Drug-drug interaction prediction method based on drug flow subgraph

Abstract

The invention discloses a drug-drug interaction prediction method based on a drug flow sub-graph, and relates to the technical field of bioinformatics, and the method comprises the steps: data preparation: collecting a reference data set including drug-drug interaction, and introducing an external knowledge graph for adaptability preprocessing; constructing a model: constructing a drug-drug interaction prediction model, and training by using the preprocessed reference data set; and effect prediction: inputting a target drug into the drug-drug interaction prediction model, and outputting a drug-drug interaction prediction result through the drug-drug interaction prediction model. The structure and semantic information of the drug flow sub-graph are fully utilized, accurate prediction of drug interaction is realized, and the prediction efficiency is improved. And the interpretability of the drug-drug interaction prediction model is improved.

Description

Technical Field

[0001] The present application relates to the technical field of bioinformatics, and in particular to a method for predicting drug-drug interactions based on a drug flow subgraph. Background Art

[0002] Chronic diseases, infectious diseases, and the elderly are often complex, and single-drug treatments are often ineffective. Combination therapy not only significantly improves efficacy but also effectively reduces drug resistance, offering a more optimal solution for the treatment of recurrent illnesses and multiple comorbidities. However, the combined use of multiple drugs also poses significant challenges to drug safety management, increasing the risk of unexpected drug-drug interactions (DDIs). For example, certain drugs can induce significant changes in pharmacokinetics or pharmacodynamics by inducing key enzymes or competitively binding to specific targets, ultimately leading to serious clinical consequences. Therefore, identifying potential DDIs is crucial for drug research and protecting patient health. With the accelerating pace of drug development and the increasing demand for personalized medicine, the global drug population is rapidly expanding. While traditional in vitro assays and clinical trials provide important insights for identifying DDIs, they lack comprehensive coverage of all potential interaction scenarios due to their lengthy and costly testing cycles. Furthermore, the potential DDI risk of novel drugs is particularly challenging to predict due to limited clinical data. Therefore, developing more efficient DDI prediction methods to address complex drug interaction networks has become a key issue in contemporary pharmaceutical science. This research direction not only has far-reaching significance for improving the efficiency of drug research and development, but also provides a scientific basis for achieving safe and accurate combination drug use.

[0003] In early research on DDIs, researchers used traditional machine learning methods to predict DDIs. These methods are generally based on the similarity hypothesis. The underlying assumption is that if drug A interacts with drug B, drugs similar to drug A are likely to interact with drug B, and vice versa. Although traditional machine learning methods have made significant progress in DDI prediction, they still have several limitations. First, machine learning models typically rely on high-quality feature inputs, and the feature construction process is often complex and dependent on domain knowledge. Furthermore, model performance can be significantly affected when data is high-dimensional or missing. Furthermore, traditional machine learning methods struggle to fully explore potential interaction patterns when integrating multi-source, heterogeneous data. To address these issues, custom strategies or specific assumptions are often employed for DDI prediction. These methods are not limited to a fixed similarity framework and therefore exhibit greater adaptability in model design and application scenarios. Although DDI prediction methods based on custom strategies have overcome some of the limitations of traditional machine learning to some extent, they still face numerous challenges. First, these methods are limited in their ability to integrate multi-source, heterogeneous data, making it difficult to fully explore potential information. Second, they perform poorly when faced with sparse or missing data, which affects prediction accuracy. In addition, due to the lack of strong modeling capabilities for complex nonlinear relationships, the predictive effect of such methods is limited, and model design and optimization are often constrained by established frameworks and lack flexibility.

[0004] To address these issues, researchers have gradually introduced deep learning technology into the field of DDI prediction in recent years. As a pioneering computational model using deep learning technology in the field of DDI prediction, most deep learning-based methods only consider the physicochemical properties of molecules, ignoring the potential correlation between other modal features of drugs and DDI events. To effectively integrate multimodal or multi-perspective features of drugs, many effective methods have been developed. Although methods based on multimodal or multi-perspective features have demonstrated powerful feature extraction and multimodal fusion capabilities in the field of DDI prediction, they are still limited in their ability to capture the complex logic and reasoning of drug relationships. Most of these methods rely on data-driven feature learning, which makes it difficult to fully explore the explicit semantic associations and contextual reasoning between drugs. In addition, model interpretability remains a challenge, which is crucial for the study of drug-drug interaction mechanisms. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a drug-drug interaction prediction method based on a drug flow subgraph to solve the problems existing in the prior art:

[0006] (1) The problem of being unable to effectively integrate biomedical knowledge graphs, DDI networks, and drug internal structural characteristics;

[0007] (2) Traditional models lack interpretability when predicting drug-drug interactions;

[0008] (3) Traditional models lack certain generalization when facing complex networks and large knowledge graph networks, and the positive and negative samples generated by traditional contrastive learning methods lack diversity.

[0009] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a drug-drug interaction prediction method based on a drug flow subgraph, comprising:

[0010] S1. Data preparation: Collect a benchmark dataset including drug-drug interactions and introduce an external knowledge graph for adaptive preprocessing;

[0011] S2. Model construction: Build a drug-drug interaction prediction model and train it using the preprocessed benchmark dataset;

[0012] The drug-drug interaction prediction model includes: a drug substructure feature extraction module for extracting local embeddings of drug substructures; a combination network universal embedding construction module for obtaining universal embeddings of combination networks through GraphSAGE; a knowledge subgraph construction module for constructing a drug flow subgraph and optimizing it using the universal embedding of the combination network and the local embeddings of drug substructures to obtain a knowledge subgraph; and a prediction module for obtaining drug-drug interaction prediction results based on the knowledge subgraph.

[0013] S3. Effect prediction: The target drug is input into the drug-drug interaction prediction model, and the drug-drug interaction prediction model outputs the drug-drug interaction prediction result.

[0014] Furthermore, the step of extracting the final representation of the drug by the drug substructure feature extraction module includes:

[0015] S201. Considering the atomic representations of the drug molecule graph as nodes and the chemical bond representations of the drug molecules as edges, and extracting the bond-level features of all edges in the drug molecule graph through a graph neural network;

[0016] S202, using an activation function to calculate the importance score of each bond-level feature in the drug molecule graph, and sending the obtained importance score to a pooling layer to obtain a pooled feature of the drug molecule graph;

[0017] S203. Based on the pooled features of the drug molecule graph, a multi-layer attention mechanism and feature aggregation are used to capture important information about the substructure in the drug molecule graph and obtain the final bond-level feature representation.

[0018] S204. Introducing the edge features of the drug molecule graph and combining them with the final bond-level feature representation, further capture the relationship between the substructures of the drug molecule graph and obtain the local embedding of the drug substructure.

[0019] Further: In S201, the bond-order characteristics of the edges in the drug molecule graph Expressed as:

[0020]

[0021] in, Represents the characteristics of the starting node, Represents the characteristics of the terminal node, is a node and Characteristics of the bond between them;

[0022] In S202, the pooling features of the drug molecule graph Expressed as:

[0023]

[0024] Where, represents the pooling layer, Indicates the number of layers of the current graph neural network, Represents the matrix composed of all key-level features in the current layer t of the graph neural network, Represents the adjacency matrix of the current layer graph;

[0025] In S203, the final bond-level feature representation Expressed as:

[0026]

[0027] in, represents the learnable matrix, represents the bias term, represents the tanh activation function, Representative The weight vector of the layer, is the number of network layers of the substructure extraction module, Representative Node The original feature representation of Representative Node The set of all neighbor nodes of is the nonlinear function of the multilayer perceptron; Representative Layer Node and The bond-level feature representation between ;

[0028] In S204, local embedding of drug substructures Expressed as:

[0029]

[0030]

[0031] in, and Both represent learnable matrices, Normalization of the representative layer, represents the edge features of the introduced drug molecule graph, express Dimensions, Represents multi-head splicing, represents the number of attention heads, Representative Layer Node The substructure representation of

[0032] In S205, the final representation of the drug Expressed as:

[0033]

[0034] in, Representative Layer Node The substructure of and Both represent feedforward neural networks, is a learnable matrix, is layer normalization, Represents the activation function.

[0035] Furthermore, the universal embedding representation of the combined network obtained by the combined network universal embedding building module is:

[0036]

[0037] in, Representation node No. The embedding representation of the layer, Representation node Features in the combined graph, Representative Node In the The neighbor set of the layer, represents the learnable weight matrix, represents the bias term, represents the activation function, Represents an aggregate function.

[0038] Further: Drug flow subgraph constructed by knowledge subgraph construction module Expressed as };in, Represents a collection of nodes, represents a set of edge types, Represents the set of edges connecting nodes, the drug flow subgraph Indicated by point to and the length does not exceed A directed subgraph of Composition, among which, ( ... ) represents the edge of the drug flow subgraph, ( ... ) represents a node of the drug flow subgraph.

[0039] Furthermore, the steps of constructing the knowledge subgraph by the knowledge subgraph construction module include:

[0040] S211, concatenating the local embedding of the drug substructure with the universal embedding of the combined network to initialize the drug features of the drug flow subgraph;

[0041] S212, evaluating the correlation between nodes of the drug flow subgraph according to the drug characteristics of the drug flow subgraph, and obtaining a correlation score;

[0042] S213, updating the connection strength of the edges in the drug flow subgraph according to the correlation score;

[0043] S214. Aggregate features of neighbor nodes by connecting edges to obtain a knowledge subgraph.

[0044] Further: In S212, relevance score Expressed as:

[0045]

[0046] in, Representing relationships Learnable relation embeddings, represents the exponential function, Indicates taking the absolute value, represents a multilayer perceptron, represents the connection along the last dimension, and Represent nodes in the drug flow subgraph and In the Node embedding of the layer;

[0047] In S213, the connection strength of the edge Expressed as:

[0048]

[0049] in, represents the hyperparameter, represents the activation function, represents the initial connection strength of the edges in the drug flow subgraph;

[0050] In S214, Embedding of layer knowledge subgraphs Expressed as:

[0051]

[0052] in, represents the learnable parameters, represents the operation of feature aggregation for each node in the drug flow subgraph, Indicates the Layer-wise knowledge graph embedding.

[0053] Furthermore: the prediction module obtains the final embedding representation of the drug pair based on the knowledge subgraph , as the drug-drug interaction prediction result, the final prediction result of the drug pair Expressed as:

[0054]

[0055] in, Respectively represent the nodes after the embedding transformation of the knowledge subgraph and nodes The final expression of the drug, Represents a multilayer perceptron.

[0056] Furthermore: The drug-drug interaction prediction model also includes a contrastive learning positive and negative sample construction module, which is used to generate positive samples through subgraph node masking and node perturbation strategies, generate negative samples by randomly selecting nodes in the combined graph to replace the original graph nodes and randomly changing the original graph node connections, and bring similar samples closer and push different samples farther away to learn the embedding space.

[0057] Further: Loss function of drug-drug interaction prediction model Expressed as:

[0058] =

[0059]

[0060]

[0061]

[0062] in, Represents label losses for different prediction tasks, including multi-class drug-drug interaction prediction losses and multi-label prediction loss , is a hyperparameter, is the loss of positive and negative sample pairs; and are the true label and the predicted label respectively, and Represent the similarity scores of positive samples and negative samples respectively, represents the temperature parameter, and represent the number of positive samples and negative samples respectively, is the positive sample count identifier, is the negative sample count identifier.

[0063] The beneficial effects of the present invention are:

[0064] 1. Achieved efficient integration and utilization of biomedical knowledge graphs, DDI networks, and drug substructure features, improving the predictive capability of DDI models;

[0065] 2. By fusion of multi-scale drug flow subgraph features, the model fully utilizes the structural and semantic information of drugs. Furthermore, through contrastive learning strategies, the model's ability to represent complex drug networks is enhanced, significantly improving the accuracy, reliability, and interpretability of DDI model predictions.

[0066] 3. By introducing the contrastive learning strategy, the representation ability and generalization performance of the DDI model are further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Flowchart of the drug-drug interaction prediction method based on drug flow subgraph;

[0068] Figure 2 An example diagram of a drug-drug interaction prediction model;

[0069] Figure 3 This is the ablation experiment result of the drug-drug interaction prediction model on the DrugBank benchmark dataset;

[0070] Figure 4 This is the result diagram of hyperparameter sensitivity analysis of drug-drug interaction prediction model;

[0071] Figure 5 Pathway-based interpretable graphs for drug-drug interaction prediction models;

[0072] Figure 6 A case study of drug structure for drug-drug interaction prediction models. DETAILED DESCRIPTION

[0073] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0074] like Figure 1 As shown, in one embodiment of the present invention, a drug-drug interaction prediction method based on a drug flow subgraph is provided, comprising:

[0075] S1. Data preparation: Collect a benchmark dataset including drug-drug interactions and introduce an external knowledge graph for adaptive preprocessing;

[0076] S2. Model construction: Build a drug-drug interaction prediction model and train it using the preprocessed benchmark dataset;

[0077] The drug-drug interaction prediction model includes: a drug substructure feature extraction module for extracting local embeddings of drug substructures; a combination network universal embedding construction module for obtaining universal embeddings of combination networks through GraphSAGE; a knowledge subgraph construction module for constructing a drug flow subgraph and optimizing it using the universal embedding of the combination network and the local embeddings of drug substructures to obtain a knowledge subgraph; and a prediction module for obtaining drug-drug interaction prediction results based on the knowledge subgraph.

[0078] S3. Effect prediction: The target drug is input into the drug-drug interaction prediction model, and the drug-drug interaction prediction model outputs the drug-drug interaction prediction result.

[0079] The drug molecular graph is composed of many basic substructures that express various physical and chemical properties. By extracting and integrating the drug substructure features, the final feature expression is made more comprehensive and rich. Specifically, the steps of the drug substructure feature extraction module to extract the final representation of the drug include:

[0080] S201. Considering the atomic representations of the drug molecule graph as nodes and the chemical bond representations of the drug molecules as edges, and extracting the bond-level features of all edges in the drug molecule graph through a graph neural network;

[0081] Bond-order characteristics of edges in drug molecule graphs Expressed as:

[0082]

[0083] S202, using an activation function to calculate the importance score of each bond-level feature in the drug molecule graph, and sending the obtained importance score to a pooling layer to obtain a pooled feature of the drug molecule graph;

[0084] Pooling features of drug molecule graphs Expressed as:

[0085]

[0086] Where, represents the pooling layer, Indicates the number of layers of the current graph neural network, Represents the matrix composed of all key-level features in the current layer t of the graph neural network, Represents the adjacency matrix of the current layer graph;

[0087] S203. Based on the pooled features of the drug molecule graph, a multi-layer attention mechanism and feature aggregation are used to capture important information about the substructure in the drug molecule graph and obtain the final bond-level feature representation.

[0088] Final bond-level feature representation Expressed as:

[0089]

[0090] in, represents the learnable matrix, represents the bias term, represents the tanh activation function, Representative The weight vector of the layer, is the number of network layers of the substructure extraction module, Representative Node The original feature representation of Representative Node The set of all neighbor nodes of is the nonlinear function of the multilayer perceptron, Representative Layer Node and The bond-level feature representation between ;

[0091] S204, introducing the edge features of the drug molecule graph and combining them with the final bond-level feature representation to further capture the relationship between the substructures of the drug molecule graph and obtain the local embedding of the drug substructure;

[0092] node Local embedding of drug substructures Expressed as:

[0093]

[0094]

[0095] in, and Both represent learnable matrices, Normalization of the representative layer, represents the edge features of the introduced drug molecule graph, express Dimensions, Represents multi-head splicing, Represents the number of attention heads; Representative Layer Node The substructure representation of

[0096] In particular, in S204, the final representation of the drug can be further obtained, specifically: S205, according to the final bond-level feature representation and the local embedding of the drug substructure, the final representation of the drug is obtained through a feedforward neural network;

[0097] In S205, the final representation of the drug Expressed as:

[0098]

[0099] in, Representative The substructure representation of the layer, and Both represent feedforward neural networks, is a learnable matrix, is layer normalization, Represents the activation function.

[0100] Furthermore, the universal embedding representation of the combined network obtained by the combined network universal embedding building module is:

[0101]

[0102] in, Representation node No. The embedding representation of the layer, Representation node Features in the combined graph, Representative Node In the The neighbor set of the layer, represents the learnable weight matrix, represents the bias term, represents the activation function, Represents an aggregate function, For aggregation nodes Neighbor nodes Features ;

[0103] In the combined network, since each drug pair depends on different contexts, and the contexts are represented as entities and relations, we construct a Construct a specific drug flow subgraph; specifically, the drug flow subgraph constructed by the knowledge subgraph construction module Expressed as };in, Represents a collection of nodes, represents a set of edge types, Represents the set of edges connecting nodes, the drug flow subgraph Indicated by point to and the length does not exceed A directed subgraph of Composition, among which, ( ... ) represents the edge of the drug flow subgraph, ( ... ) represents a node of the drug flow subgraph.

[0104] After obtaining the drug flow subgraph, the steps of constructing the knowledge subgraph by the knowledge subgraph construction module include:

[0105] S211, concatenating the local embedding of the drug substructure with the universal embedding of the combined network to initialize the drug features of the drug flow subgraph;

[0106] S212, evaluating the correlation between nodes of the drug flow subgraph according to the drug characteristics of the drug flow subgraph, and obtaining a correlation score;

[0107] Relevance score Expressed as:

[0108]

[0109] in, Representing relationships Learnable relation embeddings, represents the exponential function, Indicates taking the absolute value, represents a multilayer perceptron, represents the connection along the last dimension, and Represent nodes in the drug flow subgraph and In the The node embedding of the layer represents the features of any two nodes in the drug flow subgraph;

[0110] S213, updating the connection strength of the edges in the drug flow subgraph according to the correlation score;

[0111] Edge connection strength Expressed as:

[0112]

[0113] in, represents the hyperparameter, represents the activation function, represents the initial connection strength of the edges in the drug flow subgraph;

[0114] S214, performing feature aggregation on neighbor nodes through connecting edges to obtain a knowledge subgraph;

[0115] Further, the Embedding of layer knowledge subgraphs Expressed as:

[0116]

[0117] in, represents the learnable parameters, represents the operation of feature aggregation for each node in the drug flow subgraph, Indicates the Knowledge subgraph embedding of layers.

[0118] The prediction module obtains the final embedding representation of the drug pair based on the knowledge subgraph , as the drug-drug interaction prediction result, the final prediction result of the drug pair Expressed as:

[0119]

[0120] in, Respectively represent the nodes after the embedding transformation of the knowledge subgraph and nodes The final representation of the drugs, after being transformed, can better express the DDI prediction logic between the two drugs.

[0121] Optionally, in practical applications, the number of iterations can be set to repeatedly execute S212-S214. The continuous learning and updating of the edge connection strength can effectively filter out irrelevant information, capture the similarity relationship between drugs, and obtain the final knowledge subgraph.

[0122] In particular, as the edge connection strength is continuously learned and updated, the node features and edge weights between nodes are continuously updated based on the message passing mechanism of edge connections. When the weight is lower than the preset threshold, the edge is masked; conversely, if the weight exceeds the threshold, the unidirectional connection is upgraded to a bidirectional connection. In addition, after the message is passed, if the embeddings of the two nodes are highly similar, the model will automatically add a new bidirectional edge. After obtaining the final knowledge subgraph, the prediction module also obtains the final prediction result of the drug pair based on the final knowledge subgraph. .

[0123] In particular, the drug-drug interaction prediction model also includes a contrastive learning positive and negative sample construction module, which is used to generate positive samples through subgraph node masking and node perturbation strategies, generate negative samples by randomly selecting nodes in the combined graph to replace the original graph nodes and randomly changing the original graph node connections, and bring similar samples closer and push different samples farther away to learn the embedding space, thereby helping the model retain more information structure in high-dimensional space, so that the potential relationship between drug pairs can be better reflected.

[0124] In this application, the DDI prediction task is defined as a link prediction problem based on a knowledge subgraph. Furthermore, we transform the model optimization problem into a loss function minimization problem. Specifically, the loss function of the drug-drug interaction prediction model is Expressed as:

[0125] =

[0126]

[0127]

[0128]

[0129] in, Represents label losses for different prediction tasks, including multi-class drug-drug interaction prediction losses and multi-label prediction loss , is a hyperparameter, is the loss of positive and negative sample pairs; and are the true label and the predicted label respectively, and Represent the similarity scores of positive samples and negative samples respectively, represents the temperature parameter, and represent the number of positive samples and negative samples respectively, is the positive sample count identifier, is the negative sample count identifier.

[0130] In one application of the present invention, two public benchmark datasets, DrugBank and TWOSIDES, were used. DrugBank is a multi-class DDI prediction dataset that covers the pharmacological relationships between 86 drugs; TWOSIDES is a multi-label DDI prediction dataset that records a variety of side effects between drugs; in this application, Hetionet was introduced as an external knowledge graph (KG). Hetionet is a comprehensive biomedical KG used to represent complex biological and medical entities and their relationships. For the two benchmark datasets, DrugBank and TWOSIDES, data preprocessing follows the following process: In the DrugBank dataset, since the vast majority of edges contain only a single edge type, we filtered out a very small number of edges with multiple types. In the TWOSIDES dataset, since each edge may be associated with multiple labels, we screened 200 common DDI types. By sorting in descending order by side effect type, 200 relationships ranked between 600 and 800 were selected to ensure that each side effect corresponds to at least 900 pairs of drug combinations. For the external KG Hetionet, to avoid information leakage, we removed duplicate edges from the DDI benchmark dataset, ensuring data independence. The processed DDI graph and the external KG were then merged into a combined graph. To ensure experimental fairness, the triplets of the DDI benchmark dataset were divided into training, validation, and test sets in a ratio of 7:1:2. Information on the three datasets is shown in Table 1:

[0131] Table 1 Statistics of three datasets

[0132]

[0133] The example diagram of the drug-drug interaction prediction model provided in this application is as follows Figure 2 As shown, Figure 2 As shown in A, we first integrate the DDI graph and the external KG into a composite network, where all drug and entity nodes contain latent semantic and structural information. Then, we use GraphSAGE to extract the universal embedding of the composite network to capture the global semantic information. Figure 2 As shown in Figure 2, the model extracts and refines drug substructures to obtain local embeddings. Given a drug pair to be predicted, it extracts the corresponding drug flow subgraph from the combined network and fuses the universal embedding with the local embedding of the drug substructure based on its node information. This embedding is initialized using a weighted mechanism. To fully exploit the structure and embedding information of the drug flow subgraph while removing irrelevant interference factors, the model further iteratively optimizes the drug flow subgraph.

[0134] In particular, during this process, the features of the nodes and the edge weights between nodes are continuously updated based on the message passing mechanism of edge connections. When the weight is lower than the preset threshold, the edge is masked; conversely, if the weight exceeds the threshold, the unidirectional connection is upgraded to a bidirectional connection. In addition, after the message is passed, if the embeddings of the two nodes are highly similar, the model will automatically add a new bidirectional edge. After multiple rounds of iterative optimization, the drug flow subgraph will eventually retain only the key information. Figure 2 As shown in Figure C, a contrastive learning strategy is introduced into the final optimized drug flow subgraph to improve the model's representation and generalization capabilities. Furthermore, this optimized subgraph can derive one or more interpretable paths from high-weight paths, providing a traceable reasoning basis for DDI prediction.

[0135] In this application, in order to demonstrate the feasibility, practicality and advancement of the method provided by this application, the following experiments were conducted:

[0136] Performance comparison experiments with baseline methods:

[0137] In this experiment, the proposed method is compared with eight baseline methods: DeepWalk-DDI, node2vec-DDI, GAT-DDI, Decagon-DDI, KGNN-DDI, KGNN-DDI, DeepLGF-DDI, AdaProp-DDI, and SumGNN-DDI.

[0138] DeepWalk-DDI is a DDI prediction method based on random walk and network embedding. It captures the structural information of drug nodes by performing random walks on the DDI network and uses embedding vectors to predict the interaction type between drugs.

[0139] node2vec-DDI is a DDI prediction method based on biased random walk and flexible network embedding. It adjusts the random walk strategy to better capture the local and global structural properties of the drug network, thereby generating more refined drug embedding vectors.

[0140] GAT-DDI is a DDI prediction method based on graph attention network, which uses the attention mechanism to dynamically assign the importance of neighbor nodes and capture more critical interaction characteristics between drugs.

[0141] Decagon-DDI is a DDI prediction method based on a multi-relational graph neural network. It captures the complex and diverse types of drug interactions while also taking into account phenotypic prediction, thereby improving the comprehensiveness and accuracy of the prediction.

[0142] KGNN-DDI is a graph neural network DDI prediction method based on knowledge graph. It enhances the embedding representation by integrating drug-related biomedical knowledge and improves the accuracy of relationship prediction.

[0143] DeepLGF-DDI is a method for predicting DDIs using a biomedical knowledge graph (BKG) that fully integrates local and global information. This method utilizes the semantic information of drug SMILES sequences as local information, extracts global information from the BKG using a KG embedding method, and finally uses an MLP module for feature fusion.

[0144] AdaProp-DDI is a DDI prediction method based on an incremental sampling mechanism, which also eliminates unreasonable interactions and maintains consistent semantics through structure learning and semantics-preserving modules.

[0145] SumGNN-DDI is a DDI prediction method based on summary learning. It enhances the graph embedding representation capability by extracting summary information of knowledge graph subgraphs, thereby improving the prediction performance.

[0146] Table 2 shows the performance comparison of the proposed method Mifs-DDI and all the baseline methods on two datasets:

[0147] Table 2. Evaluation results on the DrungBank and TWOSID benchmark datasets

[0148]

[0149] According to the results in Table 2, Mifs-DDI achieves state-of-the-art performance on both datasets. Compared to the best baseline method on each dataset, it achieves an average improvement of 3.34% in ACC, 6.21% in F1 score, 3.14% in Cohen's kappa, 1.75% in AUROC, 2.05% in AUPRO, and 1.80% in AP@50. These results demonstrate the effectiveness of Mifs-DDI in the DDI prediction task. Overall, GNN methods that enhance semantic representation by incorporating an external KG perform the best. In contrast, traditional network embedding methods such as DeepWalk and node2vec rely primarily on local structural information, resulting in significantly lower performance and stability than Mifs-DDI on the DrugBank and TWOSIDES benchmark datasets. This is because these methods ignore global graph structure and node features, limiting their ability to capture complex interactions. GAT-DDI and Decagon-DDI, which are based on GNNs but do not incorporate external graphs, show improvement over traditional two-step methods, but their performance remains suboptimal. This is because while GNNs can effectively propagate and aggregate domain information and utilize attention mechanisms to capture more complex node embedding representations, these methods fail to effectively capture cross-domain knowledge due to a lack of semantic information provided by external knowledge graphs. Further research has shown that combining GNNs with knowledge graphs (KGs) can effectively alleviate the data scarcity problem. Among these methods, KGNN-DDI performs poorly. Although it can mine associations in the knowledge graph to capture drugs and their potential neighborhoods, the presence of noise in the KG negatively impacts drug-drug interaction prediction. In contrast, SumGNN-DDI combines the DDI graph and an external KG into a combined network. After node embedding, it extracts subgraphs to further encode semantic and structural information, effectively filtering out noisy information. However, these methods often overlook the expression of internal features of drug nodes and overly rely on the topological structure of nodes in the network. Mifs-DDI fuses drug substructure features within subgraphs, further filters out noisy data by updating relevance scores, and enhances the model's robustness and information interaction capabilities through a subgraph comparative learning module. Compared with existing methods, the performance improvement of Mifs-DDI verifies its effectiveness.

[0150] Ablation experiment:

[0151] In order to study the importance of each module to the model performance, the present invention designed the following variants to compare with Mifs-DDI:

[0152] Mifs-DDI (w / o SL): removes the drug substructure feature extraction module and directly predicts through the universal embedding in the combined graph;

[0153] Mifs-DDI (without CL): Contrasting learning of positive and negative sample building blocks, retaining common embedding features and drug substructure features for prediction;

[0154] Mifs-DDI (w / o KSW): removes the update of edge weights during knowledge subgraph information transfer and sets the weights of all edges to 1;

[0155] Mifs-DDI (w / o %20KG): randomly delete 20% of the entity pairs in the external KG and retain the remaining 80% to form a combined graph with the DDI graph;

[0156] Mifs-DDI (w / o %50KG): randomly delete 50% of the entity pairs in the external KG and retain the remaining 50% to form a combined graph with the DDI graph;

[0157] Mifs-DDI (w / o %80KG): randomly delete 80% of the entity pairs in the external KG and retain the remaining 20% ​​to form a combined graph with the DDI graph;

[0158] like Figure 3 As shown in Figure 2, all variants of Mifs-DDI show performance degradation compared to the original model, further verifying the importance of each module for the DDI prediction task. Figure 3 As shown in Figure 2, as the proportion of entity pairs deleted from the external KG increases, the model performance gradually decreases, indicating the effectiveness of KG in model learning. When the deletion ratio reaches 80%, the model performance decreases most significantly, because the remaining 20% ​​of entities are no longer sufficient to provide sufficient context and topology information. Figure 3 As shown in b, Mifs-DDI ( SL) retains the contextual information of KG, but the performance still drops significantly. Although the model can still learn information from the rich KG, it loses the understanding of the details of the drug molecular structure. KSW) also showed a certain performance degradation. The reason is that after removing the edge weight update, the model lost the ability to model the strength of the node relationship in the graph, resulting in poor information transmission. Finally, although Mifs-DDI ( Although the performance of the contrastive learning module has declined, the contrastive learning module can guide the model to focus on the differences between positive and negative samples, thereby extracting more discriminative features and improving the robustness of the model. This result has been further verified in the hyperparameter sensitivity analysis.

[0159] Hyperparameter sensitivity analysis experiment:

[0160] In this experiment, we conducted a hyperparameter sensitivity analysis on the DrugBank dataset and studied the effects of several key hyperparameters on the performance of Mifs-DDI while keeping other parameters unchanged: embedding dimension d, drug flow subgraph size , the number of combined graph network layers , the number of subgraph network layers .like Figure 4 As shown in Figure a, the model performance is optimal when the embedding dimension d = 32. Too small an embedding dimension may not fully capture the complex features of the data, resulting in performance degradation; while too large an embedding dimension may cause the model to overfit or increase computational complexity, thereby reducing performance. Figure 4 As shown in b, when When , the model performance reaches the best; and when Exceed When the size of the subgraph increases, the performance of Mifs-DDI begins to show an unstable downward trend. The increase of increases exponentially, which may lead to a significant increase in the noise contained in the drug flow subgraph. Figure 4 As shown in c, the number of combined network layers The best performance is achieved when the number of subgraph network layers is set to 2. Too small or too large a number of network layers cannot achieve the best performance. This is because larger views may bring useless data for DDI prediction, while smaller views cannot fully utilize the rich context in the external KG. In addition, Performance degradation is shown in The trend is faster because after the multi-layer information aggregation of the combined network, each drug in the drug flow subgraph has aggregated information from a wider range of fields. Figure 3 c, we further explore the contribution of the contrastive learning module. By removing the contrastive learning module and then gradually adding ,from Figure 4 dIt can be observed that the model performance index is This shows the contribution of this module to model performance and generalization.

[0161] Interpretability experiments:

[0162] In this experiment, Mifs-DDI uses similarity scores to evaluate the importance of edges in subgraphs to predict DDI between drugs. By analyzing the weight relationship of edges in subgraphs, interpretable paths can be obtained to reveal the types of interactions between drugs. Figure 5 The relationship between the points and edges of the three drug pairs in the subgraph is shown, where the thickness of the edge indicates the importance of the edge. Figure 5All nodes in the figure are drug nodes, and star nodes are drug pairs to be predicted. Gray solid lines represent the reactions between nodes and the edges with weak similarity types generated. Red and green solid lines highlight edges with strong similarity relationships. Figure 5 As shown in a, nodes 299 and 124 have strong similarities with the intermediate node 352, and there is a DDI relationship between nodes 124 and 352. Therefore, it can be inferred that this pair of drugs may have similar effects when used together. Mifs-DDI does not rely on a single path for DDI prediction. Figure 5 b shows the situation of DDI prediction through multiple interpretable paths. Drug nodes 716 and 726 will also have DDI with node 828. Mifs-DDI believes that node 725 has a high similarity with both nodes 716 and 726, so it further infers that node 725 will also have DDI with node 828. In addition, Mifs-DDI can also infer through indirect DDI inference paths. Figure 5 As shown in Figure c, nodes 484 and 1329, as well as nodes 60 and 939, have strong connections. By learning the reaction type between nodes 939 and 1329, the model further infers that a similar DDI relationship may exist between nodes 484 and 60. In summary, these explanatory pathways are not only reasonable but also consistent with the model's predictions. By continuously updating edge weights, Mifs-DDI learns subgraphs that are more closely aligned with the target predictions, making it easier to identify key pathways. These pathways precisely reflect the relevance and interpretability of the interaction types between drug nodes.

[0163] In order to further evaluate the reliability of the interpretable paths generated based on similarity scores and verify the application value of Mifs-DDI in actual scenarios, this experiment conducted a case analysis experiment. Figure 6 As shown in the figure, all nodes are drug nodes, star nodes are drug pairs to be predicted, and they are connected by gray dashed lines. Gray solid lines represent the reactions between nodes and the edges with weak similarity types generated. Red and green solid lines highlight edges with strong similarity relationships. The right side is a comparison diagram of key drug structures. We checked the SMILES structures of drug pairs and their corresponding two drugs with the highest connection weight scores in the knowledge subgraph and analyzed them. Figure 6For example, DB01357 (mestranol) is a synthetic estradiol found in oral contraceptives and used to treat other female reproductive system disorders, such as dysmenorrhea and dysfunctional uterine bleeding. DB00860 (prednisolone) is a cortisol-like glucocorticoid with anti-inflammatory, immunosuppressive, anti-tumor, and vasoconstrictor effects. Related studies have shown that prednisolone and mestranol increase serum concentrations when used in combination. Mifs-DDI identified DB01357 (mestranol) and DB00860 (prednisolone) through a strong similarity pathway with DB04573 (estriol) and DB00443 (betamethasone), and further speculated that estriol and betamethasone may also exhibit similar DDIs. Estriol is a weak estrogen used to treat vaginal dryness and estrogen deficiency disorders, such as vaginitis and vulvar pruritus. Betamethasone is a systemic corticosteroid used to relieve inflammation in various conditions, including but not limited to allergic conditions, dermatological conditions, gastrointestinal disorders, and hematological disorders. To further support this reasoning, we analyzed the molecular structures of these drugs. Figure 6 As shown in the molecular structure diagram on the right of a, estriol and mestranol have the same basic core structure, both consisting of a four-membered ring and a six-membered ring; at the same time, betamethasone and prednisolone also contain very similar ring structures and functional groups. These similarities suggest that they may play similar roles in biology. Figure 6 As shown in b, DB00816 (meproterenol) has a strong similar pathway to drugs such as DB00871 (terbutaline) and DB01064 (isoproterenol). Figure 6 The drug molecular structures on the right side of b further verify the reliability of the similarity paths generated by the model. Similar drugs such as DB00871 (terbutaline) and DB01064 (isoproterenol) will interact with DB00476 (duloxetine). Therefore, the model infers that DB00816 (meproterenol) and DB00476 (duloxetine) will produce similar types of interactions. Related studies have shown that the co-use of meproterenol and duloxetine will lead to the enhancement of each other's activity. Figure 6 As shown in Figure c, DB01190 (clindamycin) and DB01627 (lincomycin) have strong similar pathways, and DB00483 (galaminium iodide) and DB00202 (succinylcholine) also have similar pathways. Figure 6 The molecular diagram to the right of c further demonstrates the similarity in their core structures. Furthermore, DB00202 and DB01627 can enhance each other's effects when used together. Therefore, the model predicts that the effects of DB01190 (clindamycin) and DB00483 (galaminium iodide) will also be enhanced when used together, confirming the real-world synergistic effect of two drugs when used together.

[0164] Through this case study, we demonstrate the ability of Mifs-DDI to identify potential DDIs, as well as its effectiveness and reliability in practical applications. Furthermore, the results of this case study further demonstrate the advantages of Mifs-DDI. It not only accurately predicts potential DDIs based on drug molecular structure and similarity information, but also provides valuable clinical insights through interpretable pathways, helping researchers and medical professionals make more informed decisions when using drug combinations. Therefore, Mifs-DDI has broad application prospects in the biopharmaceutical field and holds significant practical significance.

[0165] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A drug-drug interaction prediction method based on drug flow subgraph, characterized in that: include: S1. Data preparation: Collect a benchmark dataset including drug-drug interactions and introduce an external knowledge graph for adaptive preprocessing; S2. Model construction: Build a drug-drug interaction prediction model and train it using the preprocessed benchmark dataset; The drug-drug interaction prediction model includes: a drug substructure feature extraction module to extract local embeddings of drug substructures; The general embedding construction module for the combined network is used to obtain the general embedding of the combined network through GraphSAGE. The knowledge subgraph construction module is used to construct the drug flow subgraph and optimize it using the general embedding of the combined network and the local embedding of the drug substructure to obtain the knowledge subgraph. The prediction module is used to obtain the drug-drug interaction prediction results based on the knowledge subgraph. S3. Effect prediction: The target drug is input into the drug-drug interaction prediction model, and the drug-drug interaction prediction model outputs the drug-drug interaction prediction result.

2. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 1, characterized in that: The steps of extracting the final representation of the drug by the drug substructure feature extraction module include: S201. Considering the atomic representations of the drug molecule graph as nodes and the chemical bond representations of the drug molecules as edges, and extracting the bond-level features of all edges in the drug molecule graph through a graph neural network; S202, using an activation function to calculate the importance score of each bond-level feature in the drug molecule graph, and sending the obtained importance score to a pooling layer to obtain a pooled feature of the drug molecule graph; S203. Based on the pooled features of the drug molecule graph, a multi-layer attention mechanism and feature aggregation are used to capture important substructure information in the drug molecule graph and obtain the final bond-level feature representation. S204. Introducing the edge features of the drug molecule graph and combining them with the final bond-level feature representation, further capture the relationship between the substructures of the drug molecule graph and obtain the local embedding of the drug substructure.

3. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 2, characterized in that: Bond-order characteristics of edges in drug molecule graph in S201 Expressed as: in, Represents the characteristics of the starting node, Represents the characteristics of the terminal node, is a node and Characteristics of the bond between them; In S202, the pooling features of the drug molecule graph Expressed as: Where, represents the pooling layer, Indicates the number of layers of the current graph neural network, Represents the matrix composed of all key-level features in the current layer t of the graph neural network, Represents the adjacency matrix of the current layer graph; In S203, the final bond-level feature representation Expressed as: in, represents the learnable matrix, represents the bias term, represents the tanh activation function, Representative The weight vector of the layer, is the number of network layers of the substructure extraction module, Representative Node The original feature representation of Representative Node The set of all neighbor nodes of is the nonlinear function of the multilayer perceptron; Representative Layer Node and The bond-level feature representation between ; In S204, local embedding of drug substructures Expressed as: in, and Both represent learnable matrices, Normalization of the representative layer, represents the edge features of the introduced drug molecule graph, express Dimensions, Represents multi-head splicing, represents the number of attention heads, Representative Layer Node The substructure representation of .

4. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 1, characterized in that: The general embedding representation of the combined network obtained by the combined network general embedding building module is: in, Representation node No. The embedding representation of the layer, Representation node Features in the combined graph, Representative Node In the The neighbor set of the layer, represents the learnable weight matrix, represents the bias term, represents the activation function, Represents an aggregate function.

5. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 1, characterized in that: Drug flow subgraph constructed by the knowledge subgraph construction module Expressed as };in, Represents a collection of nodes, represents a set of edge types, Represents the set of edges connecting nodes, the drug flow subgraph Indicated by point to and the length does not exceed A directed subgraph of Composition, among which, ( ... ) represents the edge of the drug flow subgraph, ( ... ) represents a node of the drug flow subgraph.

6. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 5, characterized in that: The steps of constructing a knowledge subgraph by the knowledge subgraph construction module include: S211, concatenating the local embedding of the drug substructure with the universal embedding of the combined network to initialize the drug features of the drug flow subgraph; S212, evaluating the correlation between nodes of the drug flow subgraph according to the drug characteristics of the drug flow subgraph, and obtaining a correlation score; S213, updating the connection strength of the edges in the drug flow subgraph according to the correlation score; S214. Aggregate features of neighbor nodes through connecting edges to obtain a knowledge subgraph.

7. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 6, wherein in step S212, the correlation score Expressed as: in, Representing relationships Learnable relation embeddings, represents the exponential function, Indicates taking the absolute value, represents a multilayer perceptron, represents the connection along the last dimension, and Represent nodes in the drug flow subgraph and In the Node embedding of the layer; In S213, the connection strength of the edge Expressed as: in, represents the hyperparameter, represents the activation function, represents the initial connection strength of the edges in the drug flow subgraph; In S214, Embedding of layer knowledge subgraphs Expressed as: in, represents the learnable parameters, represents the operation of feature aggregation for each node in the drug flow subgraph, Indicates the Layer-wise knowledge graph embedding.

8. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 1, characterized in that: The prediction module obtains the final embedding representation of the drug pair based on the knowledge subgraph , as the drug-drug interaction prediction result, the final prediction result of the drug pair Expressed as: in, Respectively represent the nodes after the embedding transformation of the knowledge subgraph and nodes The final expression of the drug, Represents a multilayer perceptron.

9. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 1, characterized in that: The drug-drug interaction prediction model also includes a contrastive learning positive and negative sample construction module, which is used to generate positive samples through subgraph node masking and node perturbation strategies, generate negative samples by randomly selecting nodes in the combined graph to replace the original graph nodes and randomly changing the original graph node connections, and bring similar samples closer and push different samples farther away to learn the embedding space.

10. The drug-drug interaction prediction method based on drug flow subgraph features according to claim 9, characterized in that: Loss function for drug-drug interaction prediction models Expressed as: = in, Represents label losses for different prediction tasks, including multi-class drug-drug interaction prediction losses and multi-label prediction loss , is a hyperparameter, is the loss of positive and negative sample pairs; and are the true label and the predicted label respectively, and Represent the similarity scores of positive samples and negative samples respectively, represents the temperature parameter, and represent the number of positive samples and negative samples respectively, is the positive sample count identifier, is the negative sample count identifier.

Images