A method, device, equipment and storage medium for predicting drug interactions

Abstract

The embodiment of the present invention provides a method, device, equipment and storage medium for predicting drug interactions, which relates to the field of medical technology. Among them, the prediction method includes steps S1 to S9. S1 obtains the characteristics of multiple drugs and the reaction types between drugs. S2 constructs a drug feature matrix based on the characteristics of multiple drugs. S3 constructs a drug adjacency graph based on the reaction types between multiple drugs. S4 extracts drug feature information of different levels from the drug feature matrix through a deep neural network. S5 extracts neighbor topological structure information of different levels from the drug adjacency graph through a convolutional neural network. S6 fuses the feature vector and the structure vector to obtain the final embedded representation of each drug. S7 combines the final embedded representation of each drug to obtain multiple drug pair vectors. S8 uses the reaction type of the drug pair as a label and the drug pair vector as a training data set to train a classification model. S9 predicts the interaction type of unknown drug pairs through the classification model.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method, device, equipment and storage medium for predicting drug interactions. Background Art

[0002] When multiple drugs are taken at the same time, unexpected interactions (Drug-Drug Interaction, DDI) may occur between the drugs. DDI may have a beneficial or harmful effect on treatment. Among them, beneficial DDI can produce a synergistic effect of "1+1>2" and can become a safe and effective treatment for complex diseases. It has been successfully applied to the treatment of complex diseases such as cancer and AIDS. Harmful DDI may also cause serious adverse drug reactions, thereby threatening the patient's life safety. Therefore, identifying drug interactions is an urgent and practical task.

[0003] Currently, an important method for identifying DDI is to conduct long-term in vivo and in vitro clinical trials. However, this requires a lot of money and time. As the number of drugs continues to grow, existing methods are difficult to meet actual application needs.

[0004] In view of this, the applicant filed this application after studying the existing technology. Summary of the invention

[0005] The present invention provides a drug interaction prediction method, device, equipment and storage medium to improve the above technical problems.

[0006] First,

[0007] The embodiment of the present invention provides a method for predicting drug interactions, which comprises steps S1 to S9.

[0008] S1. Obtain the characteristics of multiple drugs and the types of interactions between drugs.

[0009] S2. Construct a drug feature matrix based on the characteristics of multiple drugs.

[0010] S3. Construct a drug adjacency graph based on the reaction types between multiple drugs, wherein the drug adjacency graph uses drugs as nodes and reaction types as edges.

[0011] S4. Extract feature vectors of different levels from the drug feature matrix through deep neural network.

[0012] S5. Extracting structural vectors of different levels from the drug adjacency graph through a convolutional neural network. The feature vectors extracted from each layer of the deep neural network are respectively transmitted to the corresponding layer of the convolutional neural network for fusion.

[0013] S6. The feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network are fused by weight through the attention mechanism to obtain the final embedded representation of each drug.

[0014] S7. Combine the final embedding representations of the drugs to obtain multiple drug pair vectors, wherein the reaction types of some drug pair vectors are known.

[0015] S8. Use the reaction types between drug pairs as labels and the drug pair vectors with known reaction types as training data sets to train a classification model.

[0016] S9. Predict the response type of a drug to a vector with unknown response type through a classification model.

[0017] Second,

[0018] An embodiment of the present invention provides a drug interaction prediction device, which comprises:

[0019] The raw data processing module obtains the characteristics of multiple drugs and the reaction types between drugs, and constructs a drug characteristic matrix based on the characteristics of multiple drugs, and constructs a drug adjacency graph based on the reaction types between multiple drugs. The drug adjacency graph uses drugs as nodes and reaction types as edges.

[0020] The drug feature information extraction module is used to extract feature vectors of different levels from the drug feature matrix through a deep neural network.

[0021] The neighbor topology structure information extraction module is used to extract structure vectors of different levels from the drug adjacency graph through a convolutional neural network. The feature vectors extracted by each layer of the deep neural network are transmitted to the corresponding layer of the convolutional neural network for fusion.

[0022] The fusion module is used to fuse the feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network by weight through the attention mechanism to obtain the final embedded representation of each drug.

[0023] The combination module is used to combine the final embedding representations of each drug to obtain multiple drug pair vectors, among which the reaction types of some drug pair vectors are known.

[0024] The model training module is used to train a classification model using the reaction type between drug pairs as labels and the drug pair vectors with known reaction types as training data sets.

[0025] The link prediction module is used to predict the response type of a drug pair vector of unknown response type through a classification model.

[0026] Thirdly,

[0027] An embodiment of the present invention provides a drug interaction prediction device, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the drug interaction prediction method as described in any paragraph of the first aspect.

[0028] Fourthly,

[0029] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the drug interaction prediction method as described in any paragraph of the first aspect.

[0030] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0031] By effectively and deeply fusing drug features with topological relationships, we can capture rich information and learn representative and robust drug vector representations, thereby ensuring good prediction performance in downstream tasks. By using the attention mechanism, different weights are assigned to different model components based on their importance to the prediction task, and a lower weight is assigned to model components that contribute less to the prediction, which can effectively solve the interference of noise on the model. This can achieve more efficient fusion, and the distribution of attention values ​​can also provide a certain degree of explainability for the experimental results. The predicted results are multi-category, and the prediction results are more detailed while ensuring higher prediction accuracy in the previous case. It not only indicates whether there is an interaction, but also indicates a specific pathological description, which helps us understand the potential mechanisms behind adverse drug reactions, so that we can take appropriate measures to avoid them.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are 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 creative work.

[0034] Figure 1It is a flowchart of the prediction method provided by the first embodiment of the present invention.

[0035] Figure 2 It is a logic block diagram of the prediction method provided by the first embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 creative work are within the scope of protection of the present invention.

[0037] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0038] It should be clear that the described embodiments are only 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 creative work are within the scope of protection of the present invention.

[0039] 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.

[0040] 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.

[0041] 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.

[0042] 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 in a 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.

[0043] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0044] Embodiment 1:

[0045] See also Figure 1 and Figure 2 The first embodiment of the present invention provides a drug interaction prediction method, which can be performed by a drug interaction prediction device. In particular, it can be performed by one or more processors in the drug interaction prediction device to implement steps S1 to S9.

[0046] S1. Obtain the characteristics of multiple drugs and the types of interactions between drugs.

[0047] Specifically, the characteristics of drugs and the types of reactions between drugs can be obtained by searching in databases such as known adverse drug reaction event reports, scientific literature, clinical trials, etc. through natural language processing technology NLP, or by conducting experiments by oneself. The present invention does not specifically limit how to obtain the original data.

[0048] The characteristics of the drug may include chemical structure, target, enzyme, channel, characterization, side effect, etc. The reaction types between drugs may be various known reaction types, and the present invention does not limit the specific types and quantities of the characteristics and reaction types.

[0049] It is understandable that the drug interaction prediction device can be an electronic device with computing capabilities, such as a portable notebook computer, a desktop computer, a server, a smart phone, or a tablet computer.

[0050] S2. Construct a drug feature matrix based on the characteristics of multiple drugs.

[0051] The feature matrix can represent the characteristics of various drugs in the same format, allowing the computer to perform subsequent processing efficiently.

[0052] Based on the above embodiment, in an optional embodiment of the present invention, step S2 specifically includes steps S21 to S23:

[0053] S21, obtaining features of multiple drugs, and selecting a predetermined number of features to construct a predetermined number of initial feature matrices. Specifically, the multiple features include chemical structure, target, enzyme and channel. Preferably, the predetermined number is three. The predetermined number of features are chemical structure, target, enzyme, respectively.

[0054] S22, respectively calculating the similarities of a predetermined number of initial feature matrices to obtain a predetermined number of similarity matrices. Specifically, the similarity is Jaccard similarity.

[0055] S23, concatenating a predetermined number of similarity matrices to obtain drug feature matrices of multiple drugs.

[0056] S3. Construct a drug adjacency graph based on the reaction types between multiple drugs, wherein the drug adjacency graph uses drugs as nodes and reaction types as edges.

[0057] The following is a detailed description of step S2 and step S3 using a specific example.

[0058] The data provided by the paper DDIMDL were selected, including 572 drugs and four characteristics, including chemical substructures, targets, enzymes, pathways, 65 reaction types and 37,264 drug interaction events.

[0059] For the processing of neighbor relationships, we take the drugs involved in the 37,264 events as nodes and the reaction types between drugs as edges to construct the network graph G.

[0060] For the processing of drug features, we selected the three features of chemical substructure, target, and enzyme with the best experimental results in the DDIMDL model, and concatenated the three feature matrices obtained after similarity calculation as the final feature X.

[0061] The specific feature similarity calculation process is introduced below:

[0062] Taking the characteristic chemical structure as an example, the chemical structure characteristic values ​​of 572 drugs are traversed, and the chemical substructures obtained after each drug is split are appended to the set, and finally 583 non-repeating chemical substructure types can be obtained. Therefore, each drug can be represented by 583 bits of binary, "1" means that the drug has the corresponding chemical substructure feature, and "0" means that it does not, and finally a binary chemical structure matrix with a shape of (572, 583) can be obtained.

[0063] The same treatment is performed on the characteristic targets and enzymes. However, since the dimension of the characteristic matrix is ​​too high and there are many zero values, we calculate the Jaccard similarity using formula (1) and use the obtained similarity matrix to replace the previous binary matrix to alleviate the sparsity problem caused by too many zero values.

[0064]

[0065] After similarity calculation, we can get a chemical structure similarity matrix of size (572, 572) to represent X s ∈R Nd ×k , the target similarity matrix X is obtained after the same feature target and enzyme are processed. t ∈R Nd×k , the enzyme similarity matrix represents X e ∈R Nd×k , the final drug feature matrix X is represented by the concatenation of three feature matrices, where the feature representation of drug di is shown in formula (2). Where ⊕ represents the concatenate connection method.

[0066]

[0067] With the data foundation obtained in the first two steps, the required features need to be extracted from it. In the present invention, the characteristic information of the drug itself is extracted from the drug feature matrix through a deep neural network (DNN). The structural topological information between drugs is extracted from the drug adjacency graph through a convolutional neural network (GCN).

[0068] The deep neural network includes multiple self-compiler models AE connected layer by layer for feature extraction. The convolutional neural network includes multiple graph convolution models connected layer by layer for extracting topological structures. The multiple self-compiler models AE correspond to the multiple graph convolution models GCN one by one. The specific steps of extracting features are described in S4 and S5.

[0069] S4. Extract feature vectors of different levels from the drug feature matrix through deep neural network.

[0070] Specifically, the DNN layer uses autoencoders (AE) to learn potential feature representations from raw data. It is an unsupervised learning method that learns from the data itself and uses the minimum reconstruction loss for training. AE can not only quickly capture the nonlinear changes between input and output, but also effectively extract feature vectors with hierarchical structures through multi-layer neural networks.

[0071] like Figure 2As shown, based on the above embodiment, in an optional embodiment of the present invention, step S4 is specifically used to: extract node attribute information layer by layer from the drug feature matrix through multiple self-compiler models AE of the deep neural network until the drug feature information is extracted in the last layer. Among them, the node attribute information is used to transmit to the next layer of self-compiler model AE and the corresponding graph convolution model.

[0072] Preferably, AE is also an ideal method for dimensionality reduction and feature fusion to solve the problem of fusion of three types of heterogeneous feature similarity matrices and the problem of too high dimension after fusion. It mainly consists of two parts: encoder and decoder. The encoder uses GCN to achieve dimensionality reduction, and we regard the vector obtained after dimensionality reduction as the potential feature of the drug.

[0073] Assume that AE has l layers, use H (l) Represents the embedding learned at layer l, then the expression of the self-compiler model AE is shown in formula (3):

[0074]

[0075] Among them, H (l) is the potential feature information extracted from the lth layer, is the Sigmoid activation function, and are the weight matrix and bias matrix of the lth layer, H (l-1) is the potential feature information extracted from the l-1th layer, H (0) Represents the drug feature matrix X.

[0076] Specifically, the number of layers of the self-compiler model AE of the deep neural network and the number of nodes in each layer can be set according to actual needs, and the present invention does not make specific limitations on this.

[0077] The following is a detailed description of step S4 and step S5 based on the specific example described in step S3.

[0078] In this embodiment, the deep neural network setting includes a five-layer self-compiler model AE, and the number of nodes of the five-layer self-compiler model AE is 500, 500, 2000, 256 and 65 respectively. The input feature dimension is 1716, which is obtained by connecting three features with a dimension of 572. The number of nodes in the last layer is fixed to 65, because the final prediction result has 65 types of DDI events, and the vector representation dimension of the drug pair must be consistent with the number of label types. Then gradually increase the number of layers to explore the impact of the number of layers on the experimental results.

[0079] The learning rate of the model hyperparameters is set to 0.003, and the epoch is set to 200, because the loss curve basically does not change when the epoch is 200. Then the Adam optimizer is selected and a five-fold cross validation is performed.

[0080] S5. Extracting structural vectors of different levels from the drug adjacency graph through a convolutional neural network. The vectors extracted from each layer of the deep neural network are respectively transmitted to the corresponding layer of the convolutional neural network for fusion.

[0081] Specifically, AE can learn important information of each layer from its own feature data, such as H (1) , H (2) ...H (l) , but it ignores the structural relationship between samples. Graph convolutional neural network (GCN) has achieved good results in aggregating neighbor information. Influenced by this, we also use the GCN model to extract the topological structure information of drugs, and then fuse it with the transferred drug feature representation.

[0082] The intra-layer fusion is specifically as follows: the current graph convolution model fuses the structure vector extracted by the previous graph convolution model and the feature vector learned by the self-compiler model AE corresponding to the previous graph convolution model; wherein the feature vector learned by the self-compiler model AE corresponding to the previous graph convolution model is transmitted to the current graph convolution model through the transmission layer.

[0083] like Figure 2 As shown, based on the above embodiment, in an optional embodiment of the present invention, step S5 is specifically used to: extract structure vectors layer by layer from the drug adjacency graph through multiple graph convolution models of the convolutional neural network until the final structure vector is extracted at the last layer. Among them, the current graph convolution model is used to fuse the structure vector extracted by the graph convolution model of the previous layer and the received feature vector, and convolve the fused vector to obtain the structure vector of the current layer.

[0084] Assume that the graph convolution model has l layers, and use Z (l) Represents the potential vector learned in the lth layer, then the expression of the graph convolution model is:

[0085]

[0086] in,

[0087]

[0088] Among them, Z (l) is the neighbor topology information of layer l, is the Sigmoid activation function, is the degree matrix, A～ is the self-loop matrix, W (l-1) is the weight transformation matrix of the l-1th layer, ε is the balance factor, Z (l-1) is the neighbor topology information of the l-1th layer, H (l-1) is the potential feature information extracted from the l-1th layer.

[0089] Specifically, the potential vector representation of the lth layer of the GCN module can be obtained by convolution using formula (4), where is a self-loop matrix, That is, add a unit matrix I on the basis of the adjacency matrix A, Z is the characteristic matrix, and the self-loop matrix multiplied by the characteristic matrix will obtain the sum of all the characteristics of the neighboring nodes and itself. is the degree matrix, W is the weight transformation matrix.

[0090] The information after fusing neighbors in the (l-1)th layer is transferred to the lth layer as output.

[0091]

[0092] Considering that the autoencoder AE can obtain different information from the features, we combine the two parts through formula (5) to obtain a more expressive drug vector representation Z.

[0093] Among them, Z (0) represents the drug adjacency graph, H (0) represents the drug feature matrix.

[0094]

[0095] ε is a balancing factor, which is used to control the fusion ratio of the vectors learned by the GCN model and the AE model. The embedding representation of the l-1 layer is obtained by fusion according to the ratio. Then input it into the GCN of the lth layer to get the representation vector Z (l) , as shown in formula (6).

[0096]

[0097] like Figure 2 As shown in the figure, it should be noted that between the layers of GCN and DNN, the node attribute information learned by AE and the potential vector representation learned by the GCN module are fused according to the proportion of the balance factor through the convey operation. In this way, the representation obtained will be able to adapt to two different types of information, namely the data itself and the relationship between the data. After layer-by-layer fusion, the information of node features and topological relationships at different levels can be retained.

[0098] Specifically, through the convey operation in the middle, the upper and lower modules can promote each other. 1. GCN can provide second-order constraints for AE and better learn the abstract representation of features. 2. AE extracts information of different orders and transmits it to GCN, which can alleviate the over-smoothing problem of GCN when performing multi-layer fusion.

[0099] S6. The feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network are fused by weight through the attention mechanism to obtain the final embedded representation of each drug.

[0100] Specifically, GCN directly takes features and adjacency matrix as input. During the aggregation process, GCN may be interfered by certain noise neighbors, resulting in the model not being fully effective. The attention mechanism can assign corresponding weights according to the importance of model components to the prediction results, and assign lower weights to model components with small prediction contributions, thereby effectively solving the problem of noise interference on the model.

[0101] like Figure 2 As shown, based on the above embodiment, in an optional embodiment of the present invention, step S6 is specifically used to adopt an attention mechanism to fuse feature information and neighbor topology information to obtain the final embedding representation of each drug. The attention mechanism model is:

[0102]

[0103] Among them, E i is the final embedding representation of the i-th drug, Z is the structure vector extracted by the last layer of graph convolution model GCN, is the attention value corresponding to the structure vector of the i-th drug, H is the feature vector extracted by the last layer of the self-compiler model AE, is the attention value corresponding to the feature vector of the i-th drug.

[0104] The specific implementation process of the attention mechanism is as follows:

[0105] For example, in the lth layer before the output layer, the vector representation Z after GCN fusion has been obtained (l) The vector representation H obtained by the AE module (l) , the calculation of the attention mechanism is shown in formula (9):

[0106] (α z ,α h )=att(Z,H) (9)

[0107] α z ,α h They represent the attention coefficients corresponding to the embeddings obtained by the GCN module and the AE module respectively.

[0108] Taking node i as an example, its embedding in Z is represented by Z i We first transform the embedding through a nonlinear transformation and use a shared attention vector q to obtain the attention value of the topological structure embedding As shown in formula (10):

[0109]

[0110] Among them, w is the weight matrix, b is the bias vector,

[0111] Similarly, we can get the attention value of our own feature embedding

[0112] Then, the Softmax regularized attention value is used to obtain the final weight coefficient of the topological structure embedding, as shown in formula (11):

[0113]

[0114] Similarly, through the formula We can get the attention coefficient of our own feature embedding. Specifically,

[0115]

[0116] in,

[0117]

[0118] For all nodes n, the weight coefficient α is learned z =[α z ], α h =[α h ], which is correspondingly expressed as α z =diag(α z ), α h =diag(α h ), and then combine the two embeddings to obtain its final drug embedding representation vector E, as shown in formula (12):

[0119] E=α Z ·Z+α h ·H (12)

[0120] It can be understood that the implementation process of the above-mentioned attention mechanism is only a brief description. How to obtain the attention coefficient belongs to the prior art and will not be repeated in the present invention.

[0121] It should be noted that the embodiment of the present invention innovatively proposes deep fusion of the structural features and neighbor features of drugs, so that the final embedded representation of the drug can contain more features, so as to more accurately predict the reaction type of the drug pair, which has great practical significance. Among them, deep fusion includes intra-layer fusion, inter-layer fusion and final fusion. Among them, intra-layer fusion means that the feature vectors extracted from each layer of the deep neural network are respectively transmitted to the corresponding layers of the convolutional neural network for intra-layer fusion. Inter-layer fusion means that the structural vector obtained after the intra-layer fusion of the previous layer of the graph convolution model GCN is used as the potential feature input of the next layer of the graph convolution model GCN, so as to realize the transmission of the drug's own characteristics and topological information between layers. The final fusion means that the feature vector extracted from the last layer of the deep neural network and the structural vector extracted from the last layer of the convolutional neural network are fused through the attention mechanism.

[0122] S7. The response types of some drugs to vectors are known.

[0123] Specifically, the embedding representation E of all drugs has been obtained through the previous deep fusion. Assume that the vector representations of two drugs i and j are Φ(di) and Φ(dj) respectively. In this embodiment, four methods are combined to form a drug pair vector representation, as shown in the following table:

[0124] Table 1. Four different drug pair combinations

[0125]

[0126]

[0127] The symbol ⊙ in the table represents the Hadamard product, and the symbol Indicates Concatenate connection.

[0128] It is understandable that in other embodiments, other combinations may be used without being limited to the above four combinations.

[0129] S8. Use the reaction types between drug pairs as labels and the drug pair vectors with known reaction types as training data sets to train a classification model.

[0130] S9. Predicting the reaction type of a drug of unknown reaction type to a vector by using the classification model.

[0131] Based on the above embodiment, in an optional embodiment of the present invention, S8 specifically includes steps S81 to S83.

[0132] S81. Obtain the interaction type of the drug pair.

[0133] S82. Obtain a drug pair vector of known reaction type.

[0134] S83. Using the reaction types between drug pairs as labels and drug pair vectors with known reaction types as training data sets, input them into a cross entropy loss function, and perform supervised training to obtain the classification model.

[0135] Specifically, the drug pair vector representation is obtained by combination, and the drug pair reaction type is used as a label. Because this patent studies multi-category DDI prediction, the cross entropy loss function is used for training. Assuming that the training set is L, the number of categories is C, and a drug pair i∈L, the true label is surface in represents the probability that drug pair i belongs to category c, and the category prediction of n drug pairs is expressed as Y=[y i c ]∈R Nd×k Then the cross entropy loss calculation for all nodes in the training set is shown in formula (13):

[0136]

[0137] The embodiment of the present invention can capture rich information and learn representative and robust drug vector representations by effectively and deeply fusing drug features and topological relationships, thereby ensuring good prediction performance in downstream tasks. The attention mechanism is adopted to assign different weights to different model components based on their importance to the prediction task, and assign a lower proportion to model components with small prediction contributions, which can effectively solve the interference of noise on the model. This can achieve more efficient fusion, and the distribution of attention values ​​can also provide a certain degree of explainability for the experimental results. The predicted results are multi-category, and the prediction results are more detailed while ensuring higher prediction accuracy in the previous case. It not only indicates whether there is an interaction, but also indicates a specific pathological description, which is helpful to help us understand the potential mechanism behind adverse drug reactions, so as to take corresponding measures to avoid them.

[0138] The present invention proposes to combine drug features with topological relationships and obtain an abstract and advanced drug representation through a deep fusion strategy. In addition, unlike the traditional DDI method in which the prediction results are binary types, the results predicted by the present invention are multi-category, each category corresponds to a specific pharmacokinetic description, which helps users understand the mechanism behind DDI and provides medical insights.

[0139] This paper proposes a drug interaction prediction method, which aggregates features and neighbor topology information through convey operations, and performs layer-by-layer fusion, and finally obtains an informative drug embedding representation through end-to-end training. When link prediction tasks and classification tasks are performed on multiple data sets, the proposed model continues to show optimal performance, proving that it has good robustness.

[0140] Embodiment 2:

[0141] A second embodiment of the present invention provides a drug interaction prediction device, which comprises:

[0142] The raw data processing module obtains the characteristics of multiple drugs and the reaction types between drugs, and constructs a drug characteristic matrix based on the characteristics of multiple drugs, and constructs a drug adjacency graph based on the reaction types between multiple drugs. The drug adjacency graph uses drugs as nodes and reaction types as edges.

[0143] The drug feature information extraction module is used to extract feature vectors of different levels from the drug feature matrix through a deep neural network.

[0144] The neighbor topology structure information extraction module is used to extract structure vectors of different levels from the drug adjacency graph through a convolutional neural network. The feature vectors extracted by each layer of the deep neural network are transmitted to the corresponding layer of the convolutional neural network for fusion.

[0145] The fusion module is used to fuse the feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network by weight through the attention mechanism to obtain the final embedded representation of each drug.

[0146] The combination module is used to combine the final embedding representations of each drug to obtain multiple drug pair vectors, among which the reaction types of some drug pair vectors are known.

[0147] The model training module is used to train a classification model using the reaction type between drug pairs as labels and the drug pair vectors with known reaction types as training data sets.

[0148] The link prediction module is used to predict the response type of a drug pair vector of unknown response type through a classification model.

[0149] By effectively and deeply fusing drug features with topological relationships, we can capture rich information and learn representative and robust drug vector representations, thereby ensuring good prediction performance in downstream tasks. By using the attention mechanism, different weights are assigned to different model components based on their importance to the prediction task, and a lower weight is assigned to model components that contribute less to the prediction, which can effectively solve the interference of noise on the model. This can achieve more efficient fusion, and the distribution of attention values ​​can also provide a certain degree of explainability for the experimental results. The predicted results are multi-category, and the prediction results are more detailed while ensuring higher prediction accuracy in the previous case. It not only indicates whether there is an interaction, but also indicates a specific pathological description, which helps us understand the potential mechanisms behind adverse drug reactions, so that we can take appropriate measures to avoid them.

[0150] Based on the above embodiment, in an optional embodiment of the present invention, the drug feature extraction module specifically includes:

[0151] The initial feature unit is used to obtain the features of multiple drugs and select a predetermined number of features to construct a predetermined number of initial feature matrices.

[0152] The similarity unit is used to calculate the similarities of a predetermined number of initial feature matrices respectively to obtain a predetermined number of similarity matrices.

[0153] The splicing unit is used to splice a predetermined number of similarity matrices to obtain drug feature matrices of multiple drugs.

[0154] Based on the above embodiments, in an optional embodiment of the present invention, the characteristics of the drug include chemical structure, target, enzyme and channel, and the predetermined number is three.

[0155] Based on the above embodiment, in an optional embodiment of the present invention, the similarity is Jaccard similarity.

[0156] Based on the above embodiments, in an optional embodiment of the present invention, the deep neural network includes multiple self-compiler models AE for feature extraction; the convolutional neural network includes multiple graph convolution models for extracting topological structures; wherein the multiple self-compiler models AE and the multiple graph convolution models correspond one to one.

[0157] On the basis of the above embodiment, in an optional embodiment of the present invention, the intra-layer fusion is specifically: the current graph convolution model fuses the structure vector extracted by the previous graph convolution model and the feature vector learned by the self-compiler model AE corresponding to the previous graph convolution model; wherein the feature vector learned by the self-compiler model AE corresponding to the previous graph convolution model is transmitted to the current graph convolution model through the transmission layer. .

[0158] On the basis of the above embodiment, in an optional embodiment of the present invention, the self-compiler model AE is:

[0159]

[0160] Among them, H (l) is the potential feature information extracted from the lth layer, is the Sigmoid activation function, and are the weight matrix and bias matrix of the lth layer, H (l-1) is the potential feature information extracted from the l-1th layer, H (0) Represents the drug feature matrix X.

[0161] Based on the above embodiment, in an optional embodiment of the present invention, the structure vector extracted by the l-th layer graph convolution model is expressed as:

[0162]

[0163] in,

[0164]

[0165] in, is the Sigmoid activation function, is the degree matrix, is the self-loop matrix, W (l-1) is the weight transformation matrix of the l-1th layer, ε balance factor, Z (l-1) is the neighbor topology information of the l-1th layer, H (l-1) is the potential feature information extracted from the l-1th layer.

[0166] Based on the above embodiment, in an optional embodiment of the present invention, the fusion module is specifically used for:

[0167] The attention mechanism is used to fuse feature information and neighbor topology information to obtain the final embedding representation of each drug.

[0168] The attention mechanism model is:

[0169]

[0170] Among them, E i is the final embedding representation of the i-th drug, Z is the structure vector extracted by the last layer of graph convolution model GCN, is the attention coefficient corresponding to the structure vector of the i-th drug, H is the feature vector extracted by the last layer of the self-compiler model AE, is the attention coefficient corresponding to the feature vector of the i-th drug.

[0171] Based on the above embodiment, in an optional embodiment of the present invention, the combination of the final embedded representation includes Average, Hadamard, L1 norm and Concatenate.

[0172] Based on the above embodiment, in an optional embodiment of the present invention, the model training module specifically includes:

[0173] A first acquisition unit is used to acquire the interaction type of the drug pair;

[0174] A second acquisition unit is used to acquire a drug pair vector of known reaction type;

[0175] The training unit is used to use the reaction type between the drug pairs as labels and the drug pair vectors of known reaction types as training data sets, input them into the cross entropy loss function, and perform supervised training to obtain the classification model.

[0176] Embodiment three:

[0177] The third embodiment of the present invention provides a drug interaction prediction device, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the drug interaction prediction method as described in any paragraph of the first embodiment.

[0178] Embodiment 4:

[0179] A fourth embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the drug interaction prediction method as described in any paragraph of the first embodiment.

[0180] 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 and block diagrams 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 the flowchart, and the combination of boxes in the block diagram and / or the flowchart, 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.

[0181] 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.

[0182] 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 such an 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, including 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.

[0183] 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 drug interactions, characterized in that: Include: Obtain characteristics of multiple drugs and interaction patterns between drugs; constructing a drug feature matrix according to the features of the multiple drugs; Constructing a drug adjacency graph according to the reaction types between the multiple drugs; wherein the drug adjacency graph uses drugs as nodes and reaction types as edges; Extracting feature vectors of different levels from the drug feature matrix through a deep neural network; Extracting structural vectors of different levels from the drug adjacency graph through a convolutional neural network; wherein the feature vectors extracted from each layer of the deep neural network are respectively transmitted to the corresponding layer of the convolutional neural network for fusion; The feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network are fused by weight through the attention mechanism to obtain the final embedding representation of each drug. Combining the final embedded representations of the respective drugs to obtain a plurality of drug pair vectors; wherein the reaction types of some of the drug pair vectors are known; The reaction type between drug pairs is used as a label, and the drug pair vectors with known reaction types are used as a training data set to train a classification model; Predicting the reaction type of a drug of unknown reaction type to a vector by using the classification model; According to the characteristics of the multiple drugs, a drug characteristic matrix is ​​constructed, which specifically includes: Acquiring features of the plurality of drugs, and selecting a predetermined number of features to construct a predetermined number of initial feature matrices; Calculating the similarities of the predetermined number of initial feature matrices respectively to obtain a predetermined number of similarity matrices; splicing the predetermined number of similarity matrices to obtain the drug feature matrices of the multiple drugs; The deep neural network includes multiple self-compiler models AE for feature extraction; the convolutional neural network includes multiple graph convolution models GCN for extracting topological structures; wherein the multiple self-compiler models AE and the multiple graph convolution models GCN correspond one to one.

2. The method for predicting drug interactions according to claim 1, characterized in that: The features include chemical structure, target, enzyme and channel, and the predetermined number is three; The similarity is Jaccard similarity.

3. The method for predicting drug interactions according to claim 1, characterized in that: The self-compiler model AE is: in, For the The potential feature information extracted by the layer, is the Sigmoid activation function, and They are The weight and bias matrices of the layers, For the -1 layer of potential feature information extracted, represents the drug feature matrix X; No. The structure vector extracted by the layer graph convolution model is expressed as: in, in, is the Sigmoid activation function, is the degree matrix, is a self-loop matrix, For the The weight transformation matrix of the layer, Balance factor, For the The neighbor topology information of the layer, For the -1 layer of potential feature information extracted, Represents the drug adjacency graph.

4. The method for predicting drug interactions according to claim 1, characterized in that: The feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network are fused by weight through the attention mechanism to obtain the final embedded representation of each drug. The attention mechanism is: in, For the The final embedding representation of the drugs, It is the structure vector extracted by the last layer of graph convolution model GCN. For the The attention coefficient corresponding to the structure vector of the drug, is the feature vector extracted by the last layer of the self-compiler model AE, For the The attention coefficient corresponding to the feature vector of each drug.

5. The method for predicting drug interactions according to claim 1, characterized in that: The combination of the final embedding representation includes Average, Hadamard, L1 norm and Concatenate; The reaction type between drug pairs is used as a label, and the drug pair vectors with known reaction types are used as training data sets to train a classification model, which specifically includes: Get the interaction type of drug pairs; Get a vector of drug pairs with known reaction types; The reaction types between drug pairs are used as labels, and drug pair vectors with known reaction types are used as training data sets, which are input into a cross entropy loss function for supervised training to obtain the classification model.

6. A drug interaction prediction device, characterized in that: Include: A raw data processing module, which obtains the characteristics of multiple drugs and the reaction types between the drugs, and constructs a drug characteristic matrix according to the characteristics of the multiple drugs, and constructs a drug adjacency graph according to the reaction types between the multiple drugs; wherein the drug adjacency graph uses drugs as nodes and reaction types as edges; A drug feature information extraction module, used to extract feature vectors of different levels from the drug feature matrix through a deep neural network; A neighbor topology structure information extraction module, used to extract structure vectors of different levels from the drug adjacency graph through a convolutional neural network; wherein the feature vectors extracted from each layer of the deep neural network are respectively transmitted to the corresponding layer of the convolutional neural network for fusion; The fusion module is used to fuse the feature vector extracted from the last layer of the deep neural network and the structure vector extracted from the last layer of the convolutional neural network by weight through the attention mechanism to obtain the final embedded representation of each drug; A combination module, used for combining the final embedded representations of the respective drugs to obtain a plurality of drug pair vectors; wherein the reaction types of some drug pair vectors are known; A model training module is used to train a classification model using the reaction type between drug pairs as labels and the drug pair vectors with known reaction types as training data sets; A link prediction module, used for predicting the reaction type of a drug of unknown reaction type to a vector by using the classification model; The raw data processing module specifically includes: An initial feature unit, used to obtain the features of the plurality of drugs, and select a predetermined number of features to construct a predetermined number of initial feature matrices; A similarity unit, used to respectively calculate the similarities of the predetermined number of initial feature matrices to obtain a predetermined number of similarity matrices; A splicing unit, used for splicing the predetermined number of similarity matrices to obtain the drug feature matrices of the multiple drugs; The deep neural network includes multiple self-compiler models AE for feature extraction; the convolutional neural network includes multiple graph convolution models for extracting topological structures; wherein the multiple self-compiler models AE correspond to the multiple graph convolution models one by one.

7. A drug interaction prediction device, characterized in that: It comprises a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement the drug interaction prediction method as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the drug interaction prediction method according to any one of claims 1 to 5.

Images