Drug interaction prediction method and system based on meta-path length and type

Abstract

The application discloses a drug interaction prediction method and system based on meta-path length and type, and the method comprises the following steps: constructing a heterogeneous graph; the heterogeneous graph mainly comprises a node set and an edge set; different node types in the node set and different edge types in the edge set are combined to form a sequence to describe a meta-path structure; each meta-path is coded, and a length factor of each meta-path is set to obtain a meta-path length; according to the meta-path length, the length of the lth meta-path is k, and the fusion meta-path features of each node are obtained; the fusion meta-path feature sets of all nodes are combined to obtain a fusion meta-path feature set; an interaction prediction model between drugs is set, the fusion meta-path features of two nodes in the fusion meta-path feature set are selected as input, and the probability of interaction between the two drugs is obtained through the interaction prediction model between the drugs. Through fine coding of the path length and type, the accuracy of DDI prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug-drug interaction prediction, in particular to a drug-drug interaction prediction method and system based on meta-path length and type. BACKGROUND

[0002] Drug-drug interaction (DDI) refers to the change in the expected effect when one drug is used in combination with another. DDI can lead to unexpected side effects, harm to patients, or weaken the efficacy of drugs, and even endanger the lives of patients. However, in clinical situations, doctors often need to prescribe multiple drugs to treat patients' diseases. Therefore, accurate prediction of DDI is very important for drug development and drug safety monitoring.

[0003] Traditionally, DDI is determined by wet laboratory experiments, but this method is costly and time-consuming, increasing the cost of drug development for pharmaceutical companies. With more and more drug-related information and data being made public, researchers have been working to design computational methods to predict DDI. Existing DDI prediction methods can be mainly divided into two categories: similarity-based methods and network-based methods.

[0004] Similarity-based methods predict DDI by calculating the similarity between drugs. These methods first extract drug features from drug-related information, and then calculate the similarity between each pair of drugs. Network-based methods, on the other hand, construct a network based on drug-related information, where nodes represent drugs and related biomedical entities, and edges represent the association between corresponding nodes. Then matrix operations or graph neural networks are performed on the constructed network to complete the task of DDI prediction.

[0005] At present, we are increasingly building heterogeneous graph neural networks for drug-drug interaction (DDI) prediction, which has many advantages.

[0006] First, heterogeneous graph neural networks can better handle the complex relationships between different types of nodes and edges, such as the interaction between drugs, proteins, and other heterogeneous nodes, as well as different types of associations between them. This ability allows the model to more comprehensively capture the relationships between different types of data, improving the model's ability to integrate diverse information. Second, heterogeneous graph neural networks can effectively utilize the information in the heterogeneous network structure, thereby improving the accuracy and generalization ability of DDI prediction, especially suitable for scenarios that require consideration of multiple data types and complex associations. In addition, by combining information from different data sources, heterogeneous graph neural networks can better uncover the underlying patterns and rules hidden in the data, providing deeper insights for DDI prediction. Finally, heterogeneous graph neural networks also have good scalability and flexibility, making it easy to adapt to different types of data and task requirements, providing a more powerful and flexible tool for data analysis and prediction in the field of drug research.

[0007] However, there are also many disadvantages in building heterogeneous graph neural networks for DDI prediction, such as Figure 1 In the middle, the shortest path of d1 to d3 is d1-p1-p3-d3, and the shortest path of d1 to d2 is d1-p4-d2, the lengths of the two paths are different, the types are different, and the traditional graph neural network may not be able to well distinguish and encode the features between the two different paths. In DDI prediction, this situation may lead to insufficient information extraction and representation ability of the model for different paths, affecting the accuracy and generalization ability of the model for drug interactions. Such limitations make it difficult for the model to fully utilize the rich information in the heterogeneous graph, and cannot accurately capture the importance and features of different paths, thereby limiting the modeling ability of the model for complex drug interaction relationships.

[0008] In the prior art, the patent with the patent number CN115512761A discloses a drug-drug interaction prediction framework based on meta-path, mainly discloses: a heterogeneous information network construction module, a representation learning module based on meta-path, and a prediction module; the heterogeneous information network construction module: combines and represents multiple drug-drug interaction related data sets as a heterogeneous information network, wherein the drugs and proteins are nodes of the heterogeneous information network, and the semantic association between the two nodes is an edge; the representation learning module based on meta-path: based on the information fusion mechanism of meta-path, to capture the complex semantic association in the heterogeneous information network, so as to learn high-quality drug and protein representation; the prediction module: finally obtains the corresponding predicted interaction by using the learned drug representation. However, in DDI prediction, the model has insufficient information extraction and representation ability for different paths. SUMMARY

[0009] The technical problem to be solved by the present application is to solve the problem that the prior art cannot capture the rich semantic information between drug features, has insufficient information extraction and representation ability for different paths, and affects the accuracy and generalization ability of the prediction model for drug interactions.

[0010] To solve the above technical problems, the present application provides the following technical solutions:

[0011] A drug interaction prediction method based on meta-path length and type, comprising:

[0012] S10, constructing a heterogeneous graph; wherein the heterogeneous graph mainly consists of a node set V and an edge set E;

[0013] S20, forming a sequence of different node types in the node set V and different edge types in the edge set E to describe the meta-path structure;

[0014] S30, encode each meta-path and set the length factor of each meta-path to obtain the meta-path length;

[0015] S40, according to the meta-path length, obtain the lth meta-path length k, and the fusion meta-path feature of each node;

[0016] S50, obtain the fusion meta-path feature set by collecting the fusion meta-path features of all nodes; set the drug interaction prediction model, select the fusion meta-path features of two nodes from the fusion meta-path feature set as input, and obtain the interaction probability between the two drugs through the drug interaction prediction model.

[0017] In an embodiment of the present application, the node set V includes a drug node set and a protein node set; the edge type in the edge set E includes: interaction between drugs; interaction between proteins; protein as drug target; protein as drug carrier; protein as drug enzyme; protein as drug transporter.

[0018] In an embodiment of the present application, the fusion meta-path feature of each node is obtained by the following formula:

[0019]

[0020] In the formula, indicates the fusion meta-path feature of node d i , ReLU is an activation function, and respectively indicate the representation of drug neighbor d t and protein neighbor p t of the l-1th meta-path after information fusion based on the meta-path; indicates the encoding information of different meta-paths, W ii is the length factor of different meta-paths, is the trainable parameter matrix of the drug in the lth meta-path, is the trainable parameter matrix of the protein in the lth meta-path, is the drug node set in all neighbor nodes of node d i , is the protein node set in all neighbor nodes of node d i .

[0021] In an embodiment of the present application, the drug interaction prediction model is obtained by the following formula:

[0022]

[0023] In the formula, where h is the interaction probability between two drugs, Sigmoid is a sigmoid function, W and b are different training parameters, and h i and h j are two nodes selected from the fusion meta-path feature set, respectively.

[0024] In an embodiment of the present application, the cross-entropy loss function is used to optimize the drug interaction prediction model, and the cross-entropy loss function is defined as:

[0025]

[0026] where S is a training sample, is a predicted label, y s is a real label, and log is a logarithmic function.

[0027] The present application also provides a system using the drug interaction prediction method based on meta-path length and type, comprising:

[0028] a heterogeneous graph module for constructing a heterogeneous graph, wherein the heterogeneous graph mainly comprises a node set V and an edge set E;

[0029] a meta-path type module for forming a sequence of different node types in the node set V and different edge types in the edge set E to describe a meta-path structure;

[0030] a meta-path length module for encoding each meta-path and setting a length factor of each meta-path to obtain a meta-path length;

[0031] a fusion meta-path feature module for obtaining a fusion meta-path feature of each node according to the meta-path length, wherein the meta-path length of the lth meta-path is k;

[0032] a prediction module for obtaining a fusion meta-path feature set by collecting the fusion meta-path features of all nodes, selecting fusion meta-path features of two nodes from the fusion meta-path feature set as inputs, and obtaining an interaction probability between two drugs by a drug interaction prediction model.

[0033] In an embodiment of the present application, the node set V in the heterogeneous graph module comprises a drug node set and a protein node set, and the edge type in the edge set E comprises an interaction between drugs, an interaction between proteins, a protein acting as a drug target, a protein acting as a drug carrier, a protein acting as a drug enzyme, and a protein acting as a drug transporter.

[0034] In an embodiment of the present application, the fusion meta-path feature of each node is obtained by the following formula:

[0035]

[0036] In the formula, is represented as the fusion meta-path feature of node d i , ReLU is an activation function, and respectively represent the representation of drug neighbors d t and protein neighbors p t after information fusion based on the l-1th meta-path; represents the encoding information of different meta-paths, W ii is the length factor of different meta-paths, is the trainable parameter matrix of the lth drug, is the trainable parameter matrix of the lth protein, is the set of drug nodes among all neighbor nodes of node d i , is the set of protein nodes among all neighbor nodes of node d i .

[0037] In an embodiment of the present application, in the prediction module, the interaction prediction model between drugs is obtained by the following formula:

[0038]

[0039] In the formula, is the probability of interaction between two drugs, Sigmoid is a sigmoid function, W and b are different training parameters, h i and h j are respectively the fusion meta-path features of two nodes selected from the fusion meta-path feature set.

[0040] In an embodiment of the present application, in the prediction module, the interaction prediction model between drugs is optimized using a cross-entropy loss function; the cross-entropy loss function is defined as:

[0041]

[0042] In the formula, S is a training sample, is a predicted label, y s is a real label, and log is a logarithmic function.

[0043] Compared with the prior art, the present application has the beneficial effects that: the present application defines different types and lengths of meta-paths, and forms new adjacency matrices for each meta-path. By encoding the path length and type in these new adjacency matrices, the present application can better capture the rich semantic information between drug features, thereby improving the accuracy and interpretability of DDI prediction.

[0044] Unlike traditional similarity-based or network-based methods, this method explicitly utilizes biomedical knowledge related to DDI (such as PK and PD) and captures complex semantic associations between nodes through encoding paths. Specifically, a meta-path is a path of a specific type and length defined in a heterogeneous information network, which can capture specific relationships and complex semantic associations between nodes, including transformations and interactions between nodes.

[0045] By encoding different types and lengths of paths, the present application can represent drug features as a set of highly fine-grained vectors that better reflect the interactions and complexity between drug features, thereby improving the accuracy and interpretability of DDI prediction.

[0046] Encoding path types and lengths provides a new way for the present application to learn drug features, making full use of biomedical knowledge related to DDI and complex semantic associations between nodes in HIN.

[0047] The present application proposes a method based on meta-path to encode the length and type of the path, and the present application uses an improved heterogeneous information network-based framework for the prediction task of drug-drug interactions. By constructing a HIN containing drug and protein nodes and explicitly utilizing the pharmacological and pharmacokinetic biomedical knowledge related to DDI, the present application can effectively capture the complex interrelationships between drugs. Not only the type and length of the path are considered, but also they are fine-tuned through heterogeneous distance encoding and node type encoding. This design enables the model to more flexibly capture and integrate information from different biological entities and relationships, thereby generating higher quality drug and protein representations. This method not only improves the accuracy of DDI prediction, but also enhances the interpretability of the model. This method not only improves the accuracy of prediction, but also provides a new perspective for understanding and explaining the complex interactions between drugs, which is of great significance for guiding clinical drug use and new drug development. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A meta-path diagram.

[0049] Figure 2 A meta-path length and type-based drug interaction prediction method flowchart according to an embodiment of the present application.

[0050] Figure 3 A meta-path length and type-based drug interaction prediction system block diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0051] For the person skilled in the art to understand the technical scheme of the present application, the technical scheme of the present application will be further described in conjunction with the drawings of the specification.

[0052] The terms "first", "second", "third", etc. are used only for descriptive purposes and should not be construed as implying or suggesting relative importance or an indicated number of the technical features. Thus, the features defined with "first", "second", etc. can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0053] Please refer to Figure 2 As shown in the drawings, the present application provides a drug interaction prediction method based on meta-path length and type, comprising:

[0054] S10, constructing a heterogeneous graph; wherein the heterogeneous graph mainly consists of a node set V and an edge set E.

[0055] In an embodiment of the present application, the heterogeneous graph consists of a node set V and an edge set E, wherein the node set V can be divided into a drug node set S D and a protein node set S P . Assuming that the drug node set S D ={d1,…,d i ,…,d E}, the protein node set S P ={p1,…,p j ,…,p F}, wherein E and F represent the total number of drugs and the total number of proteins, respectively.

[0056] In the heterogeneous graph, proteins can play multiple different roles, such as targets, carriers, enzymes, and transport proteins, etc. In order to simplify the representation of proteins with different roles and reduce data storage redundancy, the present application uses the same type of node (i.e. protein) to represent these different roles. For example, serum transferrin can act as both an enzyme for ferrous sulfate and a carrier for bismuth subsalicylate. This design makes it more convenient to integrate information about proteins with different roles.

[0057] According to different types of edges, the present application defines six kinds of undirected edges to represent the different roles of proteins: (1) the interaction between drugs, represented by drug DDI-drug; (2) the interaction between proteins, represented by protein PPI-protein; (3) the protein acting as a drug target, denoted as Protein A-drug; (4) the protein acting as a drug carrier, denoted as Protein C-drug; (5) the protein acting as a drug enzyme, denoted as Protein E-drug; (6) the protein acting as a drug transport protein, denoted as Protein T-drug.

[0058] S20, a sequence of different node types in the node set V and different edge types in the edge set E is formed to describe the meta-path structure.

[0059] In an embodiment of the present application, the relationship between drugs in the heterogeneous graph is described by defining different types of meta-paths. A meta-path is a sequence of node types and edge types, which is used to describe the structure of the path. A meta-path of length K can be formally represented as where each V k (k∈{1,…,k+1}) represents a node type, and each R k (k∈{1,…,k+1}) represents an edge type. Suppose that node i and node j are the starting node and the ending node of a meta-path instance, respectively. According to the instance of the meta-path, the present application refers to node j as the neighbor of node i. Since the edges in the heterogeneous information network (HIN) in the research of the present application are undirected, according to the same instance of the meta-path, node i is also the neighbor of node j. Given a node i and a meta-path ρ, the neighbor set of node i according to the meta-path ρ is defined as the set of all neighbor nodes according to all instances of the meta-path, denoted as N ρ (i).

[0060] S30, encode each meta-path and set the length factor of each meta-path to obtain the meta-path length.

[0061] In an embodiment of the present application, each given meta-path is encoded and defined, such as for the meta-path d-d, the encoding result is W DD , for the meta-path d-p, the encoding result is W DP , for the meta-path d-d-p, the encoding result is W DD *W DP , for the meta-path d-p-d, the encoding result is W DP *W DP , and so on. At the same time, the meta-paths of different lengths are multiplied by the corresponding length factors W1, W2, W3 to realize the encoding of the influence of the length of the meta-path. For example, for the meta-path d-d, we multiply the length factor W1, for the meta-path d-d-p, we multiply the length factor W2, and for the meta-path d-d-p-d, we multiply the length factor W3.

[0062] S40, according to the meta-path length, obtain the fusion meta-path feature of each node in the lth meta-path with a length of k.

[0063] In one embodiment of the present invention, the learning module consists of L information fusions based on meta-paths. This application uses the l-th layer as an example to describe the idea of ​​information fusion based on meta-paths. We take the drug node d as an example. i Let's take an example to explain the node update process in HIN.

[0064] In the implementation of this embodiment, in order to reduce the complexity of the proposed model, information fusion based on meta-paths is performed according to the length of the meta-paths, and the maximum length of the meta-paths under consideration is represented by K.

[0065] Given length k∈{1,…,K} and drug node d i Drug node d i All neighboring nodes can be divided into two subgroups based on the metapath of length k, namely the set of drug nodes. Represents node d i The set of drug nodes in all neighboring nodes and The node d represents i The set of protein nodes among all neighboring nodes.

[0066] In the l-th layer, there is a node d with path length k. i Meta-path-based information fusion is formalized as follows:

[0067]

[0068] In the formula, Represented as node d i The fusion metapath features are used, with ReLU as the activation function. and The drug neighbor d is the result of information fusion based on the (l-1)th meta-path. t and protein neighbor p t The representation of; W represents the encoding information of different meta-paths. ii For different metapath length factors, Let be the trainable parameter matrix of the drug in the l-th clause. Let be the trainable parameter matrix of the protein in the l-th clause. For node d i The set of drug nodes among all neighboring nodes, For node d i The set of protein nodes among all neighboring nodes.

[0069] S50: Collect the fusion meta-path features of all nodes to obtain the fusion meta-path feature set; set up a drug interaction prediction model, select the fusion meta-path features of two nodes from the fusion meta-path feature set as input, and obtain the probability of interaction between the two drugs through the drug interaction prediction model.

[0070] In one embodiment of the present invention, given drug H D The final representation matrix, i.e., the fused meta-path feature set, is used to derive the DDI prediction probability for each drug pair using a fully connected layer with a sigmoid function. Formally, given a drug pair d i and d j It can be derived from the fusion meta-path feature set H D Obtain the final representation h i and h j Then drug d i and drug d j The probability that there is an interaction between them is defined as follows:

[0071]

[0072] In the formula, Let h be the probability of interaction between two drugs, sigmoid be a sigmoid function, and W and b be different training parameters. i and h j These are the fusion meta-path features of two nodes selected from the fusion meta-path feature set.

[0073] In this embodiment, a cross-entropy loss function is used to optimize the drug-drug interaction prediction model; the cross-entropy loss function is defined as:

[0074]

[0075] In the formula, S represents the training samples. To predict the label, y s To output the true label of the drug node, log is a logarithmic function. The training samples consist of multiple subsamples {(X1,y1),(X2,y2),......,(X... s ,y s )}, as well as, and These are two nodes with labels.

[0076] Please see Figure 3 As shown, the present invention also provides a system for predicting drug interactions based on metapath length and type, comprising:

[0077] The heterogeneous graph module constructs a heterogeneous graph, wherein the heterogeneous graph mainly comprises a node set V and an edge set E.

[0078] The meta-path type module describes a meta-path structure by combining different node types in the node set V and different edge types in the edge set E into a sequence.

[0079] The meta-path length module encodes each meta-path and sets a length factor of each meta-path to obtain a meta-path length.

[0080] The fusion meta-path feature module obtains the fusion meta-path feature of each node according to the meta-path length, wherein the meta-path length of the lth meta-path is k.

[0081] The prediction module obtains a fusion meta-path feature set by combining the fusion meta-path features of all nodes; and sets a drug interaction prediction model, selects the fusion meta-path features of two nodes from the fusion meta-path feature set as inputs, and obtains the probability of interaction between the two drugs by the drug interaction prediction model.

[0082] In the heterogeneous graph module, the node set V comprises a drug node set and a protein node set; and the edge types in the edge set E comprise: interaction between drugs; interaction between proteins; protein as a drug target; protein as a drug carrier; protein as a drug enzyme; and protein as a drug transporter.

[0083] In the fusion meta-path feature module, the fusion meta-path feature of each node is obtained by the following formula:

[0084]

[0085] In the formula, f represents the fusion meta-path feature of the node d i , ReLU is an activation function, and f t and f t respectively represent the representations of the drug neighbor d ii and the protein neighbor p i of the (l-1)th meta-path after information fusion based on the meta-path; represents the encoding information of different meta-paths, W i is a length factor of different meta-paths, is a trainable parameter matrix of the drug in the lth meta-path, is a trainable parameter matrix of the protein in the lth meta-path, is a drug node set in all neighbor nodes of the node d i , and is a protein node set in all neighbor nodes of the node d j .

[0086] In the prediction module, the interaction prediction model between drugs is obtained by the following formula:

[0087]

[0088] In the formula, is the probability of interaction between two drugs, Sigmoid is a sigmoid function, W and b are different training parameters, h i and h j are two node fusion meta-path features selected from the fusion meta-path feature set.

[0089] The cross-entropy loss function is used to optimize the interaction prediction model between drugs, and the cross-entropy loss function is defined as:

[0090]

[0091] In the formula, S is the training sample, is the predicted label, y s is the true label, and log is the logarithmic function.

[0092] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims should be included in the present application, and any reference signs in the claims should not be regarded as limiting the claims.

[0093] The above-described embodiments only represent the implementation of the present application, and the protection scope of the present application is not limited to the above-described embodiments. For those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. A drug interaction prediction method based on metapath length and type, characterized in that, To address the insufficient information extraction and representation capabilities of different meta-paths, including: S10, Construct a heterogeneous graph; where the heterogeneous graph mainly consists of a set of nodes V and a set of edges E; S20, which combines different node types in node set V with different edge types in edge set E to form a sequence, describing the meta-path structure; S30: Encode each metapath and set the length factor for each metapath to obtain the metapath length. S40, based on the metapath length, obtain the first... l The length of the metapath is [number]. k The fusion meta-path features of each node are obtained using the following formula: ； In the formula, Represented as nodes The fusion meta-path features, For activation function, and The first Drug Neighbors Based on Meta-Path Information Fusion and protein neighbors The representation of; Encoding information representing different meta-paths, For different metapath length factors, For the first The trainable parameter matrix of the drugs in the sample. For the first The trainable parameter matrix of the proteins in the sample. For nodes The set of drug nodes among all neighboring nodes, For nodes The set of protein nodes in all neighboring nodes; S50: Collect the fusion meta-path features of all nodes to obtain the fusion meta-path feature set; set up a drug interaction prediction model, select the fusion meta-path features of two nodes from the fusion meta-path feature set as input, and obtain the probability of interaction between the two drugs through the drug interaction prediction model.

2. The drug interaction prediction method based on metapath length and type according to claim 1, characterized in that, The node set V includes the drug node set and the protein node set; the edge set E includes the following edge types: interactions between drugs; interactions between proteins; proteins that are drug targets; proteins that are drug carriers; proteins that act as drug enzymes; and proteins that act as drug transport proteins.

3. The drug interaction prediction method based on metapath length and type according to claim 1, characterized in that, The drug-drug interaction prediction model is obtained using the following formula: ； In the formula, This represents the probability of an interaction between two drugs. It is a sigmoid function. and For different training parameters, and These are two nodes selected from the fusion meta-path feature set.

4. The drug interaction prediction method based on metapath length and type according to claim 3, characterized in that, The cross-entropy loss function is used to optimize the drug-drug interaction prediction model; the cross-entropy loss function is defined as: ； In the formula, As training samples, For predicting labels, For real labels, It is a pair of functions.

5. A system applying the drug interaction prediction method based on metapath length and type as described in any one of claims 1-4, characterized in that, include: The heterogeneous graph module constructs heterogeneous graphs; these graphs primarily consist of a set of nodes V and a set of edges E. The meta-path type module combines different node types in node set V with different edge types in edge set E to form a sequence, describing the meta-path structure. The metapath length module encodes each metapath, sets a length factor for each metapath, and obtains the metapath length. The fusion meta-path feature module obtains the first [item] based on the meta-path length. l The length of the metapath is [number]. k The fusion meta-path features of each node; The prediction module obtains the fusion meta-path feature set by collecting the fusion meta-path features of all nodes; it sets up a drug interaction prediction model, selects the fusion meta-path features of two nodes from the fusion meta-path feature set as input, and obtains the probability of interaction between the two drugs through the drug interaction prediction model.

6. The system according to claim 5, characterized in that, In the heterogeneous graph module, the node set V includes the drug node set and the protein node set; the edge set E includes the following edge types: interactions between drugs; interactions between proteins; proteins that are drug targets; proteins that are drug carriers; proteins that act as drug enzymes; and proteins that act as drug transporters.

7. The system according to claim 5, characterized in that, In the prediction module, the drug-drug interaction prediction model is obtained using the following formula: ； In the formula, This represents the probability of an interaction between two drugs. It is a sigmoid function. and For different training parameters, and These are the fusion meta-path features of two nodes selected from the fusion meta-path feature set.

8. The system according to claim 7, characterized in that, In the prediction module, the cross-entropy loss function is used to optimize the drug-drug interaction prediction model; the cross-entropy loss function is defined as: ； In the formula, As training samples, For predicting labels, For real labels, It is a pair of functions.

Images