Method, device, equipment and medium for predicting asymmetric drug interaction relationships

Abstract

The present invention provides a method, device, equipment and medium for predicting asymmetric drug interaction relationships, and relates to the field of drug interaction prediction. The present invention obtains the directed DDI network and drug feature matrix of the source drug and the target drug to be predicted, and inputs them into a dual-channel encoder; based on the directed DDI network, the drug topological vectors of the source drug and the target drug are obtained by the first encoder respectively; the drug feature matrix is ​​reduced in dimension by the second encoder to generate multi-scale embedding, and then fused and spliced ​​respectively to obtain the drug embedding vector of the corresponding drug; the drug topological vector and the drug embedding vector are input into the decoder for fusion to obtain the fused embedding vector of the corresponding drug; then the fusion is performed again based on the Hadamard product to generate the final fused drug pair representation; DDI prediction is performed according to the fused drug pair representation to obtain the prediction result. The present invention can effectively capture the directional information of asymmetric DDI and realize the accurate prediction of asymmetric DDI relationship.

Description

Technical Field

[0001] The present invention relates to the technical field of drug interaction prediction, and in particular to a method, device, equipment and medium for predicting asymmetric drug interaction relationships. Background Art

[0002] With the development of drug research and development and precision medicine, predicting drug-drug interactions (DDI) has gradually become a key issue in drug design and clinical medication. Drug interaction prediction technology can help researchers and doctors understand the effects of enhanced efficacy and increased toxicity that may occur when different drugs are used together, thereby effectively reducing medication risks and improving drug efficacy. At present, mainstream drug interaction prediction methods usually rely on the following steps: (1) characterizing drug properties through information such as molecular features and chemical fingerprints; (2) using existing drug interaction data to train prediction models; (3) inputting drug characteristics into the prediction model to obtain interaction results.

[0003] However, the prediction of drug interactions is not static, especially in practical application scenarios, the interactions between different drugs may have directional asymmetric characteristics. For example, drug A acts as a trigger factor to affect drug B, while drug B may have no significant effect on drug A. Therefore, how to introduce directional asymmetric factors into the prediction process and improve the model's ability to identify asymmetric drug interactions has become a new research focus.

[0004] Some current drug interaction prediction models are insufficient in capturing asymmetric characteristics. Conventional models mostly adopt the assumption of symmetric relationships, ignoring the role differences in drug interaction relationships, resulting in limited accuracy and reliability of prediction results. In addition, drug interaction networks and other drug feature data (such as Morgan fingerprints) are usually stored independently. How to achieve efficient data fusion in model training and prediction to ensure consistency between different data sources is still an urgent problem to be solved in drug interaction prediction technology.

[0005] In this context, designing a prediction model that can effectively learn the directional characteristics of drug interactions and maintain consistency during multi-source data fusion is of great significance to improving the accuracy and robustness of drug interaction prediction.

[0006] In view of this, the applicant filed this application after studying the existing technology. Summary of the invention

[0007] The present invention aims to provide a method, device, equipment and medium for predicting asymmetric drug interaction relationships, which can effectively learn the directional characteristics of drug interactions and maintain consistency during multi-source data fusion, thereby improving the accuracy and robustness of drug interaction prediction.

[0008] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0009] A method for predicting asymmetric drug interaction relationships, comprising:

[0010] Obtain the directed DDI network and drug feature matrix of the source drug and target drug to be predicted;

[0011] The directed DDI network and the drug feature matrix are input into a dual-channel encoder; based on the directed DDI network, the multi-level entry and exit neighborhood topological relationship of the drug is learned through the first encoder of the dual-channel encoder to obtain the drug topological vectors of the source drug and the target drug respectively; at the same time, the second encoder of the dual-channel encoder is used to reduce the dimension of the drug feature matrix to generate a multi-scale embedding, and the multi-scale fusion and splicing of the multi-scale embeddings of the source drug and the target drug are respectively performed to obtain the drug embedding vectors of the corresponding drugs;

[0012] The drug topology vector and the drug embedding vector are input into a decoder for fusion to obtain a fused embedding vector of the corresponding drug; and then, according to the fused embedding vector, fusion is performed again based on the Hadamard product to generate a final fused drug pair representation;

[0013] DDI prediction is performed based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

[0014] Preferably, the first encoder is a graph attention network; through the graph attention network, the topological information of the source drug and the target drug in the multi-layer outgoing neighbors and incoming neighbors are respectively extracted, specifically:

[0015] For the source drug, the topological representation of the out-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows:

[0016] ;

[0017] ;

[0018] in, Indicates the source drug Layer node features; A node representing the source drug in the directed graph; A node representing the target drug in the directed graph; A set of outbound neighbor nodes representing the source drug; Represents the current node in the source drug The corresponding attention coefficient; A weight matrix representing the source drug in the graph attention network; Indicates the target drug Layer node features, i.e., outbound neighbor node features of the source drug;

[0019] For splicing operation; is the weight matrix of the source drug attention coefficient; is a single-layer feed-forward network of the source drug, σ is a nonlinear activation function, i.e., LeakyReLU; for The neighborhood set of Neighborhood set The nodes in; T is the transpose; a drug embedding vector representing the source drug; a drug embedding vector representing the target drug;

[0020] For the target drug, the topological representation of the in-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows:

[0021] ;

[0022] ;

[0023] in, Indicates the target drug Layer node features; A set of inbound neighbor nodes representing the target drug; Indicates the current node in the target drug The corresponding attention coefficient; represents the weight of the target drug; Indicates the source drug Layer node features, i.e., neighbor node features of the target drug; for The neighborhood set of Neighborhood set Nodes in is a weight matrix of the target drug attention coefficient; is a single-layer feed-forward network of the target drug.

[0024] Preferably, the drug feature matrix includes a Morgan fingerprint feature matrix of the drug; the second encoder is an autoencoder; the Morgan fingerprint feature matrix is ​​layer-by-layer reduced in dimension using the autoencoder to obtain key feature embeddings of drugs at different scales, and the specific formula for reducing the dimension is:

[0025] ;

[0026] in, Indicates the source drug Layer feature embedding; Indicates the source drug Layer feature embedding; represents the activation function; , They represent the dimensionality reduction coefficients respectively.

[0027] Preferably, the specific formula of the multi-scale fusion splicing is:

[0028] ;

[0029] ;

[0030] ;

[0031] in, A drug topology vector representing the concatenated source drug; Indicates the source drug Layer node features; Indicates the source drug Layer node features; A drug topology vector representing the target drug after splicing; Indicates the target drug Layer node features; Indicates the target drug Layer node features; A drug embedding vector representing the source drug after concatenation; Indicates the source drug Layer feature embedding; Indicates the source drug Layer feature embedding.

[0032] Preferably, when the drug topology vector and the drug embedding vector are input into a decoder for fusion, an additive fusion method is adopted, and the formula is:

[0033] = + ;

[0034] = + ;

[0035] in, a fused embedding vector representing the source drug; is the drug topology vector of the source drug; is a drug embedding vector of the source drug;

[0036] a fused embedding vector representing the target drug; is the drug topology vector of the target drug; is the drug embedding vector of the target drug.

[0037] Preferably, the method further comprises: introducing a nonlinear transformation processing module into the decoder to optimize the drug pair representation and enhance the model performance.

[0038] Preferably, the method further comprises: when training the model, using a binary cross entropy loss function to evaluate the error between the true label and the predicted label; at the same time, inputting the drug embedding vectors of the two drugs into the decoder for decoding and reconstruction, and using a mean square error loss function to evaluate the error between the two drugs and the original drug features, and the expression is as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] in, represents the binary cross entropy loss; represents mean square error loss; a drug pair interaction relationship label representing the source drug; a drug pair interaction relationship label representing the target drug; An edge representing the source drug and the target drug; Represents a library of drug pair interaction relationships; Indicates the prediction result; represents the true label;

[0044] a fused embedding vector representing the source drug; A fusion embedding vector for the target drug; represents the Hadamard product; represents a multilayer perceptron; N represents the number of samples;

[0045] represents the drug embedding vector for the source drug The feature matrix after decoding and reconstruction; , They represent the decoded reconstruction coefficients respectively; a drug characteristic representing the source drug; A node representing the source drug in the directed graph; A drug feature matrix dimension representing the source drug; represents the L2 norm.

[0046] The present invention also provides a device for predicting asymmetric drug interaction relationships, comprising:

[0047] An acquisition unit, used for acquiring a directed DDI network and a drug feature matrix of a source drug and a target drug to be predicted;

[0048] A dual-channel encoding unit is used to input the directed DDI network and the drug feature matrix into a dual-channel encoder; based on the directed DDI network, the multi-level entry and exit neighborhood topological relationship of the drug is learned through the first encoder of the dual-channel encoder to obtain the drug topological vectors of the source drug and the target drug respectively; at the same time, the second encoder of the dual-channel encoder is used to reduce the dimension of the drug feature matrix to generate a multi-scale embedding, and the multi-scale fusion and splicing of the multi-scale embeddings of the source drug and the target drug are respectively performed to obtain the drug embedding vector of the corresponding drug;

[0049] A fusion unit, used for fusing the drug topological vector and the drug embedding vector into a decoder to obtain a fused embedding vector of the corresponding drug; and then fusing again based on the Hadamard product according to the fused embedding vector to generate a final fused drug pair representation;

[0050] The prediction unit is used to perform DDI prediction based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

[0051] The present invention also provides a device for predicting asymmetric drug interaction relationships, including a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a method for predicting asymmetric drug interaction relationships as described above.

[0052] The present invention also provides a computer-readable storage medium having computer-readable instructions stored thereon. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the method for predicting an asymmetric drug interaction relationship as described above is implemented.

[0053] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention utilizes a dual-channel encoder to effectively capture the directional information of asymmetric DDI, achieves accurate prediction of asymmetric DDI, and effectively solves the problem of missing directional information in the current DDI prediction model. The method model is simple and robust.

[0055] The present invention extracts the in-and-out neighborhood information between drug pairs through a multi-layer graph attention network, effectively capturing the directional interaction relationship between drug pairs. This method avoids the information loss caused by ignoring directionality in traditional methods, helps predict the order of drug administration in multi-drug combination therapy, thereby improving patient safety and treatment effects.

[0056] The present invention compresses the characteristics of the drug itself through an automatic encoder, modularizes the acquisition process of drug multi-scale embedding, realizes the dimensionality reduction of drug characteristics, and reduces the model complexity and system coupling. This structure of multi-scale embedding splicing and Hadamard product combination reduces the complexity of the encoder, makes the drug pair representation more compact, thereby reducing the amount of data transmitted by the feature and improving the computational efficiency of the model.

[0057] The decoder design of the present invention avoids a large number of feature interaction calculation processes by adding and fusing the topological information of the source drug and the target drug with their own features, further reducing the coupling of the model, thereby reducing the computational overhead and improving the prediction efficiency. When predicting the DDI task, the drug pair embedding is combined through the Hadamard product, which effectively reduces the model's dependence on external data and ensures efficient prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 A schematic diagram of a method for predicting asymmetric drug interaction relationships provided in Example 1.

[0060] Figure 2 The present invention is a flowchart of a method for predicting asymmetric drug interaction relationships provided in Example 1.

[0061] Figure 3 This is a schematic diagram of a device for predicting asymmetric drug interaction relationships provided in Example 2.

[0062] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0064] Embodiment 1

[0065] Embodiment 1 of the present invention provides a method for predicting an asymmetric drug interaction relationship, which can be implemented by a device for predicting an asymmetric drug interaction relationship (hereinafter referred to as a prediction device), and in particular, executed by one or more processors in the device.

[0066] In this embodiment, the construction device may be an electronic device equipped with a processor, the processor having a computer program of the prediction method of the asymmetric drug interaction relationship and the computer program can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc., which is not limited here.

[0067] In this embodiment, drug-drug interaction (DDI) refers to the effect change that may be caused by the interaction of two or more drugs in pharmacology, metabolism, etc. when used together. Drug interactions may lead to enhanced or weakened efficacy or side effects, so special attention should be paid to the application of drug combinations.

[0068] Asymmetric Drug-Drug Interaction: refers specifically to the directional asymmetry of the interaction relationship in drug combinations due to the different roles of different drugs in the interaction (for example, as a trigger or responder). This characteristic has an important impact on the prediction of the effect of drug combinations.

[0069] The existing technical solutions are insufficient in capturing the complex interactions and directional interactions between drugs, and are also weak in robustness in dealing with data imbalance and network noise. To address these shortcomings, the present invention proposes a DDI prediction model that not only improves the prediction accuracy but also effectively deals with directional interactions between drugs through multi-scale information fusion and directed graph feature extraction.

[0070] like Figure 1 As shown, a method for predicting asymmetric drug interaction relationships includes steps S1 to S5.

[0071] S1, obtain the directed DDI network and drug feature matrix of the source drug and target drug to be predicted.

[0072] In this embodiment, DDI network refers to drug-drug interaction network (DDI Network), which is a graph structure built based on the relationship between drugs, where nodes represent drugs and edges represent the interactions between drugs. This network is used to model the complex interactions between drugs and help understand the physiological or pharmacodynamic changes that may be caused by drug combinations.

[0073] When obtaining the directed DDI network of two drugs, DDI data can be obtained from public databases such as DrugBank, KEGG DRUG, etc. These databases provide rich information on drug interactions, including the type of interaction between drugs, the intensity of action, etc. A directed graph is constructed according to the direction of the interaction (such as the source drug affects the target drug or the target drug affects the source drug).

[0074] The drug feature matrix can be extracted from drug data using existing algorithms or tools, or obtained from public databases or literature.

[0075] S2, inputting the directed DDI network and the drug feature matrix into a dual-channel encoder; based on the directed DDI network, learning the multi-level entry and exit neighborhood topological relationship of the drug through the first encoder of the dual-channel encoder, and obtaining the drug topological vectors of the source drug and the target drug respectively; at the same time, using the second encoder of the dual-channel encoder, reducing the dimension of the drug feature matrix, generating a multi-scale embedding, and performing multi-scale fusion and splicing on the multi-scale embeddings of the source drug and the target drug respectively, to obtain the drug embedding vectors of the corresponding drugs.

[0076] like Figure 2 As shown, the first encoder is a GAT graph attention network; the directed graph topology of the DDI network is encoded through the graph attention network, and the topological information of the two drugs in the multi-layer outgoing neighbors and incoming neighbors is extracted respectively.

[0077] In this step, GAT is used to extract the relationship features of drugs in neighborhoods of different scales in order to more comprehensively capture the interaction patterns of drugs, specifically:

[0078] For the source drug, the topological representation of the out-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows:

[0079] ;

[0080] ;

[0081] in, Indicates the source drug Layer node features; A node representing the source drug in the directed graph; A node representing the target drug in the directed graph; A set of outbound neighbor nodes representing the source drug; Represents the current node in the source drug The corresponding attention coefficient; A weight matrix representing the source drug in the graph attention network; Indicates the target drug Layer node features, i.e., outbound neighbor node features of the source drug;

[0082] For splicing operation; is the weight matrix of the source drug attention coefficient; is a single-layer feedforward network of the source drug, σ is a nonlinear activation function, i.e., LeakyReLU function; for The neighborhood set of Neighborhood set The nodes in; T is the transpose; a drug embedding vector representing the source drug; a drug embedding vector representing the target drug;

[0083] For the target drug, the topological representation of the in-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows:

[0084] ;

[0085] ;

[0086] in, Indicates the target drug Layer node features; A set of inbound neighbor nodes representing the target drug; Indicates the current node in the target drug The corresponding attention coefficient; represents the weight of the target drug; Indicates the source drug Layer node features, i.e., neighbor node features of the target drug; for The neighborhood set of Neighborhood set Nodes in is a weight matrix of the target drug attention coefficient; is a single-layer feed-forward network of the target drug.

[0087] In this embodiment, the Graph Attention Network (GAT) is a neural network model for processing graph structure data. It uses the attention mechanism to assign different weights to each node in the graph, thereby capturing the complex relationship between nodes. For the directed graph topology of the DDI (drug-drug interaction) network, GAT can effectively extract topological information in multiple layers of outgoing and incoming neighbors.

[0088] The core idea of ​​GAT is to use the attention mechanism to calculate the correlation between nodes and assign a weight to each neighbor node. This weight reflects the importance of the neighbor node to the current node. In GAT, each node updates its feature representation based on the features of its neighbor nodes.

[0089] Of course, you can also use graph convolutional networks (GCN), graph isomorphism networks (GIN) or graph sampling networks (GraphSAGE) instead of GAT graph attention networks. These models also have the ability to learn topological relationships on directed graphs and can better capture the directed relationships between drugs.

[0090] In this embodiment, multi-scale embedding is an embedding method that learns features at different scales, and obtains richer node representations by capturing graph structure information at different levels. Multi-scale embedding helps to integrate local and global information of drugs, thereby improving the accuracy of interaction prediction.

[0091] Furthermore, the drug feature matrix includes a Morgan fingerprint feature matrix of the drug.

[0092] In this embodiment, the Drug Morgan Fingerprint is a fixed-length binary vector used to describe the characteristics of drug molecules. The molecules are hash-encoded through a ring substructure and are widely used in the characterization and similarity analysis of compounds.

[0093] Furthermore, the second encoder is an autoencoder.

[0094] In this embodiment, the autoencoder (AE) is a neural network model for unsupervised learning, which is mainly used for data dimensionality reduction, feature extraction and data reconstruction. It is usually composed of an encoder and a decoder and is used for low-dimensional representation learning of data. The Morgan fingerprint feature matrix of the drug is reduced layer by layer using the autoencoder AE, and data compression and reconstruction can be achieved through the structure of the encoder and decoder. The encoder compresses the input data (i.e., the Morgan fingerprint feature matrix of the drug) into a low-dimensional representation, also called "encoding" or "implicit features". This part of the network gradually reduces the dimension of the input, thereby extracting the key features of the data. The decoder reconstructs the encoded low-dimensional features back to the high-dimensional space of the original data.

[0095] Specifically, the Morgan fingerprint feature matrix of the two drugs was obtained using an autoencoder. Perform layer-by-layer dimensionality reduction to obtain key drug feature embeddings at different scales , the specific formula for dimensionality reduction is:

[0096] ;

[0097] in, Indicates the source drug Layer key feature embedding; Indicates the source drug Layer key feature embedding; σ is a nonlinear activation function; , They represent the dimensionality reduction coefficients respectively.

[0098] Indicates that two drugs are Feature embedding matrix after layer encoding dimensionality reduction.

[0099] Then the features after encoding dimension reduction are decoded and reconstructed to obtain the feature matrix of the two drugs after decoding and reconstruction. , the formula is:

[0100] ;

[0101] in, represents the drug embedding vector for the source drug The feature matrix after decoding and reconstruction; , represent the decoded reconstruction coefficients respectively.

[0102] In this embodiment, an autoencoder (AE) is used to perform multi-scale dimensionality reduction on the drug's own features. Of course, a variational autoencoder (VAE) or a graph neural network (such as GCN, GIN) can also be used to reduce the dimensionality of the drug features and generate high-quality embeddings.

[0103] Next, at the output layer of the encoder, the multi-scale embedding vectors of the intermediate layer of each channel (AE and GAT) are concatenated to form the final encoding of the drug pair. The specific formula for the multi-scale fusion concatenation is:

[0104] ;

[0105] ;

[0106] ;

[0107] in, A drug topology vector representing the concatenated source drug; Indicates the source drug Layer node features; Indicates the source drug Layer node features; A drug topology vector representing the target drug after splicing; Indicates the target drug -1 layer node features; Indicates the target drug Layer node features; A drug embedding vector representing the source drug after concatenation; Indicates the source drug Layer feature embedding.

[0108] S3, the drug topology vector and the drug embedding vector are input into the decoder for fusion to obtain a fused embedding vector of the corresponding drug; then, according to the fused embedding vector, fusion is performed again based on the Hadamard product to generate a final fused drug pair representation.

[0109] This step provides a more representative feature representation for drug interaction prediction by fusing drug features and graph topology information at different scales.

[0110] Furthermore, when the drug topology vector and the drug embedding vector are input into the decoder for fusion, direct addition fusion is adopted, or weighted splicing fusion can be used, or a self-attention mechanism is introduced after splicing to dynamically allocate the weights of each feature, which is not limited here.

[0111] The expression of direct addition fusion is:

[0112] = + ;

[0113] = + ;

[0114] in, a fused embedding vector representing the source drug; is the drug topology vector of the source drug; is a drug embedding vector of the source drug;

[0115] a fused embedding vector representing the target drug; is the drug topology vector of the target drug; is the drug embedding vector of the target drug.

[0116] Then, the fused embedding vectors of the two drugs are fused again based on the Hadamard product. Other methods of combination and fusion can also be used, such as combining the fused embedding vectors of the two drugs through a bidirectional LSTM to generate a unified representation of the drug pair, which is not limited here.

[0117] Specifically, a nonlinear transformation processing module such as a fully connected layer or a multi-layer perceptron (MLP) may be introduced into the decoder to further optimize the drug pair representation and enhance the model performance.

[0118] In this embodiment, a multi-layer perceptron (MLP) is a feed-forward neural network consisting of multiple fully connected layers, which is usually used for learning and predicting nonlinear features. MLP can be used to capture complex nonlinear relationships in drug feature encoding.

[0119] S4, performing DDI prediction based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

[0120] Specifically, the DDI prediction task is performed on the fusion drug pair to obtain a prediction result, which includes whether the two drugs have an interaction relationship, a symmetric interaction relationship, and an asymmetric interaction relationship.

[0121] Furthermore, the prediction formula is:

[0122] ;

[0123] in, Indicates the prediction result; represents the Hadamard product; represents a multi-layer perceptron; a fused embedding vector representing the source drug; is the fusion embedding vector of the target drug.

[0124] In another preferred embodiment, the drug interaction relationship prediction model is also trained. During training, the error between the true label and the predicted label is evaluated using a binary cross entropy loss function; the drug embedding vectors of the two drugs are input into the decoder for decoding and reconstruction, and the mean square error loss function is used to train the error expression between the drug feature matrix and the drug feature matrix as follows:

[0125] ;

[0126] ;

[0127] in, represents the binary cross entropy loss; represents mean square error loss; a drug pair interaction relationship label representing the source drug; a drug pair interaction relationship label representing the target drug; An edge representing the source drug and the target drug; Represents a library of drug pair interaction relationships; Indicates the prediction result; Represents the true label; N represents the number of samples;

[0128] represents the drug embedding vector for the source drug The feature matrix after decoding and reconstruction; a drug characteristic representing the source drug; A node representing the source drug in the directed graph; A drug feature matrix dimension representing the source drug; represents the L2 norm.

[0129] Similarly, the mean square error loss function is used to train the drug embedding vector of the target drug The error between the reconstructed feature matrix and the drug feature matrix is ​​decoded.

[0130] When the training evaluation results meet the preset requirements, a trained prediction model of the asymmetric drug interaction relationship is obtained, which is used for the subsequent prediction of the interaction relationship of asymmetric drug pairs.

[0131] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0132] The present invention utilizes a dual-channel encoder to effectively capture the directional information of asymmetric DDI, achieves accurate prediction of asymmetric DDI, and effectively solves the problem of missing directional information in the current DDI prediction model. The method model is simple and robust.

[0133] The present invention extracts the in-and-out neighborhood information between drug pairs through a multi-layer graph attention network, effectively capturing the directional interaction relationship between drug pairs. This method avoids the information loss caused by ignoring directionality in traditional methods, helps predict the order of drug administration in multi-drug combination therapy, thereby improving patient safety and treatment effects.

[0134] The present invention compresses the characteristics of the drug itself through an automatic encoder, modularizes the acquisition process of drug multi-scale embedding, realizes the dimensionality reduction of drug characteristics, and reduces the model complexity and system coupling. This structure of multi-scale embedding splicing and Hadamard product combination reduces the complexity of the encoder, makes the drug pair representation more compact, thereby reducing the amount of data transmitted by the feature and improving the computational efficiency of the model.

[0135] The decoder design of the present invention avoids a large number of feature interaction calculation processes by adding and fusing the topological information of the source drug and the target drug with their own features, further reducing the coupling of the model, thereby reducing the computational overhead and improving the prediction efficiency. When predicting the DDI task, the drug pair embedding is combined through the Hadamard product, which effectively reduces the model's dependence on external data and ensures efficient prediction performance.

[0136] Embodiment 2

[0137] like Figure 3 As shown, the second embodiment of the present invention further provides a prediction device for asymmetric drug interaction relationships, comprising:

[0138] An acquisition unit, used for acquiring a directed DDI network and a drug feature matrix of a source drug and a target drug to be predicted;

[0139] A dual-channel encoding unit is used to input the directed DDI network and the drug feature matrix into a dual-channel encoder; based on the directed DDI network, the multi-level entry and exit neighborhood topological relationship of the drug is learned through the first encoder of the dual-channel encoder to obtain the drug topological vectors of the source drug and the target drug respectively; at the same time, the second encoder of the dual-channel encoder is used to reduce the dimension of the drug feature matrix to generate a multi-scale embedding, and the multi-scale fusion and splicing of the multi-scale embeddings of the source drug and the target drug are respectively performed to obtain the drug embedding vector of the corresponding drug;

[0140] A fusion unit, used for fusing the drug topological vector and the drug embedding vector into a decoder to obtain a fused embedding vector of the corresponding drug; and then fusing again based on the Hadamard product according to the fused embedding vector to generate a final fused drug pair representation;

[0141] The prediction unit is used to perform DDI prediction based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

[0142] Embodiment 3

[0143] The third embodiment of the present invention further provides a prediction device for asymmetric drug interaction relationships, which includes a memory and a processor. The memory stores a computer program, and the computer program can be executed by the processor to implement the prediction method for asymmetric drug interaction relationships as described above.

[0144] Embodiment 4

[0145] The fourth embodiment of the present invention further provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the above-mentioned method for predicting asymmetric drug interaction relationships is implemented.

[0146] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0147] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0148] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0149] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0150] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0151] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0152] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting asymmetric drug interaction relationships, characterized in that: include: Obtain the directed DDI network and drug feature matrix of the source drug and target drug to be predicted; Inputting the directed DDI network and the drug feature matrix into a dual-channel encoder; Based on the directed DDI network, the multi-level in-and-out neighborhood topological relationship of the drug is learned through the first encoder of the dual-channel encoder, and the drug topological vectors of the source drug and the target drug are obtained respectively; at the same time, the second encoder of the dual-channel encoder is used to reduce the dimension of the drug feature matrix, generate multi-scale embedding, and perform multi-scale fusion splicing on the multi-scale embedding of the source drug and the target drug respectively, to obtain the drug embedding vector of the corresponding drug; wherein, the first encoder is a graph attention network; through the graph attention network, the topological information of the source drug and the target drug in the multi-layer out-neighborhood and in-neighborhood is extracted respectively, specifically: For the source drug, the topological representation of the out-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows: ； ； in, Indicates the source drug Layer node features; A node representing the source drug in the directed graph; A node representing the target drug in the directed graph; A set of outbound neighbor nodes representing the source drug; Represents the current node in the source drug The corresponding attention coefficient; A weight matrix representing the source drug in the graph attention network; Indicates the target drug Layer node features, i.e., outbound neighbor node features of the source drug; For splicing operation; is the weight matrix of the source drug attention coefficient; is a single-layer feed-forward network of the source drug, σ is a nonlinear activation function; for The neighborhood set of Neighborhood set The nodes in; T is the transpose; a drug embedding vector representing the source drug; a drug embedding vector representing the target drug; For the target drug, the topological representation of the in-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows: ； ； in, Indicates the target drug Layer node features; A set of inbound neighbor nodes representing the target drug; Indicates the current node in the target drug The corresponding attention coefficient; represents the weight of the target drug; Indicates the source drug Layer node features, i.e., neighbor node features of the target drug; for The neighborhood set of Neighborhood set Nodes in is a weight matrix of the target drug attention coefficient; A single-layer feed-forward network for the target drug; The drug topology vector and the drug embedding vector are input into a decoder for fusion to obtain a fused embedding vector of the corresponding drug; and then, according to the fused embedding vector, fusion is performed again based on the Hadamard product to generate a final fused drug pair representation; DDI prediction is performed based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

2. The method for predicting asymmetric drug interaction relationships according to claim 1, characterized in that , the drug feature matrix includes the Morgan fingerprint feature matrix of the drug; the second encoder is an automatic encoder; the Morgan fingerprint feature matrix is ​​reduced layer by layer using the automatic encoder to obtain the key feature embedding of drugs at different scales, and the specific formula for dimensionality reduction is: ； in, Indicates the source drug Layer feature embedding; Indicates the source drug Layer feature embedding; represents a nonlinear activation function; , They represent the dimensionality reduction coefficients respectively.

3. The method for predicting asymmetric drug interaction relationships according to claim 1, characterized in that ,The specific formula of the multi-scale fusion splicing is: ； ； ； in, A drug topology vector representing the concatenated source drug; Indicates the source drug Layer node features; Indicates the source drug Layer node features; A drug topology vector representing the target drug after splicing; Indicates the target drug Layer node features; Indicates the target drug Layer node features; A drug embedding vector representing the concatenated source drug; Indicates the source drug Layer feature embedding; Indicates the source drug Layer feature embedding; Represents a concatenation operation.

4. The method for predicting asymmetric drug interaction relationships according to claim 1, characterized in that When the drug topology vector and the drug embedding vector are input into the decoder for fusion, an additive fusion method is adopted, and the formula is: = + ； = + ； in, a fused embedding vector representing the source drug; is the drug topology vector of the source drug; is a drug embedding vector of the source drug; a fused embedding vector representing the target drug; is the drug topology vector of the target drug; is the drug embedding vector of the target drug.

5. The method for predicting asymmetric drug interaction relationships according to claim 1, characterized in that , also includes: introducing a nonlinear transformation processing module into the decoder to optimize the drug pair representation and enhance the model performance.

6. The method for predicting asymmetric drug interaction relationships according to claim 1, characterized in that , also includes: when training the model, the binary cross entropy loss function is used to evaluate the error between the true label and the predicted label; at the same time, the drug embedding vectors of the two drugs are input into the decoder for decoding and reconstruction, and the mean square error loss function is used to evaluate the error between them and the original drug features. The expression is as follows: ； ； ； ； in, represents the binary cross entropy loss; represents mean square error loss; a drug pair interaction relationship label representing the source drug; a drug pair interaction relationship label representing the target drug; An edge representing the source drug and the target drug; Represents a library of drug pair interaction relationships; Indicates the prediction result; Represents the true label; N represents the number of samples; a fused embedding vector representing the source drug; A fusion embedding vector for the target drug; represents the Hadamard product; represents a multi-layer perceptron; Represents the drug embedding vector for the source drug The feature matrix after decoding and reconstruction; , They represent the decoded reconstruction coefficients respectively; a drug characteristic representing the source drug; Non-linear activation function; A node representing the source drug in the directed graph; A drug feature matrix dimension representing the source drug; represents the L2 norm.

7. A prediction device for asymmetric drug interaction relationships, characterized in that: include: An acquisition unit, used for acquiring a directed DDI network and a drug feature matrix of a source drug and a target drug to be predicted; A dual-channel encoding unit is used to input the directed DDI network and the drug feature matrix into a dual-channel encoder; based on the directed DDI network, the multi-level entry and exit neighborhood topological relationship of the drug is learned through the first encoder of the dual-channel encoder to obtain the drug topological vectors of the source drug and the target drug respectively; at the same time, the second encoder of the dual-channel encoder is used to reduce the dimension of the drug feature matrix, generate a multi-scale embedding, and perform multi-scale fusion and splicing on the multi-scale embedding of the source drug and the target drug respectively to obtain the drug embedding vector of the corresponding drug; wherein the first encoder is a graph attention network; through the graph attention network, the topological information of the source drug and the target drug in the multi-layer exit and entry neighbors is extracted respectively, specifically: For the source drug, the topological representation of the out-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows: ； ； in, Indicates the source drug Layer node features; A node representing the source drug in the directed graph; A node representing the target drug in the directed graph; A set of outbound neighbor nodes representing the source drug; Represents the current node in the source drug The corresponding attention coefficient; A weight matrix representing the source drug in the graph attention network; Indicates the target drug Layer node features, i.e., outbound neighbor node features of the source drug; For splicing operation; is the weight matrix of the source drug attention coefficient; is a single-layer feed-forward network of the source drug, σ is a nonlinear activation function; for The neighborhood set of Neighborhood set The nodes in; T is the transpose; a drug embedding vector representing the source drug; a drug embedding vector representing the target drug; For the target drug, the topological representation of the in-degree neighbors in the directed graph is learned, and the learning propagation process in the graph attention network is as follows: ； ； in, Indicates the target drug Layer node features; A set of inbound neighbor nodes representing the target drug; Indicates the current node in the target drug The corresponding attention coefficient; represents the weight of the target drug; Indicates the source drug Layer node features, i.e., neighbor node features of the target drug; for The neighborhood set of Neighborhood set Nodes in is a weight matrix of the target drug attention coefficient; A single-layer feed-forward network for the target drug; A fusion unit, used for fusing the drug topological vector and the drug embedding vector into a decoder to obtain a fused embedding vector of the corresponding drug; and then fusing again based on the Hadamard product according to the fused embedding vector to generate a final fused drug pair representation; The prediction unit is used to perform DDI prediction based on the fusion drug pair representation to obtain a predicted drug pair interaction relationship.

8. A prediction device for asymmetric drug interaction relationships, characterized in that: It comprises a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a method for predicting an asymmetric drug interaction relationship as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, a method for predicting an asymmetric drug interaction relationship as described in any one of claims 1 to 6 is implemented.

Images