A method for predicting drug interaction types by integrating multi-source features across graph structures

Abstract

The present invention relates to a method for predicting drug interaction types by fusing multi-source feature cross-graph structures, and belongs to the field of deep learning technology. The method comprises the following steps: obtaining drug data, obtaining drug data sets, and preprocessing; the drug data is subjected to Jaccard similarity calculation to obtain multidimensional data representations of the drug's chemical structure, biological target, and enzyme action, thereby obtaining the initial feature representation of the drug; the initial feature representation of the drug is updated through a heterogeneous graph to obtain a drug representation of a fused topological structure; the drug representation of the fused topological structure fuses drug features of different sources through an attention mechanism to obtain a final drug pair representation; the drug representation of the fused topological structure corresponding to the two drugs in the drug pair passes through a feature interleaving layer to obtain a final comprehensive representation; the final drug pair representation and the final comprehensive representation pass through a prediction module to obtain a drug interaction prediction probability vector. The present invention optimizes the accuracy and efficiency of drug interaction prediction.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a method for predicting drug interaction types by integrating multi-source features across graph structures. Background Art

[0002] Accurate prediction of drug interactions is critical to ensuring the safety of patient treatment. Existing prediction methods face several major challenges: Insufficient integration of topological information. Most methods fail to effectively integrate and utilize topological information between drugs. This information is critical because it can help reveal how drugs interact in polypharmacy. Ignoring this information limits the model's ability to reveal potential interaction mechanisms between drugs. Insufficient integration of biochemical properties. Existing models perform poorly in integrating key biochemical properties of drugs, such as chemical structure, target, and enzyme activity. These properties are the basis for understanding the mechanism of drug action, but are often not effectively incorporated into the prediction process by existing models, resulting in the inability to fully capture the complex interactions between drugs. Summary of the invention

[0003] In order to improve the prediction accuracy of drug interaction, the present invention provides a drug interaction type prediction method integrating multi-source features across graph structures.

[0004] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0005] In a first aspect, the present invention provides a method for predicting drug interaction types by integrating multi-source features across graph structures, comprising the following steps:

[0006] S1. Obtain drug data, obtain drug data set, and perform preprocessing;

[0007] S2. The drug data is subjected to Jaccard similarity calculation to obtain a multidimensional data representation of the drug's chemical structure, biological target, and enzyme action, and the multidimensional data representation is spliced ​​to obtain an initial feature representation of the drug;

[0008] S3. constructing a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure;

[0009] S4. The drug representation of the fused topological structure is fused with drug features from different sources through an attention mechanism to obtain a final drug pair representation;

[0010] S5. construct a feature interleaving layer, wherein the drug representations of the fusion topological structure corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain a final comprehensive representation;

[0011] S6. The final drug pair representation and the final comprehensive representation are subjected to a prediction module to obtain a drug interaction prediction probability vector.

[0012] Furthermore, step S1 is specifically as follows:

[0013] The drug data comes from the DrugBank database, which includes drugs, drug pairs, and interaction event types; each drug has four types of features, including: chemical substructure, target, pathway, and enzyme;

[0014] The drug dataset is preprocessed. According to the drug pairs and drug interaction event types, the names of the two drugs in the drug pairs are extracted respectively, a unique index is assigned to each drug, and the drug names and corresponding positions are stored in the dictionary.

[0015] Furthermore, step S2 is specifically as follows:

[0016] S21. Feature extraction: Convert the three biological features of the drug, namely, chemical substructure, target and enzyme, into binary feature vectors, with each biological feature corresponding to a binary feature vector;

[0017] S22. Feature representation: Based on the binary feature vector corresponding to the biological feature of the drug obtained in step S21, the Jaccard similarity metric is used to calculate the similarity between the three biological features of chemical substructure, enzyme and target between the drugs. Each drug is given a chemical substructure similarity feature representation. , an enzyme similarity feature representation and a target similarity feature representation , and then the , as well as After splicing, the initial characteristic representation of each drug is obtained, and the calculation formula is as follows:

[0018] ,

[0019] in, Indicates drug The initial feature representation of .

[0020] Furthermore, step S3 is specifically as follows:

[0021] S31. Constructing a heterogeneous graph ,node represents a collection of drugs, , Indicates the number of drugs, Represents a set of drug interaction event types: using the adjacency matrix Indicates the connection relationship between nodes. Indicates the interaction relationship If there is an interaction between drug i and drug j, If yes, it is 1, otherwise it is 0;

[0022] S32. Update drug feature representation: Update drug Initial drug characterization With the adjacency matrix Multiply, get medicine The drug representation of the fusion topology is calculated as follows:

[0023] ,

[0024] in, Representation Node The degree, Indicates drug With drugs There is interaction , It's medicine The initial drug characteristics calculated in step S2, Indicates the relationship type The weight matrix of is the bias, is the activation function , Indicates drug Drug representation of fused topology.

[0025] Furthermore, step S4 is specifically as follows:

[0026] Get the drug by searching the index The two drugs are paired to obtain a drug pair, and the drug representation of the fusion topological structure of the two drugs in the drug pair is recorded as and ,Will Input scaled dot product self-attention to get optimized drugs Characteristic expression of The specific calculation formula is as follows:

[0027] ,

[0028] in, , , Respectively represent queries ,key ,value The weight matrix of The feature of the drug is the dimension of the key vector. Similarly, the optimized scaled dot product attention is obtained. Characteristic expression of associated drugs ;Will Input bilinear self-attention for processing to obtain the drug optimized by bilinear self-attention Characteristic expression of , the calculation formula is as follows:

[0029] ,

[0030] in, The characteristic of the drug is the dimension of the key vector; similarly, the bilinear attention optimization and drug Characteristic expression of associated drugs ; Optimize the scaled dot product attention after optimizing the drug Characteristic expression of and drugs Characteristic expression of associated drugs Input to the scaled dot product cross attention to get the first drug pair vector representation , the calculation formula is as follows:

[0031] ,

[0032] The drug after optimizing bilinear attention Characteristic expression of and drugs Characteristic expression of associated drugs Input to bilinear cross attention to get the second drug pair vector representation , the calculation formula is as follows:

[0033] ,

[0034] The first drug pair vector representation and the second drug pair vector representation are concatenated to obtain the final drug pair representation. .

[0035] Furthermore, step S5 is specifically as follows:

[0036] Constructing feature interweaving layers: Based on the Jaccard similarities of the drugs obtained in step S22 under the three biological features of chemical structure, target and enzyme, three homogeneous graphs at different levels are obtained, in which the nodes represent the drugs and the edges represent the Jaccard similarities of the drugs under the current attributes. The node features of the drugs are updated by the following formula:

[0037] ,

[0038] ,

[0039] ,

[0040] in, , , They represent the normalized adjacency matrices of drugs at the three levels of chemical structure, enzyme and target, respectively. , , Represent the trainable weight matrices of drugs in chemical structure, enzyme and target node features, respectively. It is The feature representation after order propagation, the initial value is the paired drug Drug representation of fusion topology , is the activation function, , , Respectively represent drugs The feature update results based on neighbor information in three homogeneous graphs; finally, by Add the results of features at different levels to get the drug Comprehensive feature representation of , the calculation formula is as follows:

[0041] ,

[0042] Get the same medicine Comprehensive feature representation of associated drugs ;

[0043] According to the drug Comprehensive feature representation and drug The comprehensive feature representation of the associated drugs is used to obtain the final comprehensive representation , the formula is as follows:

[0044] .

[0045] Further, step S6 is specifically as follows:

[0046] The prediction module includes a fully connected layer and Activation function;

[0047] By expressing the final drug pair And the final comprehensive representation Splice to get the comprehensive feature vector of drug pairs , the formula is as follows:

[0048] ,

[0049] in, Represents a concatenation operation; the drug pair comprehensive feature vector is linearly transformed and normalized through the first fully connected layer with a batch normalization layer to obtain the transformed first feature vector , the formula is as follows:

[0050] ,

[0051] in, represents the first fully connected layer operation, represents the normalization operation, represents the activation function, represents the regularization operation. The input dimension of the first fully connected layer is 3500 and the output dimension is 858. The first eigenvector after conversion After linear transformation and normalization processing by the second fully connected layer with batch normalization layer, the transformed second eigenvector is obtained. , processing process and the first eigenvector after conversion The processing process is the same as that of the second fully connected layer. The input dimension of the second fully connected layer is 858 and the output dimension is 463. The second eigenvector after conversion is After processing in the third fully connected layer, the drug interaction prediction probability vector is obtained , the formula is as follows:

[0052] ,

[0053] in, Represents the operation of the third fully connected layer; the input dimension of the third fully connected layer is 463, and the output dimension is 68.

[0054] In a second aspect, a drug interaction type prediction device integrating multi-source feature cross-graph structure includes:

[0055] Data acquisition unit: used to acquire drug data, obtain drug data set, and perform preprocessing;

[0056] Similarity calculation unit: used for calculating the drug data by Jaccard similarity to obtain a multi-dimensional data representation of the drug's chemical structure, biological target and enzyme action, and splicing the multi-dimensional data representation to obtain an initial feature representation of the drug;

[0057] The first construction unit is used to construct a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure;

[0058] Feature fusion unit: The drug representation of the fusion topology structure fuses drug features from different sources through an attention mechanism to obtain a final drug pair representation;

[0059] The second construction unit: constructs a feature interleaving layer, and the drug representations of the fusion topological structures corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain the final comprehensive representation;

[0060] Prediction unit: The final drug pair representation and the final comprehensive representation are processed through a prediction module to obtain a drug interaction prediction probability vector.

[0061] In a third aspect, an electronic device comprises: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the drug interaction type prediction method integrating multi-source features across graph structures described in the first aspect is performed.

[0062] In a fourth aspect, a computer-readable storage medium stores a computer program, which, when executed by a processor, executes the drug interaction type prediction method of fusing multi-source features across graph structures described in the first aspect.

[0063] The advantages of the present invention are:

[0064] First, the present invention effectively integrates multidimensional data such as the chemical structure, biological targets and enzyme action of drugs by constructing heterogeneous and homogeneous graph structures, thereby improving the characterization ability of drug features. This structure not only optimizes the capture of direct and indirect interactions between drugs, but also provides a more comprehensive information framework for understanding complex interaction mechanisms. Secondly, by introducing self-attention and cross-attention mechanisms, the present invention realizes dynamic weighting and optimization between features in the feature fusion module, enhancing the model's ability to capture complex drug interactions. This mechanism not only improves the response accuracy of the prediction model, but also further integrates features from different biological dimensions through the "feature interweaving layer", making the drug representation more comprehensive and accurate. In addition, the present invention uses graph convolutional networks to deeply mine the multidimensional characteristics of drugs, combined with batch normalization and GELU activation functions, effectively improving the accuracy and efficiency of classification and prediction. Therefore, the technical solution of the present invention not only optimizes the accuracy and efficiency of drug interaction prediction, but also provides strong technical support for drug development and clinical safety, with important economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0066] Figure 1 is a flow chart of the method of the present invention;

[0067] Figure 2 This is a visualization result image of drug-drug interaction prediction performed in the present invention. DETAILED DESCRIPTION

[0068] 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.

[0069] Example 1

[0070] In this embodiment, Figure 1 As shown, the present invention provides a method for predicting drug interaction types by integrating multi-source feature cross-graph structures, and the specific steps include:

[0071] S1. Obtain drug data, obtain drug data set, and perform preprocessing;

[0072] Specifically, the drug data comes from the DrugBank database, which includes drugs, drug pairs, and interaction event types; each drug has four types of features, including: chemical substructure, target, pathway, and enzyme; the drug data set is preprocessed, and the names of the two drugs in the drug pair are extracted according to the drug pair and drug interaction event type, a unique index is assigned to each drug, and the drug name and the corresponding position are stored in the dictionary.

[0073] S2. The drug data is subjected to Jaccard similarity calculation to obtain a multidimensional data representation of the drug's chemical structure, biological target, and enzyme action, and the multidimensional data representation is spliced ​​to obtain an initial feature representation of the drug;

[0074] Specifically, S21. Feature extraction: The three biological features of the drug, namely the chemical substructure, target and enzyme, are converted into binary feature vectors. The chemical substructure feature of each drug is a string separated by a vertical bar (|), and each string represents an independent feature descriptor. The chemical substructure features of each drug are traversed, all features are extracted, and stored in the list all_feature. A two-dimensional zero matrix feature_matrix is ​​initialized, whose number of rows is equal to the number of drugs and the number of columns is equal to the number of feature descriptors in the all_feature list. Next, each drug is traversed and feature_matrix is ​​updated according to its feature data. Specifically, if the drug has a specific descriptor in the list all_feature, the corresponding matrix element is set to 1, indicating that the feature exists in the drug. Otherwise, the corresponding matrix element is 0, indicating that the feature exists in the drug. The calculation of the target and enzyme biological features is the same as above. Each biological feature corresponds to a binary feature vector. S22. Feature representation: Based on the binary feature vector corresponding to the biological feature of the drug obtained in step S21, the Jaccard similarity metric is used to calculate the similarity between the three biological features of chemical substructure, enzyme and target between the drugs. Each drug is given a chemical substructure similarity feature representation. , an enzyme similarity feature representation and a target similarity feature representation , and then the , as well as After splicing, the initial characteristic representation of each drug is obtained, and the calculation formula is as follows:

[0075] ,

[0076] in, Indicates drug The initial feature representation of .

[0077] S3. constructing a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure;

[0078] Specifically, construct a heterogeneous graph ,node represents a collection of drugs, , Indicates the number of drugs, Represents a set of drug interaction event types: using the adjacency matrix Indicates the connection relationship between nodes. Indicates the interaction relationship If there is an interaction between drug i and drug j, If the drug is Initial drug characterization With the adjacency matrix Multiply, get medicine The drug representation of the fusion topology is calculated as follows:

[0079] ,

[0080] in, Representation Node The degree, Indicates drug With drugs There is interaction , It's medicine The initial drug characteristics calculated in step S2, Indicates the relationship type The weight matrix of is the bias, is the activation function , Indicates drug Drug representation of fused topology.

[0081] S4. The drug representation of the fused topological structure is fused with drug features from different sources through an attention mechanism to obtain a final drug pair representation;

[0082] Specifically, by searching the index, we can get the The two drugs are paired to obtain a drug pair, and the drug representation of the fusion topological structure of the two drugs in the drug pair is recorded as and ,Will Input scaled dot product self-attention to get optimized drugs Characteristic expression of The specific calculation formula is as follows:

[0083] ,

[0084] in, , , Respectively represent queries ,key ,value The weight matrix of The feature of the drug is the dimension of the key vector. Similarly, the optimized scaled dot product attention is obtained. Characteristic expression of associated drugs ;Will Input bilinear self-attention for processing to obtain the drug optimized by bilinear self-attention Characteristic expression of , the calculation formula is as follows:

[0085] ,

[0086] in, The characteristic of the drug is the dimension of the key vector; similarly, the bilinear attention optimization and drug Characteristic expression of associated drugs ; Optimize the scaled dot product attention after optimizing the drug Characteristic expression of and drugs Characteristic expression of associated drugs Input to the scaled dot product cross attention to get the first drug pair vector representation , the calculation formula is as follows:

[0087] ,

[0088] The drug after optimizing bilinear attention Characteristic expression of and drugs Characteristic expression of associated drugs Input to bilinear cross attention to get the second drug pair vector representation , the calculation formula is as follows:

[0089] ,

[0090] The first drug pair vector representation and the second drug pair vector representation are concatenated to obtain the final drug pair representation. .

[0091] S5. construct a feature interleaving layer, wherein the drug representations of the fusion topological structure corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain a final comprehensive representation;

[0092] Specifically, construct a feature interweaving layer: based on the Jaccard similarity of the drug under the three biological characteristics of chemical structure, target and enzyme obtained in step S22, three homogeneous graphs at different levels are obtained, in which the nodes represent the drugs, the edges represent the Jaccard similarity of the drugs under the current attributes, and the node features of the drugs are updated by the following formula:

[0093] ,

[0094] ,

[0095] ,

[0096] in, , , They represent the normalized adjacency matrices of drugs at the three levels of chemical structure, enzyme and target, respectively. , , Represent the trainable weight matrices of drugs in chemical structure, enzyme and target node features, respectively. It is The feature representation after order propagation, the initial value is the paired drug Drug representation of fusion topology , is the activation function, , , Respectively represent drugs The feature update results based on neighbor information in three homogeneous graphs; finally, by Add the results of features at different levels to get the drug Comprehensive feature representation of , the calculation formula is as follows:

[0097] ,

[0098] Get the same medicine Comprehensive feature representation of associated drugs ;

[0099] According to the drug Comprehensive feature representation and drug The comprehensive feature representation of the associated drugs is used to obtain the final comprehensive representation , the formula is as follows: .

[0100] S6. The final drug pair representation and the final comprehensive representation are subjected to a prediction module to obtain a drug interaction prediction probability vector.

[0101] Specifically, the prediction module includes a fully connected layer and activation function; the prediction module includes a fully connected layer and Activation function; by expressing the final drug pair And the final comprehensive representation Splice to get the comprehensive feature vector of drug pairs , the formula is as follows:

[0102] ,

[0103] in, Represents a concatenation operation; the drug pair comprehensive feature vector is linearly transformed and normalized through the first fully connected layer with a batch normalization layer to obtain the transformed first feature vector , the formula is as follows:

[0104] ,

[0105] in, represents the first fully connected layer operation, represents the normalization operation, represents the activation function, represents the regularization operation. The input dimension of the first fully connected layer is 3500 and the output dimension is 858. The first eigenvector after conversion After linear transformation and normalization processing by the second fully connected layer with batch normalization layer, the transformed second eigenvector is obtained. , processing process and the first eigenvector after conversion The processing process is the same as that of the second fully connected layer. The input dimension of the second fully connected layer is 858 and the output dimension is 463. The second eigenvector after conversion is After processing in the third fully connected layer, the drug interaction prediction probability vector is obtained , the formula is as follows:

[0106] ,

[0107] in, Represents the operation of the third fully connected layer; the input dimension of the third fully connected layer is 463, and the output dimension is 68.

[0108] Example 2

[0109] In this embodiment, the method of the present invention is tested on a data set collected by DDIMDL, and the results are shown in Table 1. The data set contains 572 drugs, 74,528 drug pairs, and 65 types of interaction events. The evaluation indicators are: accuracy, area under the precision-recall curve AUPR, area under the curve AUC, F1 score, precision and recall. The results in Table 1 show that compared with several advanced DDI prediction methods MATT-DDI, MDF-SA-DDI, DDIMDL, DeepDDI and three traditional machine learning models, the solution of the present invention has a more obvious advantage in predictive ability. Table 1 shows in detail the performance of the seven baseline models and MFCGS-DDI in multi-classification tasks. The results show that compared with the baseline model, MFCGS-DDI shows superior performance in six key performance indicators.

[0110] Table 1 Comparison of experimental results of the method of the present invention and the prior art

[0111]

[0112] In this embodiment, Figure 2 The performance comparison of multiple models is shown, among which the MFCGS-DDI model performs well in various indicators, especially in core indicators such as ACC (accuracy), AUPR (area under the precision-recall curve) and F1_score, which surpasses other models. For example, in ACC, MFCGS-DDI reaches 0.9354, the highest among all models, and is also close to the best in AUPR and AUC indicators, indicating that it has strong reliability and effectiveness in recognition and classification tasks.

[0113] Example 3

[0114] This embodiment provides a drug interaction type prediction device integrating multi-source feature cross-graph structure, including:

[0115] Data acquisition unit: used to acquire drug data, obtain drug data set, and perform preprocessing;

[0116] Similarity calculation unit: used for calculating the drug data by Jaccard similarity to obtain a multi-dimensional data representation of the drug's chemical structure, biological target and enzyme action, and splicing the multi-dimensional data representation to obtain an initial feature representation of the drug;

[0117] The first construction unit is used to construct a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure;

[0118] Feature fusion unit: The drug representation of the fusion topology structure fuses drug features from different sources through an attention mechanism to obtain a final drug pair representation;

[0119] The second construction unit: constructs a feature interleaving layer, and the drug representations of the fusion topological structures corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain the final comprehensive representation;

[0120] Prediction unit: The final drug pair representation and the final comprehensive representation are processed through a prediction module to obtain a drug interaction prediction probability vector.

[0121] Example 4

[0122] This embodiment provides an electronic device, including: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the drug interaction type prediction method of fusing multi-source features across graph structures in Example 1 is performed.

[0123] Example 5

[0124] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting drug interaction types by fusing multi-source features across graph structures in Embodiment 1 is executed.

[0125] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. 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 interaction types by integrating multi-source features across graph structures, characterized in that: The following steps are involved: S1. Obtain drug data, obtain drug data set, and perform preprocessing; S2. The drug data is subjected to Jaccard similarity calculation to obtain a multidimensional data representation of the drug's chemical structure, biological target, and enzyme action, and the multidimensional data representation is spliced ​​to obtain an initial feature representation of the drug; S3. constructing a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure; S4. The drug representation of the fused topological structure is fused with drug features from different sources through an attention mechanism to obtain a final drug pair representation; The specific steps are: Get the drug by searching the index The two drugs are paired to obtain a drug pair, and the drug representation of the fusion topological structure of the two drugs in the drug pair is recorded as and ,Will Input scaled dot product self-attention to get optimized drugs Characteristic expression of The specific calculation formula is as follows: ， in, , , Respectively represent queries ,key ,value The weight matrix of The feature of the drug is the dimension of the key vector. Similarly, the optimized scaled dot product attention is obtained. Characteristic expression of associated drugs ;Will Input bilinear self-attention for processing to obtain the drug optimized by bilinear self-attention Characteristic expression of , the calculation formula is as follows: ， in, The characteristic of the drug is the dimension of the key vector; similarly, the bilinear attention optimization and drug Characteristic expression of associated drugs ; Optimize the scaled dot product attention after optimizing the drug Characteristic expression of and drugs Characteristic expression of associated drugs Input to the scaled dot product cross attention to get the first drug pair vector representation , the calculation formula is as follows: ， The drug after optimizing bilinear attention Characteristic expression of and drugs Characteristic expression of associated drugs Input to bilinear cross attention to get the second drug pair vector representation , the calculation formula is as follows: ， The first drug pair vector representation and the second drug pair vector representation are concatenated to obtain the final drug pair representation. ; S5. construct a feature interleaving layer, wherein the drug representations of the fusion topological structure corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain a final comprehensive representation; S6. The final drug pair representation and the final comprehensive representation are subjected to a prediction module to obtain a drug interaction prediction probability vector.

2. The method for predicting drug interaction types by integrating multi-source features across graph structures according to claim 1, characterized in that: The step S1 is specifically as follows: The drug data comes from the DrugBank database, and the drug data includes drugs, drug pairs, and interaction event types; Each drug has four types of features, including: chemical substructure, target, pathway, and enzyme; The drug dataset is preprocessed. According to the drug pairs and drug interaction event types, the names of the two drugs in the drug pairs are extracted respectively, a unique index is assigned to each drug, and the drug names and corresponding positions are stored in the dictionary.

3. The method for predicting drug interaction types by integrating multi-source features across graph structures according to claim 2, characterized in that: The step S2 is specifically as follows: S21. Feature extraction: Convert the three biological features of the drug, namely, chemical substructure, target and enzyme, into binary feature vectors, with each biological feature corresponding to a binary feature vector; S22. Feature representation: Based on the binary feature vector corresponding to the biological feature of the drug obtained in step S21, the Jaccard similarity metric is used to calculate the similarity between the three biological features of chemical substructure, enzyme and target between the drugs. Each drug is given a chemical substructure similarity feature representation. , an enzyme similarity feature representation and a target similarity feature representation , and then the , as well as After splicing, the initial characteristic representation of each drug is obtained, and the calculation formula is as follows: ， in, Indicates drug The initial feature representation of .

4. The method for predicting drug interaction types by integrating multi-source features across graph structures according to claim 3, characterized in that: The step S3 is specifically as follows: S31. Constructing a heterogeneous graph ,node represents a collection of drugs, , Indicates the number of drugs, Represents a set of drug interaction event types: using the adjacency matrix Indicates the connection relationship between nodes. Indicates the interaction relationship If there is an interaction between drug i and drug j, If yes, it is 1, otherwise it is 0; S32. Update drug feature representation: Update drug Initial drug characterization With the adjacency matrix Multiply, get medicine The drug representation of the fusion topology is calculated as follows: ， in, Representation Node The degree, Indicates drug With drugs There is interaction , It's medicine The initial drug characteristics calculated in step S2, Indicates the relationship type The weight matrix of is the bias, is the activation function , Indicates drug Drug representation of fused topology.

5. The method for predicting drug interaction types by integrating multi-source features across graph structures according to claim 4, characterized in that: The step S5 is specifically as follows: Constructing feature interweaving layers: Based on the Jaccard similarities of the drugs obtained in step S22 under the three biological features of chemical structure, target and enzyme, three homogeneous graphs at different levels are obtained, in which the nodes represent the drugs and the edges represent the Jaccard similarities of the drugs under the current attributes. The node features of the drugs are updated by the following formula: ， ， ， in, , , They represent the normalized adjacency matrices of drugs at the three levels of chemical structure, enzyme and target, respectively. , , Represent the trainable weight matrices of drugs in chemical structure, enzyme and target node features, respectively. It is The feature representation after order propagation, the initial value is the paired drug Drug representation of fusion topology , is the activation function, , , Respectively represent drugs The feature update results based on neighbor information in three homogeneous graphs; finally, by Add the results of features at different levels to get the drug Comprehensive feature representation of , the calculation formula is as follows: ， Get the same medicine Comprehensive feature representation of associated drugs ; According to the drug Comprehensive feature representation and drug The comprehensive feature representation of the associated drugs is used to obtain the final comprehensive representation , the formula is as follows: 。 6. The method for predicting drug interaction types by integrating multi-source features across graph structures according to claim 5, characterized in that: The step S6 is specifically as follows: The prediction module includes a fully connected layer and Activation function; By expressing the final drug pair And the final comprehensive representation Splice to get the comprehensive feature vector of drug pairs , the formula is as follows: ， in, Represents a concatenation operation; the drug pair comprehensive feature vector is linearly transformed and normalized through the first fully connected layer with a batch normalization layer to obtain the transformed first feature vector , the formula is as follows: ， in, represents the first fully connected layer operation, represents the normalization operation, represents the activation function, represents the regularization operation. The input dimension of the first fully connected layer is 3500 and the output dimension is 858. The first eigenvector after conversion After linear transformation and normalization processing by the second fully connected layer with batch normalization layer, the transformed second eigenvector is obtained. , processing process and the first eigenvector after conversion The processing process is the same as that of the second fully connected layer. The input dimension of the second fully connected layer is 858 and the output dimension is 463. The second eigenvector after conversion is After processing in the third fully connected layer, the drug interaction prediction probability vector is obtained , the formula is as follows: ， in, Represents the operation of the third fully connected layer; the input dimension of the third fully connected layer is 463, and the output dimension is 68.

7. A drug interaction type prediction device integrating multi-source feature cross-graph structure, which executes the drug interaction type prediction method integrating multi-source feature cross-graph structure as claimed in claim 1, characterized in that: include: Data acquisition unit: used to acquire drug data, obtain drug data set, and perform preprocessing; Similarity calculation unit: used for calculating the drug data by Jaccard similarity to obtain a multi-dimensional data representation of the drug's chemical structure, biological target and enzyme action, and splicing the multi-dimensional data representation to obtain an initial feature representation of the drug; The first construction unit is used to construct a heterogeneous graph of drug interactions, wherein the initial feature representation of the drug is updated through the heterogeneous graph to obtain a drug representation of a fused topological structure; Feature fusion unit: The drug representation of the fusion topology structure fuses drug features from different sources through an attention mechanism to obtain a final drug pair representation; The second construction unit: constructs a feature interleaving layer, and the drug representations of the fusion topological structure corresponding to the two drugs in the drug pair are passed through the feature interleaving layer to obtain the final comprehensive representation; Prediction unit: The final drug pair representation and the final comprehensive representation are processed through a prediction module to obtain a drug interaction prediction probability vector.

8. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the drug interaction type prediction method for fusing multi-source features across graph structures according to any one of claims 1 to 6 is performed.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the drug interaction type prediction method of fusing multi-source features across graph structures as described in any one of claims 1 to 6.

Images