug and the target are specifically as follows: Construct a DTP correlation graph to discover the potential relationships between DTPs, where each node is a DTP; By concatenating the drug feature representation and the target feature representation, the comprehensive representation fdt of each node in the reconstructed DTP graph is obtained, that is, fdt = fd|ft, where d∈VD, t∈VT; When processing a batch of DTPs, the representation z of each DTP calculated in the above way is combined into a representation matrix Z, in which the b-th row of Z is the eigenvector of the b-th DTP in the batch. The adjacency matrix A is constructed from the DTP representation and is expressed as: A ij =softmax((ZZ T )) ij (9); Among them, A ij is the element in the i-th row and j-th column of A, which represents the normalized similarity between the i-th drug and the j-th target in the same batch of data. It can be seen that the matrix A is a symmetric matrix. ij The larger the value, the stronger the association between the i-th drug and the j-th target DTP; The DTP matrix Z is input into the two-layer GCN, and the final output y is used to predict DTI. The calculation process of y is: y = A(ReLU(AZW1))W2 (10); Where W1 and W2 represent the learnable parameters of the first and second layers of GCN respectively; Binary cross entropy loss is used to evaluate the difference between the predicted DTI and the true DTI, and the formula is: Among them, B represents the number of DTPs in a batch training, y b represents the actual value of DTP. When D and P interact, y b =1, otherwise, y b =0,y′ b Represents the predicted value of DTP, which is normalized to y′ b ∈[0, 1].

7. A drug target prediction model based on graph neural network, used to implement the drug target prediction method based on graph neural network as described in any one of claims 1 to 6, characterized in that: include: The subgraph construction module is used to automatically extract local subgraphs with drug and target nodes as the core to capture the local context information of nodes in complex biological networks; The attention fusion module is used to consider different types of edges as feature information and integrate them into the feature learning process. It dynamically adjusts the contribution of different types of edges to node representation through the attention mechanism, introduces residual connection and normalization processing strategies, and fuses node features. The prediction analysis module is used to conduct in-depth mining of the fused node features and calculate the potential interaction prediction scores between drugs and targets.

Images