Drug pair interaction prediction method and device based on soft mask dual-view learning

Abstract

The present invention provides a method and device for predicting drug pair interactions based on soft mask dual view learning, and relates to the technical field of drug interaction prediction. The present invention obtains the molecular view data of two drugs to be predicted and their interaction relationship types to form a triple, and generates a bipartite graph of the drug pair; the triple and the bipartite graph are input into the MSMDL encoder, and the new node embedding of the corresponding drug is output; wherein the MSMDL encoder is K layers, and the output results of each layer are respectively input into the next layer of MSMDL encoder and self-attention graph pooling layer, and the graph-level representation is generated through the self-attention graph pooling layer to obtain the substructure feature embedding; each layer of substructure feature embedding is used together as the input of the common attention mechanism layer to obtain the importance of the interaction between the substructure feature embeddings of each layer; finally, the DDI prediction probability is output through the decoder. The present invention integrates the local and global information of drugs and improves the accuracy of drug pair interaction prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of drug interaction prediction, and in particular to a method and device for drug pair interaction prediction based on soft mask dual-view learning. Background Art

[0002] When two different drugs are taken at the same time, the presence of one drug may cause the efficacy of the other drug to change. This phenomenon is called drug-drug interaction (DDI). When the synergistic effect occurs between drugs, it has a positive impact on the patient's treatment effect, but if the antagonistic effect occurs between drugs, it may cause adverse drug reactions (ADRs), reduce the treatment effect, and even threaten the patient's life safety.

[0003] In the study of DDI prediction, integrating the complementary information of multiple views is crucial to improving the prediction performance. Scholars have conducted in-depth exploration of multi-view representation learning in the field of DDI prediction. By introducing multi-view methods, DDI prediction models can simultaneously capture different aspects of drug molecules, which helps to optimize the accuracy and reliability of prediction. Although multi-view methods have been widely used in DDI prediction, existing studies mostly focus on obtaining drug representations from multiple perspectives and predicting through feature fusion. This method fails to fully consider the atomic-level information inside drug molecules and the relationship between molecules. This deep information is crucial for understanding the mechanism of DDI. Secondly, the existing DDI prediction algorithms mainly encode the information of the drug itself, and fail to effectively combine the deep features of drug molecules and the interaction information between drug pairs, which limits the accuracy of prediction. Other methods extract information from the inside of drug molecules, but they do not consider the chemical complementarity and interaction mechanism between drug pairs enough, and focus too much on the internal features of drug molecules. Therefore, in order to fully capture DDI information, it is necessary not only to deeply explore the deep features of drug molecules, but also to combine the global interaction information between drug pairs.

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

[0005] The present invention aims to provide a method and device for predicting drug pair interactions based on soft mask dual-view learning, so as to solve the limitations of existing methods in extracting key substructures and drug pair interaction information.

[0006] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0007] A drug pair interaction prediction method based on soft mask dual-view learning, comprising:

[0008] S1, obtain the molecular view data of the two drugs to be predicted and their interaction relationship types, and form a triple;

[0009] S2, generating a bipartite graph of drug pairs based on the triples;

[0010] S3, inputting the triplet and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output new node embeddings corresponding to the drug; wherein the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the new node embeddings into the next layer of the MSMDL encoder and the self-attention graph pooling layer, respectively, and generates a graph-level representation through the self-attention graph pooling layer to obtain a substructure feature embedding;

[0011] S4, the substructure feature embeddings of each layer are used together as the input of the common attention mechanism layer to evaluate the contribution of different substructures to the prediction of drug-pair interaction and obtain the importance of the interaction between the substructure feature embeddings of each layer;

[0012] S5, according to the embedding of the substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type, the decoder performs probability calculation and outputs the predicted probability of interaction between the two drugs, namely, the DDI predicted probability, to obtain the final prediction result.

[0013] Preferably, the MSMDL encoder includes a soft mask graph neural network ASMGNN, an internal view learning module and an external view learning module; wherein,

[0014] The soft mask graph neural network ASMGNN extracts key substructure features of the corresponding drugs from the molecular view data of the two drugs in the input triplet through a shared weight strategy, and updates the feature representation of each node in the molecular view;

[0015] The internal view learning module, by combining the graph attention mechanism, evaluates the contribution of the key substructure features of the two drugs in the DDI prediction task, mines the drug feature information that is crucial to the prediction results, and obtains the molecular internal information of the drugs related to DDI, i.e., the internal features of the drugs, by deeply extracting the key substructure features of the drugs;

[0016] The external view learning module, based on the internal drug features of the two drugs and the bipartite graph, captures the intermolecular view information of the two drugs through a common attention mechanism, learns the potential associations between different drug atoms, and thus obtains the intermolecular information of the drugs related to DDI, namely, the external drug features.

[0017] Preferably, S3 is specifically:

[0018] Step 1: Each node of the molecular view data of the two drugs in the input triplet is passed through the soft mask graph neural network of the MSMDL encoder to learn the structural feature representation of its neighbor nodes and the previous layer based on the shared weights, and update the node. The expression is:

[0019] ;

[0020] ;

[0021] in, Representation Node In the Structural feature representation of layers; Represents neighbor nodes In the Structural feature representation of the layer; Representation Node Structural feature representation in the previous layer; Represents neighbor nodes Structural feature representation at the previous layer; represents a nonlinear activation function; Indicates a connection operation; represents the trainable parameter matrix, represents a real number, Represents the dimension of the trainable parameter matrix, i.e., the node Dimensions; Representation Node The set of neighbor nodes; Representation Node The first layer mask value; Represents neighbor nodes The first layer mask value;

[0022] and are the linear mappings in updating node representation and message passing process respectively; Representation Node The learned mask value of the previous layer; Represents neighbor nodes The learned mask value of the previous layer; MLP is a two-layer perceptron with an output feature dimension of 1. The last layer of the perceptron is Activation function;

[0023] Step 2: The interaction information between atoms in the drug molecule view is obtained through the internal view learning module of the MSMDL encoder. Different weights are assigned to each atom in combination with the graph attention mechanism. The attention information of the neighboring nodes of each node in the drug molecule view is aggregated to update the representation of each node itself. The expression is:

[0024] ;

[0025] ;

[0026] in, For the inner view Node in layer The updated structural feature representation, i.e. Internal characteristics of drugs in layers; For nodes in the inner view No. The trainable weight matrix of the layer; Neighbor nodes in the inner view No. The trainable weight matrix of the layer; , For the inner view The trainable bias of the layer; is the normalization constant; is the attention coefficient in the inner view, i.e., the nodes in the same drug and The correlation of ; T is the transpose;

[0027] Step 3: In the external view learning module of the MSMDL encoder, the bipartite graph is combined to capture the view information between two drug molecules through a joint attention mechanism to learn the potential associations between different drug atoms, i.e., drug The atoms in the The atoms in send or receive attention messages, and the process expression is:

[0028] ;

[0029] ;

[0030] in, Indicates the exterior view Node in layer The updated representation is External characteristics of the drugs in the layer; For nodes in the external view No. The trainable weight matrix of the layer; Neighbor nodes in the external view No. The trainable weight matrix of the layer; and is the trainable bias term in the exterior view, Represents a bipartite graph Midpoint The set of neighbor nodes of , i.e., all atoms in another drug; is the attention coefficient in the inter-view layer, i.e., the atomic distance between two drugs in the outer view and relevance;

[0031] Step 4: Fuse the internal features of the drug with the external features of the drug to obtain a new node embedding of the corresponding drug, that is, embedding the node in the internal view The structural features of Nodes in the external view Representation Connect them, integrate them through MLP, and output the new node embedding, expressed as:

[0032] ;

[0033] in, For the Node in layer The new node embedding, MLP is a multi-layer perceptron;

[0034] Step 5: Take the medicine and medications The new node embedding output by each layer is input into the next layer MSMDL encoder and self-attention graph pooling layer respectively. The self-attention graph pooling layer is used as the readout function to generate a graph-level representation and obtain the substructure feature embedding of each layer. The formula is:

[0035] ;

[0036] ;

[0037] in, is an adjacency matrix obtained through the bipartite graph; For drugs No. The embedding representation of all nodes in the layer; n is the The number of nodes in the layer; To pass The normalization function passes the adjacency matrix A and X through the graph neural network The calculated embedding is normalized to obtain the result, Indicates drug The result after the k-th layer node aggregation and self-attention map pooling layer, that is, the drug Substructure feature embedding at the kth layer.

[0038] Preferably, the S4 is specifically:

[0039] Two drugs and The outputs after the three-layer MSMDL encoder and the self-attention graph pooling layer are expressed as and ,in, 、 Represents two drugs and The corresponding layer number of the MSMDL encoder;

[0040] The substructure feature embeddings corresponding to each layer are used as the input of the common attention mechanism layer, and the substructure feature embeddings are measured by calculating the attention weights of the substructure feature embeddings of each layer. and The importance of interactions , to evaluate the contribution of different substructures to the prediction of drug-drug interaction, i.e., DDI, and to obtain the importance of the interaction between substructure feature embeddings ,importance The expression is:

[0041] ;

[0042] in, and For two drugs and In the The substructure feature embedding of the layer, is the trainable weight vector, and are the trainable weight matrices corresponding to the two drugs, is the activation function and T is the transpose.

[0043] Preferably, the S5 is specifically:

[0044] The decoder is a common attention scoring function, and its output layer is a nonlinear activation function for converting the output into a probability value between 0 and 1;

[0045] Embedding the substructure features of each layer corresponding to the two drugs and , and the corresponding importance , trainable representation matrix , get the prediction result, the formula is:

[0046] ;

[0047] in, Represents a triple ( )’s predicted probability; represents a nonlinear activation function; K represents the total number of layers of the MSMDL encoder; T represents transposition; , Drugs In the Substructure feature embedding of layers; is the trainable representation matrix.

[0048] The present invention also provides a drug pair interaction prediction device based on soft mask dual view learning, comprising:

[0049] An acquisition unit, used for acquiring the molecular view data of two drugs to be predicted and the types of their interaction relationships, and forming a triplet;

[0050] A bipartite graph unit, for generating a bipartite graph of drug pairs according to the triples;

[0051] An MSMDL encoder unit is used to input the triple and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output a new node embedding corresponding to the drug; wherein the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the new node embedding into the next layer of the MSMDL encoder and the self-attention graph pooling layer, respectively, and generates a graph-level representation through the self-attention graph pooling layer to obtain a substructure feature embedding;

[0052] The common attention mechanism unit is used to embed the substructure features of each layer together as the input of the common attention mechanism layer, evaluate the contribution of different substructures to the prediction of drug-pair interaction, and obtain the importance of the interaction between the substructure feature embeddings of each layer;

[0053] The decoding prediction unit is used to perform probability calculation based on the embedding of substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type through the decoder, and output the predicted probability of interaction between the two drugs, that is, the DDI predicted probability, to obtain the final prediction result.

[0054] The present invention also provides a drug pair interaction prediction device based on soft mask dual-view learning, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a drug pair interaction prediction method based on soft mask dual-view learning as described above.

[0055] The present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, a drug pair interaction prediction method based on soft mask dual-view learning as described above is implemented.

[0056] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0057] The proposed multi-layer soft mask dual-view learning method for drug-drug interaction prediction (MSMDL-DDI) extracts substructure features related to DDI prediction through a multi-layer soft mask neural network (ASMGNN), and uses a dual-view learning strategy to integrate local and global information of drugs to form a comprehensive drug pair representation. This method makes full use of the representation information generated by the neural network at different levels, and adaptively focuses on the key substructures in the drug molecule graph through a multi-layer soft mask mechanism, while ignoring noise, thereby improving the model's sensitivity and robustness to drug molecular structures.

[0058] The present invention adopts a dual-view learning strategy to effectively integrate local and global information, significantly improving the accuracy of drug interaction prediction. By focusing on the internal characteristics of individual drug molecules and the interactions between drugs, the generalization ability of the model is enhanced. MSMDL-DDI not only provides new technical means for drug safety assessment and new drug development, but also provides new perspectives and ideas for research in the field of DDI prediction.

[0059] The present invention uses a multi-layer soft mask graph neural network to extract substructure related information from the original drug molecule graph, and learns the sequence representation of the substructure through a multi-layer adaptive soft mask graph neural network, thereby obtaining the drug chemical substructure that is critical to DDI prediction, and can improve the drug molecule representation capability without increasing additional training costs. This solves the problem of "critical substructure identification challenges", achieves accurate integration of drug features without being disturbed by a large amount of irrelevant information, and effectively combines the intrinsic properties of drug molecules with the interaction information between drug pairs. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0061] Figure 1 A flowchart of a drug pair interaction prediction method based on soft mask dual-view learning provided in Example 1.

[0062] Figure 2 A schematic diagram of the structure of a drug pair interaction prediction method based on soft mask dual-view learning provided in Example 1.

[0063] Figure 3 This is a schematic diagram of the structure of the MSMDL encoder provided in Example 1.

[0064] Figure 4 A schematic diagram of a drug pair interaction prediction device based on soft mask dual-view learning provided in Example 2.

[0065] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0067] Embodiment 1

[0068] Embodiment 1 of the present invention provides a drug pair interaction prediction method based on soft mask dual-view learning, which can be implemented by a drug pair interaction prediction device based on soft mask dual-view learning (hereinafter referred to as interaction prediction device), and in particular, executed by one or more processors in the interaction prediction device.

[0069] In this embodiment, the interaction prediction device may be an electronic device equipped with a processor, the processor having a computer program of the drug pair interaction prediction method based on soft mask dual-view learning and the computer program can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc., which is not limited here.

[0070] In DDI prediction research, integrating complementary information from multiple views is crucial to improving prediction performance. Scholars have conducted in-depth exploration of multi-view representation learning in the field of DDI prediction. By introducing multi-view methods, DDI prediction models can simultaneously capture different aspects of drug molecules, which helps to optimize the accuracy and reliability of predictions. Specifically, multi-view methods can integrate multiple information sources such as the chemical structure, pharmacological properties, and drug usage patterns of drugs, thereby providing richer and more accurate data support for DDI prediction. This comprehensive analysis strategy not only enhances the model's ability to recognize the complexity of drug interactions, but also provides a new perspective for personalized medicine and precision drug design.

[0071] In the prior art, Ma et al. proposed a method to integrate multi-view similarity features through an attention mechanism, in which the view references a similarity matrix constructed by features and uses a graph autoencoder to learn drug representation. Zhang et al. proposed the HTCL-DDI model, starting from the aspects of drug molecules, structures, and semantic information. The model uses a hierarchical three-view contrastive learning method to achieve DDI prediction, simulates the complex information involved in DDI prediction through molecular views, structural views, and semantic views, and uses a dual attention network to aggregate the composition and structural information of molecules in the molecular view. Different from the method that considers the drug's own information, Niu et al. proposed the DAS-DDI model, which adopts a dual-view method to build a model from two perspectives: drug association and drug structure, and dynamically fuses multi-view drug embeddings through an attention mechanism, thereby improving the ability to distinguish data from different views. Considering deeper drug molecular information from different perspectives, Shen et al. designed different dual-view strategies for known and new drugs by fusing information on atomic nodes and chemical bonds to improve the adaptability and performance of DDI prediction. To make up for the defect that previous papers failed to consider the topological information between drugs, Pan et al. proposed the MSDF model, which constructs a comprehensive drug embedding representation by building topological views and feature views of drugs.

[0072] Although multi-view methods have been widely used in DDI prediction, existing studies have mostly focused on obtaining drug characterization from multiple perspectives and making predictions through feature fusion. This approach may not fully consider the atomic-level information within drug molecules and the relationship between their molecules. This deep-level information is crucial for understanding the mechanism of DDI.

[0073] At present, some methods have begun to focus on the role of functional groups and chemical substructures in DDI. However, the noise information in the drug molecule graph makes it more complicated to accurately extract key substructures. Secondly, many existing DDI prediction algorithms mainly encode the information of the drug itself, and fail to effectively combine the deep characteristics of the drug molecule and the interaction information between drug pairs, which limits the accuracy of the prediction. To address this problem, the GMPNN-CS model improves DDI prediction by learning chemical substructures of different sizes and shapes in the drug molecule graph representation. Compared with the traditional GNN model, GMPNN-CS adopts a message passing neural network with a gating mechanism, in which the edges are regarded as gates that control the message passing process, thereby defining substructures in a learnable way. GMPNN-CS considers the interaction between drug substructures in DDI prediction by assigning a relevance score to each pair of substructure interactions. This method can weightedly evaluate the interaction between substructures to generate more accurate DDI prediction results. MAMSMDDI also considers the importance of chemical substructures in DDI. This method is based on a multi-layer adaptive soft mask graph neural network to mine the substructure feature information of drug molecule graphs. The latent features are updated through the attention mechanism to identify the importance of substructures to DDI and optimize the DDI prediction accuracy.

[0074] Although the above methods have made progress in extracting information from drug molecules, they have not given enough consideration to the chemical complementarity and interaction mechanisms between drug pairs and have placed too much emphasis on the internal characteristics of drug molecules.

[0075] Starting from different perspectives of drug molecules, the present invention conducts dual-view learning for the intra-drug molecule view and the intermolecular view, aiming to discover the potential relationship between drug molecule information and DDI information, thereby obtaining the key substructures of drug molecules that are important for DDI prediction. In order to fully capture DDI information, not only the deep-level characteristics of drug molecules are deeply mined, but also the global interaction information between drug pairs is combined to provide a more solid foundation for DDI prediction.

[0076] like Figure 1-Figure 2 As shown, a drug pair interaction prediction method based on soft mask dual-view learning includes steps S1 to S5.

[0077] S1, obtain the molecular view data of the two drugs to be predicted and their interaction relationship types, and form a triple.

[0078] In this embodiment, the molecular view data of two drugs are defined as follows: and , the interaction relationship type between the two drugs is R, and the triplet is , The value is 1 or 0, 1 means that both drugs , 0 means there is an interaction relationship, and 0 means there is no interaction relationship.

[0079] like Figure 2 As shown, each drug molecule is regarded as a graph, the atoms of each drug are regarded as nodes of the graph, the chemical bonds connecting the atoms are regarded as edges of the graph, and the molecular view of each drug contains a node set V and an edge set E.

[0080] S2: Generate a bipartite graph of drug pairs based on the triples.

[0081] In this embodiment, a bipartite graph between two drug views is generated. It is expressed as: .

[0082] In the constructed bipartite graph, all atoms in one drug are considered as one node set, while all atoms in the other drug form another node set.

[0083] In medicine , If the view A node in the view If there is an interaction relationship between a node in The node and view in Draw an edge between corresponding nodes in .

[0084] S3, input the triplet and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output new node embeddings corresponding to the drugs; wherein, the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the next layer of the MSMDL encoder and the self-attention map pooling layer respectively, generates a graph-level representation through the self-attention map pooling layer, and obtains the substructure feature embedding.

[0085] like Figure 3As shown, in this embodiment, in order to achieve focusing on a specific substructure or ignoring irrelevant parts, the MSMDL encoder of this embodiment is set to three layers (it can be set according to actual conditions, such as 4 layers, 5 layers, etc., which are not limited here), and a soft mask mechanism is used in the MSMDL encoder to learn local atomic representations. The MSMDL encoder includes a soft mask graph neural network ASMGNN, an internal view learning module, and an external view learning module. The soft mask graph neural network ASMGNN composed of MaskConv (mask layer) and EmbeddingConv (embedding layer) in the MSMDL encoder is responsible for extracting the key substructure features of the corresponding drugs from the molecular view data of the two drugs in the input triplet through a shared weight strategy, and updating the feature representation of each node in the molecular view;

[0086] Next, the obtained key substructure features are input into the internal view learning module and the external view learning module of the MSMDL encoder.

[0087] The internal view learning module, combined with the graph attention mechanism, evaluates the contribution of the key substructure features of the two drugs in the DDI prediction task, mines the drug feature information that is crucial to the prediction results, and obtains the molecular internal information of the drugs related to DDI, namely the internal features of the drugs, by deeply extracting the key substructure features of the drugs.

[0088] The external view learning module focuses on the interaction between the two drugs to learn the potential associations between different drug atoms. Based on the internal drug features of the two drugs, combined with the bipartite graph, the intermolecular view information of the two drugs is captured through a joint attention mechanism, and the potential associations between different drug atoms are learned, thereby obtaining the intermolecular information of drugs related to DDI, namely the external drug features.

[0089] Specifically, it is achieved through steps one to five.

[0090] In step 1, the molecular view data of the two drugs in the input triplet are passed through the soft mask graph neural network, and the key substructure features of the corresponding drugs are extracted using a shared weight strategy. The weights of the nodes are dynamically adjusted during the updating process, thereby focusing on specific substructures or ignoring irrelevant parts.

[0091] In this embodiment, soft masking is a mechanism for dynamically emphasizing or suppressing specific nodes or features in a neural network. Using the same graph convolution layer weights for both drugs helps the model learn the key substructure features common to both drug molecules. A mask value is initialized for each node, and after each graph convolution operation, the mask value is updated according to the node representation and attention mechanism; when updating the node representation, the mask value is applied to the updated representation of the node to emphasize or suppress specific nodes. Based on the final mask value, key nodes that contribute significantly to the properties of the drug are identified. Based on these key nodes, the corresponding substructures are extracted. These substructures represent the key functional parts or active sites of the drug, i.e., key substructures.

[0092] Specifically, the nodes in each layer represent The weights are updated by considering the representations of its neighbor nodes and its own previous layer representation. and a trainable weight matrix Specifically, nodes In the Structural feature representation of layers As shown in the formula:

[0093] ;

[0094] in, Representation Node In the Structural feature representation of layers; Represents neighbor nodes In the Structural feature representation of the layer; Representation Node Structural feature representation in the previous layer; Represents neighbor nodes Structural feature representation at the previous layer; represents a nonlinear activation function, such as Activation function; Indicates a connection operation; represents the trainable parameter matrix, represents a real number, Represents the dimension of the trainable parameter matrix, i.e., the node Dimensions; Representation Node The set of neighbor nodes; Representation Node The first layer mask value; Represents neighbor nodes The first layer mask value;

[0095] Mask value Determines whether the target node and edge should be included in the subgraph of the current layer. The calculation method is:

[0096] ;

[0097] in, and They are respectively the linear mapping in updating node representation and message passing process; Representation Node The learned mask value of the previous layer; Represents neighbor nodes The learned mask value of the previous layer; MLP is a two-layer perceptron with an output feature dimension of 1. The last layer of MLP is The activation function determines The value at each node.

[0098] in addition, To model the message passing process, the initial node representation The original image The node features in the . The drug molecule representation obtained by the soft mask graph neural network ASMGNN contains rich local and global interaction information at different levels.

[0099] Step 2: Use the graph attention mechanism to evaluate the contribution of the key substructure features of the two drugs in the DDI prediction task, and mine the drug feature information that is crucial to the prediction results. Combined with the drug feature information, the key substructure features of the drugs are deeply extracted to obtain the molecular internal information of the corresponding drugs related to DDI, that is, the drug internal features.

[0100] Specifically, the internal view learning module is responsible for learning local representations from the molecular graph of a single drug. This embodiment uses the GAT model to further obtain the interaction information between atoms in the drug molecule view. The graph attention mechanism allows the model to assign different weights to each atom, reflecting their relative importance in drug interactions. Specifically, the GAT model of the internal view learning module obtains the interaction information between atoms in the drug molecule view, and the graph attention mechanism is combined to assign different weights to each atom. By aggregating the attention information of the neighboring nodes of each node in the drug molecule view, the representation of each node itself is updated. The process formula is:

[0101] ;

[0102] ;

[0103] in, It is a nonlinear activation function (such as RELU function, Sigmoid function, Tanh function, Leaky ReLU function, etc.), For the inner view Node in layer The updated structural feature representation, i.e. Internal characteristics of drugs in layers; For nodes in the inner view No. The trainable weight matrix of the layer; Neighbor nodes in the inner view No. The trainable weight matrix of the layer; , For the inner view The trainable bias of the layer; is the normalization constant; is the attention coefficient in the inner view, i.e., the nodes in the same drug and The correlation of ; T is the transpose.

[0104] Step three, based on the internal drug characteristics of the two drugs and combined with the bipartite graph, the intermolecular view information of the two drugs is extracted, and the potential associations between different drug atoms are learned, so as to obtain the intermolecular information of the corresponding drugs related to DDI, that is, the external drug characteristics.

[0105] Different from the internal view learning module, the external view learning module uses the inter-view adjacency matrix A, which is a bipartite graph composed of two drugs. The interaction between the two drugs is defined and captured through the joint attention mechanism. In the constructed bipartite graph, each atom of one drug is regarded as a node, while all atoms of the other drug form another set of nodes. Through inter-view learning, this embodiment is able to identify and encode the interaction between two different drugs, which is crucial for understanding the joint effect of drug combinations. Using this method, the model is able to learn and reveal the intrinsic connections and potential relationships between drug pairs. Specifically, drug The atoms in the The atoms in send or receive attention messages, and the process formula is:

[0106] ;

[0107] ;

[0108] in, is a nonlinear activation function, Indicates the exterior view Node in layer The updated representation is External characteristics of the drugs in the layer; For nodes in the external view No. The trainable weight matrix of the layer; Neighbor nodes in the external view No. The trainable weight matrix of the layer; and is the trainable bias term in the exterior view, Represents a bipartite graph Midpoint The set of neighbor nodes of , i.e., all atoms in another drug; is the attention coefficient in the inter-view layer, i.e., the atomic distance between two drugs in the outer view and of relevance.

[0109] is a normalization constant. In this formula, its value is . Attention coefficient Consider two drug atoms and Through the inter-view learning method, the model can not only understand the effects of individual drugs, but also gain insight into the synergistic or antagonistic effects of different drug combinations, providing a deeper perspective on the interaction relationship.

[0110] Step 4: After obtaining the feature representation learned in the dual-view interaction, the feature representation is used to aggregate at the node level and the graph level to fuse the internal features of the drug with the external features of the drug. The structural features of Nodes in the external view Representation Connect them, integrate them through MLP, and output the new node embedding. The node-level representation of the layer is a fusion of the inner view representation and the outer view representation. The formula is:

[0111] ;

[0112] in, For the Node in layer The new node embedding is MLP, which is a multi-layer perceptron. The integrated node representation is is fed to the next layer, providing multi-view information for each node.

[0113] Step 5: The new node embeddings corresponding to the two drugs output by each layer of MSMDL encoder are input into the next layer of MSMDL encoder and self-attention graph pooling layer respectively. The self-attention graph pooling layer is used as the readout function to generate a graph-level representation and obtain the substructure feature embedding of each layer.

[0114] Specifically, when aggregating node representations layer by layer, a graph-level representation is also generated through the readout layer. This embodiment uses the Self-Attention Graph Pooling (SAGPooling) layer as the readout function. The importance of each node in the drug molecule graph is determined by node embedding and its topological information. The formula is:

[0115] ;

[0116] ;

[0117] in, It is a nonlinear activation function, such as the sigmoid activation function; is the adjacency matrix, obtained through the bipartite graph; For drugs No. The embedding representation of all nodes in the layer; n is the The number of nodes in the layer; is used The normalization function passes the adjacency matrix A and X through the graph neural network The result obtained by normalizing the calculated embedding; Indicates drug The result after the k-th layer node aggregation and self-attention map pooling layer, that is, the drug Substructure feature embedding at the kth layer.

[0118] The substructure feature embedding of each layer corresponding to the two drugs output by the self-attention graph pooling layer is expressed as and ;in, 、 Respectively represent the layer number variables of the MSMDL encoder corresponding to the two drugs; for example, the first layer is , , the second layer is: , The third layer is , .

[0119] S4, the substructure feature embeddings of each layer are taken together as the input of the common attention mechanism layer to evaluate the contribution of different substructures to the prediction of drug-pair interaction and obtain the importance of the interaction between the substructure feature embeddings of each layer.

[0120] In this embodiment, the substructure feature embeddings output by each layer are used as the input of the common attention mechanism layer. and The attention weight between , to measure the importance of the interaction between substructure feature embeddings, to evaluate the contribution of different substructures to the prediction of drug-drug interaction, i.e., the contribution of DDI, and to obtain the importance of the interaction between substructure feature embeddings , whose expression is:

[0121] ;

[0122] in, and There are two drugs and exist The substructure feature embedding of the layer, is the trainable weight vector, and The trainable weight matrices corresponding to the two drugs are shown in Figure 1, and T is the transpose. Through different weight matrices, the model can prevent similar substructures from getting too high scores, which helps to simplify the learning process and avoid unnecessary complexity. Considering that non-interactive drugs may get negative scores, we use As an activation function, it produces positive and negative outputs.

[0123] S5, according to the embedding of the substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type, the decoder performs probability calculation and outputs the predicted probability of interaction between the two drugs, namely, the DDI predicted probability, to obtain the final prediction result.

[0124] Specifically, in this embodiment, the trainable representation matrix Used to adjust the joint embedding representation in order to learn better features during the training process. A trainable representation matrix means that in a machine learning model, a matrix is ​​initialized and set to be trainable, and then automatically adjusted through machine learning.

[0125] Embedding based on the fusion structural features of the two drugs and ,importance With the trainable representation matrix , through the decoder, the DDI prediction probability is obtained, the formula is:

[0126] ;

[0127] in, Represents a triple ( ), represents a nonlinear activation function, K represents the total number of layers of the MSMDL encoder, and T represents the transpose.

[0128] Through the above steps, a drug pair interaction prediction model based on multi-layer soft mask dual-view learning is obtained. Finally, the model is trained, tested and evaluated. When the evaluation result reaches the preset indicator, a trained drug pair interaction prediction model based on multi-layer soft mask dual-view learning, namely, the MSMDL-DDI model, is obtained.

[0129] The triple consisting of the molecular view data of the two drugs to be predicted and their interaction relationship types is input into the MSMDL-DDI model to obtain the DDI prediction probability of the two drugs, thereby obtaining the interaction prediction results.

[0130] In another preferred embodiment, as shown in Table 1, the DDI prediction performance of the model MSMDL-DDI of the present invention and other existing models on DrugBank and Twosides datasets in the direct push scenario is listed. Due to limited computing resources, we cannot run the MHCADDI model on the Twosides dataset, so we refer to the AUROC results of the original article.

[0131] This example uses four common evaluation indicators: prediction accuracy (ACC), receiver operating characteristic curve (Area Under the Receiver Operating Characteristic Curve, AUROC), area under the average precision-recall curve (Area Under the Precision-Recall Curve, AUPRC) and F1 score (indicating the average of accuracy and recall) to comprehensively evaluate the prediction performance of the model.

[0132] Table 1. Comparative experimental results of the proposed MSMDL-DDI and other methods in the direct push scenario on the DrugBank and Twosides datasets

[0133]

[0134] Dataset Description:

[0135] DrugBank is a database that integrates bioinformatics and cheminformatics resources, providing detailed drug data and comprehensive molecular information on drug targets and their mechanisms. The database is maintained by the University of Alberta and aims to help researchers better understand and predict the mechanism of action of drugs. DrugBank contains information on drug entries, drug targets (such as protein targets), chemical structures of drugs, pharmacological characteristics, transporter-related information, drug side effects, etc. In addition, it also provides drug classification information to which the drug belongs and links to other databases (such as the PDB database and the KEGG database) to further analyze the detailed information of the drug.

[0136] The Twosides dataset is a dataset focused on drug-drug interactions (DDIs), which contains a large amount of information on drug pairs and their interactions, such as synergy and antagonism between drugs. These data are of great significance for studying drug combination therapy, predicting drug interactions, and optimizing drug treatment regimens.

[0137] Model Description:

[0138] MR-GNN: Multi-Resolution and Dual Graph Neural Network for Predicting Structured Entity Interactions, a multi-resolution and dual graph neural network, is used to predict interactions between structured entities. It improves the limitations of traditional graph neural networks by introducing multi-resolution architectures and dual graph-state LSTMs.

[0139] MHCADDI: multi-head co-attentive drug-drug interactions. MHCADDI constructs a drug interaction network and uses a graph collaborative attention mechanism to capture the complex associations between drugs, while considering drug characteristics and their relationships in the network.

[0140] SSI-DDI: Substructure-Substructure Interactions for Drug-DrugInteraction Prediction, is a substructure interaction-based deep learning framework for predicting drug-drug interactions. The model decomposes the DDI prediction task between two drugs into identifying pairwise interactions between their respective substructures.

[0141] GAT-DDI: Graph attention networks for Drug-Drug Interaction Prediction, GAT-DDI directly uses graph attention networks for drug modeling and DDI prediction. This model effectively captures the interactions between drug molecules by assigning different attention weights to different nodes (atoms in drug molecules).

[0142] GMPNN-CS: Gated message passing neural network-CS. This study designs a gated message passing neural network that focuses on learning chemical substructure features of different scales from the graph structure of drug molecules to predict DDI between drugs.

[0143] SA-DDI: substructure attention mechanism for Drug-Drug Interaction Prediction. SA-DDI adopts the attention mechanism to propose a substructure-aware graph neural network, which can adaptively extract key substructures in drug molecules and emphasize the most important molecular fragments for DDI prediction, thereby improving the accuracy and efficiency of DDI prediction.

[0144] MASM-DDI: Multi layer Adaptive Soft Mask Graph Neural Network for Drug-Drug Interaction Prediction. MASM-DDI uses a multi-layer soft mask GNN to adaptively extract drug substructures, and uses the attention mechanism to explore the potential features of substructures.

[0145] MSMDL-DDI: the method of the present invention.

[0146] It can be found that: MSMDL-DDI (the method of the present invention) performs particularly well on the Twosides dataset, and compared with the second-ranked SA-DDI method, the ACC and F1 scores are relatively improved by 10.89% and 9.05%, respectively. In addition, the AUROC values ​​of MSMDL-DDI on the DrugBank and Twosides datasets reach 99.85% and 99.28%, respectively, and the AUPRC reaches 99.83% and 99.10%, respectively.

[0147] This result fully proves that MSMDL-DDI can effectively distinguish interacting drugs from non-interacting drugs in direct-induction experiments and can achieve DDI prediction with high accuracy. The remarkable experimental performance further confirms the powerful representation ability of MSMDL-DDI and proves the effectiveness of the innovative fusion soft mask mechanism and dual-view substructure learning module. In order to display the experimental results more intuitively, we visualized the experimental results.

[0148] As shown in Table 2, in another preferred embodiment, in the inductive scenario, because the Twosides dataset has false positive samples, that is, some drug pairs do not actually interact with each other, which may affect the evaluation accuracy of the inductive experiment, this embodiment only chooses to use the DrugBank dataset for the experiment.

[0149] Table 2. Experimental results of MSMDL-DDI and other models in S1 and S2 tasks in the inductive scenario

[0150]

[0151] It can be observed from Table 2 that in the inductive experiment scenario, the scores of all evaluation indicators are significantly lower than those in the direct experiment scenario. Although the overall performance of the model has declined, the performance of MSMDL-DDI in the inductive experiment is similar to that in the direct experiment, and the method of the present invention also achieves the best performance in the three evaluation indicators.

[0152] In addition, the method of the present invention significantly outperforms the second-ranked algorithm in performance, and the AUROC indicators on the S1 and S2 schemes achieved relative improvements of 5.28% and 2.07%, respectively.

[0153] The results show that the MSMDL-DDI strategy is very effective. The method of the present invention not only focuses on the substructure characteristics of a single drug, but also learns the interactions between drug molecular views from drug combinations, making up for the lack of existing prior knowledge and unknown drug interaction information.

[0154] In another embodiment, in order to study the source of the model performance improvement, this study further conducts a comprehensive ablation study to observe the importance of each component of MSMDL-DDI. To this end, as shown in Table 3, this embodiment performs ablation experiments by removing different components in MSMDL-DDI as variants.

[0155] w / o_ASMGNN: The MSMDL-DDI model removes the multi-layer soft mask network with shared weights. This ablation variant verifies the effectiveness of learning the shared substructure of the two drugs.

[0156] w / o_Inter: The MSMDL-DDI model removes the inter-view learning module. The substructure representation is derived only from the intra-view representation. This is used to verify the importance of inter-view learning for substructure extraction and DDI prediction.

[0157] w / o_Intra: The MSMDL-DDI model removes the intra-view learning module. The substructure representation is derived only from the inter-view representation. This demonstrates the importance of inter-view learning and the importance of performing both inter-view and intra-view feature extraction simultaneously.

[0158] w / o_CoAtt: A model that removes the co-attention operation from the decoder of the MSMDL-DDI model. DDI predictions are still calculated based on pairwise interactions between drug-pair substructures, but the combination of all interactions contributes equally to the DDI prediction.

[0159] Table 3. Comparison results of MSMDL-DDI ablation experiments on DrugBank dataset

[0160]

[0161] As shown in Table 3, on the DrugBank dataset, the full MSMDL-DDI model performs best among all variants, indicating that all proposed modules are effective. Among them, the performance of the w / o_Inter and w / o_Intra variants has declined to some extent, indicating that jointly learning drug-drug representations from a dual-view perspective is more beneficial than learning from each view separately. The soft mask module contributes the most to MSMDL-DDI, because the w / o_ASMGNN ablation variant has the most significant performance degradation. The prediction performance of the soft mask network variant w / o_ASMGNN without shared weights has declined, indicating that the main network learns a common representation of all drugs by sharing parameters, which is quite beneficial to the improvement of model performance. Although the w / o_CoAtt ablation variant has little impact on the performance of the model, it is a key component of the model decoding and an indispensable module for MSMDL-DDI.

[0162] Table 4. Comparison results of MSMDL-DDI ablation experiments on the Twosides dataset

[0163]

[0164] As shown in Table 4, on the Twosides dataset, the full MSMDL-DDI model performs best among all variants, indicating that all proposed modules are effective. Different from the ablation results on the DrugBank dataset, the performance of the two performance variants w / o_Inter and w / o_Intra has dropped most significantly, indicating that joint learning of drug-drug representation from a dual-view perspective is most important for the Twosides dataset. Although, probably due to the large amount of data in the Twosides dataset, the soft mask network variant w / o_ASMGNN without shared weights has a smaller impact on the performance of the model on the Twosides dataset, as a key component of substructure extraction, it is an indispensable module for MSMDL-DDI. The w / o_CoAtt ablation variant shows a more important role on the Twosides dataset, which fully demonstrates the effectiveness of this component for MSMDL-DDI. The above experimental analysis shows the importance of each module in MSMDL-DDI and their contribution to the model performance. These findings not only verify the effectiveness of the proposed method, but also provide valuable references for future research in the field of drug interaction prediction.

[0165] To verify the predictive ability of the model, we used part of the data in DrugBank as a training set to train the MSMDL-DDI model. Subsequently, we selected drug pairs that were not included in the training set to verify the model's prediction results for these drug pairs. According to the scores predicted by the MSMDL-DDI model, the possibility of DDI between drug pairs was quantified, and the higher the score, the greater the probability of DDI. We sorted the predicted probabilities of these new drug pairs in descending order and selected the top 20 drug pairs with the highest prediction scores for further verification.

[0166] As shown in Table 5, we queried the interaction information between these new drug pairs through the DrugBank database (go.drugbank.com) and the Drug InteractionChecker website (Drugs.com), and recorded the corresponding interaction descriptions. Among the first 20 new drug pairs, we found that the interaction information of 16 drug pairs was recorded in the database. For example, the interaction between the drugs "levochlorobenzenethiol" and "sulpiride" was described as "levochlorobenzenethiol may enhance the antipsychotic activity of sulpiride".

[0167] For those drug pairs for which no interaction information was found in the validation resources, we marked them as "N / A". This does not mean that there is no interaction between these drug pairs, but it indicates that the possibility of their interaction needs to be further verified by experimental methods.

[0168] Table 5. Top 20 drug pairs data for new drug prediction using MSMDL-DDI

[0169]

[0170]

[0171] The results of this case study show that the MSMDL-DDI model can effectively predict potential DDIs between new drugs. In addition, among these 20 new drug pairs, 5 pairs contain the drug "phenylephrine". The interactions between these 5 drug pairs have been experimentally confirmed, and this finding suggests that "phenylephrine" may have a tendency to interact with other drugs.

[0172] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0173] The multi-layer soft mask dual-view learning method MSMDL-DDI proposed in this paper extracts substructure features related to DDI prediction through a multi-layer soft mask neural network, and uses a dual-view learning strategy to integrate local and global information to form a comprehensive drug pair representation. This method makes full use of the representation information generated by the neural network at different levels, and adaptively focuses on the key substructures in the drug molecule graph through a multi-layer soft mask mechanism, while ignoring noise, thereby improving the model's sensitivity and robustness to drug molecular structures.

[0174] In addition, MSMDL-DDI adopts a dual-view learning strategy to effectively integrate local and global information, significantly improving the accuracy of drug interaction prediction. By focusing on the internal characteristics of individual drug molecules and the interactions between drugs, the generalization ability of the model is enhanced. Multiple experiments have demonstrated the effectiveness of MSMDL-DDI in the DDI prediction task. The results show that this method exhibits strong generalization ability when dealing with unseen drug molecules and significantly improves the accuracy of DDI prediction. MSMDL-DDI not only provides new technical means for drug safety assessment and new drug development, but also provides new perspectives and ideas for research in the field of DDI prediction.

[0175] The present invention improves the accuracy of DDI prediction by mining deep-level information and comprehensive drug molecular structure information in GNN.

[0176] The present invention uses a multi-layer soft mask graph neural network to extract substructure related information from the original drug molecule graph, and learns the sequence representation of the substructure through a multi-layer adaptive soft mask graph neural network, thereby obtaining the drug chemical substructure that is critical to DDI prediction, and can improve the drug molecule representation capability without increasing additional training costs. This directly solves the "key substructure identification challenge" problem and achieves accurate integration of drug features without being disturbed by a large amount of irrelevant information.

[0177] The present invention solves the problem of not effectively combining the intrinsic properties of drug molecules with the interaction information between drug pairs. Starting from the perspective of drug molecule information, in order to obtain global and local drug interaction information, MSMDL-DDI adopts a dual-view learning strategy to learn the drug intermolecular view and the intramolecular view, combining the information inside the drug molecule and between the drug molecule pairs, and then obtains DDI-related information.

[0178] The embodiment of the present invention provides two task scenarios, direct induction and inductive induction, and conducts a large number of experiments on real-world drug datasets to verify the effectiveness of the MSMDL-DDI model proposed in this paper in DDI prediction task applications. The experimental results show that the MSMDL-DDI model has superior performance compared with other baseline models.

[0179] Embodiment 2

[0180] like Figure 4 As shown, the second embodiment of the present invention further provides a drug pair interaction prediction device based on soft mask dual-view learning, comprising:

[0181] An acquisition unit, used for acquiring the molecular view data of two drugs to be predicted and the types of their interaction relationships, and forming a triplet;

[0182] A bipartite graph unit, for generating a bipartite graph of drug pairs according to the triples;

[0183] An MSMDL encoder unit is used to input the triple and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output a new node embedding corresponding to the drug; wherein the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the new node embedding into the next layer of the MSMDL encoder and the self-attention graph pooling layer, respectively, and generates a graph-level representation through the self-attention graph pooling layer to obtain a substructure feature embedding;

[0184] The common attention mechanism unit is used to embed the substructure features of each layer together as the input of the common attention mechanism layer, evaluate the contribution of different substructures to the prediction of drug-pair interaction, and obtain the importance of the interaction between the substructure feature embeddings of each layer;

[0185] The decoding prediction unit is used to perform probability calculation based on the embedding of substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type through the decoder, and output the predicted probability of interaction between the two drugs, that is, the DDI predicted probability, to obtain the final prediction result.

[0186] Embodiment 3

[0187] The third embodiment of the present invention also provides a drug pair interaction prediction device based on soft mask dual-view learning, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the drug pair interaction prediction method based on soft mask dual-view learning as described above.

[0188] Embodiment 4

[0189] The fourth embodiment of the present invention further provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the above-mentioned drug pair interaction prediction method based on soft mask dual-view learning is implemented.

[0190] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0191] 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.

[0192] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0193] 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.

[0194] 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.

[0195] 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.

[0196] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0197] 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 drug pair interaction prediction method based on soft mask dual-view learning, characterized in that: include: S1, obtain the molecular view data of the two drugs to be predicted and their interaction relationship types, and form a triple; S2, generating a bipartite graph of drug pairs based on the triples; S3, inputting the triplet and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output new node embeddings corresponding to the drug; wherein the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the new node embeddings into the next layer of the MSMDL encoder and the self-attention graph pooling layer, respectively, and generates a graph-level representation through the self-attention graph pooling layer to obtain a substructure feature embedding; The MSMDL encoder includes a soft mask graph neural network ASMGNN, an internal view learning module and an external view learning module; wherein, The soft mask graph neural network ASMGNN extracts key substructure features of the corresponding drugs from the molecular view data of the two drugs in the input triplet through a shared weight strategy, and updates the feature representation of each node in the molecular view; The internal view learning module, by combining the graph attention mechanism, evaluates the contribution of the key substructure features of the two drugs in the DDI prediction task, mines the drug feature information that is crucial to the prediction results, and obtains the molecular internal information of the drugs related to DDI, i.e., the internal features of the drugs, by deeply extracting the key substructure features of the drugs; The external view learning module captures the intermolecular view information of the two drugs through a common attention mechanism based on the internal drug features of the two drugs and the bipartite graph, and learns the potential associations between different drug atoms, thereby obtaining the intermolecular information of the drugs related to DDI, i.e., the external drug features; S4, the substructure feature embeddings of each layer are used together as the input of the common attention mechanism layer to evaluate the contribution of different substructures to the prediction of drug-pair interaction and obtain the importance of the interaction between the substructure feature embeddings of each layer; S5, according to the embedding of the substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type, the decoder performs probability calculation and outputs the predicted probability of interaction between the two drugs, namely, the DDI predicted probability, to obtain the final prediction result.

2. A drug pair interaction prediction method based on soft mask dual view learning according to claim 1, characterized in that , the S3 is specifically: Step 1: Each node of the molecular view data of the two drugs in the input triplet is passed through the soft mask graph neural network of the MSMDL encoder to learn the structural feature representation of its neighbor nodes and the previous layer based on the shared weights, and update the node. The expression is: ； ； in, Representation Node In the Structural feature representation of layers; Represents neighbor nodes In the Structural feature representation of the layer; Representation Node Structural feature representation in the previous layer; Represents neighbor nodes Structural feature representation at the previous layer; represents a nonlinear activation function; Indicates a connection operation; Representation Node No. The trainable weight matrix of the layer, represents a real number, Represents the dimension of the trainable parameter matrix, i.e., the node Dimensions; Representation Node The set of neighbor nodes of Representation Node The first layer mask value; Represents neighbor nodes The first layer mask value; and are the linear mappings in updating node representation and message passing process respectively; Representation Node The learned mask value of the previous layer; Represents neighbor nodes The learned mask value of the previous layer; MLP is a two-layer perceptron with an output feature dimension of 1. The last layer of the perceptron is Activation function; Step 2: The interaction information between atoms in the drug molecule view is obtained through the internal view learning module of the MSMDL encoder. Different weights are assigned to each atom in combination with the graph attention mechanism. The attention information of the neighboring nodes of each node in the drug molecule view is aggregated to update the representation of each node itself. The expression is: ； ； in, For the Node in layer The structural characteristics of Internal characteristics of drugs in layers; For Node No. The trainable weight matrix of the layer; Neighbor nodes in the inner view No. The trainable weight matrix of the layer; , For the inner view The trainable bias of the layer; is the normalization constant; is the attention coefficient in the inner view, i.e., the nodes in the same drug and The correlation of ; T is the transpose; Step 3: In the external view learning module of the MSMDL encoder, the bipartite graph is combined to capture the view information between two drug molecules through a joint attention mechanism to learn the potential associations between different drug atoms, i.e., drug The atoms in the The atoms in send or receive attention messages, and the process expression is: ； ； in, Indicates the first Node in layer The updated representation is External characteristics of the drugs in the layer; For nodes in the external view No. The trainable weight matrix of the layer; Neighbor nodes in the external view No. The trainable weight matrix of the layer; and is the trainable bias term in the exterior view, Represents a bipartite graph Midpoint The set of neighbor nodes of , i.e., all atoms in another drug; is the attention coefficient in the inter-view layer, i.e., the atomic distance between two drugs in the outer view and relevance; Step 4: Fuse the internal features of the drug with the external features of the drug to obtain a new node embedding of the corresponding drug, that is, embedding the node in the internal view The structural features of Nodes in the external view Representation Connect them, integrate them through MLP, and output the new node embedding, expressed as: ； in, For the Node in layer The new node embedding, MLP is a multi-layer perceptron; Step 5: Take the drug and medications The new node embedding output by each layer is input into the next layer MSMDL encoder and self-attention graph pooling layer respectively. The self-attention graph pooling layer is used as the readout function to generate a graph-level representation and obtain the substructure feature embedding of each layer. The formula is: ； ； in, is an adjacency matrix obtained through the bipartite graph; For drugs No. The embedding representation of all nodes in the layer; n is the The number of nodes in the layer; To pass The normalization function passes the adjacency matrix A and X through the graph neural network The calculated embedding is normalized to obtain the result, Indicates drug The result after the k-th layer node aggregation and self-attention map pooling layer, that is, the drug Substructure feature embedding at the kth layer.

3. The method for predicting drug pair interactions based on soft mask dual-view learning according to claim 1, characterized in that , the S4 is specifically: Two drugs and The outputs after the three-layer MSMDL encoder and the self-attention graph pooling layer are expressed as and ,in, , Represents two drugs and The corresponding layer number of the MSMDL encoder; The substructure feature embeddings corresponding to each layer are used as the input of the common attention mechanism layer, and the substructure feature embeddings are measured by calculating the attention weights of the substructure feature embeddings of each layer. and The importance of interactions , to evaluate the contribution of different substructures to the prediction of drug-drug interaction, i.e., DDI, and to obtain the importance of the interaction between substructure feature embeddings ,importance The expression is: ； in, and For two drugs and In the The substructure feature embedding of the layer, is the trainable weight vector, and are the trainable weight matrices corresponding to the two drugs, is the activation function and T is the transpose.

4. The method for predicting drug pair interactions based on soft mask dual-view learning according to claim 1, characterized in that , the S5 is specifically: The decoder is a common attention scoring function, and its output layer is a nonlinear activation function for converting the output into a probability value between 0 and 1; Embedding the substructure features of each layer corresponding to the two drugs and , and the corresponding importance , trainable representation matrix , get the prediction result, the formula is: ； in, Represents a triple ( )’s predicted probability; represents a nonlinear activation function; K represents the total number of layers of the MSMDL encoder; T represents transposition; , Drugs In the Substructure feature embedding of layers; is the trainable representation matrix.

5. A drug pair interaction prediction device based on soft mask dual view learning, characterized in that: include: An acquisition unit, used for acquiring the molecular view data of two drugs to be predicted and the types of their interaction relationships, and forming a triplet; A bipartite graph unit, for generating a bipartite graph of drug pairs according to the triples; The MSMDL encoder unit is used to input the triple and the bipartite graph into a multi-layer soft mask dual-view learning encoder, i.e., an MSMDL encoder, to extract and learn features within and between molecules, and output a new node embedding corresponding to the drug; wherein the MSMDL encoder is K layers, each layer of the MSMDL encoder outputs a new node embedding corresponding to the drug, and inputs the new node embedding of the next layer of the MSMDL encoder and the self-attention graph pooling layer respectively, and generates a graph-level representation through the self-attention graph pooling layer to obtain a substructure feature embedding; the MSMDL encoder includes a soft mask graph neural network ASMGNN, an internal view learning module, and an external view learning module; wherein, The soft mask graph neural network ASMGNN extracts key substructure features of the corresponding drugs from the molecular view data of the two drugs in the input triplet through a shared weight strategy, and updates the feature representation of each node in the molecular view; The internal view learning module, by combining the graph attention mechanism, evaluates the contribution of the key substructure features of the two drugs in the DDI prediction task, mines the drug feature information that is crucial to the prediction results, and obtains the molecular internal information of the drugs related to DDI, i.e., the internal features of the drugs, by deeply extracting the key substructure features of the drugs; The external view learning module captures the intermolecular view information of the two drugs through a common attention mechanism based on the internal drug features of the two drugs and the bipartite graph, and learns the potential associations between different drug atoms, thereby obtaining the intermolecular information of the drugs related to DDI, i.e., the external drug features; The common attention mechanism unit is used to embed the substructure features of each layer together as the input of the common attention mechanism layer, evaluate the contribution of different substructures to the prediction of drug-pair interaction, and obtain the importance of the interaction between the substructure feature embeddings of each layer; The decoding prediction unit is used to perform probability calculation through the decoder according to the embedding of the substructure features of each layer corresponding to the two drugs and their importance and interaction relationship type, and output the interaction prediction probability of the two drugs, that is, the DDI prediction probability, to obtain the final prediction result.

Images