A method for predicting drug interactions

Abstract

The application provides a drug interaction prediction method, comprising: obtaining a group of drug entities to be predicted for interaction; inputting the group of drug entities to be predicted for interaction into a prediction model to obtain a prediction result; wherein the prediction model is constructed by: constructing a multi-relation directed knowledge graph containing drug-related information; obtaining the representation of a source node and a target node in each message propagation process in the knowledge graph respectively; based on a double-aggregation message passing layer, aggregating the feature information of the drug nodes and the feature information of the neighborhood set of the drug nodes in the multi-relation directed knowledge graph to obtain the final node features of the drug nodes; reconstructing a relation matrix based on the aggregated node features; and predicting the interaction between drugs according to the reconstructed relation matrix. The method makes the drug node features more comprehensive and capable of reflecting complex correlations, and also updates the node features by means of message propagation, fully considering the influence of multi-level relationships on drug interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug interaction, and specifically provides a prediction method of drug interaction. BACKGROUND

[0002] In clinical practice, the combination of multiple drugs can bring significant therapeutic effect on one hand, and can cause unexpected drug-drug interaction (DDI) on the other hand, and even lead to adverse reactions (ADR). Therefore, it is crucial to identify potential DDIs in the drug development stage. The dependence of traditional in vivo and in vitro experiments on environment, dose and usage is strong, and combined with the strict drug approval process, it is difficult and delayed to identify DDIs. In contrast, information science methods show obvious advantages in DDI prediction. However, how to effectively organize, extract and operate these data using computational methods has become a realistic and urgent problem in the field of drug development and clinical treatment.

[0003] In recent years, knowledge graph has been widely used in many research fields, and its downstream tasks can be used for knowledge representation and knowledge integration, or for matrix completion and link prediction. The ideal assumption of using knowledge graph embedding as features makes it possible to obtain detailed and complete information of drugs at one time and immediately judge whether there is DDI. Most of the existing methods model complex multi-type interactions as separate data instances, and this method only uses shallow relationship to construct embedding function and does not consider the correlation between different entities. Therefore, the embedding function of this way lacks explicit coding of key interaction information, so that they are not enough to extract complete interaction features from the collective behavior of entities.

[0004] Therefore, there is a need in the art for a new prediction method of drug interaction to solve the above problems. SUMMARY

[0005] In order to overcome the above defects, the present application is proposed to provide a solution or at least partially solve the problem of being unable to extract complete interaction features from the collective behavior of entities.

[0006] The application provides a drug interaction prediction method, comprising: obtaining a group of drug entities to be predicted for interaction; inputting the group of drug entities to be predicted for interaction into a prediction model to obtain a prediction result; wherein the process of constructing the prediction model comprises: constructing a multi-relation directed knowledge graph containing drug-related information; obtaining the representation of a source node and a target node in each message propagation process in the multi-relation directed knowledge graph respectively; based on a double-aggregation message passing layer, aggregating the feature information of the drug nodes and the feature information of the neighborhood set of the drug nodes obtained in the multi-relation directed knowledge graph to obtain the final node features of the drug nodes; reconstructing a relation matrix based on the aggregated node features; and predicting the interaction between drugs according to the reconstructed relation matrix.

[0007] In one of the technical solutions of the above drug interaction prediction method, the neighborhood set is obtained by at least the following method: obtaining the neighbor nodes of the drug nodes and the nodes adjacent to the neighbor nodes.

[0008] In one of the technical solutions of the above drug interaction prediction method, based on the double-aggregation message passing layer, the process of aggregating the feature information of the drug nodes and the feature information of the neighborhood set of the drug nodes obtained in the multi-relation directed knowledge graph to obtain the final node features of the drug nodes at least comprises: obtaining the messages from the neighbor nodes of the drug nodes to the drug nodes; setting the corresponding message propagation mechanism for the relationship between the drug and different types of entities; based on the message propagation mechanism, the messages propagate along the edges of different relationship types in the multi-relation directed knowledge graph, so that the drug nodes can obtain the information of their neighbor nodes and the neighbor nodes of the neighbor nodes; in each message passing process, the information collected from other nodes is aggregated with the information of the drug nodes themselves; after multiple message passing, the aggregation results corresponding to different relationship types are obtained, and the aggregation results of different levels and types are combined in a tensor stacking manner to generate the final node features.

[0009] In one of the technical solutions of the above drug interaction prediction method, based on the double-aggregation message passing layer, the process of aggregating the feature information of the drug nodes and the feature information of the neighborhood set of the drug nodes obtained in the multi-relation directed knowledge graph to obtain the final node features of the drug nodes further comprises: performing smoothing processing on the features after tensor stacking.

[0010] In one of the technical solutions of the above drug interaction prediction method, the multi-relation is at least the relationship between drugs, targets, enzymes, structures and pathways.

[0011] In one of the above technical solutions of the drug interaction prediction method, the method comprises at least the following steps of obtaining a message propagation mechanism: obtaining an adjacency matrix corresponding to different types of edge relationships and a corresponding weight matrix; obtaining the self feature information of the drug node; and calculating the message propagation process based on the adjacency matrix corresponding to different types of edge relationships and the corresponding weight matrix and the self feature information of the drug node.

[0012] In one of the above technical solutions of the drug interaction prediction method, the process of reconstructing the relationship matrix based on the aggregated node features comprises: passing the embedding vector of the source node, the embedding vector of the target node and the relationship matrix between the source node and the target node through different layers to reconstruct the relevance of drug-target, drug-structure, drug-enzyme and drug-pathway.

[0013] In one of the above technical solutions of the drug interaction prediction method, the method further comprises at least the following steps of optimizing the prediction model: calculating the difference between the prediction result and the true result based on the constructed joint loss optimization function, and updating the trainable parameters in the prediction model through the back propagation algorithm.

[0014] In one of the above technical solutions of the drug interaction prediction method, the method comprises at least the following steps of obtaining a joint loss optimization function: obtaining a real adjacency matrix between different relationships; obtaining the mean square error between the real adjacency matrix between different relationships and the reconstructed relationship matrix, and obtaining the loss function of the mean square error matrix; and combining all the loss functions of the mean square error matrix by summation to construct a total loss function.

[0015] In one of the above technical solutions of the drug interaction prediction method, the method further comprises the following steps of obtaining a joint loss optimization function: introducing a regularization term, which contains the norm of the node embedding and the relationship matrix; and obtaining the joint loss optimization function based on the total loss function and the regularization term.

[0016] The above one or more technical solutions of the present application have at least one or more of the following advantages:

[0017] In the technical scheme of the present application, the present application provides a drug interaction prediction method, comprising: obtaining a set of drug entities to be predicted for interaction; inputting the set of drug entities to be predicted for interaction into a prediction model to obtain a prediction result; wherein the process of constructing the prediction model comprises: constructing a multi-relation directed knowledge graph containing drug-related information; obtaining the representation of the source node and the target node in each message propagation process in the multi-relation directed knowledge graph respectively; based on the double-aggregated message passing layer, aggregating the feature information of the drug nodes and the feature information of the neighborhood set of the drug nodes obtained in the multi-relation directed knowledge graph to obtain the final node features of the drug nodes; reconstructing the relationship matrix based on the aggregated node features; and predicting the interaction between drugs according to the reconstructed relationship matrix. Compared with the prior art, the drug interaction prediction method provided by the present application has the following beneficial effects:

[0018] The present method constructs a multi-relation directed knowledge graph containing drug-related information, which can comprehensively integrate complex relationship information between drugs and various entities. This relationship can truly reflect the directionality of drug action and the actual situation of the sequence of mutual influence between entities. By respectively obtaining the representation of the source node and the target node in each message propagation process in the multi-relation directed knowledge graph, and using the double-aggregated message passing layer to aggregate the feature information of the drug nodes and their neighborhood set, the features of the drug nodes are no longer limited to their initial attributes. The drug nodes can integrate their local structure information, making the node features more comprehensive and reflecting the complex association in the knowledge graph. The message propagation process will proceed along the edges of different relationship types and continuously update the node features at different levels (multiple message passing), which means that the influence of multi-level relationships between drugs and different entities on drug interaction can be fully considered.

[0019] Further, the representation of different levels of classification information in multiple types of relationships is comprehensively considered, including interaction protein information, related pathway information, drug chemical structure information, and related enzyme information.

[0020] Further, the joint loss function is designed by reconstructing the relationship matrix to optimize the model, which guides the path selection in the message propagation process in the reverse direction through optimization training, avoiding labor-intensive path materialization. BRIEF DESCRIPTION OF DRAWINGS

[0021] The disclosure of the present application will become more readily understood by referring to the accompanying drawings. As those skilled in the art will readily appreciate, the drawings are not intended to limit the present application in any way. Further, like reference numerals are intended to represent like parts throughout the various figures. In the drawings:

[0022] Figure 11 is a flow chart showing the main steps of a method for predicting drug interactions according to one embodiment of the present invention;

[0023] Figure 2 are the predicted top 50 relationship pairs according to one embodiment of the present invention. DETAILED DESCRIPTION

[0024] Some embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] Example 1

[0026] like Figure 1 As shown, the method for predicting drug interactions in the embodiment of the present invention mainly includes the following steps S1 to S2.

[0027] Step S1, obtaining a group of drug entities whose interactions are to be predicted;

[0028] Step S2: inputting the set of drug entities whose interactions are to be predicted into a prediction model to obtain a prediction result;

[0029] The process of constructing the prediction model includes: constructing a multi-relationship directed knowledge graph containing drug-related information; respectively obtaining the representations of the source node and the target node in the multi-relationship directed knowledge graph during each message propagation process; based on a dual-aggregation message transmission layer, aggregating the feature information of the drug node and the feature information of the neighborhood set of the drug node obtained in the multi-relationship directed knowledge graph to obtain the final node features of the drug node; reconstructing the relationship matrix based on the aggregated node features; and predicting the interaction between drugs based on the reconstructed relationship matrix.

[0030] In this embodiment, the purpose of constructing a directed knowledge graph is to capture the complex relationships in the data, and the representation of different levels of classification information in various types of relationships will be comprehensively considered, such as drugs, targets, enzymes, structures, and pathways.

[0031] Specifically, a directed knowledge graph is composed of a network of triplets. Entities are abstracted as nodes in mathematics, and relationships are abstracted as edges in mathematics, thereby converting the knowledge graph into a directed knowledge graph for knowledge computing. The directed knowledge graph can be represented as:

[0032]

[0033] Among them, D represents the set of drug entities, E represents the set of other biomolecule entities, and R is the relationship subset, which indicates whether there is a mutual connection between the head entity and the tail entity.

[0034] In the constructed directed knowledge graph, each entity usually involves multiple triples, representing that the entity has multiple different interaction relationships. The head entity in the triple is called the target node of the learning feature, and the tail entity is called the source node adjacent to the target node. Therefore, there are four different edge relationships in the directed knowledge graph, respectively, the edge relationship between the drug head entity and the structure tail entity, denoted as {r1 | r1 e R}, the edge relationship between the drug head entity and the target tail entity, denoted as {r2, | r2 e R}, the edge relationship between the drug head entity and the enzyme tail entity, denoted as {r3, | r3 e R}, and the edge relationship between the drug head entity and the pathway tail entity, denoted as {r4, | r4 e R}. These four types of edges help the drug to collect the neighbor features of the nodes in the DKG, introduce initial attributes to each node, and pass them to the nonlinear layer, and the features of each node are calculated by convolution.

[0035] In the constructed multi-relation directed knowledge graph, information propagates between nodes (message propagation process), and the source node is the node that initiates the propagation of information, and the target node is the node to which the information propagates. In each information transmission, the two nodes will exhibit different feature representations (representations), and these representations will be obtained separately. The double-aggregated message passing layer is the key link to realize feature aggregation. The drug node itself has feature information such as chemical structure and pharmacodynamic characteristics, and the nodes (neighborhood set) associated with it also have their own features, such as the active characteristics of a certain biomolecule node in the neighborhood. Using the double-aggregated message passing layer, the features of the drug node itself and the features of its neighborhood set are integrated and processed, and after this integration process, the final node features of the drug node that are more comprehensive and can reflect its status and characteristics in the entire graph can be obtained, making the description of the drug more perfect and accurate, facilitating subsequent calculation and analysis. The original relationship matrix is a mathematical representation of the relationship between nodes in the knowledge graph. After obtaining the aggregated more accurate and more comprehensive node features through the previous steps, the original relationship matrix needs to be reconstructed based on these new features. Because the previous relationship matrix may not be accurate enough after the update of the node features, the reconstructed relationship matrix can better fit the actual association of the current nodes and more accurately reflect the real mutual relationship between drugs and other biomolecule entities. With the reconstructed relationship matrix, corresponding methods and algorithms can be used to carry out the prediction of the interaction between drugs. For example, based on a specific prediction function wherein represents the prediction score of the connection edge between the two entities, and Θ is the function The model parameters are combined with the relationship matrix and other related information such as the entire knowledge graph, and vector, matrix, and tensor calculation methods are used to realize link prediction, that is, to infer whether different drugs interact with each other and the specific interaction, such as whether a synergistic effect is produced or whether an adverse reaction is triggered.

[0036] In one embodiment, the neighborhood set is obtained by at least the following method: obtaining the neighbor nodes of the drug node and the nodes adjacent to the neighbor nodes.

[0037] In this embodiment, the neighbor nodes of the drug node are nodes directly connected to the drug node, and there is a direct relationship between the drug node and the neighbor nodes (such as being directly connected by a certain edge in the knowledge graph). For example, in a drug knowledge graph, if a drug node represents drug A, the nodes directly connected to drug A may be nodes representing its target points, nodes corresponding to enzymes participating in its metabolism, etc. These are the neighbor nodes of drug A. The nodes adjacent to the neighbor nodes: on the basis of the neighbor nodes found, other nodes directly connected to the neighbor nodes are found. Still taking drug A as an example, a neighbor node of drug A is a node corresponding to target point B, and the nodes directly connected to the node of the target point B may be nodes corresponding to other proteins participating in the regulation of the target point, etc. These are the nodes adjacent to the neighbor nodes.

[0038] In one embodiment, based on a double aggregation message passing layer, the process of aggregating the feature information of the drug node and the feature information of the neighborhood set of the drug node in the multi-relation directed knowledge graph to obtain the final node feature of the drug node at least includes: obtaining the message from the neighbor nodes of the drug node to the drug node; setting a corresponding message propagation mechanism for the relationship between the drug and different types of entities; based on the message propagation mechanism, the message is propagated along the edges of different relationship types in the multi-relation directed knowledge graph, so that the drug node can obtain the information of its neighbor nodes and the neighbor nodes of the neighbor nodes; in each message passing process, the information collected from other nodes is aggregated with the information of the drug node itself; after multiple message passing, the aggregation results corresponding to different relationship types are obtained, and the aggregation results of different levels and types are combined again in a tensor stacking manner to generate the final node feature.

[0039] Specifically, for a triple (d, r, v) related to a drug, the message from the source node v to the target node d is defined as: m d←v = f(d i , v j , p dv ), where m d←v represents the embedding of the drug node, di Represents drug entity v j Refers to other entity types related to drugs, p dv Represents the impact factor of different edge relationships on drug node embedding, f(.) is used as the message encoding mechanism function, and d i , v j As the input of the function. Set f(.) to: in, is a normalization parameter determined by the number of adjacent nodes with a relationship r to the tail entity. From the perspective of message passing, It means that the message being propagated decreases as the number of neighbor nodes increases, which can be regarded as a decay coefficient. represents the transpose of the adjacency matrix with a relationship r with the tail entity v, is a trainable weight matrix that refers to the influence of messages transmitted by different types of edges.

[0040] In the above drug entity representation, after each node undergoes l message transfers, its embedded code becomes the complete features of its l-hop neighborhood. This process integrates the node's local structural information into its features, ensuring that each node's features reflect not only its own information but also that of its neighboring nodes. The message propagation mechanism for the drug entity in the diagram is modeled, and the message aggregation process at each layer is modeled as follows:

[0041]

[0042] in, Represents the feature vector of the target node d after aggregation at layer l, Represents the feature vector of the source node v at layer l-1, Represents the feature vector of the target node d at layer l-1, and are the weight matrices of different relationship types and self-relationships in the l-1th layer, σ is a nonlinear activation function, and |N(v, r)| represents the set of neighbor nodes whose relationship with node v is r.

[0043] By combining its own feature vector and the changed feature vectors of neighboring nodes, different edge types from other nodes to the drug node are processed independently, achieving the purpose of naturally encoding drug node features through the message propagation mechanism.

[0044] To capture the complex interactions between multiple relationship types and reflect the complex relationship structure and dependencies between nodes, we use tensor stacking. The processing is as follows:

[0045]

[0046] in, is the feature vector of drug node d after aggregating the structure-related information at the lth layer. is the feature vector of drug node d after aggregating the target-related information at the lth layer. is the feature vector of drug node d after aggregating the enzyme-related information at the lth layer. is the feature vector of drug node d after aggregating the pathway-related information at the lth layer.

[0047] wherein the elements of E are represented as:

[0048]

[0049] wherein n is the element number, and j, k represent the position of the element in the matrix.

[0050] In one embodiment, based on the double-aggregation message passing layer, the process of aggregating the feature information of the drug node and the feature information of the neighborhood set of the drug node in the multi-relation directed knowledge graph to obtain the final node feature of the drug node further includes: performing smoothing processing on the features after tensor stacking.

[0051] In this embodiment, one drug node can be connected to four different types of neighbors, and different types of neighbors have different influences on node attributes. The edge relationships of each drug node are different, and some nodes can only contain one or two edge relationships. In the process of matrix stacking, in order to reduce the noise caused by such uncertain relationships, the stacked matrix E is subjected to smoothing processing, and the processing process of other nodes is the same as that of the drug node. The process is as follows:

[0052]

[0053] In one embodiment, the method at least includes the following method for obtaining the message propagation mechanism: obtaining the adjacency matrix corresponding to different types of edge relationships and the corresponding weight matrix; obtaining the self-feature information of the drug node; and calculating the message propagation process based on the adjacency matrix corresponding to different types of edge relationships and the corresponding weight matrix, and the self-feature information of the drug node.

[0054] Specifically, it is assumed that the adjacency matrix corresponding to different types of edge relationships is A s，d , A t，d , A e，d , A p，d , the trainable weight matrix is w s，d , w t，d , w e，d , w p，d , and the initial feature of the node in the relationship network is m d , m s , mt , m e , m p . The drug node related to the substructure is denoted as m d←s The drug node related to the target is denoted as m d←t The drug node related to the enzyme is denoted as m d←e The drug node related to the pathway is denoted as m d←t , a controls the mixing ratio of new features and original features in the message propagation process. Therefore, the message propagation process of different relationship types can be symbolized as:

[0055]

[0056]

[0057] In one embodiment, the process of reconstructing the relationship matrix based on the aggregated node features includes: passing the embedding vector of the source node, the embedding vector of the target node, and the relationship matrix between the source node and the target node into different layers respectively to reconstruct the association of drug-target, drug-structure, drug-enzyme, and drug-pathway.

[0058] In this embodiment, it is assumed that the embedding of the node is where v represents different types of source nodes, and our goal is to reconstruct the relationship between these triplets. The complexity of the reconstruction process lies in the fact that the decoder needs to reconstruct the edges of the graph according to the node representation. In order to solve this problem, the input of the decoder is set to be able to receive a pair of node embeddings and a relationship type, i.e. The association of drug-target, drug-structure, drug-enzyme, and drug-pathway is reconstructed by passing the embedding vector of the source node, the embedding vector of the target node, and the relationship matrix between the two nodes into different layers respectively.

[0059] The drug node features, target node features, and the association between the two nodes are sent to the input layer of the encoder, and the association between the drug and the target is calculated, and the calculation formula is as follows:

[0060]

[0061] wherein represents the reconstructed relationship matrix of drug-target in the lth layer, represents the embedding feature of the drug node in the (l-1)th layer, represents the current trainable parameter, an initial diagonal matrix representing the relationship matrix between the two nodes, represents the embedding feature of the target node in the (l-1)th layer.

[0062] Specifically, the relationships among drugs-structures, drugs-enzymes, and drugs-pathways are reconstructed in turn using a pair of node embeddings and the relationships of the nodes, and the reconstructed relationship matrices are represented as follows:

[0063]

[0064] wherein are the reconstructed relationship matrices of the drugs-structures, drugs-enzymes, and drugs-pathways at the lth layer, respectively, are the embedding features of the drug nodes at the (l-1)th layer, respectively, are the current trainable parameters, respectively, are initialized as a diagonal matrix representing the relationship matrix between two nodes, respectively, are the embedding features of the structure nodes, enzyme nodes, and pathway nodes at the (l-1)th layer, respectively.

[0065] In one embodiment, the method further comprises the following steps of optimizing the prediction model: calculating the difference between the prediction result and the true result based on the constructed joint loss optimization function, and updating the trainable parameters in the prediction model through a back propagation algorithm.

[0066] In this embodiment, the features and context information of the nodes can be shared by different types of relationships. For example, proteins or enzymes participating in the same biological pathway usually have interrelated biological functions; similar chemical structures usually mean similar biological activities and targets; drugs with similar mechanisms of action or interactions may have similar targets or affect the same biological pathways. According to the above content, different types of relationships are considered to be introduced, multiple loss functions are set, and the shared information between tasks is used to jointly adjust the learnable parameters. This way of using shared information to guide the model for joint optimization not only realizes the information complementation of the knowledge graph, but also solves the problem of insufficient model generalization ability caused by data sparsity.

[0067] In one embodiment, the method comprises the following steps of obtaining the joint loss optimization function: obtaining the true adjacency matrix between different relationships; obtaining the mean square error between the true adjacency matrix and the reconstructed relationship matrix between different relationships, and obtaining the loss function of the mean square error matrix; combining all the loss functions of the mean square error matrix by summation to construct the total loss function.

[0068] Specifically, for the relationship {r i |r i ∈R}, the reconstruction loss is the mean square error between the true relationship matrix and the reconstructed relationship matrix, and the loss function of the difference matrix can be represented as:

[0069]

[0070] wherein, and denote the real adjacency matrix between different relations, and denote the reconstructed relation matrix, L dd denote the drug-drug loss, L ds denote the drug-structure loss, L dt denote the drug-target loss, L de denote the drug-enzyme loss, L dp denote the drug-pathway loss. The above loss functions are combined by summation to form a total loss function:

[0071]

[0072] Specifically, when making a prediction, the threshold of the prediction score is set to 0.5, and if the element of the difference matrix is lower than 0.5, it is considered that there is no potential drug interaction, otherwise it is considered that there is a potential drug interaction.

[0073] In an embodiment, the method further comprises the following steps of obtaining the joint loss optimization function: introducing a regularization term, the regularization term including the norm of node embedding and relation matrix; obtaining the joint loss optimization function based on the total loss function, the regularization term.

[0074] In this embodiment, in order to prevent model overfitting, a regularization term is introduced. The regularization term includes the norm of node embedding and relation matrix, and the specific form is:

[0075]

[0076] wherein, E is the double-aggregated feature vector in the graph, R r is the edge relation matrix in the graph, ||·||F F denotes the Frobenius norm, and λ is a regularization coefficient, representing a trade-off parameter between accuracy and complexity. After introducing the regularization term, the final optimization objective is:

[0077]

[0078] By jointly training the loss functions between the above different linear combinations, the learnable parameters can be optimized together, thereby promoting the knowledge complementation between multiple biological information and making up for the influence of data sparsity on model performance.

[0079] The training process of the model is in the form of mini-batch. For a batch of randomly sampled (d, r, v), the node representation and relation matrix after l-layer message propagation are input into the model, and the overall loss is calculated. Then, the gradient is calculated by backpropagation, and the model parameters are updated using the Adam optimizer. At the same time, the early stopping strategy is used to prevent overfitting and terminate the training in advance.

[0080] When starting the prediction task, the model complexity mainly comes from two aspects: the complexity of the message propagation and aggregation process and the complexity of calculating the scoring function. For the former, the time complexity of each epoch of training is O (|Q| · F 2 + |V| · F), and the time complexity of n epochs is O (Q (|Q| · F 2 + |V| · F)). For the latter, the time is mainly spent on the inner product operation of the prediction layer, with a complexity of O (V 2 ), and the overall time complexity is where Q is the number of relationship edges, F is the feature dimension, V is the number of nodes, and l is the number of message propagation layers.

[0081] In clinical practice, chloroquine has been proven to be effective against malaria. In addition, chloroquine has also been found to have some effect in the treatment of some autoimmune diseases such as systemic lupus erythematosus and rheumatoid arthritis, possibly by suppressing the overactive immune system to reduce inflammation. In terms of drug interactions, chloroquine produces complex interactions with many other drugs through mechanisms such as affecting intracellular pH, affecting the metabolism of pathogen heme, etc. For example, chloroquine can affect the activity of cytochrome P450 enzymes, thereby affecting the metabolic pathways and plasma concentrations of certain drugs.

[0082] Table 1 Top 50 predicted drug pairs

[0083]

[0084]

[0085] Reasonable management of the combination of chloroquine and other drugs is essential. In order to further verify the prediction performance of the model on drug interactions, the approved drug chloroquine was selected for case study. From the unknown drug interaction pairs, 122 pairs of data containing chloroquine were selected, and the model was used for case study experiments on this data. According to the prediction score, the top 50 prediction results were biologically verified, of which 48 pairs were verified in the database, Figure 2The 48 interactions are shown using a knowledge graph. The specific data is shown in Table 1. Also, in the data of the FDA, the use of chloroquine and hydroxychloroquine can have potential interactions, and after checking, it was found that in a recent clinical experiment, chloroquine appeared serious side effects in the combination therapy of COVID-19, so the FDA revoked the emergency use authorization of chloroquine and hydroxychloroquine in the treatment of COVID-19.

[0086] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the original technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will all fall within the protection scope of the present application.

Claims

1. A method for predicting drug interactions, characterized in that: include: Obtaining a set of drug entities for which interactions are to be predicted; Inputting the set of drug entities whose interactions are to be predicted into a prediction model to obtain a prediction result; The process of constructing the prediction model includes: constructing a multi-relationship directed knowledge graph containing drug-related information; respectively obtaining the representations of the source node and the target node in the multi-relationship directed knowledge graph during each message propagation process; based on a dual-aggregation message transmission layer, aggregating the feature information of the drug node and the feature information of the neighborhood set of the drug node obtained in the multi-relationship directed knowledge graph to obtain the final node feature of the drug node; reconstructing the relationship matrix based on the aggregated node features, including respectively passing the embedding vector of the source node, the embedding vector of the target node, and the relationship matrix between the source node and the target node to different layers to reconstruct the association between drug-target, drug-structure, drug-enzyme, and drug-pathway. The calculation formula is as follows: in represents the reconstructed relationship matrix of drug-target at the lth layer, represents the embedding features of the drug node in the l-1 layer, Represents the current trainable parameters, Initialize a diagonal matrix to represent the relationship matrix between two nodes. Represents the embedding features of the target node in the l-1 layer; using a pair of node embeddings and the relationship between nodes, the relationships between drug-structure, drug-enzyme, and drug-pathway are reconstructed in turn. The reconstructed relationship matrices are expressed as follows: in Represent the reconstructed relationship matrices of drug-structure, drug-enzyme, and drug-pathway at the first level, Represent the embedding features of the drug nodes in the l-1 layer, Represent the current trainable parameters, Initialize a diagonal matrix to represent the relationship matrix between two nodes. Represent the embedded features of the l-1 layer structure nodes, enzyme nodes, and pathway nodes respectively; predict the interactions between drugs based on the reconstructed relationship matrix; The neighborhood set is obtained by at least the following method: obtaining neighbor nodes of the drug node and nodes adjacent to the neighbor nodes; Based on the dual aggregation message passing layer, the feature information of the drug node and the feature information of the neighborhood set of the drug node obtained in the multi-relationship directed knowledge graph are aggregated to obtain the final node feature of the drug node, which at least includes: obtaining messages from the neighbor nodes of the drug node to the drug node; setting corresponding message propagation mechanisms for the relationships between drugs and different types of entities; based on the message propagation mechanism, messages are propagated along the edges of different relationship types in the multi-relationship directed knowledge graph, so that the drug node can obtain information about its neighbor nodes and the neighbor nodes of the neighbor nodes; after each node undergoes l message transmissions, the embedded code of the node becomes all the features of its l-hop neighborhood; the message propagation mechanism of the drug entity is modeled, and the message aggregation process of each layer is modeled as follows: in, Represents the feature vector of the target node d after aggregation at layer l, Represents the feature vector of the source node v at layer l-1, Represents the feature vector of the target node d at layer l-1, and are the weight matrices of different relationship types and self-relationships in the l-1th layer, σ is a nonlinear activation function, and |N(v, r)| represents the set of neighbor nodes whose relationship with node v is r; During each message transmission process, the information collected from other nodes is aggregated with the information of the drug node itself; after multiple message transmissions, the aggregation results corresponding to different relationship types are obtained, and the aggregation results of different levels and types are secondary combined using tensor stacking to generate the final node features.

2. The method according to claim 1, characterized in that Based on the dual aggregation message passing layer, the feature information of the drug node and the feature information of the neighborhood set of the drug node obtained in the multi-relationship directed knowledge graph are aggregated to obtain the final node feature of the drug node, and the process also includes: smoothing the features after tensor stacking.

3. The method according to claim 1, characterized in that The multiple relationships are at least relationships among drugs, targets, enzymes, structures and pathways.

4. The method according to claim 1, wherein The method at least includes the following method to obtain the message propagation mechanism: Get the adjacency matrix and weight matrix corresponding to different types of edge relationships; Obtaining characteristic information of the drug node itself; The message propagation process is calculated based on the adjacency matrix and the corresponding weight matrix corresponding to the different types of edge relationships and the characteristic information of the drug node itself.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises at least the following steps of optimizing the prediction model: The difference between the predicted result and the true result is calculated based on the constructed joint loss optimization function, and the trainable parameters in the prediction model are updated through the back propagation algorithm.

6. The method according to claim 5, characterized in that The method comprises at least the following steps to obtain the joint loss optimization function: Get the true adjacency matrix between different relations; Obtain the true adjacency matrix between different relationships, the mean square error between the reconstructed relationship matrices, and obtain the loss function of the mean square error matrix; The loss functions of all mean square error matrices are combined by summing to construct the total loss function.

7. The method according to claim 6, characterized in that The method further comprises the following steps of obtaining the joint loss optimization function: Introducing a regularization term that includes the norms of node embedding and relationship matrices; Based on the total loss function and the regularization term, the joint loss optimization function is obtained.

Images