Construction method of drug interaction prediction model based on multi-view feature representation

Abstract

The invention provides a method for constructing a drug interaction prediction model based on multi-view feature representation, and the method comprises the steps: obtaining the node embedding representation of each drug from a drug knowledge graph through a heterogeneous graph neural network, and obtaining a first feature vector of each drug pair; utilizing a graph attention neural network to obtain graph embedded representation of each drug from the drug molecular graph to obtain a second feature vector of each drug pair, and obtaining a third feature vector of each drug pair from the drug bipartite graph; the first feature vector, the second feature vector and the third feature vector of each drug pair serve as input, the real DDI of each drug pair serves as output, the prediction probability of the multi-layer perceptron is infinitely close to the real DDI of each drug pair by adjusting parameters of the multi-layer perceptron, and a drug interaction prediction model is obtained. According to the method, the distinction degree of the DDI features is improved by using multi-view feature representation, so that the model prediction robustness is enhanced, and meanwhile, the problem of overfitting when the DDI of a new drug is predicted is solved.

Description

Technical Field

[0001] This application belongs to the field of pharmaceutical data analysis technology, and in particular relates to a method for constructing a drug interaction prediction model based on multi-perspective feature representation. Background Art

[0002] Drug-drug interaction (DDI) refers to the phenomenon where the simultaneous use of two or more drugs leads to mutual influence, resulting in enhanced or weakened drug efficacy or increased adverse reactions. When predicting DDI for known drugs, existing models exhibit poor robustness due to insufficient representation of drug characteristics. Conversely, when predicting DDI for new drugs, most methods have limited sensitivity to new drugs, leading to significantly lower prediction accuracy compared to DDI predictions for known drugs. Therefore, DDI prediction results are unsatisfactory in both scenarios.

[0003] To overcome the aforementioned problems, drug molecule graphs and drug knowledge graphs are commonly used for drug interaction prediction (DDI). Accurately reading the graph-level representation of drug molecules is crucial for enhancing the predictive power of the model. Researchers have proposed several methods to read the graph-level representation of drug molecules, including: 1) using multi-layer graph attention networks to model the molecular structure of drugs based on the pharmacological principle that chemical reactions may occur between different drug substructures, and extracting substructure information at different granularities in a hierarchical manner; 2) using granularity calculations to identify key drug substructures, thereby gaining a deeper understanding of the intrinsic mechanisms of drug interactions; and 3) integrating information from drug molecule graphs and DDI event graphs, and using graph contrastive learning methods to predict DDI events.

[0004] However, most of the aforementioned drug molecule graph-based models neglect the potential connections between drugs and other important entities in the body (targets, enzymes, carriers), which are crucial for modeling pharmacodynamic interactions. To address this deficiency, node-level representations in drug knowledge graphs are typically used to focus on the biomedical connections between drugs and related entities. Specific methods include: 1) developing a novel graph autoencoder to predict links on multimodal graphs; and 2) using graph neural networks to model the relationships between drugs and related entities such as targets to capture rich neighborhood information about the drug. While these methods demonstrate excellent predictive performance, they often focus on a single perspective of drug analysis and fail to fully consider the complementarity between drug molecule structure and knowledge graphs.

[0005] As research deepens, more and more studies are attempting to combine graph-level and node-level representations to fully utilize the information from both. Specific methods include: 1) addressing the heterogeneity of drug knowledge graphs and drug molecule graphs through architecture, by combining a relation-aware self-attention mechanism to learn multiple relationships between node pairs in heterogeneous graphs, thereby obtaining rich drug contextual information in both channels; 2) designing a four-channel graph neural network to extract high-level semantic information of drugs from different drug knowledge graphs and integrating drug features from different channels using a multilayer perceptron, thereby obtaining rich drug features. However, most existing models combining graph-level and node-level representations only contain a single drug molecule graph representation, lacking fine-grained modeling of the intrinsic substructural relationships between drug pairs, resulting in insufficient accuracy in DDI prediction. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a method for constructing a drug interaction prediction model based on multi-view feature representation. The method uses multi-view feature representation to improve the discriminativeness of DDI features, thereby enhancing the model's prediction robustness and solving the overfitting problem when predicting the DDI of new drugs.

[0007] This application provides a method for constructing a drug interaction prediction model based on multi-view feature representation, including:

[0008] Heterogeneous graph neural networks are used to obtain the node embedding representations of each drug from a drug knowledge graph to obtain the first feature vector of each drug pair; wherein, the drug knowledge graph is used to characterize the multi-type relationships between drugs and related entities, as well as the known DDIs between each drug;

[0009] A graph attention neural network is used to obtain the graph embedding representation of each drug from the drug molecule graph to obtain the second feature vector of each drug pair, and the third feature vector of each drug pair is obtained from the drug bipartite graph; wherein, the drug molecule graph is used to characterize the structural features of a single drug molecule; the drug bipartite graph is used to characterize the interaction relationships of each substructure in the drug pair;

[0010] Using the first, second, and third feature vectors of each drug pair as inputs and the true DDI of each drug pair as outputs, the parameters of the multilayer perceptron are adjusted so that the predicted probability of the multilayer perceptron is infinitely close to the true DDI of each drug pair, thereby obtaining a drug interaction prediction model.

[0011] Furthermore, the step of using a heterogeneous graph neural network to obtain the node embedding representation of each drug from the drug knowledge graph includes:

[0012] Each drug node in the drug knowledge graph is taken as the target node, related entities are taken as neighbor nodes, the relationship type between the drug and each related entity is taken as the relationship edge, and the known DDI between each drug is taken as the DDI edge; wherein, the related entities include: enzymes, targets, and carrier proteins.

[0013] When the drug interaction prediction model is used to explore unknown DDIs of known drugs, the dynamic attention mechanism of the heterogeneous graph neural network is used to obtain the node embedding representation of each drug in the drug knowledge graph.

[0014] When the drug interaction prediction model is used to predict the DDI of a new drug, each batch of subgraphs is extracted from the drug knowledge graph as input, and the node embedding representation of each drug in the subgraph is obtained by using the dynamic attention mechanism of the heterogeneous graph neural network.

[0015] Furthermore, the step of extracting the sub-graph for each batch as input from the drug knowledge graph includes:

[0016] A certain number of drug nodes are identified from the drug knowledge graph as pseudo-new drugs;

[0017] The identified pseudo-novel drugs, the DDI edges between each pseudo-novel drug, and the DDI edges connected to each pseudo-novel drug are hidden to obtain a sub-map for each batch as input.

[0018] Furthermore, the heterogeneous graph neural network obtains the node embedding representation of each drug using the following formula.

[0019]

[0020] In the formula, A-Linear τ(t) σ represents a linear transformation related to the type of the target node, typically a ReLU or ELU, where s, e, and t represent the source node, edge, and target node, respectively. The node feature representation of the (l-1)th layer;

[0021] The importance weight α of the neighbor node relative to the target node is obtained by the following formula. s,t :

[0022]

[0023] In the formula, Softmax represents the normalization function, and ATT-head i (s,e,t) represents the feature representation of each attention head output, and μ<τ(s),φ(e),τ(t)> are scalars for adjusting the attention score. K is the scaling factor.i (s) and Q i (t) represents the feature representation after linear transformation of the source node and the target node. This represents the weight matrix parameterized by edge type φ(e). and H represents the linear transformation matrix between the source node and the target node, where H is the number of attention heads.

[0024] Furthermore, the message passing mechanism m is characterized by the following formula. s,t :

[0025]

[0026] In the formula, MSG-head i (s,e,t) represents the feature representation of the source node output by a single attention head after message passing. This represents the linear mapping matrix corresponding to the source node. This represents the parameter matrix corresponding to the edge.

[0027] Furthermore, the step of using a graph attention neural network to obtain graph embedding representations of each drug from the drug molecule graph includes:

[0028] For each drug, the atoms that make up the drug molecule are used as nodes, and the chemical bonds connecting the atoms are used as edges to obtain the drug molecule diagram.

[0029] Atoms and chemical bonds are encoded using their respective feature information to obtain atomic feature vectors and chemical bond feature vectors, which serve as feature representations of nodes and edges in the drug molecule graph. The feature information of atoms is used to characterize the chemical properties and local environment of atoms, while the feature information of chemical bonds is used to characterize the topological and electronic properties of chemical bonds.

[0030] A multi-layered graph attention neural network is used to learn the topological structure of the drug molecule graph in order to capture the long-distance dependencies of each node in the drug molecule graph and obtain the graph embedding representation of the drug.

[0031] Furthermore, the multi-layered graph attention neural network obtains the graph embedding representation of the drug through the following method:

[0032] Each attention head calculates the attention coefficient of each neighbor node to the current node in the drug molecule graph using the following formula.

[0033]

[0034] In the formula, v i Represents the current node, v j Represents a neighboring node, vj ∈N e (v i ),in, Ne Let a be the number of edges. k Let W be the learnable attention vector of the k-th attention head, σ be the non-linear activation function, and W be the learningable attention vector of the k-th attention head. k This is a trainable weight matrix used to perform linear transformations on the feature vectors of nodes. For the current node v i Feature representation, For neighboring nodes vj Feature representation, The feature representation of other neighboring nodes; || represents the concatenation operation.

[0035] The output of each attention point is obtained using the following formula.

[0036]

[0037] In the formula, Indicates the current node v i The feature representation of the k-th attention head in the l-th layer. Represents neighbor node v j Feature representation at layer l-1;

[0038] The output of each attention head is aggregated using the following formula to obtain the current node v. i Updated embedding representation

[0039]

[0040] The graph embedding representation of the drug is obtained through the Readout function.

[0041]

[0042] In the formula, Indicates the number of atoms in a drug molecule. This indicates the atomic features of a drug molecule.

[0043] Furthermore, the graph attention neural network obtains the third feature vector of each drug pair from the drug bipartite graph in the following manner:

[0044] For each drug pair, obtain the drug bipartite graph of that drug pair and the corresponding vertex feature vectors;

[0045] The features of each vertex in the drug bipartite graph are updated using the following formula:

[0046]

[0047] In the formula, Let α be the feature representation of vertex w at the k-th attention head. k vw Let v and w be the importance weights of the vertices at the k-th attention head. h is the feature representation of vertex v at layer l. w Let w be the vertex eigenvector of vertex w. The feature representation of vertex w at the (l-1)th layer;

[0048] Using the SAGPooling technique, the importance β of each vertex feature vector in the drug bipartite graph is determined by the following formula:

[0049] β = softmax(GNN(A,X));

[0050] In the formula, β is determined by the coefficient β k Composition: A is the adjacency matrix of the drug bipartite graph, and X is the feature matrix of the drug bipartite graph;

[0051] The third eigenvector of the drug pair can be obtained using the following formula:

[0052]

[0053] In the formula, This represents the number of atoms in the drug bipartite diagram. This represents the characteristic representation of atom k in the drug bipartite diagram.

[0054] Furthermore, the drug bipartite graph of the drug pair and the corresponding vertex feature vectors are obtained in the following way:

[0055] Each node is extracted from its respective drug molecule graph and used as two vertex clusters. The Cartesian product of the two vertex clusters is calculated as an edge. At the same time, the edges in each drug molecule graph are hidden to obtain the drug bipartite graph of the drug pair.

[0056] Graph neural networks are used to perform graph convolution on the respective drug molecule graphs. By aggregating the information of neighboring nodes, the substructure information centered on the current node is obtained and used as the vertex feature vector of the drug bipartite graph.

[0057] The method for constructing a drug interaction prediction model based on multi-view feature representation provided in this application obtains the node embedding representation of drugs from the drug knowledge graph. It extracts graph embedding representations from the drug molecule graph, which characterizes the structural features of a single drug molecule, and from the drug bipartite graph, which characterizes the interaction relationships between substructures in a drug pair. Using multi-view feature representation improves the discriminative power of the drug interaction prediction (DDI) features, thereby enhancing the model's predictive robustness. Simultaneously, by extracting sub-graphs from the drug knowledge graph, targeted edge trimming of pseudo-novel drugs is achieved during the model training phase. This effectively avoids the problem of insufficient model sensitivity to new drugs and solves the overfitting problem caused by bias towards older drugs when predicting the DDI of new drugs. Attached Figure Description

[0058] Figure 1 A flowchart is shown illustrating the method for constructing a drug interaction prediction model based on multi-view feature representation provided in an embodiment of this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this technical solution clearer, the following detailed description, in conjunction with specific embodiments, further illustrates this technical solution. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this technical solution.

[0060] First, the application scenarios of this application are introduced. This application does not change the basic framework of the drug interaction prediction model, enabling it to explore unknown DDIs of known drugs and predict DDIs of new drugs.

[0061] Secondly, the dataset used in this application to construct the drug interaction prediction model is introduced. 1) 941,218 drug pairs with interactions were collected from DrugBank Version 5.1.3, involving a total of 2,755 drugs, and these drug pairs were labeled as positive samples; 2) Two negative sample sets, PU-Learn negatives and Randomnegatives, were generated using Positive-Unlabeled Learning (PU Learning) and random sampling methods, respectively; 3) The PU-dataset (containing PU-Learn negatives and positive samples) and the RA-dataset (containing Random negatives and positive samples) were used for model training and evaluation.

[0062] Next, please refer to... Figure 1 The flowchart shown illustrates the construction method of a drug interaction prediction model based on multi-view feature representation. Figure 1 As shown, the method includes:

[0063] S101. Use a heterogeneous graph neural network to obtain the node embedding representation of each drug from the drug knowledge graph, so as to obtain the first feature vector of each drug pair.

[0064] The drug knowledge graph is used to characterize the various types of relationships between drugs and related entities, as well as the known DDIs between drugs.

[0065] In this step, the drug knowledge graph integrates multiple types of relationships between drugs and related entities (such as targets, enzymes, pathways, and carrier proteins). Therefore, by analyzing the drug knowledge graph, potential drug interactions can be inferred (such as certain drug combinations reducing each other's efficacy due to competition for the same receptor or enzyme). The dynamic attention mechanism in heterogeneous graph neural networks can effectively model the relationships between different types of nodes and edges in the drug knowledge graph and adaptively learn the importance weights of different types of nodes and edges, thereby generating high-quality node embedding representations.

[0066] In practical implementation, the node embedding representations of each drug can be obtained from the drug knowledge graph in the following ways:

[0067] Step 1011: Take each drug node in the drug knowledge graph as the target node, the related entities as the neighbor nodes, the relationship type between the drug and each related entity as the relationship edge, and the known DDI between each drug as the DDI edge.

[0068] The related entities include: enzymes, targets, and carrier proteins.

[0069] Step 1012: When the drug interaction prediction model is used to explore unknown DDIs of known drugs, the dynamic attention mechanism of the heterogeneous graph neural network is used to obtain the node embedding representation of each drug in the drug knowledge graph.

[0070] Step 1013: When the drug interaction prediction model is used to predict the DDI of a new drug, the sub-graph of each batch as input is extracted from the drug knowledge graph, and the node embedding representation of each drug in the sub-graph is obtained by using the dynamic attention mechanism of the heterogeneous graph neural network.

[0071] The drug interaction prediction model of this application is known to be applicable to two scenarios: exploring unknown drug interactions (DDI) of known drugs (Scenario 1) and predicting the DDI of new drugs (Scenario 2). These two scenarios differ significantly: in Scenario 1, there are edges (DDI edges) between known drugs and between known drugs and related entities (relationship edges); in Scenario 2, there are edges (relationship edges) between new drugs and related entities, but the edges between new drugs and known drugs are unknown (to be predicted). Therefore, addressing the issue that most existing technologies focus on Scenario 1 while neglecting Scenario 2, this application proposes a sub-graph generation method that can make the drug interaction prediction model applicable to both scenarios simultaneously without altering its framework.

[0072] In practice, the sub-maps of each batch of input can be extracted in the following ways:

[0073] Step 201: Identify a certain number of drug nodes from the drug knowledge graph as pseudo-new drugs.

[0074] Step 202: Hide the identified pseudo-new drugs, the DDI edges between each pseudo-new drug, and the DDI edges connected to each pseudo-new drug to obtain the sub-map for each batch as input.

[0075] During the training phase of the drug interaction prediction model, in order to simulate the real-world scenario of predicting the drug interaction incidence (DDI) of a new drug, a sub-map is constructed for each batch to ensure that the new drug is not disclosed during the training phase, thereby preventing model overfitting.

[0076] In practical implementation, the heterogeneous graph neural network obtains the node embedding representation of each drug using the following formula.

[0077]

[0078] In the formula, A-Linear τ(t) σ represents a linear transformation related to the type of the target node, typically a ReLU or ELU, where s, e, and t represent the source node, edge, and target node, respectively. The node feature representation of the (l-1)th layer;

[0079] The importance weight α of the neighbor node relative to the target node is obtained by the following formula. s,t :

[0080]

[0081] In the formula, Softmax represents the normalization function, and ATT-head i(s,e,t) represents the feature representation of each attention head output, and μ<τ(s),φ(e),τ(t)> is a scalar for adjusting the attention score. K is the scaling factor. i (s) and Q i (t) represents the feature representation after linear transformation of the source node and the target node. This represents the weight matrix parameterized by edge type φ(e). and H represents the linear transformation matrix between the source node and the target node, where H is the number of attention heads.

[0082] Furthermore, the message passing mechanism m is characterized by the following formula. s,t :

[0083]

[0084] In the formula, MSG-head i (s,e,t) represents the feature representation of the source node output by a single attention head after message passing. This represents the linear mapping matrix corresponding to the source node. This represents the parameter matrix corresponding to the edge.

[0085] Furthermore, after obtaining the node embedding representations of each drug, the first feature vector of each drug pair can be obtained based on the node embedding representations of each drug.

[0086] S102. Use a graph attention neural network to obtain the graph embedding representation of each drug from the drug molecule graph to obtain the second feature vector of each drug pair, and obtain the third feature vector of each drug pair from the drug bipartite graph.

[0087] The drug molecule diagram is used to characterize the structural features of a single drug molecule; the drug bipartite diagram is used to characterize the interaction relationships between the substructures of a drug pair.

[0088] In practical implementation, the graph embedding representation of each drug can be obtained from the drug molecule diagram in the following ways:

[0089] Step 1021: For each drug, take each atom that makes up the drug molecule as a node and the chemical bond connecting each atom as an edge to obtain the drug molecule diagram.

[0090] Step 1022: Encode atoms and chemical bonds using their respective feature information to obtain atomic feature vectors and chemical bond feature vectors, which serve as feature representations of nodes and edges in the drug molecule graph.

[0091] Among them, the characteristic information of atoms is used to characterize the chemical properties and local environment of atoms, and the characteristic information of chemical bonds is used to characterize the topological and electronic properties of chemical bonds.

[0092] In this step, the characteristic information of atoms includes: the atom symbol (44-dimensional one-hot encoding), the degree of atom (1-dimensional representation), the implicit valence (1-dimensional representation), the formal charge (1-dimensional representation), the number of free radical electrons (1-dimensional representation), the hybridization (5-dimensional one-hot encoding), whether the atom has aromaticity (1-dimensional representation), and the total number of hydrogen atoms on the atom (1-dimensional representation); the characteristic information of chemical bonds includes: the type of bond (5-dimensional one-hot encoding), the atomic number of the atoms at both ends of the bond (2-dimensional representation), whether the bond is in a ring (1-dimensional representation), whether the bond participates in a conjugated system (1-dimensional representation), and whether the bond belongs to an aromatic system (1-dimensional representation).

[0093] Step 1023: Use a multi-level graph attention neural network to learn the topological structure of the drug molecule graph in order to capture the long-distance dependencies of each node in the drug molecule graph and obtain the graph embedding representation of the drug.

[0094] In practical implementation, the multi-layered graph attention neural network obtains the graph embedding representation of the drug through the following method:

[0095] Each attention head calculates the attention coefficient of each neighbor node to the current node in the drug molecule graph using the following formula.

[0096]

[0097] In the formula, v i Represents the current node, v j Represents a neighboring node, v j ∈N e (v i ),in, Ne Let a be the number of edges. k Let W be the learnable attention vector of the k-th attention head, σ be the non-linear activation function, and W be the learningable attention vector of the k-th attention head. k This is a trainable weight matrix used to perform linear transformations on the feature vectors of nodes. For the current node v i Feature representation, For neighbor node v j Feature representation, The feature representation of other neighboring nodes; || represents the concatenation operation.

[0098] The output of each attention point is obtained using the following formula.

[0099]

[0100] In the formula, Indicates the current node v i The feature representation of the k-th attention head in the l-th layer. Represents neighbor node v j Feature representation at layer l-1;

[0101] The output of each attention head is aggregated using the following formula to obtain the current node v. i Updated embedding representation

[0102]

[0103] The graph embedding representation of the drug is obtained through the Readout function.

[0104]

[0105] In the formula, Indicates the number of atoms in a drug molecule. This indicates the atomic features of a drug molecule.

[0106] Furthermore, after obtaining the graph embedding representation of each drug, the second feature vector of each drug pair can be obtained based on the graph embedding representation of each drug.

[0107] In practical implementation, the graph attention neural network obtains the third feature vector of each drug pair from the drug bipartite graph in the following way:

[0108] For each drug pair, obtain the drug bipartite graph of that drug pair and the corresponding vertex feature vectors;

[0109] The features of each vertex in the drug bipartite graph are updated using the following formula:

[0110]

[0111] In the formula, Let α be the feature representation of vertex w at the k-th attention head. k vw Let v and w be the importance weights of the vertices at the k-th attention head. h is the feature representation of vertex v at layer l. w Let w be the vertex eigenvector of vertex w. The feature representation of vertex w at the (l-1)th layer;

[0112] Using the SAGPooling technique, the importance β of each vertex feature vector in the drug bipartite graph is determined by the following formula:

[0113] β = softmax(GNN(A,X));

[0114] In the formula, β is determined by the coefficient β k Composition: A is the adjacency matrix of the drug bipartite graph, and X is the feature matrix of the drug bipartite graph;

[0115] The third eigenvector of the drug pair can be obtained using the following formula:

[0116]

[0117] In the formula, This represents the number of atoms in the drug bipartite diagram. This represents the characteristic representation of atom k in the drug bipartite diagram.

[0118] Here, we actually use a graph attention neural network to obtain the graph embedding representation of the drug bipartite graph. Since the drug bipartite graph is used to characterize the interaction relationship between the substructures in the drug pair, the graph embedding representation of the drug bipartite graph can be used as the third feature vector of the drug pair.

[0119] The drug bipartite graph and corresponding vertex feature vectors for this drug pair are obtained as follows:

[0120] Step 301: Extract each node from the respective drug molecule graphs as two vertex clusters, calculate the Cartesian product of the two vertex clusters as an edge, and hide the edges in the respective drug molecule graphs to obtain the drug bipartite graph of the drug pair.

[0121] In this step, the edges in the drug bipartite graph (obtained by calculating the Cartesian product) do not have specific chemical or physical meanings. Their initial feature vectors are generated through random initialization. These feature vectors are gradually adjusted by optimization algorithms during subsequent model training to capture potential interaction patterns between drug molecules.

[0122] Step 302: Perform graph convolution on the respective drug molecule graphs using a graph neural network, and obtain substructure information centered on the current node by aggregating neighbor node information, which is used as the vertex feature vector of the drug bipartite graph.

[0123] S103. Using the first feature vector, second feature vector, and third feature vector of each drug pair as input, and the true DDI of each drug pair as output, the parameters of the multilayer perceptron are adjusted so that the predicted probability of the multilayer perceptron is infinitely close to the true DDI of each drug pair, so as to obtain a drug interaction prediction model.

[0124] After obtaining the drug interaction prediction model, in order to verify the effectiveness of the drug interaction prediction model, the embodiments of this application also tested it, specifically:

[0125] When testing the performance of the drug interaction prediction model for exploring unknown drug interactions (DDIs) of known drugs, 5% of the DDI triples were randomly selected from the positive samples as positive samples in the test set, 5% of the DDI triples were randomly selected from the PU-Learn negatives as negative samples in the test set, and 5% of the DDI triples were randomly selected from the random negatives as negative samples in the test set. The extracted positive and negative samples were used to test the predictive performance of the drug interaction prediction model.

[0126] When testing the performance of the drug interaction prediction model in predicting the drug interaction discovery (DDI) of new drugs, 200 drugs were first selected as pseudo-new drugs from the drug set. All DDI triples associated with these pseudo-new drugs were selected from the positive samples as positive samples in the test set, and all DDI triples associated with these pseudo-new drugs were selected from the PU-Learn negatives as negative samples in the test set. Under the RA-dataset, the remaining drugs in the drug set and the new drugs were randomly selected and combined (combinations appearing in the positive samples of the test set were removed) to generate negative samples of the same size as the positive samples in the test set. Simultaneously, these 200 pseudo-new drugs and all associated connections were removed from the drug knowledge graph. That is, information about the new drugs was not leaked at all during the training phase. The purpose of this was to predict whether there were interactions between these new drugs and existing drugs or between new drugs.

[0127] The working mechanism of the drug interaction prediction model based on multi-view feature representation in this application is summarized as follows:

[0128] (1) Input representation: The drug pairs in the drug knowledge graph are used as input. By executing a heterogeneous graph neural network on the drug knowledge graph, information is finely aggregated from different related entities to obtain the embedding representation of drug nodes in the drug knowledge graph.

[0129] (2) Molecular graph modeling: 1) Convert each drug in the drug pair into its corresponding drug molecule graph (atoms are nodes, chemical bonds are edges); 2) Perform independent intra-node updates on each drug molecule graph and extract the molecular structure features of each drug using graph neural networks; 3) Merge the updated two drug molecule graphs into a drug bipartite graph and complete the interaction modeling between the two drug molecule graphs through the node update mechanism on the drug bipartite graph, thereby capturing the potential chemical reactions or physical interactions between the drug pairs.

[0130] (3) Feature aggregation and representation learning: Global embedding representations are extracted from the drug molecule graph and drug bipartite graph using the READOUT function. These representations contain the molecular structural features of the drugs and the interaction features of drug pairs. Drug representations at different levels are mapped to the same feature space, and the final feature representation of drug pairs is generated by integrating representations at multiple levels.

[0131] (4) Prediction and output: Based on the feature representation of the integrated drug pair, the probability of drug interaction is predicted by multilayer perceptron.

[0132] The above content is only a preferred embodiment of the present invention. For those skilled in the art, many changes can be made in the specific implementation and application scope based on the ideas of the present invention. As long as these changes do not depart from the concept of the present invention, they all fall within the protection scope of the present invention.

Claims

1. A method for constructing a drug interaction prediction model based on multi-view feature representation, characterized in that, include: Heterogeneous graph neural networks are used to obtain the node embedding representations of each drug from a drug knowledge graph to obtain the first feature vector of each drug pair; wherein, the drug knowledge graph is used to characterize the multi-type relationships between drugs and related entities, as well as the known DDIs between each drug; A graph attention neural network is used to obtain the graph embedding representation of each drug from the drug molecule graph to obtain the second feature vector of each drug pair, and the third feature vector of each drug pair is obtained from the drug bipartite graph; wherein, the drug molecule graph is used to characterize the structural features of a single drug molecule; the drug bipartite graph is used to characterize the interaction relationships of each substructure in the drug pair; Using the first, second, and third feature vectors of each drug pair as inputs and the true DDI of each drug pair as outputs, the parameters of the multilayer perceptron are adjusted so that the predicted probability of the multilayer perceptron is infinitely close to the true DDI of each drug pair, thereby obtaining a drug interaction prediction model.

2. The method as described in claim 1, characterized in that, The method of obtaining node embedding representations of each drug from a drug knowledge graph using a heterogeneous graph neural network includes: Each drug node in the drug knowledge graph is taken as the target node, related entities are taken as neighbor nodes, the relationship type between the drug and each related entity is taken as the relationship edge, and the known DDI between each drug is taken as the DDI edge; wherein, the related entities include: enzymes, targets, and carrier proteins. When the drug interaction prediction model is used to explore unknown DDIs of known drugs, the dynamic attention mechanism of the heterogeneous graph neural network is used to obtain the node embedding representation of each drug in the drug knowledge graph. When the drug interaction prediction model is used to predict the DDI of a new drug, each batch of subgraphs is extracted from the drug knowledge graph as input, and the node embedding representation of each drug in the subgraph is obtained by using the dynamic attention mechanism of the heterogeneous graph neural network.

3. The method as described in claim 2, characterized in that, The step of extracting each batch of sub-graphs as input from the drug knowledge graph includes: A certain number of drug nodes are identified from the drug knowledge graph as pseudo-new drugs; The identified pseudo-novel drugs, the DDI edges between each pseudo-novel drug, and the DDI edges connected to each pseudo-novel drug are hidden to obtain a sub-map for each batch as input.

4. The method as described in claim 3, characterized in that, The heterogeneous graph neural network obtains the node embedding representation of each drug using the following formula. In the formula, A-Linear τ(t) σ represents a linear transformation related to the type of the target node, typically a ReLU or ELU, where s, e, and t represent the source node, edge, and target node, respectively. The node feature representation of the (l-1)th layer; The importance weight α of the neighbor node relative to the target node is obtained by the following formula. s,t : In the formula, Softmax represents the normalization function, and ATT-head i (s,e,t) represents the feature representation of each attention head output, and μ<τ(s),φ(e),τ(t)> are scalars for adjusting the attention score. K is the scaling factor. i (s) and Q i (t) represents the feature representation after linear transformation of the source node and the target node. This represents the weight matrix parameterized by edge type φ(e). and H represents the linear transformation matrix between the source node and the target node, where H is the number of attention heads. Furthermore, the message passing mechanism m is characterized by the following formula. s,t : In the formula, MSG-head i (s,e,t) represents the feature representation of the source node output by a single attention head after message passing. This represents the linear mapping matrix corresponding to the source node. This represents the parameter matrix corresponding to the edge.

5. The method as described in claim 1, characterized in that, The method of obtaining graph embedding representations of each drug from the drug molecule graph using a graph attention neural network includes: For each drug, the atoms that make up the drug molecule are used as nodes, and the chemical bonds connecting the atoms are used as edges to obtain the drug molecule diagram. Atoms and chemical bonds are encoded using their respective feature information to obtain atomic feature vectors and chemical bond feature vectors, which serve as feature representations of nodes and edges in the drug molecule graph. The feature information of atoms is used to characterize the chemical properties and local environment of atoms, while the feature information of chemical bonds is used to characterize the topological and electronic properties of chemical bonds. A multi-layered graph attention neural network is used to learn the topological structure of the drug molecule graph in order to capture the long-distance dependencies of each node in the drug molecule graph and obtain the graph embedding representation of the drug.

6. The method as described in claim 5, characterized in that, The multi-layered graph attention neural network obtains the graph embedding representation of the drug through the following method: Each attention head calculates the attention coefficient of each neighbor node to the current node in the drug molecule graph using the following formula. In the formula, v i Represents the current node, v j Represents a neighboring node, v j ∈N e (v i ), where N e Let a be the number of edges. k Let W be the learnable attention vector of the k-th attention head, σ be the non-linear activation function, and W be the learningable attention vector of the k-th attention head. k This is a trainable weight matrix used to perform linear transformations on the feature vectors of nodes. For the current node v i Feature representation, For neighbor node v j Feature representation, The feature representation of other neighboring nodes; || represents the concatenation operation. The output of each attention point is obtained using the following formula. In the formula, Indicates the current node v i The feature representation of the k-th attention head in the l-th layer. Represents neighbor node v j Feature representation at layer l-1; The output of each attention head is aggregated using the following formula to obtain the current node v. i Updated embedding representation The graph embedding representation of the drug is obtained through the Readout function. In the formula, Indicates the number of atoms in a drug molecule. This indicates the atomic features of a drug molecule.

7. The method as described in claim 1, characterized in that, The graph attention neural network obtains the third feature vector of each drug pair from the drug bipartite graph in the following way: For each drug pair, obtain the drug bipartite graph of that drug pair and the corresponding vertex feature vectors; The features of each vertex in the drug bipartite graph are updated using the following formula: In the formula, Let α be the feature representation of vertex w at the k-th attention head. k vw Let v and w be the importance weights of the vertices at the k-th attention head. h is the feature representation of vertex v at layer l. w Let w be the vertex eigenvector of vertex w. The feature representation of vertex w at the (l-1)th layer; Using the SAGPooling technique, the importance β of each vertex feature vector in the drug bipartite graph is determined by the following formula: β = softmax(GNN(A,X)); In the formula, β is determined by the coefficient β k Composition: A is the adjacency matrix of the drug bipartite graph, and X is the feature matrix of the drug bipartite graph; The third eigenvector of the drug pair can be obtained using the following formula: In the formula, This represents the number of atoms in the drug bipartite diagram. This represents the characteristic representation of atom k in the drug bipartite diagram.

8. The method as described in claim 7, characterized in that, The drug bipartite graph and corresponding vertex feature vectors for this drug pair are obtained in the following way: Each node is extracted from its respective drug molecule graph and used as two vertex clusters. The Cartesian product of the two vertex clusters is calculated as an edge. At the same time, the edges in each drug molecule graph are hidden to obtain the drug bipartite graph of the drug pair. Graph neural networks are used to perform graph convolution on the respective drug molecule graphs. By aggregating the information of neighboring nodes, the substructure information centered on the current node is obtained and used as the vertex feature vector of the drug bipartite graph.

Images