Asymmetric drug-drug interaction prediction method and system based on graph network

Abstract

The invention discloses an asymmetric drug-drug interaction prediction method and system based on a graph network, and the method comprises the steps: obtaining drug data based on a comprehensive drug database, carrying out the data preprocessing, and constructing drug-DDI graph data and drug-drug graph data; introducing a graph attention network and a heterogeneous graph neural network, and constructing a heterogeneous graph neural network asymmetric DDI prediction model; and based on the heterogeneous graph neural network asymmetric DDI prediction model, performing drug interaction probability prediction on the drug-DDI graph data and the drug-drug graph data to obtain an asymmetric drug-drug interaction prediction result. The method can generate the interaction probability, achieves the precise prediction of the asymmetric DDI, and improves the accuracy and reliability of an asymmetric drug-drug interaction prediction result. The asymmetric drug-drug interaction prediction method and system based on the graph network can be widely applied to the technical field of drug interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug interaction, in particular to an asymmetric drug-drug interaction prediction method and system based on a graph network. BACKGROUND

[0002] In the field of modern drug research and development, drug-drug interaction (DDI) prediction is of great significance for reducing experimental costs, accelerating new drug discovery, and ensuring the safety of clinical drug use. With the continuous expansion and improvement of drug database resources, DDI prediction methods based on deep learning have made significant progress. However, existing methods have obvious shortcomings in dealing with asymmetric DDI problems.

[0003] Firstly, the lack of consideration of asymmetry, most existing methods treat the two drugs in the interaction as equal roles in pharmacology, ignoring the pharmacological asymmetry of drug interaction. For example, some experimental studies have shown that most drug interactions between antifungal drugs and non-antifungal drugs are unidirectional, and the asymmetric interaction of drugs is also related to the order of drug administration. However, existing methods fail to fully consider this asymmetry, resulting in limited prediction accuracy. Secondly, the inability to effectively distinguish the type tendency difference in the interaction direction, although a few studies have begun to focus on the asymmetry of DDI, these methods also have shortcomings. For example, the DGAT-DDI model does not consider the asymmetry of drug interaction and relationship type caused by the relationship information between drugs, and the DRGATAN model attempts to introduce relationship graph attention to handle DDI types, but its simple type embedding fails to distinguish the type tendency difference of drugs in different interaction directions, i.e., it fails to fully capture the type difference of drugs as the source and target of action. SUMMARY

[0004] To solve the above technical problems, the purpose of the present application is to provide an asymmetric drug-drug interaction prediction method and system based on a graph network, which can generate interaction probabilities and achieve accurate prediction of asymmetric DDI, improving the accuracy and reliability of asymmetric drug-drug interaction prediction results.

[0005] The first technical solution adopted by the present application is an asymmetric drug-drug interaction prediction method based on a graph network, comprising the following steps: Obtain drug data based on a comprehensive drug database and perform data preprocessing to construct drug-DDI graph data and drug-drug graph data; Introduce a graph attention network and a heterogeneous graph neural network to construct a heterogeneous graph neural network asymmetric DDI prediction model; Based on the asymmetric DDI prediction model of the heterogeneous graph neural network, drug interaction probability prediction is performed on the drug-DDI graph data and the drug-drug graph data, and an asymmetric drug-drug interaction prediction result is obtained.

[0006] Further, the step of obtaining drug data based on a comprehensive drug database and performing data preprocessing, and constructing drug-DDI graph data and drug-drug graph data specifically comprises: Obtaining drug data based on a comprehensive drug database and performing data screening to obtain screened drug data; Performing molecular fingerprint feature extraction processing on the screened drug data to obtain drug feature data; Performing semantic feature extraction processing on the screened drug data based on BERT embedding to obtain DDI features; Regarding the drug feature data as a node in the graph, if there is DDI between two drugs, a directed edge is established between the corresponding nodes to construct drug-drug graph data; Regarding the drug feature data and the DDI features as graph nodes, and establishing a directed edge between the drug feature data and the DDI features to construct drug-DDI graph data.

[0007] Further, the asymmetric DDI prediction model of the heterogeneous graph neural network specifically comprises an encoder and a decoder, and the output end of the encoder is connected with the input end of the decoder, wherein: The encoder comprises a first heterogeneous graph neural network, a second heterogeneous graph neural network, a first graph attention network, a second graph attention network, a first multilayer perceptron, a second multilayer perceptron, a first gating layer and a second gating layer; The decoder comprises a first linear layer, a second linear layer, a third linear layer and a fourth linear layer; The first output end of the first heterogeneous graph neural network is connected with the first input end of the first gating layer, the second output end of the first heterogeneous graph neural network is connected with the input end of the first linear layer, the output end of the first graph attention network is connected with the second input end of the first gating layer, the output end of the first gating layer is connected with the input end of the second linear layer, the output end of the second graph attention network is connected with the first input end of the second gating layer, the output end of the second gating layer is connected with the input end of the third linear layer, the first output end of the second heterogeneous graph neural network is connected with the second input end of the second gating layer, and the second output end of the second heterogeneous graph neural network is connected with the input end of the fourth linear layer.

[0008] Further, the expression of the binary cross-entropy loss function of the asymmetric DDI prediction model of the heterogeneous graph neural network is specifically as follows: in the above formula, denotes the binary cross-entropy loss function of the heterogeneous graph neural network asymmetric DDI prediction model, denotes the drug interaction probability prediction value, denotes the total number of drug interactions, denotes the true label of the drug interaction denotes the drug and the edge between the drug .

[0009] Further, the drug interaction probability prediction based on the heterogeneous graph neural network asymmetric DDI prediction model is performed on the drug-DDI graph data and the drug-drug graph data, and the asymmetric drug-drug interaction prediction result is obtained, and the specific steps are as follows: The drug-DDI graph data and the drug-drug graph data are input into the heterogeneous graph neural network asymmetric DDI prediction model; The encoder of the heterogeneous graph neural network asymmetric DDI prediction model is used to extract features of the drug-DDI graph data and the drug-drug graph data, and the multi-modal features of the drug nodes are obtained; The decoder of the heterogeneous graph neural network asymmetric DDI prediction model is used to introduce a DDI type-specific weight generation mechanism to predict the drug interaction probability based on the multi-modal features of the drug nodes, and the drug interaction probability prediction value is obtained; The asymmetric drug-drug interaction prediction result is determined according to the drug interaction probability prediction value.

[0010] Further, the encoder of the heterogeneous graph neural network asymmetric DDI prediction model is used to extract features of the drug-DDI graph data and the drug-drug graph data, and the multi-modal features of the drug nodes are obtained, and the specific steps are as follows: The drug-DDI graph data and the drug-drug graph data are input into the encoder; The first graph attention network and the second graph attention network of the encoder are used to extract features of the drug-drug graph data in a direction-aware manner, and the drug source role features and the drug target role features are obtained; The first heterogeneous graph neural network and the second heterogeneous graph neural network of the encoder are used to capture the heterogeneous relationship between the drug and the DDI type based on the drug-DDI graph data, and the source role embedding features of the drug and the target role embedding features of the drug are obtained; The first gating layer and the second gating layer of the encoder are used to respectively perform weighted splicing fusion on the drug source role features and the source role embedding features of the drug, and the drug target role features and the target role embedding features of the drug; The first and second multi-layer perceptrons based on the encoder obtain the chemical structure information of the source drug in the drug-drug graph data, and obtain the source role encoding feature of the drug and the target role encoding feature of the drug. The multi-modal feature of the drug node is obtained by combining the source role final feature of the drug, the target role final feature of the drug, the source role encoding feature of the drug, the target role encoding feature of the drug, the source role embedding feature of the drug and the target role embedding feature of the drug.

[0011] Further, the first and second graph attention networks based on the encoder perform direction-aware feature extraction on the drug-drug graph data to obtain the source role feature of the drug and the target role feature of the drug, and the specific steps include: The drug-drug graph data is input into the first and second graph attention networks respectively. According to the drug-drug graph data, the first-order out-neighbor and the first-order in-neighbor of the drug node are defined to obtain the information of the adjacent nodes of the drug; The information of the adjacent nodes of the drug is aggregated by the attention mechanism to obtain the attention score of the drug node; According to the attention score of the drug node, the source role feature of the drug and the target role feature of the drug are defined.

[0012] Further, the decoder of the asymmetric DDI prediction model based on the heterogeneous graph neural network introduces a DDI type-specific weight generation mechanism to perform drug interaction probability prediction on the multi-modal feature of the drug node to obtain the drug interaction probability prediction value, and the specific steps include: The multi-modal feature of the drug node is input into the decoder. The second and third linear layers of the decoder are used to align the feature space of the source role final feature of the drug and the target role final feature of the drug to obtain the source role final feature of the drug after dimension alignment and the target role final feature of the drug after dimension alignment; The first and fourth linear layers of the decoder are used to map the source role embedding feature of the drug and the target role embedding feature of the drug into type-specific weights through a DDI type perception mechanism to obtain the source role attention weight of the drug node and the target role attention weight of the drug node; The source role probability prediction value of the drug node is obtained by performing inner product operation on the source role final feature of the drug after dimension alignment and the source role encoding feature of the drug, and then performing inner product operation on the source role attention weight of the drug node; The target role probability prediction value of the drug node is obtained by performing inner product operation on the target role final feature of the drug after dimension alignment and the target role encoding feature of the drug, and then performing inner product operation on the target role attention weight of the drug node; The source role probability prediction value of the drug node is added to the target role probability prediction value of the drug node to obtain a drug interaction probability prediction value.

[0013] Further, the expression of the drug interaction probability prediction value is specifically as follows: In the above formula, represents the drug interaction probability prediction value, , respectively represent the source role feature of the drug and the target role feature of the drug , , represent a weight matrix, , represent the self role feature of the drug, , represent a bias term, , respectively represent the weight of the corresponding drug under the weight of the source role and the weight of the drug under the target role.

[0014] The second technical solution adopted by the present application is: an asymmetric drug-drug interaction prediction system based on a graph network, comprising: A first module for obtaining drug data based on a comprehensive drug database and performing data preprocessing, constructing drug-DDI graph data and drug-drug graph data; A second module for introducing a graph attention network and a heterogeneous graph neural network to construct a heterogeneous graph neural network asymmetric DDI prediction model; A third module for predicting the drug interaction probability based on the heterogeneous graph neural network asymmetric DDI prediction model, the drug-DDI graph data and the drug-drug graph data, and obtaining an asymmetric drug-drug interaction prediction result.

[0015] The method and system have the beneficial effects that: the method and system obtain drug data based on a comprehensive drug database, and perform data preprocessing, construct drug-DDI graph data and drug-drug graph data, further introduce a graph attention network and a heterogeneous graph neural network, construct a heterogeneous graph neural network asymmetric DDI prediction model, use a bidirectional graph attention network to extract feature representations of drugs as action sources and action targets, fully consider the differences of drugs in different interaction directions, and finally predict drug interaction probability based on the heterogeneous graph neural network asymmetric DDI prediction model, the drug-DDI graph data and the drug-drug graph data, the asymmetric drug-drug interaction prediction result can generate interaction probability, realize accurate prediction of asymmetric DDI, and improve the accuracy and reliability of the asymmetric drug-drug interaction prediction result. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a step flow chart of the asymmetric drug-drug interaction prediction method based on a graph network of the present application; Figure 2 is a structural block diagram of the asymmetric drug-drug interaction prediction system based on a graph network of the present application; Figure 3 is a schematic diagram for obtaining DDI type features using BERT provided by the embodiment of the present application; Figure 4 is a schematic diagram of a drug-drug graph network provided by the embodiment of the present application; Figure 5 is a schematic diagram of a drug-DDI heterogeneous graph network provided by the embodiment of the present application; Figure 6 is a schematic diagram of a heterogeneous graph neural network asymmetric DDI prediction model provided by the embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be further described in detail below in combination with the drawings and specific embodiments. For the step numbers in the following embodiments, only the setting is for the convenience of description, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0018] First of all, it needs to be pointed out that drug-drug interaction (DDI) is an important problem in clinical treatment, which can lead to reduced or enhanced drug efficacy, or even cause serious adverse reactions. DDI is a complex multi-factor problem involving drug pharmacological properties, chemical structure, dosage and time of administration, and many other factors. In actual clinical application, about 30%-50% of serious adverse events are caused by DDI. Although there have been some research results in the field of DDI, the existing methods have obvious deficiencies in dealing with asymmetric DDI.

[0019] Asymmetric DDI refers to the use of two drugs in combination, one drug has a significant impact on the pharmacological effect of the other drug, while the other drug has little or no impact. This asymmetry is very common in clinical practice. For example, the interaction between antifungal drugs and non-antifungal drugs usually exhibits unidirectional pharmacological antagonism. In addition, the order of administration also affects the asymmetry of DDI. For example, taking vincristine first and then taking cyclophosphamide can enhance the anti-tumor activity, while taking them at the same time has no additive effect. This asymmetry makes DDI prediction more complex.

[0020] Based on this, the embodiment of the present application aims to provide a new asymmetric drug-drug interaction prediction model based on heterogeneous graph neural network (Drug Heterogeneous Graph Neural Network for Asymmetric Drug-Drug Interaction Prediction, DHGNN-DDI). The model focuses on the problem that the existing methods in the field of drug-drug interaction (DDI) prediction do not consider the asymmetry, innovatively constructs a drug-DDI heterogeneous graph network, integrates drug molecular fingerprint features and Bert embedding features of DDI text description, fully utilizes the multi-modal information fusion capability of heterogeneous graph neural network, and accurately captures the complex heterogeneous relationship between drugs and DDI types. Through a bidirectional graph attention network, the feature representation of the drug as the source and target of the interaction is extracted, and a direction perception mechanism is introduced to dynamically adjust the drug embedding weight, thereby finely modeling the sensitivity difference of the drug in different interaction directions, aiming to significantly improve the accuracy and reliability of asymmetric DDI prediction, and providing more accurate decision support for drug research and development and clinical rational drug use.

[0021] Reference Figure 1 The present application provides a graph network-based asymmetric drug-drug interaction prediction method, which comprises the following steps: S100, obtain drug data based on a comprehensive drug database and perform data preprocessing, construct drug-DDI graph data and drug-drug graph data; S110, obtain drug data based on a comprehensive drug database and perform data screening to obtain screened drug data; In this embodiment, the data is derived from the DrugBank database, which is a comprehensive drug database developed and maintained by the Computational Systems Biology Group at the University of Alberta, Canada. It is one of the core resources in the field of pharmacy. It combines chemical, pharmacological, genomic and clinical data, covering a variety of drug information such as small molecule drugs, biological drugs, approved drugs, experimental drugs and health products. At the same time, it also provides target information such as protein, gene and pathway of drug action, as well as drug-drug interaction (DDI), drug-food interaction, drug-disease correlation and other interaction information.

[0022] DrugBank has become an important resource in the fields of drug discovery, pharmacological research and clinical drug reference due to its comprehensive, accurate and easy-to-use data, and is widely used in various aspects of drug research and development, such as drug target discovery, drug design, drug docking or screening, drug metabolism prediction, and drug interaction prediction.

[0023] S120, performing molecular fingerprint feature extraction processing on the screened drug data to obtain drug feature data; In this embodiment, first, the collected drug data was screened, and drugs with incorrect SMILES strings or unable to be represented by Morgan fingerprints were excluded. This screening process ensures that the drugs used for model training have accurate and processable chemical structure information. Finally, we retained 1752 drugs and 415194 asymmetric DDI information between them, which covers 412 different DDI types.

[0024] Drug feature extraction: Each drug is represented as a 1024-bit molecular fingerprint. Molecular fingerprint is a binary vector used to describe molecular structure, which can effectively capture the chemical properties of drugs. Through this feature extraction method, the chemical structure information of drugs can be converted into a numerical form that can be processed by the model.

[0025] S130, performing semantic feature extraction processing on the screened drug data based on text-based BERT embedding to obtain DDI features; In this embodiment, as shown in Figure 3 For DDI nodes, text-based BERT embedding is used to extract features (as shown in Figure 1As shown in FIG. 1, we obtain the text description of drug interactions from DrugBank and convert it into a 768-dimensional feature vector using the BERT model. This way can effectively extract semantic information from the text description and capture the relevant features of DDI.

[0026] In this embodiment, as shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and a drug-drug graph data is constructed. In this embodiment, as shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and a drug-drug graph data is constructed. Figure 4 As shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and the direction of the edge represents the direction of the DDI. The feature of the drug node is its 1024-bit molecular fingerprint.

[0027] In this embodiment, as shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and a drug-drug graph data is constructed.

[0028] In this embodiment, as shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and a drug-drug graph data is constructed. Figure 5 As shown in FIG. 2, in this network, the circular node represents the drug, and the triangular node represents the DDI type. The feature of the drug node is also the 1024-bit molecular fingerprint, and the feature of the DDI node is the 768-dimensional BERT embedding feature. The direction of the edge represents the role of the drug in the DDI, the incoming edge represents the sensitivity weight of the drug as the victim to the DDI, and the outgoing edge represents the sensitivity weight of the drug as the perpetrator to the DDI.

[0029] In this embodiment, as shown in FIG. 2, each drug is regarded as a node in the graph, and if there is a DDI between two drugs, a directed edge is established between the corresponding nodes, and a drug-drug graph data is constructed. Specifically, the heterogeneous graph neural network asymmetric DDI prediction model specifically comprises an encoder and a decoder, an output end of the encoder is connected with an input end of the decoder, wherein the encoder comprises a first heterogeneous graph neural network, a second heterogeneous graph neural network, a first graph attention network, a second graph attention network, a first multilayer perceptron, a second multilayer perceptron, a first gating layer and a second gating layer; the decoder comprises a first linear layer, a second linear layer, a third linear layer and a fourth linear layer; a first output end of the first heterogeneous graph neural network is connected with a first input end of the first gating layer, a second output end of the first heterogeneous graph neural network is connected with an input end of the first linear layer, an output end of the first graph attention network is connected with a second input end of the first gating layer, an output end of the first gating layer is connected with an input end of the second linear layer, an output end of the second graph attention network is connected with a first input end of the second gating layer, an output end of the second gating layer is connected with an input end of the third linear layer, a first output end of the second heterogeneous graph neural network is connected with a second input end of the second gating layer, and a second output end of the second heterogeneous graph neural network is connected with an input end of the fourth linear layer.

[0030] As shown in Figure 6 The model mainly consists of an encoder (light purple dashed box) and a decoder (red dashed box). The model explicitly captures the heterogeneous relationships between drugs and DDI types by constructing a Drug-DDI heterogeneous graph network, and uses the multi-modal information fusion ability of the graph neural network to model the sensitivity differences of drugs in different interaction directions.

[0031] The input of the model mainly has two parts. One part is the Drug-DDI graph, which explicitly represents the interaction relationship between drugs and DDI types. The features of drug nodes are based on feature vectors of molecular fingerprints, and the features of DDI nodes are based on feature vectors embedded by BERT. These features are used to capture the complex interaction relationship between drugs and DDI types. The other part is the Drug-drug graph, which represents the direct relationship between drugs. The features of drug nodes are also based on feature vectors of molecular fingerprints, which are used to capture the direct interaction relationship between drugs.

[0032] The binary cross-entropy loss function of the entire model framework training is as follows: In the above formula, represents the binary cross-entropy loss function of the heterogeneous graph neural network asymmetric DDI prediction model, represents the predicted value of drug interaction probability, represents the total number of drug interactions, Real labels representing drug interactions Representing drugs With edges between drugs .

[0033] S300, based on the heterogeneous graph neural network asymmetric DDI prediction model, the drug-DDI graph data and the drug-drug graph data are subjected to drug interaction probability prediction, and the asymmetric drug-drug interaction prediction result is obtained.

[0034] S310, inputting the drug-DDI graph data and the drug-drug graph data into the heterogeneous graph neural network asymmetric DDI prediction model respectively; S320, based on the encoder of the heterogeneous graph neural network asymmetric DDI prediction model, the drug-DDI graph data and the drug-drug graph data are subjected to feature extraction, and the multi-modal features of the drug nodes are obtained; Specifically, the drug-DDI graph data and the drug-drug graph data are input into the encoder; based on the first graph attention network and the second graph attention network of the encoder, the drug-drug graph data is subjected to directional feature extraction, and the source role features of the drugs and the target role features of the drugs are obtained; based on the first heterogeneous graph neural network and the second heterogeneous graph neural network of the encoder, the drug-DDI graph data is subjected to capture of the heterogeneous relationship between the drugs and the DDI types, and the source role embedding features of the drugs and the target role embedding features of the drugs are obtained; based on the first gate layer and the second gate layer of the encoder, the source role features of the drugs and the source role embedding features of the drugs and the target role features of the drugs and the target role embedding features of the drugs are respectively subjected to weighted splicing and fusion, and the final features of the source roles of the drugs and the final features of the target roles of the drugs are obtained; based on the first multilayer perceptron and the second multilayer perceptron of the encoder, the drug-drug graph data is subjected to acquisition of the chemical structure information of the drugs, and the source role coding features of the drugs and the target role coding features of the drugs are obtained; in combination with the final features of the source roles of the drugs, the final features of the target roles of the drugs, the source role coding features of the drugs, the target role coding features of the drugs, the source role embedding features of the drugs and the target role embedding features of the drugs, the multi-modal features of the drug nodes are obtained.

[0035] In this embodiment, the encoder module is responsible for extracting the multi-modal feature representation of the drug nodes, mainly including the graph attention network (GAT Layer), the heterogeneous graph neural network (HGNN Layer), the gate mechanism (Gate Layer) and the multilayer perceptron (MLP).

[0036] In the asymmetric interaction relationship, the two drugs are not equal, because one drug will affect the action of the other drug, so we use two different direction GAT (graph attention network) layers to extract the feature representation of the drug as the source of action and the target of action With . In the directed interaction graph , two kinds of neighborhoods of a node are defined according to its source role and target role, respectively. One is the neighborhood , which represents the first-order out-neighbors of a node . The other is , which represents the first-order in-neighbors of a node . Wherein, is the drug node, is the set of directed edges.

[0037] Then, we aggregate the information from the neighboring nodes of using GAT through attention mechanism. The attention score of with respect to the role is defined as follows: Wherein, is the concatenation operation, is the neighbor of node , for in , is the source role, is the target role, is the weight matrix, is a single-layer feedforward neural network, is a nonlinear activation function. The attention score aggregates different neighbors of , which are in the same role , discarding the self-loop of the center node when using attention aggregation, because its source role and target role should be distinguishable, and the center node cannot be both its own out-neighbor and in-neighbor in the directed graph. Finally, the embedding representation of the source role and the target role

[0038] is defined in a similar way as and , as follows: For DDI type information in different directions, we use two heterogeneous graph neural networks (HGNN Layer) to capture the heterogeneous relationship between drugs and DDI types, generating type embeddings of drugs in the drug-DDI heterogeneous graph , where ​Two types of nodes are included, drug nodes and DDI nodes, is a weighted edge, representing the impact of the DDI.

[0039] Then we use GAT to aggregate the information of adjacent nodes through attention mechanism, where the attention coefficient is defined as follows: where, is a concatenation operation, is a learnable attention vector, is the neighbor of the node , is a learnable linear transformation matrix for mapping the features of nodes and edges to a unified dimension, is a nonlinear activation function, is the feature vector of the edge .

[0040] The final type embedding representation of the drug is as follows: where, is the set of DDI types that the drug participates in, is the drug , the attention coefficient of the DDI node is the linear transformation matrix of the DDI node.

[0041] In order to make the drug's feature representation as a source or target role more informative, we use a gating mechanism to fuse the drug's source or target role and type features through a gating mechanism or weighted concatenation to obtain the final feature representation of the drug , The specific fusion method is as follows: where, is the drug's feature representation as a source or target role, and is a learnable gating vector, is the type embedding of the drug in different directions, is the activation function.

[0042] In the encoder, in order to obtain the chemical structure information of the drug, we use a multi-layer perceptron to obtain the self-role encoding of the drug , as shown below: where, is an activation function, and is a weight matrix, and is a bias vector.

[0043] S330, based on the decoder of the heterogeneous graph neural network asymmetric DDI prediction model, a DDI type specific weight generation mechanism is introduced, the multi-modal features of the drug nodes are subjected to drug interaction probability prediction, and drug interaction probability prediction values are obtained; Specifically, the multi-modal features of the drug nodes are input into the decoder; based on the second linear layer and the third linear layer of the decoder, the drug source role final feature and the drug target role final feature are subjected to feature space alignment, and the drug source role final feature after dimension unification and the drug target role final feature after dimension unification are obtained; based on the first linear layer and the fourth linear layer of the decoder, the source role embedding feature of the drug and the target role embedding feature of the drug are mapped to type specific weights through a DDI type perception mechanism, and the source role attention weight of the drug node and the target role attention weight of the drug node are obtained; the drug source role final feature after dimension unification and the source role encoding feature of the drug are subjected to inner product operation, and then subjected to inner product operation with the source role attention weight of the drug node, to obtain the source role probability prediction value of the drug node; the drug target role final feature after dimension unification and the target role encoding feature of the drug are subjected to inner product operation, and then subjected to inner product operation with the target role attention weight of the drug node, to obtain the target role probability prediction value of the drug node; the source role probability prediction value of the drug node and the target role probability prediction value of the drug node are added, to obtain the drug interaction probability prediction value.

[0044] In this embodiment, the decoder module generates drug node embedding based on the encoder, adopts a type perception role alignment decoder, introduces a DDI type specific weight generation mechanism, and captures the role features of drugs in different interaction types through directed graph structure encoding, to realize directional modeling of the interaction relationship in the graph structure.

[0045] The decoder first performs feature space alignment on the source role feature and the target role feature through a bidirectional linear transformation layer. The two linear transformation layers realize dimension unification across the role features, and the specific implementation is as follows: wherein, and are learnable weight matrices, and are the source role feature and the target role feature, respectively. is a bias term, with is the feature representation after unifying dimensions.

[0046] Then through the DDI type-aware mechanism, the node type is embedded into the type-specific weight by using a nonlinear transformation, so that the drug node can dynamically adjust the feature contribution degree of the source role and the target role according to the interaction type, and the calculation expression is as follows: wherein, and respectively correspond to the weights of the drug in the source role and the drug in the target role, and are learnable weight matrices, and are the self-role encodings of the drugs, and are bias terms.

[0047] The final interaction score is composed of a bidirectional type alignment term, and the interaction probability is generated by the Sigmoid function. The expression is as follows: wherein, and are the source role features of the drug and the target role features of the drug , and are weight matrices, and are the self-role features of the drugs, and are bias terms, and respectively correspond to the weights of the drug in the source role and the drug in the target role.

[0048] For a drug pair , if , it is determined that the drug pair will interact; otherwise, it is considered that the drug pair will not interact.

[0049] S340, determining the asymmetric drug-drug interaction prediction result according to the drug interaction probability prediction value.

[0050] In summary, the embodiment of the present application proposes an asymmetric DDI prediction model (DHGNN-DDI) based on a heterogeneous graph neural network, aiming to solve the defects of insufficient processing of asymmetric DDI problems in the prior art. In view of the deficiency that the existing method regards the two drugs in interaction as equal roles and ignores the pharmacological asymmetry, the present model collects data from databases such as DrugBank and constructs a drug-drug graph network and a drug-DDI heterogeneous graph network. In the encoder, a bidirectional graph attention network is used to extract the feature representation of the drug as the source and target of interaction, fully considering the differences of the drug in different interaction directions. At the same time, in order to overcome the problem that the existing method cannot effectively distinguish the type tendency difference in the interaction direction, the present model uses a heterogeneous graph neural network to capture the heterogeneous relationship between the drug and the DDI type, and generates the type embedding of the drug. Further, the role features and type features of the drug are fused through a gating mechanism, and a multi-layer perception machine is used to obtain the self-role encoding of the drug. In the decoder part, based on the embedding generated by the encoder, a type-aware role alignment decoder is used, which generates interaction probability after linear transformation and type-aware mechanism adjustment of feature contribution, so as to realize accurate prediction of asymmetric DDI and significantly improve the accuracy and reliability of prediction.

[0051] Finally, the embodiment of the present application improves the performance of asymmetric DDI prediction by constructing a drug-DDI heterogeneous graph network and using a graph attention mechanism to extract the features of the drug in different interaction directions. Experimental results show that the proposed DHGNN-DDI model performs excellently in the prediction task. The model achieves an AUROC (Area Under the Receiver Operating Characteristic Curve) of 99.0%, an AUPRC (Area Under the Precision-Recall Curve) of 99.0%, an accuracy of 94.2%, an F1 score of 94.1%, a precision of 95.6%, and a recall of 92.8%. Compared with other asymmetric DDI prediction models, the AUROC is improved by at least 3.9%, the AUPRC is improved by at least 5.7%, the ACC is improved by at least 5.6%, the F1 score is improved by at least 5.7%, and the precision is improved by at least 8.7%. This fully proves the advantage of the DHGNN-DDI model in asymmetric DDI prediction.

[0052] In the model construction process, a direction perception mechanism is introduced to dynamically adjust the drug embedding weight, which enables the DHGNN-DDI to accurately capture the sensitivity differences of drugs in different interaction directions. Through symmetry analysis, it is found that compared with the model without using the direction perception mechanism, the prediction accuracy of DHGNN-DDI is improved by 8.5%-10.2%, and the recall rate is improved by 5.7%-7.8%, which fully proves the importance of the direction perception mechanism for improving the asymmetric DDI prediction performance.

[0053] In the decoder, a DDI type perception mechanism is introduced to dynamically adjust the feature contribution of drug nodes according to the interaction type. This improvement not only improves the understanding ability of the model for complex DDI relationships, but also provides more accurate basis for the prediction results. Specifically, this strategy improves the AUROC index by 0.5%-0.8% and the AUPRC by 0.7%-1.2%, further proving the important role of considering interaction type features in improving prediction accuracy.

[0054] In addition, in the unknown DDI prediction experiment, the embodiment of the application adopts a cross-version dataset strategy, uses the old version data of the DrugBank database to train the model, and uses the new version data to construct the test set. The model ranks the unlabeled drug pairs, and 7 of the top 10 asymmetric DDIs are successfully predicted, 2 groups of drug pairs are not correctly predicted in the interaction direction, and 1 group of drug pairs is not found or incorrectly predicted. This reflects the high accuracy of the model in unknown DDI prediction, and verifies its generalization ability and time adaptability.

[0055] Therefore, the embodiment of the application significantly improves the accuracy, reliability and generalization ability of DDI prediction by innovative heterogeneous graph network construction and graph attention mechanism, combined with direction perception and DDI type perception mechanism, which provides strong support for drug research and development and clinical drug safety.

[0056] Reference Figure 2 , the asymmetric drug-drug interaction prediction system based on graph network comprises: The first module 201 is used for acquiring drug data based on a comprehensive drug database and performing data preprocessing to construct drug-DDI graph data and drug-drug graph data. The second module 202 is used for introducing a graph attention network and a heterogeneous graph neural network to construct a heterogeneous graph neural network asymmetric DDI prediction model. The third module 203 is used for predicting the drug interaction probability of the drug-DDI graph data and the drug-drug graph data based on the heterogeneous graph neural network asymmetric DDI prediction model, and predicting the asymmetric drug-drug interaction.

[0057] The contents in the method embodiments are applicable to the system embodiments, the system embodiments specifically implement the same functions as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0058] The above is a specific description of the preferred embodiments of the application, but the application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the application. These equivalent modifications or replacements are all included in the scope defined by the claims of the application.

Claims

1. A graph network-based method for predicting asymmetric drug-drug interactions, characterized in that, Includes the following steps: Drug data is acquired from a comprehensive drug database and preprocessed to construct drug-DDI graph data and drug-drug graph data; By introducing graph attention networks and heterogeneous graph neural networks, an asymmetric DDI prediction model of heterogeneous graph neural networks is constructed. Based on the heterogeneous graph neural network asymmetric DDI prediction model, drug interaction probability prediction is performed on drug-DDI graph data and drug-drug graph data, and the asymmetric drug-drug interaction prediction results are obtained.

2. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 1, characterized in that, The step of acquiring drug data based on a comprehensive drug database, performing data preprocessing, and constructing drug-DDI graph data and drug-drug graph data specifically includes: Drug data is obtained from a comprehensive drug database and then filtered to obtain the filtered drug data. Molecular fingerprint feature extraction was performed on the screened drug data to obtain drug feature data; Text-based BERT embedding is used to extract semantic features from the screened drug data to obtain DDI features; Treating drug feature data as a node in a graph, if there is a DDI between two drugs, a directed edge is established between the corresponding nodes to construct drug-drug graph data. Both drug feature data and DDI features are treated as graph nodes, and directed edges are established between drug feature data and DDI features to construct drug-DDI graph data.

3. The asymmetric drug-drug interaction prediction method based on graph networks according to claim 2, characterized in that, The heterogeneous graph neural network asymmetric DDI prediction model specifically includes an encoder and a decoder, wherein the output of the encoder is connected to the input of the decoder, and: The encoder includes a first heterogeneous graph neural network, a second heterogeneous graph neural network, a first graph attention network, a second graph attention network, a first multilayer perceptron, a second multilayer perceptron, a first gating layer, and a second gating layer; The decoder includes a first linear layer, a second linear layer, a third linear layer, and a fourth linear layer; The first output of the first heterogeneous graph neural network is connected to the first input of the first gating layer, the second output of the first heterogeneous graph neural network is connected to the input of the first linear layer, the output of the first graph attention network is connected to the second input of the first gating layer, the output of the first gating layer is connected to the input of the second linear layer, the output of the second graph attention network is connected to the first input of the second gating layer, the output of the second gating layer is connected to the input of the third linear layer, the first output of the second heterogeneous graph neural network is connected to the second input of the second gating layer, and the second output of the second heterogeneous graph neural network is connected to the input of the fourth linear layer.

4. The asymmetric drug-drug interaction prediction method based on graph networks according to claim 3, characterized in that, The specific expression for the binary cross-entropy loss function of the heterogeneous graph neural network asymmetric DDI prediction model is as follows: In the above formula, The binary cross-entropy loss function represents the asymmetric DDI prediction model of a heterogeneous graph neural network. This represents the predicted probability of drug interaction. Indicates the total number of drug interactions. True labeling of drug interactions Indicates drug With drugs The edges between them.

5. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 4, characterized in that, The step of predicting drug interaction probabilities and asymmetric drug-drug interaction prediction results based on the heterogeneous graph neural network asymmetric DDI prediction model specifically includes: The drug-DDI graph data and the drug-drug graph data are respectively input into the heterogeneous graph neural network asymmetric DDI prediction model; An encoder based on a heterogeneous graph neural network asymmetric DDI prediction model extracts features from drug-DDI graph data and drug-drug graph data to obtain multimodal features of drug nodes. The decoder based on the heterogeneous graph neural network asymmetric DDI prediction model introduces a DDI type-specific weight generation mechanism to predict the drug interaction probability of the multimodal features of the drug node and obtain the predicted drug interaction probability value. The prediction results of asymmetric drug-drug interactions are determined based on the predicted drug interaction probability values.

6. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 5, characterized in that, The encoder based on the heterogeneous graph neural network asymmetric DDI prediction model extracts features from drug-DDI graph data and drug-drug graph data to obtain multimodal features of drug nodes. This step specifically includes: Input the drug-DDI map data and drug-drug map data into the encoder; Based on the first and second graph attention networks of the encoder, direction-aware feature extraction is performed on the drug-drug graph data to obtain drug source role features and drug target role features. Based on the first heterogeneous graph neural network and the second heterogeneous graph neural network of the encoder, the heterogeneous relationship between drugs and DDI types is captured in the drug-DDI graph data, and the source role embedding feature and the target role embedding feature of the drug are obtained. Based on the first and second gating layers of the encoder, the drug source role features and the drug source role embedding features, as well as the drug target role features and the drug target role embedding features, are weighted and spliced ​​together to obtain the final features of the drug source role and the final features of the drug target role. Based on the first and second multilayer perceptrons of the encoder, the chemical structure information of the drug is obtained from the drug-drug graph data, and the source role coding features and target role coding features of the drug are obtained. By combining the final features of the drug source role, the final features of the drug target role, the source role encoding features of the drug, the target role encoding features of the drug, the source role embedding features of the drug, and the target role embedding features of the drug, the multimodal features of the drug node are obtained.

7. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 6, characterized in that, The step of extracting direction-aware features from drug-drug graph data using the first and second graph attention networks based on the encoder to obtain drug source role features and drug target role features specifically includes: The drug-drug graph data is input into the first graph attention network and the second graph attention network, respectively. Based on the drug-drug graph data, define the first-order out neighbors and the first-order in neighbors of the drug node to obtain information about the drug's neighboring nodes. The attention score of a drug node is obtained by aggregating information from drug neighboring nodes through an attention mechanism. Based on the attention score of the drug node, define the characteristics of the drug source role and the drug target role.

8. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 7, characterized in that, The decoder based on the heterogeneous graph neural network asymmetric DDI prediction model introduces a DDI type-specific weight generation mechanism to predict drug interaction probabilities based on the multimodal features of drug nodes, and obtains the predicted drug interaction probability values. This step specifically includes: The multimodal features of the drug node are input into the decoder; Based on the second and third linear layers of the decoder, the final features of the drug source role and the final features of the drug target role are aligned in the feature space to obtain the final features of the drug source role and the final features of the drug target role with unified dimensions. Based on the first and fourth linear layers of the decoder, the source role embedding features and target role embedding features of the drug are mapped to type-specific weights through the DDI type-aware mechanism, thus obtaining the source role attention weight and target role attention weight of the drug node. The final feature of the drug source role after unifying the dimensions is inner producted with the source role encoding feature of the drug, and then inner producted with the source role attention weight of the drug node to obtain the source role probability prediction value of the drug node. The final feature of the drug target role after unification of dimensions is inner producted with the target role encoding feature of the drug, and then inner producted with the target role attention weight of the drug node to obtain the probability prediction value of the target role of the drug node. The predicted probability of the source role of the drug node is added to the predicted probability of the target role of the drug node to obtain the predicted probability of drug interaction.

9. The method for predicting asymmetric drug-drug interactions based on graph networks according to claim 8, characterized in that, The expression for the predicted drug interaction probability is shown below: In the above formula, This represents the predicted probability of drug interaction. , They represent drugs The source role characteristics and drugs Target role characteristics , Represents the weight matrix. , This indicates the self-role characteristics of a drug. , Indicates the bias term. , They represent the corresponding drugs. Weights under the source role and drugs Weight under the target role.

10. A graph network-based asymmetric drug-drug interaction prediction system, characterized in that, Includes the following modules: The first module is used to acquire drug data based on a comprehensive drug database and perform data preprocessing to construct drug-DDI graph data and drug-drug graph data; The second module is used to introduce graph attention networks and heterogeneous graph neural networks to construct a heterogeneous graph neural network asymmetric DDI prediction model. The third module is used to predict the probability of drug interactions and the asymmetric drug-drug interaction prediction results based on the heterogeneous graph neural network asymmetric DDI prediction model.

Images