Drug interaction prediction method and system based on multi-view comparative learning

Abstract

The invention relates to a drug interaction prediction method and system based on multi-view comparative learning, and belongs to the technical field of natural language processing. According to the method, two channels of a drug molecular map and a biomedical knowledge map are constructed in parallel, structural and semantic features are extracted by using a pre-trained heterogeneous map neural network, and multi-view comparative learning guided by information gain is introduced for joint optimization, so that the generalization ability and robustness of the model to unknown drug pairs are enhanced. According to the method, system evaluation is carried out on the performance of the system in two types of prediction tasks (multi-type and multi-label) and three prediction scenes. Experimental results show that the method has excellent performance in all tasks and scenes. Further case analysis also verifies the effectiveness of the system in predicting the interaction type of the unseen drug pair.

Description

Technical Field

[0001] This invention belongs to the technical field of natural language processing, and relates to a drug interaction prediction technology, particularly a drug interaction prediction method and system based on multi-view contrast learning. Background Technology

[0002] With an aging population and an increasing prevalence of co-occurring diseases, the clinical demand for multidrug combination therapy is growing. However, interactions between multiple drugs can trigger serious adverse reactions, leading to drug withdrawal from the market and even public safety incidents and economic losses. Although combination therapy has shown good efficacy in treating multiple diseases, its potential interaction risks cannot be ignored. Therefore, accurate and efficient prediction of drug interactions is of great significance for ensuring drug safety, optimizing clinical decision-making, and protecting patient health.

[0003] In recent years, biological computation methods, as an efficient and scalable alternative, have gradually become a research hotspot in the field of drug-drug interaction (DDI) prediction, and have achieved remarkable results. Existing computational methods can be divided into the following four categories: text mining-based methods, drug molecule graph-based methods, biomedical knowledge graph-based methods, and dual-channel fusion-based methods.

[0004] Text mining-based methods typically utilize natural language processing techniques to extract DDI-related information from massive biomedical literature and databases. While these methods demonstrate high accuracy in DDI extraction, they primarily rely on already labeled DDI instances, making it difficult to identify unlabeled or implicit interactions within the literature. Furthermore, the continuous growth and rapid evolution of biomedical literature and database content presents significant challenges to these methods.

[0005] Drug molecular mapping (DIM)-based methods are guided by the idea that DIM, as a structural characterization method, plays a crucial role in depicting the chemical composition of drugs and their potential interactions. Some researchers argue that the essence of drug-derived interactions (DDIs) stems from the interactions between local chemical substructures, rather than the overall molecular structure. Therefore, some studies define the DDI prediction task as identifying the interactions between potential substructures of two drugs. This approach, through in-depth modeling of drug molecular maps, improves the pharmacological plausibility of DDI predictions to some extent, but it does not fully consider more biomedical characteristics.

[0006] Methods based on biomedical knowledge graphs improve the accuracy of drug detection and identification (DDI) predictions by mining the rich information contained within the knowledge graph. Researchers have proposed a series of methods for modeling knowledge graphs, such as constructing multi-hop domain receptive fields, fusing structural and semantic information, and using graph convolutional networks to comprehensively capture the relationships between drug nodes and their potential neighbors.

[0007] Dual-channel fusion methods simultaneously consider drug molecule graphs and biomedical knowledge graphs, leveraging the structured features of drugs and the rich information contained in the knowledge graph to improve the performance of DDI prediction models. In such methods, drug molecule graphs and knowledge graphs are often treated as heterogeneous graphs, and effectively integrating these two types of heterogeneous information remains one of the key challenges in DDI prediction tasks.

[0008] Therefore, how to effectively capture the complex structural and semantic features in drug molecule graphs and biomedical knowledge graphs, and explore their synergistic effect to achieve accurate DDI prediction, is a technical problem that this invention urgently needs to solve. Summary of the Invention

[0009] In view of this, the present invention provides a drug interaction prediction method and system based on multi-view contrastive learning, which can effectively capture structural and semantic information in drug molecular graphs and knowledge graphs to predict potential interactions between drugs.

[0010] Specifically, this invention designs three key modules. First, a graph data processing module converts drug SMILES sequences into molecular graphs and, combined with a biomedical knowledge graph, generates an original graph and an information gain-guided perturbation graph to construct a multi-view input. Second, a heterogeneous graph encoder introduces a pre-trained heterogeneous graph neural network to encode drug molecular graphs and knowledge graphs under different views, obtaining multi-view embedding representations. Finally, a DDI classifier fuses molecular graph embeddings and knowledge graph embeddings for DDI prediction, while also incorporating multi-view comparative learning to improve the system's consistency modeling capability and perturbation robustness on heterogeneous graph representations.

[0011] The technical solution adopted in this invention is as follows: On one hand, this invention provides a drug interaction prediction method based on multi-view contrast learning, comprising the following steps: Drug-related data are processed to construct drug molecular graphs and biomedical knowledge graphs, and perturbation views are generated based on information gain to form multi-view inputs; A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph and its perturbation view to obtain multi-view feature embedding; the heterogeneous graph neural network includes a multi-layer heterogeneous attention layer that fuses node and edge features, which can adaptively aggregate multi-relation information according to the differences in the types of nodes and edges, and introduce residual connections and batch normalization to alleviate the oversmoothing problem. A joint training module integrating supervised learning and multi-view contrastive learning is designed to predict drug interactions through feature fusion and improve consistency and robustness using contrastive learning. The joint training module includes a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters, achieving complementary fusion of structural and semantic features. The multi-view contrastive learning unit employs bidirectional normalized temperature-scaling cross-entropy loss to impose contrast constraints on the representations between the original and perturbed views, and between structural and semantic views, enhancing cross-view... Figure 1 Consistency and perturbation robustness.

[0012] Furthermore, drug-related data is processed to construct drug molecule graphs and biomedical knowledge graphs, and perturbation views are generated based on information gain to form multi-view inputs, including: Write a preprocessing program for isomorphic graph structures to clean irrelevant data and convert drugs into molecular graph form and knowledge graph data into triplet form; Based on the information gain-based graph perturbation strategy, perturbation views are generated from preprocessed heterogeneous graph data to construct multi-view inputs.

[0013] Furthermore, a pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph, and their perturbed views to obtain multi-view feature embeddings, including: A pre-training mechanism is introduced into the heterogeneous graph encoder. For drug molecule graphs, a self-supervised pre-training strategy based on node masking is adopted, and for biomedical knowledge graphs, a self-supervised pre-training strategy based on link prediction is adopted. The drug molecule graph, knowledge graph, and their perturbed views are input into a pre-trained heterogeneous graph neural network for encoding to obtain multi-view feature embeddings.

[0014] Furthermore, a joint training module integrating supervised learning and multi-view contrastive learning is designed to predict drug interactions through feature fusion and to improve consistency and robustness using contrastive learning, including: Design a joint optimization module, including a supervised classifier based on feature fusion and a multi-view contrastive learner; Drug molecule graph embedding and knowledge graph embedding are fed into a multilayer perceptron for drug interaction prediction; Multi-view embeddings are used for contrastive learning, contrastive loss is calculated, and it is jointly optimized with supervised loss.

[0015] In another aspect, the present invention also provides a drug interaction prediction system based on multi-view contrastive learning, comprising: Graph data processing module: processes drug-related data, constructs drug molecular graphs and biomedical knowledge graphs, and generates perturbation views based on information gain to form multi-view input; Heterogeneous graph encoder module: A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph and its perturbation view to obtain multi-view feature embedding; the heterogeneous graph neural network includes a multi-layer heterogeneous attention layer that fuses node and edge features, which can adaptively aggregate multi-relation information according to the differences in the types of nodes and edges, and introduce residual connections and batch normalization to alleviate the oversmoothing problem. Drug Interaction Classifier Module: A joint training module integrating supervised learning and multi-view contrastive learning is designed. It predicts drug interactions through feature fusion and leverages contrastive learning to improve consistency and robustness. The joint training module includes a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters, achieving complementary fusion of structural and semantic features. The multi-view contrastive learning unit employs bidirectional normalized temperature-scaling cross-entropy loss to impose contrast constraints on the representations between the original and perturbed views, and between structural and semantic views, enhancing cross-view... Figure 1 Consistency and perturbation robustness.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention designs a dual-channel architecture, which combines a pre-trained heterogeneous graph neural network to extract structural features from drug molecule graphs and semantic features from biomedical knowledge graphs, respectively. This solves the challenge of insufficient fusion of structural and semantic information in existing methods and improves the generalization of the system and its ability to identify potential DDIs.

[0017] (2) In order to enhance the consistency modeling and robustness against disturbances of the system for heterogeneous graph representations, this invention proposes a graph structure perturbation strategy based on information gain, and generates perturbation views in two channels respectively, and introduces a multi-view contrast learning objective, thereby obtaining a more discriminative and robust drug embedding representation, which effectively addresses the problems caused by data perturbation and view differences.

[0018] (3) This invention further explores more challenging DDI prediction scenarios, including the identification of specific DDI types between drug pairs and the prediction of DDI types between new drugs, in order to support more refined clinical medication decisions and new drug discovery processes, and expand the application scope of DDI prediction.

[0019] This invention effectively extracts and integrates structural and semantic information from drug molecule graphs and biomedical knowledge graphs, and performs DDI prediction tasks in multiple scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a drug interaction prediction method based on multi-view contrast learning in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a drug interaction prediction method based on multi-view contrast learning in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] This invention presents a drug interaction prediction method based on multi-view contrastive learning. First, the drug's SMILES sequences are converted into a molecular graph. This graph, combined with a biomedical knowledge graph, is used to generate a perturbation view through information gain guidance, constructing a multi-view input. Subsequently, pre-trained parameters are introduced into a heterogeneous graph neural network to encode features for both the original and perturbation views, thereby enhancing the system's representational and generalization capabilities. Finally, in the DDI classification module, this invention integrates molecular graph representation and knowledge graph representation, and performs prediction using a multilayer perceptron. Simultaneously, a multi-view contrastive learning mechanism is introduced to improve the system's consistent modeling ability and robustness against perturbations in heterogeneous graph representations.

[0026] like Figure 1 As shown in the figure, a drug interaction prediction method based on multi-view contrast learning in an embodiment of the present invention specifically includes the following steps: S1: Process drug-related data, construct drug molecular graphs and biomedical knowledge graphs, and generate perturbation views based on information gain to form multi-view input.

[0027] Specifically, firstly, data on drugs, genes, diseases, and side effects are extracted and constructed into a biomedical knowledge graph. This knowledge graph is a multi-type heterogeneous network, with nodes including entities such as drugs, diseases, genes, symptoms, and side effects. Edges represent semantic relationships between them, such as "drug-treatment-disease," "drug-effect-gene," and "drug-cause-side effect," used to describe the multidimensional biological semantics between drugs. Secondly, the drug data is processed into drug molecular graphs. Subsequently, information gain-guided graph perturbations are applied to the drug molecular graphs and the knowledge graph to construct multi-view inputs. These multi-view inputs include four forms: the original molecular structure view, the original semantic view, and the corresponding perturbed views, used for subsequent multi-view feature encoding and comparative learning.

[0028] The above node information gain The calculation process is as follows:

[0029] in, Represents a non-linear activation function. This represents the temperature coefficient. and They represent the first The embedded representation of each node and the global center of all node embeddings.

[0030] For an edge Its information gain is calculated using the information difference between adjacent nodes:

[0031] By randomly masking the features of nodes and edges with a probability proportional to the information gain, graph feature perturbation guided by information gain is achieved. This strategy can introduce more meaningful perturbation signals while preserving the overall semantics of the graph structure, thus promoting robust representation learning in contrastive learning scenarios.

[0032] S2: A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph and its perturbation view to obtain multi-view feature embedding.

[0033] Given that both drug molecule graphs and biomedical knowledge graphs possess heterogeneous graph characteristics, this system employs a heterogeneous graph neural network as an encoder. The drug molecule graph and knowledge graph, along with their respective perturbation views, are encoded by a heterogeneous graph encoder with pre-trained parameters to obtain feature embeddings for multiple views.

[0034] Heterogeneous graph neural networks (HNNs) consist of multiple heterogeneous attention layers that fuse node and edge features. They adaptively aggregate multi-relational information based on differences in node and edge types, while introducing residual connections and batch normalization to mitigate oversmoothing issues. Their working principle involves calculating interaction weights between different types of nodes through an edge-guided attention mechanism, achieving weighted aggregation of structural and semantic features. Combined with a pre-training mechanism, prior knowledge of chemical structure and biological semantics is injected, thereby generating high-quality multidimensional drug representations. Their role is to unify the modeling of molecular structural features and semantic knowledge associations of drugs, improving the completeness and generalization ability of the representation.

[0035] Specifically, taking a drug molecule diagram as an example, the input consists of an atom embedding matrix A and a chemical bond embedding matrix C.

[0036] Projecting node and edge features using the pre-trained linear mapping:

[0037] in, Indicates the first The node characteristics of each node. Indicates the first and the Edge characteristics between nodes and These are the pre-trained weight matrices for nodes and edges, respectively. Represents the first atomic embedding matrix. One atom, Represents the first in the chemical bond embedding matrix and the The characteristics of chemical bonds between atoms.

[0038] To model the interaction relationships between nodes, an edge-guided attention mechanism is adopted. For nodes... Its neighboring nodes Attention weights are calculated as follows:

[0039] in, Represents a node Its neighboring nodes Attention weights between them This represents vector concatenation. and These are the attention projection matrices for node and edge features, respectively. This represents a learnable attention vector.

[0040] We perform a weighted aggregation of all neighboring nodes to obtain the updated node representation:

[0041] in Represents the nodes after weighted aggregation The expression, Represents a node The set of neighboring nodes.

[0042] To address the common problems of oversmoothing and gradient explosion in graph neural networks, previous research has shown that introducing residual connections can alleviate these phenomena to some extent. Especially in heterogeneous graph structures, residual connections help enhance the stability and expressive power of the system. Therefore, this invention introduces a residual connection mechanism into a heterogeneous graph encoder to improve its fitting ability. Specifically, for the... The layer, feature aggregation process is as follows:

[0043] in and Represents a node In the Layer node features and weighted aggregation node representation, Represents a node In the The representation after layer aggregation Indicates batch normalization, This is a pre-trained linear transformation matrix.

[0044] S3: Design a joint training module that integrates supervised learning and multi-view contrastive learning to predict drug interactions through feature fusion and improve consistency and robustness through contrastive learning.

[0045] Design a joint training module integrating supervised learning and multi-view contrastive learning, including a supervised classifier based on feature fusion and a multi-view contrastive learner. The joint training module comprises a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters, achieving complementary fusion of structural and semantic features. The multi-view contrastive learning unit employs bidirectional normalized temperature-scaling cross-entropy loss to impose contrast constraints on the representations between the original and perturbed views, and between structural and semantic views, enhancing cross-view contrast. Figure 1 Consistency and robustness to perturbations. In the supervised classifier, cross-entropy loss is used to optimize the system; in the multi-view contrastive learner, bidirectional normalized temperature-scaling cross-entropy loss is used as the loss function.

[0046] Specifically, drug molecule graph embeddings are obtained from the heterogeneous graph encoder. and knowledge graph embedding Then, the two are separated into left and right representations corresponding to the two drugs in a drug pair, and the structural and semantic information of the two channels are fused using a splicing operation to obtain the final drug joint embedding representation:

[0047] in, and These represent two drug representations embedded in the molecular diagrams of a pair of drugs. and These represent two drug representations embedded in a knowledge graph for a pair of drugs. and These represent two drug representations resulting from the fusion of molecular graph embedding and knowledge graph embedding, respectively.

[0048] For two distinct DDI prediction tasks, the left and right drug pair representations are concatenated and fed into a multilayer perceptron to predict the DDI probability of the drug pair. For multi-class prediction tasks, this step is as follows:

[0049] in, This represents the activation function. This represents a multilayer perceptron. Indicates the drug's effect The probability of DDI between them.

[0050] For multi-label prediction tasks, this step is as follows:

[0051] in, This represents the activation function. Indicates the drug's effect Interaction probability scores between them.

[0052] During training, this invention optimizes the system by minimizing cross-entropy loss, thereby improving DDI prediction performance. For multi-class prediction tasks, the supervised loss function is defined as follows:

[0053] in, This represents the value of the supervised loss function. Indicates a collection of drugs. This indicates the total number of drug interaction types. Represents a logarithmic function. Indicates the drug's effect Does the interaction between them belong to a type? , Is it the drug's effect? The interaction between them belongs to type The probability of.

[0054] For multi-label prediction tasks, the supervised loss function is defined as follows:

[0055] in This represents one of the 200 tag types for multi-tag classification. Indicates drug triplet Does it belong to a type? . This indicates that the triple belongs to type The probability of.

[0056] To enhance the consistency between the structural view and the semantic view and improve robustness to perturbations and differences in cross-view analysis, a multi-view contrastive learning strategy based on a bidirectional contrast mechanism is introduced, using bidirectional normalized temperature scaling cross-entropy loss as the core optimization objective. Four contrastive losses are designed based on four different outputs of the heterogeneous graph encoder: (1) Comparison of the original molecular diagram and the perturbed molecular diagram; (2) Comparison between the original knowledge graph and the perturbed knowledge graph; (3) A comparison between molecular diagrams and knowledge graphs; (4) A comparison between perturbation molecular graphs and perturbation knowledge graphs.

[0057] Its general form is as follows:

[0058] in, This represents the value of the bidirectional normalized temperature-scaled cross-entropy loss function. , Indicates the current anchor point. It is a positive sample. As its negative sample, Indicates the number of samples. express and The similarity between them.

[0059] The final multi-view contrast loss is a weighted sum of the four contrast losses:

[0060] in, This represents the loss value compared to other views. This represents the loss weight coefficient. An adaptive weighting strategy is used here to adaptively weight different losses according to their importance. The overall training objective of the system is to minimize the joint supervised loss and contrastive loss.

[0061] in, This represents the final loss function value from the joint training. is a hyperparameter used to balance the weights of supervised loss and contrastive learning loss.

[0062] Corresponding to the drug interaction prediction method in the above embodiments, this invention also provides a drug interaction prediction system based on multi-view contrastive learning, including: Graph data processing module: processes drug-related data, constructs drug molecular graphs and biomedical knowledge graphs, and generates perturbation views based on information gain to form multi-view input; Heterogeneous graph encoder module: A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph and its perturbation view to obtain multi-view feature embedding; the heterogeneous graph neural network includes a multi-layer heterogeneous attention layer that fuses node and edge features, which can adaptively aggregate multi-relation information according to the differences in the types of nodes and edges, and introduce residual connections and batch normalization to alleviate the oversmoothing problem. Drug Interaction Classifier Module: A joint training module integrating supervised learning and multi-view contrastive learning is designed. It predicts drug interactions through feature fusion and leverages contrastive learning to improve consistency and robustness. The joint training module includes a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters, achieving complementary fusion of structural and semantic features. The multi-view contrastive learning unit employs bidirectional normalized temperature-scaling cross-entropy loss to impose contrast constraints on the representations between the original and perturbed views, and between structural and semantic views, enhancing cross-view... Figure 1 Consistency and perturbation robustness.

[0063] The drug interaction prediction system based on multi-view contrast learning in this embodiment of the invention is relatively simple to describe because it corresponds to the drug interaction prediction method based on multi-view contrast learning in the above embodiment. For related similarities, please refer to the description of the drug interaction prediction method based on multi-view contrast learning in the above embodiment, which will not be described in detail here.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drug interaction prediction method based on multi-view contrast learning, characterized in that, Includes the following steps: Drug-related data is processed to construct two heterogeneous graphs: a drug molecule graph and a biomedical knowledge graph. Based on an information gain-guided feature perturbation strategy, node and edge features are adaptively perturbed to generate structural views and semantic views, forming a multi-view input. A pre-trained heterogeneous graph neural network is used as an encoder to extract features from drug molecule graphs and knowledge graphs and their perturbed views to obtain multi-view feature embeddings. The heterogeneous graph neural network includes a multi-layer heterogeneous attention layer that fuses node and edge features. It can adaptively aggregate multi-relation information according to the differences in the types of nodes and edges. At the same time, residual connections and batch normalization are introduced to alleviate the oversmoothing problem. A joint training module integrating supervised learning and multi-view contrastive learning is designed to predict drug interactions through feature fusion and improve consistency and robustness through contrastive learning. The joint training module includes a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters to achieve complementary fusion of structural and semantic features. The multi-view contrastive learning unit uses bidirectional normalized temperature-scaling cross-entropy loss to impose contrast constraints on the representations between the original and perturbed views, as well as between the structural and semantic views, thereby enhancing cross-view consistency and perturbation robustness.

2. The method according to claim 1, characterized in that, Drug-related data is processed to construct drug molecule graphs and biomedical knowledge graphs. Perturbation views are then generated based on information gain to form a multi-view input, including: Write a preprocessing program for isomorphic graph structures to clean irrelevant data and convert drugs into molecular graph form and knowledge graph data into triplet form; Based on the information gain-based graph perturbation strategy, perturbation views are generated from preprocessed heterogeneous graph data to construct multi-view inputs.

3. The method according to claim 1, characterized in that, A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph, and their perturbed views to obtain multi-view feature embeddings, including: A pre-training mechanism is introduced into the heterogeneous graph encoder. For drug molecule graphs, a self-supervised pre-training strategy based on node masking is adopted, and for biomedical knowledge graphs, a self-supervised pre-training strategy based on link prediction is adopted. The drug molecule graph, knowledge graph, and their perturbed views are input into a pre-trained heterogeneous graph neural network for encoding to obtain multi-view feature embeddings.

4. The method according to claim 1, characterized in that, Design a joint training module that integrates supervised learning and multi-view contrastive learning. This module predicts drug interactions through feature fusion and leverages contrastive learning to improve consistency and robustness, including: Design a joint optimization module, including a supervised classifier based on feature fusion and a multi-view contrastive learner; Drug molecule graph embedding and knowledge graph embedding are fed into a multilayer perceptron for drug interaction prediction; Multi-view embeddings are used for contrastive learning, contrastive loss is calculated, and it is jointly optimized with supervised loss.

5. A drug interaction prediction system based on multi-view contrastive learning, characterized in that, include: Graph data processing module: processes drug-related data, constructs drug molecular graphs and biomedical knowledge graphs, and generates perturbation views based on information gain to form multi-view input; Heterogeneous graph encoder module: A pre-trained heterogeneous graph neural network is used as an encoder to encode the drug molecule graph, knowledge graph and its perturbation view to obtain multi-view feature embedding; the heterogeneous graph neural network includes a multi-layer heterogeneous attention layer that fuses node and edge features, which can adaptively aggregate multi-relation information according to the differences in the types of nodes and edges, and introduce residual connections and batch normalization to alleviate the oversmoothing problem. Drug Interaction Classifier Module: A joint training module integrating supervised learning and multi-view contrastive learning is designed to predict drug interactions through feature fusion and improve consistency and robustness using contrastive learning. The joint training module includes a gated feature fusion unit and a multi-view contrastive learning unit. The gated feature fusion unit adaptively balances the importance of molecular graph and knowledge graph embeddings through learnable gating parameters to achieve complementary fusion of structural and semantic features. The multi-view contrastive learning unit uses bidirectional normalized temperature-scaling cross-entropy loss to compare and constrain the representations between the original and perturbed views, as well as between the structural and semantic views, thereby enhancing cross-view consistency and perturbation robustness.

Images