DDI prediction method based on siamese structure and graph contrastive learning

Abstract

The application discloses a DDI prediction method based on a Siamese structure and graph contrast learning, comprising the following steps: collecting drug-drug interaction text data and physical and chemical property data files and targeting relationship data files of drugs; extracting the physical and chemical property characteristics and the targeting relationship characteristics of the drugs, fusing the characteristics to obtain initial characteristics based on the physical and chemical properties and the targeting relationship; calculating a drug-drug interaction adjacency matrix, combining the initial characteristics to construct a drug-drug interaction heterogeneous graph; inputting the heterogeneous graph into a Siamese structure-based graph contrast learning model to learn and obtain embedding characteristics of drug nodes; and using a link prediction method to calculate the score of edges between any two drug nodes. The application can alleviate the problem that the drug targeting relationship characteristics and the drug-drug interaction text data are difficult to be used alone when the data are sparse and have an impact on the performance of the model, improves the accuracy of drug-drug interaction prediction, and can be applied to identifying potential interactions between drugs.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a DDI prediction method based on Siamese structure and graph contrastive learning. Background Technology

[0002] The use of two or more drugs with synergistic effects to achieve therapeutic goals is called combination therapy. Combination therapy not only improves drug efficacy but also reduces toxicity and side effects. Therefore, predicting potential drug-drug interactions (DDIs) is crucial.

[0003] Early drug-drug interaction (DDI) prediction problems primarily focused on binary classification to predict the existence of synergistic / antagonistic effects between drugs. However, such binary classification results are insufficient to provide clear guidance for drug combination therapy in the real world. Therefore, the DDI prediction problem has gradually shifted to multi-label classification to predict the specific side effects that drugs may produce. Typical methods for predicting drug-drug interactions include DDIMDL, MDF-SA-DDI, SSI-DDI, GMPNN, 3DGT, MTDDI, Mircle, and AttentionDDI. The DDIMDL method uses drug physicochemical properties such as chemical substructure, targets, enzymes, and pathways as drug features to establish a prediction model based on a multilayer perceptron. The MDF-SA-DDI method uses drug physicochemical properties such as chemical substructure, targets, and enzymes as features, performing multi-source drug fusion based on Siamese networks, convolutional neural networks, and autoencoders, as well as multi-source feature fusion based on the Transformer's self-attention mechanism. It uses fully connected layers as classifiers to predict DDI. SSI-DDI and GMPNN use the chemical substructure of drugs divided in different ways as features. DDI prediction models based on GAT, MPNN, and Co-Attention layers were established. 3DGT uses drug substructures and DDI text relationships as features, embedding these features using CNN and BERT, and then builds a prediction model based on deep neural networks. MTDDI uses multi-relation DDI networks as features, establishing a specific DDI prediction model based on R-GCN and a tensor decoder. Mircle uses DDI networks as features, establishing a DDI prediction model based on GCN and key-aware attention message propagation. AttentionDDI uses the drug similarity matrix as features, establishing a DDI prediction model based on Transformer and Siamese encoders.

[0004] However, the vast majority of DDI relationships in the real world are unknown, especially for new drugs, where known DDI relationships are sparse. Methods limited by the size of the DDI text data struggle to achieve optimal performance in the face of this problem. Therefore, it is crucial to enable models to learn potential DDI relationship information when dealing with sparse DDI relationships.

[0005] Therefore, how to provide a drug-drug interaction prediction method that can learn potential DDI relationship information when faced with sparse DDI relationships has become an urgent problem for practitioners. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the existing technology by providing a DDI prediction method based on Siamese structure and graph comparison learning.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] This invention provides a DDI prediction method based on Siamese structure and graph contrast learning, comprising the following steps:

[0009] S1: Collect drug-drug interaction text data, drug physicochemical property data files, and target relationship data files; filter and preprocess the collected data to obtain the raw set of drug-drug interaction data to be processed.

[0010] S2: Extract the physicochemical properties based on the drug SMILES molecular format string and the targeting relationship features based on the target, enzyme, and pathway from the original drug-drug interaction data set to be processed. Then, fuse the physicochemical properties and targeting relationship features to obtain the initial features based on the drug's physicochemical properties.

[0011] S3: Obtain a drug-drug interaction adjacency matrix based on drug-drug interaction text data. Combine the drug-drug interaction adjacency matrix with the initial features obtained in S2 based on the physicochemical properties and targeting relationship of drugs to construct a drug-drug interaction heterogeneity graph.

[0012] S4: Input the drug-drug interaction heterogeneity graph into a graph contrastive learning model based on Siamese structure, and learn the embedding features of drug nodes from the graph topology;

[0013] S5: Using the embedding features of the drug nodes as input, calculate the score of the edge between any two drug nodes using the link prediction method.

[0014] Preferably, in S1:

[0015] Drug-drug interaction text data includes: drug-drug interaction text;

[0016] The drug-target interaction data files include: drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data;

[0017] The physicochemical property data file for a drug includes: the SMILES molecular format string of the drug.

[0018] Preferably, S2 includes:

[0019] S21: Extracting physicochemical property features includes: converting the SMILES molecular format string of the drug into a drug molecular fingerprint format according to its physicochemical properties, and further encoding it into a binary bit vector, and calculating a drug-drug similarity matrix based on the bit vector;

[0020] S22: Extracting target relationship features includes: converting drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data into binary bit vectors according to drug-target, drug-enzyme, and drug-pathway interaction relationships, and calculating drug-target similarity matrix, drug-enzyme similarity matrix, and drug-pathway similarity matrix based on the bit vectors respectively;

[0021] S23: The four similarity feature matrices in S21 and S22 are fused to obtain a multi-source drug similarity matrix. Each row in the matrix is ​​the two-step drug feature of the current drug after multi-source feature fusion.

[0022] Preferably, the steps for calculating the similarity matrix in S21 and S22 include:

[0023] The similarity matrix is ​​calculated using the following formula:

[0024]

[0025] Wherein, N(D) A D B ) 11 It is a bit vector D A and D B The number of attributes in a given class that are all equal to 1, N(D) A D B ) 01 It is a bit vector D A and D B D A The attribute value is 0 and D B The number of attributes with a value of 1, N(D) A D B ) 10 It is a bit vector D A and DB D A The attribute value is 1 and D B The number of attributes with a value of 0;

[0026] For calculating the drug-drug similarity matrix: D A D B The molecular fingerprint bit vectors representing drug A and drug B;

[0027] For calculating the drug-target similarity matrix: D A D B One-Hot vector representing the drug-target;

[0028] For calculating the drug-enzyme similarity matrix: D A D B One-Hot vector representing drug-enzyme;

[0029] For calculating the drug-pathway similarity matrix: D A D B One-Hot vector representing drug-pathway.

[0030] Preferably, the fusion step S23 includes:

[0031] S231: Combine the drug-drug similarity matrix, drug-target similarity matrix, drug-enzyme similarity matrix and drug-pathway similarity matrix in a tandem manner to obtain a tandem drug similarity matrix, where each row in the matrix is ​​a one-step drug feature of the current drug based on its physicochemical properties;

[0032] S232: Calculate the cascaded drug similarity matrix using the Gaussian kernel function to obtain a multi-source drug similarity matrix based on the Gaussian kernel function, where each row in the matrix is ​​the two-step drug feature of the current drug after multi-source feature fusion.

[0033] Preferably, the step of obtaining the drug-drug interaction adjacency matrix based on drug-drug interaction text data in step S3 includes:

[0034] The drug-drug interaction text data includes descriptive text describing the DDI event between each two drugs. By replacing keywords in the descriptive text with fixed words, the drug-drug interaction relationship is obtained. The keywords include drug names.

[0035] Construct a drug-drug interaction adjacency matrix based on drug-drug interaction relationships.

[0036] Preferably, the drug-drug interaction isomer diagram in S3 includes:

[0037] The nodes of the drug-drug interaction isomer diagram are drugs;

[0038] The position of 0 in the adjacency matrix determines that there is no edge between two nodes, and the position of 1 in the adjacency matrix determines that there is an edge between two nodes.

[0039] The node attributes are determined based on the two-step drug features obtained after multi-source feature fusion.

[0040] Preferably, the graph contrast learning model in S4 is a graph isomorphic network, where the nodes are drug nodes and the node features are the attributes of the nodes;

[0041] The steps for learning the embedding features of drug nodes include:

[0042] S41: Extract two isomorphic subgraphs from a heterogeneous graph structure containing multiple DDI relationships, which have the same side type and different node attributes;

[0043] S42: Input the two isomorphic subgraphs as input objects to the graph isomorphic network based on the Siam structure into the comparison model of the graph isomorphic network based on the Siam structure;

[0044] S43: Update and iterate the node features in the graph isomorphic network, wherein in each layer of the graph isomorphic network, the node features of the next layer of each drug node are aggregated from the node features of the current layer of the drug node and the node features of other drug nodes in the current layer.

[0045] S44: After the iteration is completed, the compressed drug embedding vector is obtained.

[0046] Preferably, step S5 specifically includes:

[0047] The link prediction model based on GNN calculates the score of links between drug nodes, and the calculation formula is as follows:

[0048]

[0049] in, and These are the features of drug nodes u and v at layer L, calculated from a multi-layer graph isomorphic network. It is a pre-defined prediction operation.

[0050] Preferably, step S5 further includes performance evaluation of the graph contrast learning model, specifically including the following steps:

[0051] The area under the receiver operating characteristic curve (ROC-AUC), accuracy (Acc), F1 score, precision, recall, and area under the PR curve (PR-AUC) were used as evaluation metrics.

[0052] The specific calculation formula is as follows:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] Where TP is a DDI that is predicted to be true by the model, TN is a DDI that is predicted to be false by the model, FP is a DDI that is predicted to be true by the model, and FN is a DDI that is predicted to be false by the model.

[0059] ROC_AUC is obtained by using FPR and TPR as the horizontal and vertical axes respectively to obtain the ROC curve, and then calculating the AUC under the ROC curve.

[0060] PR-AUC is obtained by calculating the AUC under the PR curve with Recall and Precision as the horizontal and vertical axes, respectively.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] (1) This invention utilizes a Gaussian kernel function to fuse the physicochemical characteristics and targeting relationship characteristics of a drug, which can alleviate the problem that the drug targeting relationship characteristics are difficult to use alone when they are sparse. This feature fusion method can provide empirical support and methodological reference for related research.

[0063] (2) This invention utilizes a learning model based on Siamese structure and graph comparison to alleviate the problem of the impact on model performance when drug-drug interaction text data is sparse, thereby improving the accuracy of drug-drug interaction prediction.

[0064] (3) The model of the present invention has good generalization performance and can be applied to identify potential interactions between drugs, providing good lead support for drug combination therapy. Attached Figure Description

[0065] Figure 1 A flowchart of a drug-drug interaction prediction method based on Siamese structure and graph contrast learning provided in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of drug feature extraction, multi-source drug feature fusion, and DDI text adjacency matrix extraction provided in this embodiment of the invention.

[0067] Figure 3 This is a schematic diagram illustrating the learning of drug node embedding features using a graph contrastive learning model based on the Siamese structure, provided in an embodiment of the present invention.

[0068] Figure 4 This is a schematic diagram of the PR curve comparing the method provided in this embodiment of the invention with similar methods;

[0069] Figure 5 A schematic diagram of ROC curves comparing the method provided in this embodiment of the invention with similar methods. Detailed Implementation

[0070] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0071] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front end," "rear end," "both ends," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0072] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0073] This invention provides a drug-drug interaction prediction method based on Siamese structure and graph contrastive learning, such as... Figure 1 As shown, it includes the following steps:

[0074] S1: Collect drug-drug interaction text data, drug physicochemical property data files, and target relationship data files; filter and preprocess the collected data to obtain the raw set of drug-drug interaction data to be processed.

[0075] S2: Extract the physicochemical properties based on the drug SMILES molecular format string and the targeting relationship features based on the target, enzyme, and pathway from the original drug-drug interaction data set to be processed. Then, fuse the physicochemical properties and targeting relationship features to obtain the initial features based on the drug's physicochemical properties.

[0076] S3: Obtain the drug-drug interaction adjacency matrix based on the drug-drug interaction text data. Combine the drug-drug interaction adjacency matrix with the initial features based on the physicochemical properties of drugs obtained in S2 to construct a drug-drug interaction heterogeneity graph.

[0077] S4: Input the drug-drug interaction heterogeneity graph into a graph contrastive learning model based on Siam structure, and learn the embedding features of drug nodes from the graph topology;

[0078] S5: Using the embedding features of drug nodes as input, calculate the score of the edge between any two drug nodes using the link prediction method.

[0079] In one embodiment, S1:

[0080] Drug-drug interaction text data includes: drug-drug interaction text;

[0081] The drug-target interaction data files include: drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data;

[0082] The physicochemical property data file for a drug includes: the SMILES molecular format string (Simplified Molecular Linear Input Specification) for the drug.

[0083] In actual execution, the two data files are obtained in the following ways:

[0084] Download drug-drug interaction text data, drug-target interaction data, drug-enzyme interaction data, and drug SMILES molecular format strings from the drug database DRUGBANK (https: / / go.drugbank.com / ).

[0085] Drug-pathway interaction data can be downloaded from KEGGPATHWAY (https: / / www.genome.jp / kegg / ) from KEGG (https: / / www.genome.jp / kegg / ), a comprehensive database that integrates genomic information.

[0086] The data obtained in S1 were integrated to form a raw dataset containing 699 drugs, 162 targets, 808 enzymes, 469 pathways, and 259,730 drug-drug interaction data to be processed. This dataset was then used for data preprocessing in S2 and S3.

[0087] In one embodiment, in S2, for targets, enzymes, and pathways, corresponding 0-1 matrices are obtained by integrating the drug-target, drug-enzyme, and drug-pathway interactions, and drug similarity matrices based on targets, enzymes, and pathways are calculated using these matrices. In S2, for the SMILES string, the SMILES string is abstracted into a One-Hot molecular fingerprint based on its physicochemical properties, and a drug similarity matrix based on the SMILES string is calculated using this molecular fingerprint. To achieve multi-source drug fusion, the above four drug similarity matrices are combined in a concatenated manner to obtain a concatenated drug similarity matrix. A Gaussian kernel function is then used to calculate the multi-source drug similarity matrix, resulting in a Gaussian kernel function-based multi-source drug similarity matrix. S2 specifically includes:

[0088] S21: Extracting physicochemical property features includes: converting the SMILES molecular format string of the drug into a drug molecular fingerprint format according to its physicochemical properties, and further encoding it into a binary bit vector, and calculating the drug-drug similarity matrix based on the bit vector.

[0089] S22: Extracting target relationship features includes: converting drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data into binary bit vectors according to drug-target, drug-enzyme, and drug-pathway interaction relationships, and calculating drug-target similarity matrices, drug-enzyme similarity matrices, and drug-pathway similarity matrices based on the bit vectors respectively.

[0090] S23: The four similarity feature matrices in S21 and S22 are fused to obtain a multi-source drug similarity matrix. Each row in the matrix is ​​the two-step drug feature of the current drug.

[0091] In this embodiment, the steps for calculating the similarity matrix in S21 and S22 include:

[0092] The similarity matrix is ​​calculated using the following formula (1):

[0093]

[0094] Wherein, N(D) A D B ) 11 It is a bit vector D A and D B The number of attributes in a given class that are all equal to 1, N(D) AD B ) 01 It is a bit vector D A and D B D A The attribute value is 0 and D B The number of attributes with a value of 1, N(D) A D B ) 10 It is a bit vector D A and D B D A The attribute value is 1 and D B The number of attributes with a value of 0;

[0095] For calculating the drug-drug similarity matrix: D A D B The molecular fingerprint bit vectors representing drug A and drug B;

[0096] For calculating the drug-target similarity matrix: D A D B One-Hot vector representing the drug-target;

[0097] For calculating the drug-enzyme similarity matrix: D A D B One-Hot vector representing drug-enzyme;

[0098] For calculating the drug-pathway similarity matrix: D A D B One-Hot vector representing drug-pathway.

[0099] When executing S21-S22, taking the SMILES format as an example, SMILES stands for Simplified Linear Input Specification for Drugs, a specification that explicitly describes the molecular structure using ASCII strings. For example, the SMILES representation of cyclohexane (C6H12) is C1CCCCC1. To convert the string structure into a mathematical form usable for analysis, we use the RDKIT package to convert the SMILES format into a drug molecular fingerprint format. A molecular fingerprint is an abstract representation of a molecule. It encodes the SMILES format into binary bit vectors. To reduce computational resource consumption and based on the assumption that drugs with similar chemical structures may interact, we do not directly use the bit vectors as input, but instead calculate the similarity between drugs based on the bit vectors.

[0100] When processing targets, enzymes, and pathways, drug-target, drug-enzyme, and drug-pathway interactions are integrated to obtain corresponding 699*162, 699*808, and 699*469 dimensional 0-1 matrices. At this point, D... A DB These are the One-Hot vectors for drug-target, drug-enzyme, and drug-pathway, respectively. These vectors are derived based on whether the drug interacts with a certain target, enzyme, or pathway. The method for further calculating the similarity matrix of the target, enzyme, and pathway is shown in formula (1).

[0101] In this embodiment, the fusion step S23 includes:

[0102] S231: Combine the drug-drug similarity matrix, drug-target similarity matrix, drug-enzyme similarity matrix and drug-pathway similarity matrix in a tandem manner to obtain a tandem drug similarity matrix, where each row in the matrix is ​​a one-step drug feature of the current drug based on its physicochemical properties;

[0103] S232: Calculate the cascaded drug similarity matrix using the Gaussian kernel function to obtain a multi-source drug similarity matrix based on the Gaussian kernel function, where each row in the matrix is ​​the two-step drug feature of the current drug after multi-source feature fusion.

[0104] When specifically executing S23, such as Figure 2 As shown, four 699*699-dimensional feature matrices based on similarity are obtained from S21. These four similarity feature matrices are combined in a concatenated manner to obtain a 699*2796-dimensional concatenated similarity feature matrix. Each row in the matrix is ​​considered a one-step drug feature based on the physicochemical properties of a certain drug; that is, each row represents a 1*2796-dimensional concatenated similarity feature vector for a certain drug. To reduce the dimensionality of the concatenated similarity feature matrix while achieving multi-source feature fusion, a Gaussian kernel function is used to calculate the Gaussian kernel similarity for every two drugs in the concatenated similarity feature matrix. The calculation method of the Gaussian kernel similarity is shown in Formula 2.

[0105]

[0106] Where x i x j Let be the 1*2796-dimensional feature vectors for the corresponding drugs i and j. h is a hyperparameter, which is set to 0.5 in this paper. Thus, the 699*2796-dimensional concatenated similarity feature matrix is ​​reduced to a 699*699-dimensional multi-source drug similarity matrix, where each row in the matrix is ​​considered as a two-step drug feature of a certain drug after multi-source feature fusion.

[0107] In one embodiment, such as Figure 2As shown, in S3, keyword substitution is used to obtain a textual pattern independent of specific drugs from the drug-drug interaction text data. The drug-drug interaction text data is then filtered based on this textual pattern to obtain a drug-drug interaction textual relationship adjacency matrix based on a specific textual pattern. The steps to obtain the drug-drug interaction adjacency matrix include:

[0108] Drug-drug interaction text data includes descriptive text describing DDI events between each pair of drugs. Keywords in the descriptive text are replaced with fixed words to obtain drug-drug interaction relationships. Keywords include drug names.

[0109] Construct a drug-drug interaction adjacency matrix based on drug-drug interaction relationships.

[0110] In practice, text substitution is used to concretize the data. For example, the sentence "Fomepizole may decrease the excretion rate of Meloxicam, which could result in a higher serum level" describes a drug-drug interaction (DDI) event between Fomepizole and Meloxicam. Since the descriptions of drug-drug interactions in DrugBank have a fixed syntax, to better organize DDI events, keyword substitution is used, replacing specific drug names with the word 'drug' to extract the fixed grammatical form. For the description "The riskor severity of adverse effects can be increased when Azacitidine is combined with Cladribine," the grammatical form "The risk or severity of adverse effects can be increased when drug is combined with drug" can be extracted. This specific grammatical form represents a drug-drug interaction relationship. Using the above method, 7 types of drug-drug interactions, totaling 74,137 records, were selected from the dataset, and an adjacency matrix containing these 7 types of drug-drug interactions was constructed.

[0111] In one embodiment, the drug-drug interaction isomer diagram in S3 includes:

[0112] The nodes of the drug-drug interaction isomer diagram are drugs;

[0113] The position of 0 in the adjacency matrix determines that there is no edge between two nodes, and the position of 1 in the adjacency matrix determines that there is an edge between two nodes.

[0114] The attributes of nodes are determined based on initial characteristics based on the physicochemical properties of the drug.

[0115] It should be noted that a position of 0 in the adjacency matrix indicates that there is no interaction between the two drugs, while a position of 1 indicates that there is an interaction between the two drugs. This is used to determine whether there are edges between the nodes of the graph structure, and the attributes of each drug node are the corresponding two-step drug features.

[0116] In practice, a drug-drug interaction heterogeneous graph is constructed using the two-step drug features obtained in S22 based on the drug's physicochemical properties and targeting relationship, and the seven drug-drug interaction adjacency matrices in S31. In the heterogeneous graph, nodes represent drugs, and the seven different edge types represent seven different types of drug-drug interactions between drugs. The attributes of the nodes are the 1*699-dimensional two-step drug features obtained in S22, and the existence of specific types of edges between nodes is determined by the seven drug-drug interaction adjacency matrices in S31.

[0117] In one embodiment, the graph contrastive learning model in S4 is a graph isomorphic network, where the nodes are drug nodes and the node features are the attributes of the nodes.

[0118] The steps for learning the embedding features of drug nodes are as follows: Figure 3 As shown, it includes:

[0119] S41: Extract two isomorphic subgraphs from a heterogeneous graph structure containing multiple DDI relationships, which have the same side type and different node attributes.

[0120] S42: Input the two isomorphic subgraphs as input objects to the graph isomorphic network with shared weight Siam structure into the comparison model of graph isomorphic networks based on Siam structure;

[0121] S43: Based on the DDI edge relationships in the graph structure and the two-step drug features in the graph structure, i.e., the drug node feature vector, the drug node feature vector is updated iteratively. The specific iterative formula is shown in formula (3):

[0122]

[0123] Where ∈ is a learnable irrational number parameter, and N(i) is the set of nodes. f is a latent feature of node i in the l-th layer of GIN. Θ The activation function can be selected. Aggregate is an optional aggregation operator, including sum, max, and min.

[0124] S44: After iteration, the compressed drug embedding vector is obtained. The low-dimensional drug embedding vector obtained by formula (3) is a compressed drug feature vector generated by fusing the drug node features in the isomorphic subgraph with information about the graph topology. This feature will be used in step S5 to calculate the drug-drug interaction score and output the model prediction structure.

[0125] It's important to note that this approach leverages the ability of graph isomorphic networks to map similar graph structures to similar embeddings and capture dependencies between them. In a heterogeneous graph of drug-drug interactions, each drug node contains two distinct node attributes, and there are multiple types of edge relationships between drug nodes. By selecting a specific type of edge relationship, two isomorphic subgraphs with different node attributes based on that specific type of edge relationship can be extracted from the heterogeneous graph. To learn the embedding features of drug nodes from the graph topology, a shared-weight Siamese graph isomorphic network is used to extract these features. Since the Siamese structure requires two input objects, and the graph isomorphic network requires the graph structure as input, two isomorphic subgraphs with different node attributes based on a specific type of edge relationship are chosen as the input objects for the shared-weight Siamese graph isomorphic network. In each layer of the graph isomorphic network, the node features of the next layer for each drug node are aggregated from the node features of the current layer and the node features of other drug nodes in the current layer.

[0126] In this embodiment, in step S4, to ensure the model performs well even with sparse DDI text relationships, two isomorphic graphs with different node attributes are used to mitigate information loss issues that may arise when edges are sparse in the graph structure. Furthermore, to intuitively quantify the feature extraction effect of the isomorphic graph network on the two graphs, and to ensure similar encoding for similar data, the model's total loss consists of cross-entropy and normalized temperature-scale cross-entropy loss. The normalized temperature-scale cross-entropy loss functions as a contrastive learning component in the model. The calculation method for the normalized temperature-scale cross-entropy loss is as follows:

[0127]

[0128] The average contrastive learning loss L for all pairs with batch size = N is obtained using the following formula:

[0129]

[0130] Among them, z i , z jTo extract different drug node embedding features of drug z from two isomorphic graphs with different node attributes, based on the Siamese structure graph contrastive learning model, τ is a manually set temperature hyperparameter, and N is the batch size.

[0131] In one embodiment, step S5 specifically includes:

[0132] The link prediction model based on GNN calculates the score of links between drug nodes, and the calculation formula is as follows:

[0133]

[0134] in, and These are the features of drug nodes u and v at layer L, calculated from a multi-layer graph isomorphic network. This refers to the predefined prediction operation. In this embodiment... This is a dot product. In this embodiment, the edge scores are normalized, normalizing the scores of positive and negative samples to the range (0, 1). A higher score indicates a higher probability of an interaction between the two drugs, and vice versa.

[0135] Preferably, S5 further includes performance evaluation of the graph contrastive learning model. After the model is trained, the performance of the drug-drug interaction prediction model based on Siamese structure and graph contrastive learning proposed in this invention is evaluated using multiple evaluation metrics. Specific steps include:

[0136] The area under the receiver operating characteristic curve (ROC-AUC), accuracy (Acc), F1 score, precision, recall, and area under the PR curve (PR-AUC) were used as evaluation metrics.

[0137] The specific calculation formula is as follows:

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] Where TP is a DDI that is predicted to be true by the model, TN is a DDI that is predicted to be false by the model, FP is a DDI that is predicted to be true by the model, and FN is a DDI that is predicted to be false by the model.

[0144] ROC_AUC is obtained by using FPR and TPR as the horizontal and vertical axes respectively to obtain the ROC curve, and then calculating the AUC under the ROC curve.

[0145] PR-AUC is obtained by calculating the AUC under the PR curve with Recall and Precision as the horizontal and vertical axes, respectively.

[0146] Among the aforementioned metrics, a higher ROC_AUC indicates better predictive model performance. F1 score and AUPR are more meaningful in evaluating model performance because, in real-world scenarios, it's impossible to recognize all potential drug-drug interactions. Therefore, drug interactions where the existence of an unknown interaction is often considered negative samples. This results in a small number of negative samples compared to positive samples in DDI prediction problems, leading to frequent false positives. AUPR is more effective at penalizing false positives of drug interactions. F1 represents the harmonic mean of precision and recall, focusing on the proportion of correctly predicted drug-drug interaction pairs. ACC focuses not only on the proportion of correctly predicted drug-drug interaction pairs but also on the proportion of correctly predicted drug-drug non-interaction pairs. For predicting drug interactions, F1 should be more effective than ACC.

[0147] The recognition effect of the present invention will be further described below with two specific experimental examples.

[0148] To evaluate the impact of each component in the model on the prediction performance, this embodiment establishes five ablation experiments to verify the effectiveness of each component in the model.

[0149] Table 1.1 Gin with different number of layers

[0150]

[0151] Table 1.2 Different Graph Neural Networks

[0152]

[0153] As shown in Tables 1.1 and 1.2, compared with single-layer GIN and three-layer GIN variants, the performance of the present invention is superior to both single-layer and three-layer GIN variants, indicating that the two-layer GIN structure may be suitable for the degree of embedding of drug features. Compared with GraphConv and Sageconv, the performance of the present invention is superior to GraphConv and Sageconv variants, indicating that the GIN structure may be more suitable for extracting drug node features from multi-graph SAM structures.

[0154] Table 1.3 Variations of the non-contrast learning component

[0155]

[0156] As shown in Table 1.3, the performance of the variant after removing the contrastive learning component (normalized temperature scale cross-entropy loss) is lower than that of the present invention, which indicates that the normalized temperature scale cross-entropy loss does indeed play a positive role in model iteration and weight update.

[0157] Table 1.4 No Siamese structural variants

[0158]

[0159] As shown in Table 1.4, the performance of the variant of the Siamese structure model is lower than that of the present invention, proving that the Siamese GIN structure with shared weights can better integrate multi-graph features than two independent GINs.

[0160] Table 1.5 Characteristic variants of drugs without multiple origins

[0161]

[0162] As shown in Table 1.5, the model variant using only the SMILES format as drug features performed slightly better than the present invention by 0.0028 in F1 score. However, the present invention significantly outperformed the variant in the Aupr metric. This may be due to the sparseness of the similarity matrix constructed from drug target, enzyme, and pathway data in the dataset.

[0163] To verify the superior performance of this invention, drug-drug interaction prediction tasks were performed on the test dataset constructed in S1 using the methods of this invention, SSI-DDI, GMPNN, DDIMDL, MDF-SA-DDI, AttentionDDI, MTDDI, MIRCLE, and 3DGT. The prediction results were analyzed, and the performance is shown in Table 2. From Table 2, and... Figure 4 , Figure 5 As can be seen, the present invention outperforms similar methods in all other evaluation metrics except Auc, demonstrating better predictive performance.

[0164] Table 2 Comparison of performance with similar methods

[0165]

[0166] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A DDI prediction method based on Siamese structure and graph contrastive learning, characterized in that, Includes the following steps: S1: Collect drug-drug interaction text data, drug physicochemical property data files, and target relationship data files; filter and preprocess the collected data to obtain the raw set of drug-drug interaction data to be processed. S2: Extract the physicochemical property features based on the drug SMILES molecular format string and the targeting relationship features based on the target, enzyme, and pathway from the physicochemical property data file of the drug to be processed, and then fuse the physicochemical property features and the targeting relationship features to obtain the initial features based on the drug. S3: Obtain a drug-drug interaction adjacency matrix based on drug-drug interaction text data. Combine the drug-drug interaction adjacency matrix with the initial features obtained in S2 based on the physicochemical properties and targeting relationship of drugs to construct a drug-drug interaction heterogeneity graph. S4: Input the drug-drug interaction heterogeneity graph into a graph contrastive learning model based on Siamese structure, and learn the embedding features of drug nodes from the graph topology; The graph contrastive learning model is a graph isomorphic network, where the nodes are drug nodes and the node features are the attributes of the nodes. The steps for learning the embedding features of drug nodes include: S41: Extract two isomorphic subgraphs from a heterogeneous graph structure containing multiple DDI relationships, which have the same side type and different node attributes; S42: Input the two isomorphic subgraphs as input objects to the graph isomorphic network with shared weight Siam structure into the comparison model of graph isomorphic networks based on Siam structure; S43: Update and iterate the node features in the graph isomorphic network, wherein in each layer of the graph isomorphic network, the node features of the next layer of each drug node are aggregated from the node features of the current layer of the drug node and the node features of other drug nodes in the current layer. S44: After the iteration is completed, the compressed drug embedding vector is obtained; S5: Using the embedding features of the drug nodes as input, calculate the score of the edge between any two drug nodes using the link prediction method.

2. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 1, characterized in that, In S1: Drug-drug interaction text data includes: drug-drug interaction text; The drug-target interaction data files include: drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data; The physicochemical property data file for a drug includes: the SMILES molecular format string of the drug.

3. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 1, characterized in that, S2 includes: S21: Extracting physicochemical property features includes: converting the SMILES molecular format string of the drug into a drug molecular fingerprint format according to its physicochemical properties, and further encoding it into a binary bit vector, and calculating a drug-drug similarity matrix based on the bit vector; S22: Extracting target relationship features includes: converting drug-target interaction data, drug-enzyme interaction data, and drug-pathway interaction data into binary bit vectors according to drug-target, drug-enzyme, and drug-pathway interaction relationships, and calculating drug-target similarity matrix, drug-enzyme similarity matrix, and drug-pathway similarity matrix based on the bit vectors respectively; S23: The four similarity feature matrices in S21 and S22 are fused to obtain a multi-source drug similarity matrix. Each row in the matrix is ​​the two-step drug feature of the current drug.

4. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 3, characterized in that, The steps for calculating the similarity matrix in S21 and S22 include: The similarity matrix is ​​calculated using the following formula: ； Wherein, N(D) A D B )11 is the bit vector D A and D B The number of attributes in a given class that are all equal to 1, N(D) A D B )01 is the bit vector D A and D B D A The attribute value is 0 and D B The number of attributes with a value of 1, N(D) A D B )10 is the bit vector D A and D B D A The attribute value is 1 and D B The number of attributes with a value of 0; For calculating the drug-drug similarity matrix: D A D B The molecular fingerprint bit vectors representing drug A and drug B; For calculating the drug-target similarity matrix: D A D B One-Hot vector representing the drug-target; For calculating the drug-enzyme similarity matrix: D A D B One-Hot vector representing drug-enzyme; For calculating the drug-pathway similarity matrix: D A D B One-Hot vector representing drug-pathway.

5. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 3, characterized in that, The fusion step of S23 includes: S231: Combine the drug-drug similarity matrix, drug-target similarity matrix, drug-enzyme similarity matrix and drug-pathway similarity matrix in a tandem manner to obtain a tandem drug similarity matrix, where each row in the matrix is ​​a one-step drug feature of the current drug based on physicochemical properties and targeting relationships; S232: Calculate the cascaded drug similarity matrix using the Gaussian kernel function to obtain a multi-source drug similarity matrix based on the Gaussian kernel function, where each row in the matrix is ​​the two-step drug feature of the current drug after multi-source feature fusion.

6. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 1, characterized in that, The step in S3 of obtaining the drug-drug interaction adjacency matrix based on drug-drug interaction text data includes: The drug-drug interaction text data includes descriptive text describing the DDI event between each two drugs. By replacing keywords in the descriptive text with fixed words, the drug-drug interaction relationship is obtained. The keywords include drug names. Construct a drug-drug interaction adjacency matrix based on drug-drug interaction relationships.

7. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 5, characterized in that, The drug-drug interaction isomer diagram in S3 includes: The nodes of the drug-drug interaction isomer diagram are drugs; The position of 0 in the adjacency matrix determines that there is no edge between two nodes, and the position of 1 in the adjacency matrix determines that there is an edge between two nodes. The attributes of nodes are determined based on the two-step drug features obtained from multi-source feature fusion.

8. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 1, characterized in that, The specific steps in S5 include: The link prediction model based on GNN calculates the score of links between drug nodes, and the calculation formula is as follows: ； in, and φ represents the features of drug nodes u and v at layer L, calculated from a multi-layer graph isomorphic network, and φ is the defined prediction operation.

9. The DDI prediction method based on Siamese structure and graph contrast learning according to claim 1, characterized in that, S5 also includes performance evaluation of the graph contrast learning model, with specific steps including: The area under the receiver operating characteristic curve (ROC-AUC), accuracy (Acc), F1 score, precision, recall, and area under the PR curve (PR-AUC) were used as evaluation metrics. The specific calculation formula is as follows: ； ； ； ； ； Where TP is a DDI that is predicted to be true by the model, TN is a DDI that is predicted to be false by the model, FP is a DDI that is predicted to be true by the model, and FN is a DDI that is predicted to be false by the model. ROC_AUC is obtained by using FPR and TPR as the horizontal and vertical axes respectively to obtain the ROC curve, and then calculating the AUC under the ROC curve. PR-AUC is obtained by calculating the AUC under the PR curve with Recall and Precision as the horizontal and vertical axes, respectively.

Images