Drug interaction prediction system and method based on time-space diagram attention network

Abstract

The invention relates to the technical field of drug interaction prediction, in particular to a drug interaction prediction system and method based on a space-time diagram attention network, and the system comprises a subgraph generation module, an enhanced space-time molecular graph representation module, a drug interaction prediction module connected with the subgraph generation module, a drug interaction prediction module connected with the enhanced space-time molecular graph representation module, and a drug interaction prediction module connected with the enhanced space-time molecular graph representation module. Receiving a drug molecule node representation vector and a temporal context vector; predicting interactions between drugs within a future time window based on the drug molecule node representation vector and the temporal context vector; the risk assessment module is connected with the drug interaction prediction module and receives an interaction prediction result between drugs in a future time window; and on the basis of the probability graph model, interpretable risk assessment is carried out on the prediction result, and accurate prediction and reliable risk assessment of drug interaction are realized by innovatively combining the time-space diagram attention network, the probability graph model and the interpretable risk assessment method.

Description

Technical Field

[0001] The present invention relates to the technical field of drug interaction prediction, in particular to a drug interaction prediction system and method based on a spatio-temporal graph attention network. Background Art

[0002] In recent years, with the popularization of polypharmacy and complex treatment regimens, the prediction and risk assessment of drug interactions have become key areas of clinical medicine and pharmaceutical research. Traditional drug interaction prediction methods mainly rely on laboratory studies and clinical trials, which are not only time-consuming and laborious but also difficult to comprehensively cover all possible drug combinations. With the development of computer technology and artificial intelligence, machine learning-based drug interaction prediction methods have gradually become a research hotspot.

[0003] Existing machine learning methods mainly focus on using static features such as the chemical structure of drugs, target information, and gene expression data to predict drug interactions. For example, some researchers use traditional machine learning algorithms such as support vector machines (SVM) or random forests (Random Forest) to predict interactions based on the molecular fingerprints or protein sequence information of drugs. Although these methods have achieved certain results in some cases, they still have obvious limitations. First, they often ignore the time-dynamic characteristics of drug actions and cannot capture the complex interaction patterns of drugs changing over time in the body. Second, these methods are difficult to effectively model the indirect interactions between drugs, especially the complex interaction networks mediated by biological macromolecules.

[0004] Recently, some researchers have begun to try using deep learning methods, such as graph neural networks (GNN), to model drug interaction networks. These methods have improved the modeling ability for complex interactions to a certain extent, but still mainly focus on static interaction patterns and are difficult to accurately predict and evaluate the risks of dynamically changing drug interactions. In addition, most of the existing methods lack interpretability of the prediction results, which is a serious defect in clinical decision-making.

[0005] In view of the deficiencies of the existing technology, there is an urgent need for a prediction system that can simultaneously consider the time-dynamic characteristics and complex network structure of drug interactions and provide interpretable risk assessment. The present invention is an innovative solution proposed in response to this urgent need. Summary of the Invention

[0006] The drug interaction prediction system and its risk assessment method based on a spatio-temporal graph attention network proposed by the present invention aim to solve the technical problems in the existing technology that it is impossible to effectively capture the time-dynamic characteristics of drug interactions, difficult to model complex interaction networks, and lack of interpretable risk assessment.

[0007] The present invention proposes a drug interaction prediction system based on a spatio-temporal graph attention network, including:

[0008] A subgraph generation module, configured to:

[0009] Obtain data of drug molecules and their metabolites;

[0010] Generate a spatio-temporal dynamic graph of drug molecules including direct interaction subgraphs and indirect interaction subgraphs;

[0011] An enhanced spatio-temporal molecular graph representation module, connected to the subgraph generation module, configured to:

[0012] Receive the spatio-temporal dynamic graph of drug molecules;

[0013] Based on the spatio-temporal dynamic graph of drug molecules, capture the time dynamic features of the drug molecular graph;

[0014] Establish an intermediate transmission relationship of interactions between drugs through biological macromolecules;

[0015] Generate drug molecule node representation vectors and time context vectors;

[0016] A drug interaction prediction module, connected to the enhanced spatio-temporal molecular graph representation module, configured to:

[0017] Receive the drug molecule node representation vectors and time context vectors;

[0018] Based on the drug molecule node representation vectors and time context vectors, predict the interactions between drugs within a future time window;

[0019] A risk assessment module, connected to the drug interaction prediction module, configured to:

[0020] Receive the interaction prediction results between drugs within the future time window;

[0021] Based on a probabilistic graph model, perform an interpretable risk assessment on the prediction results.

[0022] Preferably, the subgraph generation module includes:

[0023] A supervised learning unit, configured to:

[0024] Calculate the correlation between the vector representations of drug nodes in a training set of drug molecules;

[0025] Obtain a correlation matrix M1 between drugs;

[0026] A bi-directional graph attention unit, configured to:

[0027] Based on the correlation matrix M1, capture the direct interaction relationships between drugs;

[0028] Generate a direct interaction sub - graph H1;

[0029] A self - supervised learning unit for:

[0030] Introduce a graph attention layer to capture the direct and indirect interaction relationships between drugs;

[0031] Generate an indirect interaction sub - graph H2.

[0032] Preferably, the bidirectional graph attention unit updates the drug molecule node representation using the following calculation formula:

[0033] H1 = H1′·W1 + b1,

[0034] where H′1 is the output of the first - layer graph attention network, W1 is a trainable parameter matrix, and b1 is a bias term.

[0035] Preferably, the self - supervised learning unit updates the drug molecule node representation using the following calculation formula:

[0036] H2 = H2″·W2 + b2,

[0037] where H″2 is the output of the second - layer graph attention network, W2 is a trainable parameter matrix, and b2 is a bias term.

[0038] Preferably, the enhanced spatio - temporal molecular graph representation module includes:

[0039] A spatio - temporal feature extraction unit for:

[0040] Calculate the relationship matrix M between the t - k moment and the t moment k ;

[0041] Based on the relationship matrix M k , obtain the interaction relationship features between drugs and time;

[0042] A first graph attention update unit for:

[0043] Based on the relationship matrix M k , calculate the drug molecule node representation at the t moment;

[0044] A second graph attention update unit for:

[0045] Calculate the correlation matrix between drug molecule nodes;

[0046] Based on the correlation matrix, further update the drug molecule node representation at the t moment.

[0047] Preferably, the first graph attention update unit updates the representation of the drug molecule node at time t using the following calculation formula:

[0048] X′(i) = σ′(a i ·(M t-k ·S(i) + M t-k ·S(j))),

[0049] where X′(i) is the updated representation of the drug molecule node, σ′ is the activation function, a i is a learnable parameter, M t -k is the relationship matrix between time t-k and time t, and S(i) and S(j) are the representations of nodes i and j respectively.

[0050] Preferably, the drug interaction prediction module includes:

[0051] A global average pooling unit for:

[0052] Fusing the drug molecule node representation vector into the time context vector;

[0053] A probabilistic graph model unit for:

[0054] Constructing a probabilistic graph model of drug interactions;

[0055] Evaluating the distribution of the probabilities of drug interactions within the future T time window.

[0056] Preferably, the probabilistic graph model unit constructs the probabilistic graph model using the following formula:

[0057] P(Y|X) = Π t P(Y t |X t ,Y t-1 ),

[0058] where Y represents the prediction result of drug interactions, X represents the input features, and t represents the time step. Preferably, the risk assessment module uses the maximum likelihood estimation method for risk assessment, and its

[0059] approximate solution is:

[0060]

[0061] Preferably, the probabilistic graph model unit constructs the probabilistic graph model using the following formula where K is the total number of negative time windows, K′ is the total number of positive time windows, Z t-k is the latent variable at time t-k, Z t+1 is the latent variable at time t+1, and θ is the model parameter.

[0062] The method of using the drug interaction prediction system based on the spatio-temporal graph attention network includes the following steps:

[0063] Construct a drug molecular graph, including:

[0064] Obtain a drug molecular graph through the subgraph generation module, where the drug molecular graph includes a direct interaction subgraph and an indirect interaction subgraph;

[0065] Capture spatio-temporal features, including:

[0066] Capture the time dynamic features of the drug molecular graph through the enhanced spatio-temporal molecular graph representation module;

[0067] Establish an intermediary transmission relationship for the indirect interaction between drugs through biomacromolecules;

[0068] Obtain drug molecular node representation vectors and time context vectors;

[0069] Predict drug interactions, including:

[0070] Through the drug interaction prediction module, use the drug molecular node representation vectors and time context vectors to predict the interactions between drugs within a future time window;

[0071] Conduct risk assessment, including:

[0072] Evaluate the interaction relationship between drugs and the time window according to the prediction results of the interactions between drugs;

[0073] Construct a drug interaction prediction system based on a probabilistic graphical model;

[0074] Based on the drug interaction prediction system of the probabilistic graphical model, conduct interpretable risk assessment.

[0075] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0076] The present invention innovatively combines the spatio-temporal graph attention network, the probabilistic graphical model, and the interpretable risk assessment method to achieve accurate prediction and reliable risk assessment of drug interactions. The core advantages of this system are reflected in the following aspects:

[0077] First of all, the subgraph generation module of the present invention comprehensively captures the complex interaction relationships between drug molecules by fusing the direct interaction subgraph and the indirect interaction subgraph. This design not only considers the direct effects of drugs but also can effectively model the indirect interactions mediated by biomacromolecules, greatly improving the modeling ability of complex drug interaction networks.

[0078] Secondly, the enhanced spatio-temporal molecular graph representation module innovatively introduces the time dimension. By capturing the time-dynamic characteristics of drug molecular graphs, it achieves accurate modeling of the temporal pattern of drug interactions. This feature enables the present invention to accurately predict the changes in the interaction strength of drugs at different time points, providing an important basis for formulating clinical medication plans.

[0079] Furthermore, the drug interaction prediction module adopts an advanced probabilistic graph model, which can not only predict the strength of drug interactions but also evaluate the uncertainty of the prediction results. This probabilistic method greatly improves the reliability of the prediction, providing more comprehensive information support for clinical decision-making.

[0080] Finally, the risk assessment module of the present invention realizes the interpretability assessment of the prediction results by using the maximum likelihood estimation method. This feature enables doctors and researchers to clearly understand the basis and reliability of the prediction results, thus making more informed clinical decisions.

[0081] From the overall architecture, the modules of the present invention cooperate closely to form a complete drug interaction prediction and risk assessment process. The subgraph generation module provides rich feature representations for subsequent analysis, the enhanced spatio-temporal molecular graph representation module captures the time-dynamic characteristics, the drug interaction prediction module realizes accurate prediction, and the risk assessment module provides reliable risk assessment. This synergistic effect among the modules not only solves various technical contradictions in the prior art but also produces significant effect superposition and complementarity.

[0082] From a microscopic perspective, the present invention has innovations and breakthroughs in many technical details. For example, the bidirectional graph attention mechanism adopted in the subgraph generation process can effectively capture the asymmetric interactions between drug molecules; the multi-level graph attention update mechanism used in the enhanced spatio-temporal molecular graph representation module can make full use of the multi-scale characteristics of drug interactions. These innovations at the microscopic level jointly build the technical barrier of the present invention, giving it significant technical advantages in the field of drug interaction prediction and risk assessment.

[0083] Generally speaking, the present invention innovatively solves the key technical problems in drug interaction prediction, not only improving the accuracy and reliability of the prediction but also providing interpretable risk assessment results for clinical decision-making. This system is expected to play an important role in the fields of personalized medicine, new drug research and development, and drug safety assessment, making important contributions to improving medication safety and treatment effects. Brief Description of the Drawings

[0084] Figure 1 It is the logical block diagram of the overall system of the present invention.

[0085] Figure 2It is the logic block diagram of the sub - graph generation module of the present invention.

[0086] Figure 3 It is the logic block diagram of the enhanced spatio - temporal molecular graph representation module of the present invention.

[0087] Figure 4 It is the logic block diagram of the drug interaction prediction module of the present invention.

[0088] Figure 5 It is the logic block diagram of the risk assessment module of the present invention. Detailed implementation manners

[0089] Refer to Figures 1-5 , the present invention provides a drug interaction prediction system based on spatio - temporal graph attention network and its risk assessment method. The system mainly includes a sub - graph generation module 1, an enhanced spatio - temporal molecular graph representation module 2, a drug interaction prediction module 3, and a risk assessment module 4.

[0090] The sub - graph generation module 1 is used to obtain data of drug molecules and their metabolites, and generate a spatio - temporal dynamic graph of drug molecules including a direct interaction sub - graph and an indirect interaction sub - graph. Preferably, the sub - graph generation module 1 includes a supervised learning unit 11, a bi - directional graph attention unit 12, and a self - supervised learning unit 13.

[0091] The supervised learning unit 11 is mainly responsible for calculating the correlation between the vector representations of drug nodes in the training set of drug molecules, and obtaining a correlation matrix M1 of drugs. In an embodiment of the present invention, the calculation of the correlation matrix M1 can adopt the following formula:

[0092]

[0093] where, x i and x j respectively represent the vector representations of drug molecules i and j.

[0094] The bi - directional graph attention unit 12 captures the direct interaction relationship between drugs based on the correlation matrix M1, and generates a direct interaction sub - graph H1. The present invention uses the following calculation formula to update the drug molecule node representation:

[0095] H1 = H1′ * W1 + b1,

[0096] where, H1 ′ is the output after the first - layer graph attention network, W1 is a trainable parameter matrix, and b1 is a bias term. Specifically, the calculation of H1 ′ can adopt the following attention mechanism:

[0097]

[0098] Among them, Q, K, and V are the query, key, and value matrices respectively, and d k is the dimension of the key vector. This attention mechanism can effectively capture the complex interactions between drug molecules.

[0099] The supervised learning unit 11 is mainly responsible for calculating the correlation between the vector representations of drug nodes in the training set of drug molecules and obtaining the correlation matrix M1 of drugs. In an embodiment of the present invention, the calculation of the correlation matrix M1 can adopt the following formula:

[0100] The self-supervised learning unit 13 introduces a graph attention layer to capture the direct and indirect interaction relationships between drugs and generates an indirect interaction subgraph H2. The present invention updates the drug molecule node representation using the following calculation formula:

[0101] H2 = H2″ * W2 + b2,

[0102] where H2″ is the output after the second-layer graph attention network, W2 is a trainable parameter matrix, and b2 is a bias term. The calculation of H2″ can adopt a similar attention mechanism to $H$11, but with the addition of a self-supervised learning loss function, for example:

[0103]

[0104] where z i and z j are the representations of positive sample pairs, and τ is the temperature parameter. In this way, the subgraph generation module 1 can comprehensively capture the direct and indirect interactions between drug molecules, providing a solid foundation for subsequent drug interaction prediction.

[0105] In this way, the subgraph generation module 1 can comprehensively capture the direct and indirect interactions between drug molecules, providing a solid foundation for subsequent drug interaction prediction. This design of the present invention has the following advantages: First, it can consider both the structural information and functional information of drug molecules; Second, by introducing self-supervised learning, pre-training can be performed on unlabeled data to improve the generalization ability of the model; Finally, the bidirectional graph attention mechanism can capture the non-linear and dynamic interactions between drug molecules.

[0106] In practical applications, the parameters of the subgraph generation module 1 can be adjusted according to the specific drug dataset. For example, for small molecule drugs, a smaller hidden layer dimension (such as 64 or 128) can be selected; while for large molecule drugs or complex biological agents, a larger hidden layer dimension (such as 256 or 512) may be required to capture richer features. At the same time, the number of attention heads can also be adjusted according to the complexity of the task, usually selected between 4 and 16.

[0107] Generally speaking, the subgraph generation module 1 of the present invention provides a powerful and flexible feature extraction framework for drug interaction prediction by combining supervised learning and self-supervised learning, and innovatively applying a bidirectional graph attention mechanism. This method can not only effectively capture the complex relationships between drug molecules, but also has good interpretability and scalability, laying a solid foundation for subsequent drug interaction prediction and risk assessment. The enhanced spatio-temporal molecular graph representation module 2 of the present invention is connected to the subgraph generation module 1 and is used to receive the spatio-temporal dynamic graph of drug molecules and capture the temporal dynamic features of the drug molecular graph based on this graph. This module is also responsible for establishing the mediating transmission relationship between drugs through biomacromolecule interactions, and finally generating the drug molecular node representation vector and the time context vector.

[0108] Preferably, the enhanced spatio-temporal molecular graph representation module 2 includes a spatio-temporal feature extraction unit 21, a first graph attention update unit 22, and a second graph attention update unit 23. This modular design enables the present invention to more finely process the spatio-temporal features of drug molecules, thereby improving the prediction accuracy.

[0109] The spatio-temporal feature extraction unit 21 is mainly responsible for calculating the relationship matrix M between the t-k moment and the t moment k , and obtaining the interaction relationship features between drugs and time based on this matrix. In an embodiment of the present invention, the calculation of the relationship matrix M k can adopt the following formula:

[0110]

[0111] where Q k and K k are the query and key matrices at the t-k moment and the t moment respectively, and d k is the dimension of the key vector. This calculation method can effectively capture the drug interaction relationships between different time points. The first graph attention update unit 22 calculates the drug molecular node representation at the t moment based on the relationship matrix M k . The present invention uses the following calculation formula to update the drug molecular node representation at the t moment:

[0112] X′(i) = σ′(a i * (M t-k * S(i) + M t-k * S(j))),

[0113] where X′(i) is the updated drug molecular node representation, σ′ is the activation function, a i is the learnable parameter, and M t-kis the relationship matrix between t-k and t moments, and S(i) and S(j) are the representations of nodes i and j respectively. This update method can consider both the characteristics of the node itself and the influence of neighbor nodes, thereby more comprehensively capturing the spatio-temporal characteristics of drug molecules.

[0114] The second graph attention update unit 23 further calculates the correlation matrix between drug molecule nodes and updates the representation of drug molecule nodes at time t based on this matrix. In a preferred embodiment of the present invention, the calculation of the correlation matrix can adopt the following formula:

[0115]

[0116] where W is the weight matrix, a is the attention vector, h i and h j are the characteristics of nodes i and j respectively, is the neighbor set of node i. The use of the LeakyReLU activation function can effectively alleviate the problem of gradient disappearance and improve the training effect of the model.

[0117] The drug interaction prediction module 3 is connected to the enhanced spatio-temporal molecular graph representation module 2, and is used to receive the drug molecule node representation vector and the time context vector, and predict the interaction between drugs within the future time window based on these vectors.

[0118] Preferably, the drug interaction prediction module 3 includes a global average pooling unit 31 and a probabilistic graph model unit 32.

[0119] The global average pooling unit 31 is mainly responsible for fusing the drug molecule node representation vector into the time context vector. This fusion operation can be achieved through the following formula:

[0120]

[0121] where h G is the fused global representation, h i is the representation of the i-th node, and N is the total number of nodes. This global average pooling operation can effectively aggregate the information of all nodes and provide a global perspective for subsequent prediction tasks.

[0122] The probabilistic graph model unit 32 is responsible for constructing the probabilistic graph model of drug interactions and evaluating the distribution of the interaction probabilities of drugs within the future T time window. In an embodiment of the present invention, the probabilistic graph model is constructed using the following formula:

[0123] P(Y|X) = ∏ t P(Y t |X t ,Y t-1 ),

[0124] Among them, Y represents the prediction result of drug interaction, X represents the input feature, and t represents the time step. This probabilistic graphical model not only considers the features at the current moment but also introduces the prediction result at the previous moment, thus being able to better capture the time dependence of drug interactions.

[0125] The risk assessment module 4 is connected to the drug interaction prediction module 3, and is used to receive the prediction results of the interactions between drugs within a future time window, and perform interpretable risk assessment on these prediction results based on the probabilistic graphical model. Preferably, the risk assessment module 4 uses the maximum likelihood estimation method for risk assessment, and its approximate solution is:

[0126]

[0127] Among them, K is the total number of negative time windows, K′ is the total number of positive time windows, Z t-k is the latent variable at time t - k, Z t+1 is the latent variable at time t + 1, and θ is the model parameter. This maximum likelihood estimation method can effectively evaluate the reliability of the prediction results and provide an important reference for clinical decision-making.

[0128] In an embodiment of the present invention, the selection of the negative time window K and the positive time window K' can be adjusted according to the duration of the drug action and the specific requirements of the prediction task. For example, for short-acting drugs, smaller values of K and K' (such as 3 - 5) can be selected; while for long-acting drugs, larger values of K and K' (such as 7 - 14) can be selected. This flexible time window setting enables the system of the present invention to adapt to different types of drug interaction prediction tasks.

[0129] In addition, the present invention also provides a method for predicting drug interactions using the above system. The method includes the following steps:

[0130] First, construct a drug molecular graph. This step obtains the drug molecular graph through the subgraph generation module 1, including a direct interaction subgraph and an indirect interaction subgraph. In a preferred embodiment of the present invention, the construction of the drug molecular graph can be based on the chemical structure of the drug, target information, and known interaction data. For example, a graph neural network (GNN) or molecular fingerprint technology can be used to encode the structural information of the drug, and at the same time, a knowledge graph is combined to capture the known interaction relationships between drugs.

[0131] Secondly, capture spatio-temporal features. This step is completed by the enhanced spatio-temporal molecular graph representation module 2, which mainly includes capturing the temporal dynamic features of drug molecular graphs, establishing the mediating transmission relationship between drugs indirectly through the interaction of biological macromolecules, and obtaining the drug molecular node representation vectors and temporal context vectors. In practical applications, recurrent neural networks (RNNs) or transformer architectures can be used to model time series data to better capture the dynamic features of drug interactions.

[0132] Thirdly, predict drug interactions. This step is achieved by the drug interaction prediction module 3, which uses the drug molecular node representation vectors and temporal context vectors obtained previously to predict the interactions between drugs within a future time window. In an embodiment of the present invention, an attention mechanism can be used to adaptively adjust the importance of different features, thereby improving the accuracy of prediction.

[0133] Finally, conduct risk assessment. This step first evaluates the interaction relationship between drugs and the time window according to the prediction results of drug interactions. Then, a drug interaction prediction system based on a probabilistic graphical model is constructed, and an interpretable risk assessment is performed based on this system. In practical applications, Bayesian networks or Markov random fields can be used to model the probabilistic dependence relationship of drug interactions, thereby achieving more accurate risk assessment.

[0134] The method of the present invention realizes the accurate prediction of drug interactions and reliable risk assessment by combining graph neural networks, time series analysis, and probabilistic graphical models. This method can not only effectively utilize the structural information of drugs and historical interaction data, but also capture the temporal dynamic characteristics of drug actions, providing strong support for clinical medication decisions.

[0135] In practical applications, the method of the present invention can be flexibly adjusted according to specific datasets and prediction tasks. For example, for large-scale drug databases, distributed computing technologies can be adopted to improve the processing efficiency of the system; for rare drugs or new drugs, transfer learning or few-shot learning technologies can be introduced to improve the accuracy of prediction. In addition, the method of the present invention can also be combined with other machine learning technologies (such as deep reinforcement learning) to further optimize the prediction and risk assessment processes of drug interactions.

[0136] Generally speaking, the drug interaction prediction system and its risk assessment method based on the spatio-temporal graph attention network provided by the present invention achieve the accurate prediction of drug interactions and reliable risk assessment by innovatively combining a variety of advanced machine learning technologies. This not only provides an important reference for clinical medication decisions, but also provides strong support for new drug research and development and personalized medicine, with broad application prospects.

[0137] It should be noted that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A drug interaction prediction system based on spatiotemporal graph attention network, characterized by: include: Subgraph generation module, used to: Obtain data on drug molecules and their metabolites; Generate a spatiotemporal dynamic graph of drug molecules including a direct interaction subgraph and an indirect interaction subgraph; The enhanced spatiotemporal molecular graph representation module is connected to the subgraph generation module and is used to: Receiving the spatiotemporal dynamic graph of drug molecules; Based on the spatiotemporal dynamic graph of drug molecules, capturing the temporal dynamic characteristics of the drug molecule graph; Establish the mediated transmission relationship between drugs through the interaction of biomacromolecules; Generate drug molecule node representation vector and time context vector; The drug interaction prediction module is connected to the enhanced spatiotemporal molecular graph representation module and is used to: Receiving the drug molecule node representation vector and the time context vector; Predicting the interactions between drugs in a future time window based on the drug molecule node representation vector and the time context vector; A risk assessment module, connected to the drug interaction prediction module, is used to: receiving a prediction result of the interaction between drugs within the future time window; Based on the probabilistic graphical model, an interpretability risk assessment is performed on the prediction results.

2. The system according to claim 1, characterized in that The sub-graph generation module comprises: Supervised learning units for: Calculate the correlation between the vector representations of drug nodes in the training set of drug molecules; Obtain the correlation matrix between drugs ; Bidirectional graph attention unit for: Based on the correlation matrix , capturing the direct interaction relationship between drugs; Generating Direct Interaction Subgraphs ; Self-supervised learning unit for: Introducing a graph attention layer to capture the direct and indirect interactions between drugs; Generate indirect interaction subgraph .

3. The system according to claim 2, characterized in that The bidirectional graph attention unit uses the following calculation formula to update the drug molecule node representation: , in, is the output of the first layer of graph attention network, is the trainable parameter matrix, is the bias term.

4. The system according to claim 2, characterized in that The self-supervised learning unit updates the drug molecule node representation using the following calculation formula: , in, is the output of the second-layer graph attention network, is the trainable parameter matrix, is the bias term.

5. The system according to claim 1, characterized in that The enhanced spatiotemporal molecular graph representation module comprises: Spatiotemporal feature extraction unit, used for: calculate Moment and The relationship matrix of the moment ; Based on the relationship matrix , obtain the interactive relationship characteristics between drugs and time; The first image attention update unit is used to: Based on the relationship matrix ,calculate Drug molecule node representation at the moment; The second image attention update unit is used to: Calculate the correlation matrix between drug molecule nodes; Based on the correlation matrix, further update The drug molecule node representation at the time.

6. The system according to claim 5, characterized in that The first graph attention update unit uses the following calculation formula to update the drug molecule node representation at time t: , in, is the updated drug molecule node representation, is the activation function, is a learnable parameter, for Moment and The relationship matrix of time, and They are and Representation of a node.

7. The system according to claim 1, characterized in that The drug interaction prediction module comprises: Global average pooling unit, used for: fusing the drug molecule node representation vector into the time context vector; Probabilistic graphical model unit, used to: Construct probabilistic graphical models of drug interactions; Evaluate the distribution of drug-drug interaction probabilities over a future time window of T.

8. The system according to claim 7, characterized in that The probability graph model unit uses the following formula to construct the probability graph model: , in, represents the prediction result of drug interaction, represents the input features, Represents the time step.

9. The system according to claim 1, characterized in that The risk assessment module uses the maximum likelihood estimation method to perform risk assessment, and its approximate solution is: 。 10. The system according to claim 7, characterized in that The probability graph model unit uses the following formula to construct the probability graph model, where: is the total number of negative time windows, is the total number of positive time windows, for The latent variables at time, for The latent variables at time, is the model parameter.

11. A method for using a drug interaction prediction system based on a spatiotemporal graph attention network as described in any one of claims 1 to 10, characterized in that: The following steps are involved: Construct drug molecule graphs, including: Obtaining a drug molecule graph through the subgraph generation module, wherein the drug molecule graph includes a direct interaction subgraph and an indirect interaction subgraph; Captures spatiotemporal features, including: The temporal dynamic characteristics of the drug molecular graph are captured by the enhanced spatiotemporal molecular graph representation module; Establish the indirect transmission relationship between drugs through the interaction of biomacromolecules; Obtain drug molecule node representation vector and time context vector; Predict drug interactions, including: The drug interaction prediction module predicts the interactions between drugs in a future time window using the drug molecule node representation vector and the time context vector; Conduct a risk assessment, including: According to the predicted results of the interactions between the drugs, the interactive relationship between the drugs and the time window is evaluated; Construct a drug interaction prediction system based on probabilistic graphical models; The drug interaction prediction system based on the probabilistic graphical model performs explainable risk assessment.

Images