A Drug Interaction Prediction Method Based on Multimodal Network and Width Learning

Abstract

The present invention discloses a drug interaction prediction method based on multimodal networks and width learning, comprising: obtaining multimodal feature information of drugs and obtaining drug interaction graphs under various modes; learning the features of drugs under different modes through a graph attention network; using a width learning system to fuse multimodal features, and predicting the probability of interaction between each drug pair through the fused multimodal features of the drugs. The present invention extracts drug feature information from multimodal data and utilizes the powerful feature representation capability of graph neural networks to model and learn the different modal features of drugs. By mapping the data features of different modalities of drugs to a unified low-dimensional vector space, the characteristics of multiple features can be fused by width learning, and the information of four modalities can be fused, thereby fully utilizing the feature information and enhancing the model's expression capability. The width learning system can perform incremental learning to improve the model's predictive performance.

Description

Technical Field

[0001] The present invention relates to the technical field of drug interaction prediction, and in particular to a drug interaction prediction method based on multimodal networks and width learning. Background Art

[0002] Polypharmacy has become a common practice, increasing the likelihood of drug interactions. A drug-drug interaction (DDI) occurs when a patient takes two or more medications simultaneously and the efficacy of one drug is altered by another, resulting in decreased efficacy or toxic side effects. Therefore, predicting and identifying potential adverse drug interactions in combination medications, thereby mitigating potential risks and promoting safe drug co-prescription, has become a major challenge urgently needed in the field of bioinformatics.

[0003] In the existing technology, most DDI prediction methods first extract the topological structural features generated by drug interactions through SMILES (Simplified molecular input line entry system) sequence information, and then perform DDI prediction. In fact, drug characteristics can not only be extracted through SMILES sequence information, but there are also many types of drug data such as: chemical substructures of drugs, targets, side effects, text information describing drugs, etc. How to reasonably integrate multi-source data of drugs and fuse various characteristic information of drugs to improve the prediction ability of models is still an urgent problem to be solved in DDI research. In addition, existing DDI prediction methods lack the ability to incrementally learn for new data, and the entire model needs to be retrained. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a drug interaction prediction method based on multimodal networks and width learning that can fully utilize feature information, has strong model expression ability, and has superior prediction performance.

[0005] To achieve the above objectives, the technical solutions provided by the present invention are:

[0006] A drug interaction prediction method based on multimodal networks and width learning is proposed. Drug interactions are predicted using a drug interaction prediction model. The drug interaction prediction model includes an input layer, an embedding layer, and a width learning layer. The prediction process includes:

[0007] In the input layer, the multimodal feature information of the drug is obtained, and the drug interaction graph under each modality is obtained;

[0008] In the embedding layer, the features of drugs in different modalities are learned through the graph attention network;

[0009] In the width learning layer, the width learning system is used to perform multimodal feature fusion, and the fused drug multimodal features are used to predict the probability of interaction between each drug pair.

[0010] Furthermore, we obtain the multimodal feature information of the drug and obtain the drug interaction graph under each modality, including:

[0011] Obtain drug chemical substructure, target, enzyme and pathway modality feature information from drugbank and KEGG databases;

[0012] Obtain drug interaction graph G under chemical substructure, target, enzyme and pathway modalities s , G t , G e , G p .

[0013] Furthermore, the drug interaction graph G under the chemical substructure mode is obtained s ,include:

[0014] The similarity matrix S of drugs in chemical substructure mode is calculated by the following formula s :

[0015]

[0016] Among them, F s The binary feature vector representing the chemical substructure of the drug; M 11 Indicates the total number of elements whose corresponding positions in the two drug feature vectors are all 1, M 01 Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 0 and in F s The value in v is 1; M 10 Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 1 and in F s The value in v is 0;

[0017] According to the similarity matrix, the drug interaction graph G under the chemical substructure mode is obtained s :

[0018]

[0019] Among them, N v and N u Represent the neighbor node sets of drugs v and u respectively.

[0020] Furthermore, the graph attention network is used to learn the characteristic representation of drugs in the chemical substructure mode, including:

[0021] Assume that the drug molecular features of the input and output of the lth layer are h i and h i ';

[0022] First, the features of the l-layer input are transformed through a parameter-sharing linear transformation, and its weight matrix is set to W;

[0023] Then, in the interaction graph G s For each node, the attention mechanism is used to calculate the attention score between other nodes and the node:

[0024]

[0025] e ij represents the importance of drug j to drug i;

[0026] Next, use softmax to normalize the attention weights, use the LeakyReLU activation function for nonlinearity, and use the attention weights to aggregate neighboring node features and calculate the attention coefficient:

[0027]

[0028] Among them, a represents the weight vector of the attention network;

[0029] After calculating the attention coefficient, the drug molecular features are updated using the multi-head attention method:

[0030]

[0031] Among them, K represents the number of heads of multi-head attention, and multi-head attention learns K different attention scores in parallel subspaces;

[0032] After feature extraction by the graph attention network, the characteristic representation Z of the drug in the chemical substructure mode is obtained. s .

[0033] Furthermore, using the width learning system for multimodal feature fusion includes:

[0034] The obtained chemical substructure, target, enzyme and pathway modal features are fused to obtain Z:

[0035] Z=Z s ||Z t ||Z e ||Z p

[0036] Among them, Z sis the characteristic of the drug in the chemical substructure mode, Z t is the characteristic of the drug in the target mode, Z e is the characteristic of the drug in the enzyme mode, Z p is the characteristic of the drug in the pathway mode;

[0037] Then T is mapped through the feature layer to obtain F i :

[0038]

[0039] in, and It is randomly generated, φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the features obtained are recorded as F n =[F1,F2,…,F n ], these features are enhanced by the enhanced layer to obtain enhanced features H m , where the calculation method of the i-th enhanced feature is:

[0040]

[0041] in, and is randomly generated, ξ is the activation function, and the probability Y of interaction to be predicted is calculated as follows:

[0042] Y=[F1,F2,…,F n |H1,H2,…,H m ]W m =[F n |H m ]W m

[0043] Where W m Through pseudo-inverse operation Y is calculated.

[0044] Furthermore, in order to improve the incremental learning ability of the drug interaction prediction model, the weight parameters of the width learning system are updated through the incremental learning algorithm before the width learning system performs multimodal feature fusion. The multimodal feature fusion uses the width learning system after the weight parameters are updated.

[0045] Furthermore, the weight parameters of the width learning system are updated through an incremental learning algorithm, including:

[0046] Let the new data be X a ,use To represent the concatenation matrix of n groups of feature nodes and m groups of enhancement nodes in the latest width learning system; for new samples, the feature layer and enhancement layer are used to map new feature nodes separately and Enhanced Node And it is expressed by the following formula:

[0047]

[0048] The new input matrix is updated as:

[0049]

[0050] According to the relevant pseudo-inverse update algorithm, the following formula is derived:

[0051]

[0052] in

[0053]

[0054] Integrate the weight information of the newly added samples into the weight of the width learning system, dynamically update the width learning system, and realize the integration of new and old knowledge; the output weight of the width learning system is updated as follows:

[0055]

[0056] Among them, Y a Indicates the label of the newly added data.

[0057] Compared with the existing technology, the principles and advantages of this solution are as follows:

[0058] 1. This solution extracts drug feature information from multimodal data and leverages the powerful feature representation capabilities of graph neural networks to model and learn the different modal characteristics of drugs. By mapping the data features of different drug modalities into a unified low-dimensional vector space, wide-width learning can be used to integrate the characteristics of multiple features. By integrating information from four modalities, this approach fully utilizes the feature information and enhances the model's expressive power.

[0059] 2. This solution utilizes the incremental learning feature of the wide learning system to learn new samples and new knowledge without retraining the entire drug interaction prediction model, thereby improving the prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the services required for use in the embodiments or the prior art descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 This is a principle flow chart of a drug interaction prediction method based on multimodal networks and width learning of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with specific embodiments:

[0063] like Figure 1 As shown, the drug interaction prediction method based on multimodal network and width learning described in this embodiment predicts drug interactions using a drug interaction prediction model. The drug interaction prediction model includes an input layer, an embedding layer, and a width learning layer. The prediction process includes the following steps:

[0064] S1. In the input layer, obtain the drug's chemical substructure, target, enzyme, and pathway modality feature information, and obtain the drug interaction graph under each modality;

[0065] Specifically, the chemical substructure, target, enzyme and pathway modality feature information of drugs is obtained from drugbank and KEGG databases, and the drug interaction graph G under chemical substructure, target, enzyme and pathway modality is obtained. s , G t , G e , G p .

[0066] Among them, the drug interaction graph G under the chemical substructure mode is obtained s ,include:

[0067] The similarity matrix S of drugs in chemical substructure mode is calculated by the following formula s :

[0068]

[0069] In the above formula, F s The binary feature vector representing the chemical substructure of the drug; M 11 Indicates the total number of elements whose corresponding positions in the two drug feature vectors are all 1, M 01 Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 0 and in F s The value in v is 1; M 10Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 1 and in F s The value in v is 0;

[0070] According to the similarity matrix, the drug interaction graph G under the chemical substructure mode is obtained s :

[0071]

[0072] In the above formula, N v and N u Represent the neighbor node sets of drugs v and u respectively.

[0073] According to the above data preprocessing method, three other modalities can be obtained: drug interaction graph G under target, enzyme and pathway information t , G e , G p .

[0074] S2. In the embedding layer, the graph attention network (GAT) is used to learn the characteristics of drugs in different modalities.

[0075] Among them, the feature representation of drugs in chemical substructure mode is learned through the graph attention network, including:

[0076] Assume that the drug molecular features of the input and output of the lth layer are h i and h i ';

[0077] First, the features of the l-layer input are transformed through a parameter-sharing linear transformation, and its weight matrix is set to W;

[0078] Then, in the interaction graph G s For each node, the attention mechanism is used to calculate the attention score between other nodes and the node:

[0079]

[0080] e ij represents the importance of drug j to drug i;

[0081] Next, use softmax to normalize the attention weights, use the LeakyReLU activation function for nonlinearity, and use the attention weights to aggregate neighboring node features and calculate the attention coefficient:

[0082]

[0083] Among them, a represents the weight vector of the attention network;

[0084] After calculating the attention coefficient, the drug molecular features are updated using the multi-head attention method:

[0085]

[0086] Among them, K represents the number of heads of multi-head attention, and multi-head attention learns K different attention scores in parallel subspaces;

[0087] After feature extraction by the graph attention network, the characteristic representation Z of the drug in the chemical substructure mode is obtained. s Similarly, the above method can be used to obtain the characteristic Z under the target, enzyme and pathway mode t , Z e , Z p .

[0088] S3. Update the weight parameters of the width learning system in the width learning layer through the incremental learning algorithm, including:

[0089] Let the new data be X a ,use To represent the concatenation matrix of n groups of feature nodes and m groups of enhancement nodes in the latest width learning system; for new samples, the feature layer and enhancement layer are used to map new feature nodes separately and enhanced nodes And it is expressed by the following formula:

[0090]

[0091] The new input matrix is updated as:

[0092]

[0093] According to the relevant pseudo-inverse update algorithm, the following formula is derived:

[0094]

[0095] in

[0096]

[0097] Integrate the weight information of the newly added samples into the weight of the width learning system, dynamically update the width learning system, and realize the integration of new and old knowledge; the output weight of the width learning system is updated as follows:

[0098]

[0099] Among them, Y a Indicates the label of the newly added data.

[0100] S4. In the width learning layer, the updated width learning system is used to perform multimodal feature fusion, and the fused drug multimodal features are used to predict the probability of interaction between each drug pair.

[0101] Multimodal feature fusion using a wide learning system includes:

[0102] The obtained chemical substructure, target, enzyme and pathway modal features are fused to obtain Z:

[0103] Z=Z s ||Z t ||Z e ||Z p

[0104] Among them, Z s is the characteristic of the drug in the chemical substructure mode, Z t is the characteristic of the drug in the target mode, Z e is the characteristic of the drug in the enzyme mode, Z p is the characteristic of the drug in the pathway mode;

[0105] Then T is mapped through the feature layer to obtain F i :

[0106]

[0107] in, and It is randomly generated, φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the features obtained are recorded as F n =[F1,F2,…,F n ], these features are enhanced by the enhanced layer to obtain enhanced features H m , where the calculation method of the i-th enhanced feature is:

[0108]

[0109] in, and is randomly generated, ξ is the activation function, and the probability Y of interaction to be predicted is calculated as follows:

[0110] Y=[F1,F2,…,F n |H1,H2,…,H m ]W m =[F n |H m ]W m

[0111] Where Wm Through pseudo-inverse operation Y is calculated.

[0112] This embodiment extracts drug feature information from multimodal data and leverages the powerful feature representation capabilities of graph neural networks to model and learn the different modal features of drugs. By mapping the data features of different drug modalities into a unified low-dimensional vector space, this embodiment leverages the ability of wide learning to fuse multiple features, integrating information from four modalities, fully utilizing feature information and enhancing the model's expressive power. Furthermore, by leveraging the incremental learning capabilities of the wide learning system, new samples and new knowledge can be learned without retraining the entire drug interaction prediction model, improving the model's predictive performance.

[0113] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made based on the shape and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A drug interaction prediction method based on multimodal networks and width learning, characterized in that: Drug interactions are predicted using a drug interaction prediction model. The drug interaction prediction model includes an input layer, an embedding layer, and a width learning layer. The prediction process includes: In the input layer, the multimodal feature information of the drug is obtained, and the drug interaction graph under each modality is obtained; In the embedding layer, the features of drugs in different modalities are learned through the graph attention network; In the width learning layer, the width learning system is used to fuse multimodal features, and the fused multimodal features of drugs are used to predict the probability of interaction between each drug pair. Obtain multimodal feature information of drugs and obtain drug interaction graphs under each modality, including: Obtain drug chemical substructure, target, enzyme and pathway modality feature information from drugbank and KEGG databases; Obtain drug interaction graph G under chemical substructure, target, enzyme and pathway modalities s , G t , G e , G p ; Obtain drug interaction graph G under chemical substructure mode s ,include: The similarity matrix S of drugs in chemical substructure mode is calculated by the following formula s : Among them, F s The binary feature vector representing the chemical substructure of the drug; M 11 Indicates the total number of elements whose corresponding positions in the two drug feature vectors are all 1, M 01 Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 0 and in F s The value in v is 1; M 10 Indicates that the corresponding positions of the two eigenvectors are in F s The value in u is 1 and in F s The value in v is 0; According to the similarity matrix, the drug interaction graph G under the chemical substructure mode is obtained s : Among them, N v and N u Represent the neighbor node sets of drugs v and u respectively.

2. A drug interaction prediction method based on multimodal network and width learning according to claim 1, characterized in that: The graph attention network is used to learn the feature representation of drugs in the chemical substructure mode, including: Assume that the drug molecular features of the input and output of layer I are h i and h i '; First, the features of the I layer input are transformed through a parameter-sharing linear transformation, and its weight matrix is set to W; Then, in the interaction graph G s For each node, the attention mechanism is used to calculate the attention score between other nodes and the node: e ij represents the importance of drug j to drug i; Next, use softmax to normalize the attention weights, use the LeakyReLU activation function for nonlinearity, and use the attention weights to aggregate neighboring node features and calculate the attention coefficient: Among them, a represents the weight vector of the attention network; After calculating the attention coefficient, the drug molecular features are updated using the multi-head attention method: Among them, K represents the number of heads of multi-head attention, and multi-head attention learns K different attention scores in parallel subspaces; After feature extraction by the graph attention network, the characteristic representation Z of the drug in the chemical substructure mode is obtained. s .

3. The drug interaction prediction method based on multimodal network and width learning according to claim 2, characterized in that: Multimodal feature fusion using a wide learning system includes: The obtained chemical substructure, target, enzyme and pathway modal features are fused to obtain Z: Z=Z s ||From t ||From e ||From p Among them, Z s is the characteristic of the drug in the chemical substructure mode, Z t is the characteristic of the drug in the target mode, Z e is the characteristic of the drug in the enzyme mode, Z p is the characteristic of the drug in the pathway mode; Then Z is mapped through the feature layer to obtain F i : in, and It is randomly generated, φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the features obtained are recorded as F n =[F1,F2,…,F n ], these features are enhanced by the enhanced layer to obtain enhanced features H m , where the calculation method of the i-th enhanced feature is: in, and is randomly generated, ξ is the activation function, and the probability Y of interaction to be predicted is calculated as follows: Y=[F1,F2,…,F n |H1,H2,…,H m ]W m =[F n |H m ]W m Where W m Through pseudo-inverse operation Calculated.

4. The drug interaction prediction method based on multimodal network and width learning according to claim 1, characterized in that: Before the width learning system performs multimodal feature fusion, the weight parameters of the width learning system are updated through an incremental learning algorithm. The multimodal feature fusion adopts the width learning system after the weight parameters are updated.

5. The drug interaction prediction method based on multimodal network and width learning according to claim 4, characterized in that: The weight parameters of the width learning system are updated through an incremental learning algorithm, including: Let the new data be X a ,use To represent the concatenation matrix of n groups of feature nodes and m groups of enhancement nodes in the latest width learning system; for new samples, the feature layer and enhancement layer are used to map new feature nodes separately and Enhanced Node And it is expressed by the following formula: The new input matrix is updated as: According to the relevant pseudo-inverse update algorithm, the following formula is derived: in Integrate the weight information of the newly added samples into the weight of the width learning system, dynamically update the width learning system, and realize the integration of new and old knowledge; the output weight of the width learning system is updated as follows: Among them, Y a Indicates the label of the newly added data.

Images