Method for predicting drug pair interaction based on strong negative sample generated by multi-scale drug information

Abstract

The method for predicting drug pair interaction based on strong negative sample generated by multi-scale drug information involves the field of bio-information technology.In order to solve the problems that the existing technology only depends on the experimental data for drug interaction judgment, the model trained on weak negative samples selected by random negative sampling shows poor robustness and prediction performance, and most DDI prediction methods have poor prediction accuracy for some interaction categories under the cold start setting due to the skew distribution of data, the technical scheme provided by the present application is as follows: the method comprises the following steps: learning the representation of drug knowledge graph, drug structure and drug physical and chemical attribute information; obtaining a drug similarity matrix according to the drug multi-scale information representation, and constructing a drug similarity network; the step of fusing the drug similarity network; the step of generating strong negative samples for the target drug interaction data set.The method is suitable for application in the work of predicting the interaction between drugs.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to the construction and application of strong negative samples for training a model for predicting drug-drug interaction relationships. Background Art

[0002] Drug-drug interactions (DDIs) refer to interactions between two or more drugs during drug administration, including pharmacological changes and toxic side effects. Predicting drug-drug interactions plays a crucial role in drug discovery. Understanding the pharmacological changes of drug pairs during drug administration helps understand the mechanisms of action of drug combination therapy; understanding the toxic side effects of drug pairs is crucial for avoiding adverse reactions in patients receiving combination therapy. Drug combination therapy plays a key role in treating various complex diseases, such as cancer, diabetes, and bacterial infections, improving therapeutic efficacy while minimizing toxicity. Many DDIs cannot be identified during clinical trials, making it crucial to predict unknown DDIs before drug approval. However, detecting DDIs through biological or pharmacological methods is costly and time-consuming. Computational methods for DDI prediction offer a cost-effective alternative, helping to identify potential DDIs before clinical trials. Over the past two decades, numerous drug-related databases, such as DrugBank, TwoSides, and DDInter, have emerged, providing a rich data resource for DDI prediction. Current DDI prediction methods fall into four main categories: similarity-based methods, knowledge graph-based methods, and machine learning-based methods. However, these DDI prediction methods face two major challenges. First, existing methods primarily train models through random negative sampling, where known DDIs are set as positive samples and all unknown DDIs are labeled as negative samples. Negative sampling is necessary due to the imbalance in the number and distribution of known and unknown DDIs. However, most methods randomly select negative samples equal to the number of positive samples from all unknown DDIs, which may misselect potential positive samples, generating weak negative samples. This leads to poor robustness and predictive performance when the model is trained on these weak negative samples. Second, due to the skewed distribution of DDI data, most DDI prediction methods exhibit poor prediction accuracy for certain interaction categories under cold-start settings. Prediction performance can be improved by leveraging relationships between drugs and other entity types in the knowledge graph (such as diseases, targets, and genes). However, over-reliance on drug knowledge graph information may overlook the drug's structural and physicochemical properties, necessitating the model's ability to fully learn multi-scale information representations of the drug. Summary of the Invention

[0003] To solve the technical problems in the prior art that drug interaction judgment only relies on existing experimental data, the model trained on weak negative samples selected by random negative sampling shows poor robustness and prediction performance, and most DDI prediction methods have poor prediction accuracy for certain interaction categories in a cold start setting, and cannot meet the current research needs, the technical scheme provided by the present application is:

[0004] The method comprises the following steps:

[0005] A step of learning drug knowledge graph, drug structure and drug physical and chemical attribute information representation;

[0006] A step of obtaining a drug similarity matrix according to the drug knowledge graph, drug structure and drug physical and chemical attribute information representation;

[0007] A step of obtaining a drug similarity network according to the drug similarity matrix;

[0008] A step of fusing the drug similarity network.

[0009] Further, a preferred embodiment is provided, wherein the drug knowledge graph embedding representation is learned by using a knowledge graph embedding representation learning model through the drug-target, drug-disease and drug-side effect entity and relationship interaction information covered in the biological and medical knowledge graph.

[0010] Further, a preferred embodiment is provided, wherein the drug structure information representation is represented by a dictionary-based molecular fingerprint.

[0011] Further, a preferred embodiment is provided, wherein the drug physical and chemical attributes include drug similarity quantitative estimation (QED), synthesis accessibility (SA), lipid-water partition coefficient (LogP), molecular weight (MW), topological polar surface area (TPSA), number of hydrogen bond acceptors (HBA) and number of hydrogen bond donors (HBD).

[0012] Further, a preferred embodiment is provided, wherein the drug similarity matrix is calculated by using cosine similarity and Jaccard similarity, and is normalized to construct a drug similarity network.

[0013] Based on the same inventive concept, the present application also provides a model construction device for predicting drug pair interaction based on strong negative samples generated by multi-scale drug information, which comprises:

[0014] A module for learning drug knowledge graph, drug structure and drug physical and chemical attribute information representation;

[0015] A module for obtaining a drug similarity matrix based on the drug knowledge graph, drug structure, and drug physicochemical property information;

[0016] Obtaining a module of a drug similarity network according to the drug similarity matrix;

[0017] A module for fusing the drug similarity network.

[0018] Based on the same inventive concept, the present invention also provides a method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, the method comprising:

[0019] The step of generating strong negative samples for the target drug interaction dataset according to the model constructed by the method.

[0020] Based on the same inventive concept, the present invention also provides a device for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, the device comprising:

[0021] A module for generating strong negative samples for a target drug interaction dataset based on a model constructed by the device.

[0022] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method described above.

[0023] Based on the same inventive concept, the present invention also provides a computer, comprising a processor and a storage medium. When the processor reads the computer program stored in the storage medium, the computer executes the method described above.

[0024] Based on the same inventive concept, the present invention also provides a computer program product, which is a computer program. When the computer program is read, the method described above is implemented.

[0025] Compared with the prior art, the technical solution provided by the present invention is beneficial in that:

[0026] The method provided by the present invention for predicting drug pair interactions by using strong negative samples generated based on multi-scale drug information uses the entity and relationship interaction information such as drug-target, drug-disease and drug-side effect covered in the biomedical knowledge graph, and uses the knowledge graph embedding representation learning large model to learn the representation of drug entities. This method significantly improves the accuracy and richness of drug representation and enhances the accuracy of drug interaction prediction.

[0027] The method provided by the present invention for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information characterizes drug structural information through drug molecular fingerprints, meticulously reflects the substructure of drug molecules, and helps to more accurately calculate the similarity between drugs.

[0028] The method provided by the present invention for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, including physicochemical properties such as quantitative estimation of drug similarity (QED) and synthetic accessibility (SA), provides a detailed characterization of drugs at the physical and chemical levels, further improving the calculation accuracy of drug similarity.

[0029] The method provided by this paper generates strong negative samples for predicting drug-pair interactions based on multi-scale drug information. It calculates drug similarity matrices based on knowledge graph representation, structural representation, and physicochemical property representation, and then performs Min-Max normalization. This process comprehensively considers multiple aspects of drug characteristics, effectively calculating drug information at different scales, and making similarity calculations more comprehensive and accurate.

[0030] The method provided by this paper, which generates strong negative samples based on multi-scale drug information to predict drug pair interactions, uses a nonlinear combination approach to fuse three similarity networks. After multiple iterations, a more accurate fused similarity matrix, simF, is obtained. This approach effectively resolves the information silo problem, fusing information from different data sources and significantly improving the performance of the prediction model.

[0031] The method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information provided by the present invention is to generate a strong negative sample based on the drug fusion similarity matrix for each positive sample triple (D a ,D b ,R i ), the relationship types between drugs remain unchanged, and the drug to be replaced is selected with a certain probability. The replacement drug is guaranteed to have the minimum fusion similarity with the remaining drugs, generating strong negative samples that are unique to the positive samples and have the same relationship type distribution. This method not only solves the problem of unbalanced sampling of positive and negative samples, but also generates more diverse and challenging negative samples, significantly improving the training effect of drug-drug interaction prediction models.

[0032] The method provided by this invention for predicting drug-drug interactions by generating strong negative samples based on multi-scale drug information systematically constructs and fuses drug similarity networks by integrating multi-scale drug information and innovatively generates strong negative samples, fundamentally improving the accuracy and robustness of drug-drug interaction predictions. Existing methods typically only consider information at a single scale or have deficiencies in negative sample generation. This method, however, comprehensively enhances prediction performance by comprehensively utilizing information from multiple data sources.

[0033] The method provided by the present invention for predicting drug pair interactions by using strong negative samples generated based on multi-scale drug information is suitable for use in the work of predicting drug-drug relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the process of predicting drug pair interactions based on strong negative samples generated by multi-scale drug information;

[0035] Figure 2 The following is a framework diagram of the drug-drug similarity fusion method;

[0036] Figure 3 This is an example diagram of random negative sampling and strong negative sampling;

[0037] Figure 4 tSNE visualization diagram of the knowledge graph embedding representation of positive samples and all unknown DDI interactions;

[0038] Figure 5 Represent similarity frequency distribution histogram for drug pair knowledge graph;

[0039] Figure 6 A histogram of the frequency distribution of structural similarity for drug pairs;

[0040] Figure 7 Represent frequency distribution histograms of similarity of physicochemical properties for drug pairs;

[0041] Figure 8 Represent similarity frequency distribution histograms for drug pair fusions;

[0042] Figure 9 Radar chart comparing the prediction performance of the MSTE method trained with random negative sampling and MDNDDI sampling after perturbation;

[0043] Among them, 9(a) shows the prediction performance of the MSTE method after uniform distribution perturbation, using random negative sampling and MDNDDI sampling training, and 9(b) shows the prediction performance of the MSTE method after normal distribution perturbation, using random negative sampling and MDNDDI sampling training. DETAILED DESCRIPTION

[0044] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail with reference to the accompanying drawings, specifically:

[0045] Embodiment 1: This embodiment provides a method for constructing a model for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, the method comprising:

[0046] Steps to learn drug knowledge graphs, drug structures, and drug physicochemical property information representation;

[0047] The step of obtaining a drug similarity matrix based on the drug knowledge graph, drug structure and drug physicochemical property information representation;

[0048] A step of obtaining a drug similarity network according to the drug similarity matrix;

[0049] The step of fusing the drug similarity network.

[0050] Specifically, this embodiment provides a method for predicting drug-drug interactions by generating strong negative samples based on multi-scale drug information. This method constructs a multi-scale representation of drugs, calculates a drug similarity matrix, and selectively generates strong negative samples to improve the performance of drug-drug interaction prediction.

[0051] Step 1: Construct a multi-scale representation of the drug

[0052] Step 1.1: Drug representation based on knowledge graph

[0053] Extract entity and relationship interaction information such as drug-target, drug-disease and drug-side effect from biomedical knowledge graph.

[0054] The knowledge graph embedding representation learning model is used to learn the representation of drug entities. Nodes represent entities such as drugs, proteins, genes, diseases, and side effects, and edges represent the relationships between entities (such as drug-drug interactions).

[0055] Input the knowledge graph triple (h, t, r) and obtain the entity embedding representation by learning the semantic relationship mapping function between the triple head entity h, tail entity t, and relation r.

[0056] Step 1.2: Structure-based drug representation

[0057] The drug structure information is represented by the dictionary-based drug molecular fingerprint, and a fixed-length binary vector is generated. The positions corresponding to the substructures of the drug molecules (such as hydroxyl groups, benzene rings or nitrogen atoms) are 1, and the other positions are 0.

[0058] Step 1.3: Drug representation based on physicochemical properties

[0059] A physicochemical embedding representation of the drug is generated by calculating the quantitative estimate of similarity (QED), synthetic accessibility (SA), lipid-water partition coefficient (LogP), molecular weight (MW), topological polar surface area (TPSA), number of hydrogen bond acceptors (HBA), and number of hydrogen bond donors (HBD).

[0060] Step 2: Construct drug similarity network and integrate

[0061] Step 2.1: Calculate the drug similarity matrix

[0062] The drug similarity matrix was calculated for the knowledge graph representation and physicochemical property representation of drug pairs using cosine similarity, and the drug similarity matrix was calculated for the structural representation of drug pairs using Jaccard similarity.

[0063] The drug similarity matrices calculated by the three drug representation methods were Min-Max normalized to obtain the drug knowledge graph similarity matrix, drug structure similarity matrix and drug physicochemical property similarity matrix.

[0064] Step 2.2: Constructing a drug similarity network

[0065] According to the drug similarity matrix, a drug similarity network is constructed, the adjacency matrix and degree matrix are defined, and the global similarity matrix and local similarity matrix are calculated.

[0066] Step 2.3: Fusion of multi-scale drug similarity networks

[0067] The global similarity matrices and local similarity matrices of the three networks are fused using a nonlinear combination method. After multiple iterations, the fused similarity matrix simF is obtained.

[0068] Step 3: Generate strong negative samples

[0069] Step 3.1: Select strong negative samples

[0070] According to the drug fusion similarity matrix, for each positive sample triple, the relationship type between drugs remains unchanged, and the replaced drug is selected with a certain probability. The replaced drug ensures the minimum fusion similarity with the remaining drugs, generating strong negative samples that are non-repetitive with the positive sample and have the same relationship type distribution.

[0071] This embodiment proposes a new strong negative sampling method that can generate strong negative samples of drug interactions for training DDI prediction models.

[0072] By measuring the similarity between drugs and fusing the information between drug similarity networks, the performance of drug-drug interaction prediction methods can be improved.

[0073] This implementation proposes a method for predicting drug pair interactions using strong negative samples based on multi-scale drug information generation (MDNDDI), through three steps: multi-scale drug information representation, drug similarity network construction and fusion, and strong negative sample generation. This method effectively addresses the issues of imbalanced positive and negative sample sampling and poor cold start performance in existing methods, improving prediction accuracy and robustness.

[0074] Implementation method 2. This implementation method further limits the model construction method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information provided in implementation method 1. The drug knowledge graph covers entity and relationship interaction information such as drug-target, drug-disease and drug-side effect, and obtains entity embedding representation through semantic relationship mapping function.

[0075] Implementation method three: This implementation method further limits the model construction method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information provided in implementation method one, and the drug structure is represented by a map-based drug molecular fingerprint.

[0076] Implementation method 4. This implementation method further limits the model construction method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information provided in implementation method 1, and the drug physicochemical properties include quantitative estimation of drug similarity, synthetic accessibility, distribution coefficient, molecular weight, topological polar surface area, number of hydrogen bond acceptors and number of hydrogen bond donors.

[0077] Implementation method 5. This implementation method further limits the model construction method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information provided in implementation method 1. The drug similarity matrix is ​​calculated using cosine similarity and Jaccard similarity, and normalized to obtain a drug similarity network.

[0078] Embodiment 6: This embodiment provides a device for constructing a model for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, the device comprising:

[0079] Modules for learning drug knowledge graphs, drug structures, and drug physicochemical property information representation;

[0080] A module for obtaining a drug similarity matrix based on the drug knowledge graph, drug structure, and drug physicochemical property information;

[0081] Obtaining a module of a drug similarity network according to the drug similarity matrix;

[0082] A module for fusing the drug similarity network.

[0083] Embodiment 7: This embodiment provides a method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, the method comprising:

[0084] The model constructed according to the method provided in the first embodiment generates strong negative samples for the target drug interaction data set.

[0085] Embodiment 8: This embodiment provides a device for predicting drug pair interactions based on strong negative samples generated based on multi-scale drug information, the device comprising:

[0086] According to the model constructed by the device provided in the first embodiment, a module is used to generate strong negative samples for the target drug interaction data set.

[0087] Implementation method 9: This implementation method provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method provided in implementation method 1.

[0088] Implementation 10: This implementation provides a computer, including a processor and a storage medium. When the processor reads the computer program stored in the storage medium, the computer executes the method provided in Implementation 1.

[0089] Implementation Method 11: Combination Figure 1-9 This embodiment further describes the above technical solution in detail through specific examples, specifically:

[0090] This embodiment provides a method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information.

[0091] The method described in this embodiment specifically includes the following steps:

[0092] Step 1: Construct a multi-scale representation of the drug based on its knowledge graph information, structure information, and physical and chemical property information.

[0093] Step 2: Calculate the drug similarity matrix based on the multi-scale representation of the drug, construct a drug similarity network, and then fuse several drug similarity matrices based on the similarity network fusion method.

[0094] In step 3, based on the generated drug similarity matrix that integrates multi-scale drug information, strong negative samples are generated for the target drug interaction dataset to improve the performance of the drug-drug interaction prediction model.

[0095] Furthermore, in step 1,

[0096] Step 1.1: Based on the interaction information between multiple entities and relationships in the biomedical knowledge graph, such as drug-target, drug-disease, and drug-side effect, the knowledge graph representation of drug entities is learned using the knowledge graph embedding representation learning model. Nodes in the knowledge graph represent entities such as drugs, proteins, genes, diseases, and side effects, and edges represent relationships between entities, such as drug-drug interactions, drug-disease interactions, and drug-side effect interactions. A knowledge graph triple (h, t, r) is input, and the entity embedding representation is obtained by learning the semantic relationship mapping function between the triple's head entity h, tail entity t, and relationship r. The embedding representation of a correct triple should satisfy h + r = t; conversely, if the triple is incorrect, the embedding representation between them should not satisfy this relationship. Therefore, we need to learn a representation so that the d(h+r, t) distance of the positive sample is as small as possible, and the d(h+r, t) distance of the negative sample is as large as possible. The distance is defined as the L2 norm between the vectors. The loss function is defined as follows, where S is the positive sample set, S' is the negative sample set, y is a constant representing the distance between positive and negative samples, [x] + =max(0,x).

[0097] minΣ (h,t,r∈S) Σ (h`，t`，r∈S`) [d(h+r,t)-d(h`+r,t`)+γ] +

[0098] In step 1.2, the drug structure information is represented according to the dictionary-based drug molecular fingerprint, and a fixed-length binary vector is generated, in which each position is defined as a substructure of the drug molecule (e.g., hydroxyl group, benzene ring, or nitrogen atom). If the drug contains the substructure, the value at the corresponding position of the substructure is 1, otherwise it is 0.

[0099] In step 1.3, a physicochemical embedding representation of the drug is generated based on multiple physicochemical properties of the drug, including quantitative estimate of drug likeness (QED), synthetic accessibility (SA), lipid-water partition coefficient (LogP), molecular weight (MW), topological polar surface area (TPSA), number of hydrogen bond acceptors (HBA), and number of hydrogen bond donors (HBD).

[0100] Furthermore, in step 2,

[0101] Step 2.1, according to the drug i and D j Different embedding representations of d i and d j , calculate the drug similarity matrix. For the knowledge graph topology representation and physicochemical property representation of the drug, cosine similarity is used for calculation, and for the drug structure representation, Jaccard similarity is used for calculation.

[0102]

[0103] The drug similarity matrices calculated by the three drug representation methods are normalized by Min-Max. After calculation, three drug similarity matrices are obtained, namely, drug knowledge graph similarity matrix A KG , drug structure similarity matrix A S Similarity matrix A with drug physicochemical properties PC .

[0104] Step 2.2: Construct three drug similarity networks based on the three drug similarity matrices. Define the adjacency matrix A of each network as the corresponding drug similarity matrix. Define the degree matrix D of each network as a diagonal matrix, where the diagonal elements are the sum of the corresponding rows of the similarity matrix. The initialized global similarity matrix P represents the global similarity of each node. Drawing on the idea of ​​calculating the probability transfer matrix of random walk, it is defined as

[0105] P=(I+D)-1(I+A)

[0106] Where I is the identity matrix, D is the degree matrix, and A is the adjacency matrix. The sum of each row element of P is 1, that is, ∑ j P(i, j) = 1. In order to consider the local similarity of drugs, the K nearest neighbor similarity matrix is ​​used to identify the local similarity of drugs. The initialized local similarity matrix S is defined as:

[0107] S=K-1A

[0108] Where K is a diagonal matrix, each diagonal element is the sum of the K largest values ​​on that row, and A is the adjacency matrix.

[0109] Step 2.3, fuse the multi-scale drug similarity network. Use the nonlinear combination method to combine the global similarity matrix (P KG , P S , P PC ) and the local similarity matrix (S KG , S S , S PC ) fusion, after t iterations, the global similarity matrix of each network is updated to

[0110]

[0111] set up Denotes the initial matrix at t = 0. The global similarity matrices of the three networks after t iterations are averaged to obtain the drug fusion similarity matrix simF, which is defined as

[0112]

[0113] In the formula and Represent the global similarity matrices of the three drug similarity networks after t iterations.

[0114] Furthermore, in step 3,

[0115] Step 3.1, according to the drug fusion similarity matrix SimF, for each positive sample triple (D a , D b , R i ), choose drug D with a certain probability a (or drug D b ) is replaced by drug D c , where drug D is replaced c and retain drug D b (or drug D a ) has a minimum drug fusion similarity, defined as:

[0116]

[0117] A higher fusion similarity score for a drug pair indicates that the two drugs are more similar and the distance between them is smaller. Conversely, a lower fusion similarity score indicates that the two drugs are less similar and the distance between them is larger. Ensure that the negative sample triplets do not exist in the positive samples, that the negative samples are not repeated, and that the ratio of positive to negative samples is 1:1.

[0118] The technical solution provided in this embodiment is different from most existing methods that rely on random negative sampling. It introduces a new strong negative sampling method that can generate strong negative samples of drug interactions for training DDI prediction.

[0119] The similarity between drugs is measured by calculating three drug representation methods (drug knowledge graph information, drug structure information and drug physicochemical property information).

[0120] The three drug similarities are fused through information transfer between drug similarity networks.

[0121] Figure 1 The following is a flowchart of the method of this embodiment. This embodiment comprises four steps: 1) Constructing a multi-scale representation of a drug based on its knowledge graph topology, structure, and physicochemical properties. 2) Fusion of three drug-drug similarity matrices through information transfer between drug similarity networks. 3) Based on the generated drug similarity matrix fused with multi-scale drug information, generating strong negative samples for the target drug interaction dataset. 4) Using these strong negative samples to improve the performance of the drug-drug interaction prediction method.

[0122] Figure 2 The framework of the drug-drug similarity fusion method is shown below. It consists of three steps: 1) Calculate drug similarity matrices based on the multi-scale representation of each drug. 2) Construct a drug similarity network. 3) Fusion of multiple drug similarity matrices using the similarity network fusion method.

[0123] Figure 3 Figures show examples of random negative sampling and strong negative sampling. a) For drug pairs representing similar spaces, the number of positive samples is significantly smaller than the number of negative samples. To reduce the impact of the imbalance between positive and negative data, downsampling is used to balance the data. b) Randomly selected negative samples may contain potential positive samples, that is, undetected positive samples. The prediction boundary of the model trained using these negative samples is close to the positive samples, resulting in reduced prediction performance. c) The negative samples in strong negative sampling are learned non-interacting drug pairs. The prediction boundary of the trained model is between the positive and negative samples, which can improve prediction performance.

[0124] In a specific embodiment, step 1, three ways of expressing the drug:

[0125] In order to learn the embedding representation of large-scale drug knowledge graphs, we use the DRKG knowledge graph with 5,874,261 triples covering 107 relations consisting of 17 entity pairs, and use the TransE-l2 method widely used in the DGL-KE toolkit to generate a 400-dimensional knowledge graph embedding representation of drug entities. The tSNE visualization of the embedding representation of positive samples and all unknown interactions (negative samples) is as follows Figure 4 As shown, it can be seen that this representation method is helpful in distinguishing whether there is an interaction between drug pairs.

[0126] We characterize drug structure information through dictionary-based drug molecular fingerprints. We use the RDKit tool to generate a 166-dimensional drug MACCS molecular fingerprint binary vector representation. Each position is defined as a substructure of the drug molecule (for example, a hydroxyl group, a benzene ring, or a nitrogen atom). If the drug contains the substructure, the value at the position corresponding to the substructure is 1, otherwise it is 0.

[0127] The RDKit tool was used to calculate the quantitative estimate of drug similarity (QED), synthetic accessibility (SA), lipid-water partition coefficient (LogP), molecular weight (MW), topological polar surface area (TPSA), number of hydrogen bond acceptors (HBA) and number of hydrogen bond donors (HBD) values, and the 7-dimensional physicochemical property characterization of the drug was generated.

[0128] Step 2: Drug-drug similarity fusion

[0129] Compared with simple mean or weighted methods, SNF can better capture and integrate the associations between drug information at different scales through the random walk process, thereby more accurately reflecting the similarity between drugs. Considering the transfer matrix P = D -1 A does not take into account the probability of a node staying on itself and modifies it to P = (I + D) -1 (I+A), thereby enhancing the similarity between the drug and itself and weakening the similarity transmission between drugs. Figure 5-8 The similarity frequency distributions of the embedding representations of drug pairs at three scales are shown for positive samples and all unknown interactions (negative samples). A rank sum test shows that the distributions of the three groups of positive and negative samples are significantly different (p < 0.05). Overall, drug information obtained through knowledge graphs plays a more important role in DDI prediction using negative sampling.

[0130] Step 3: Generate strong negative samples

[0131] Our method is evaluated using DrugBank, a widely used real-world DDI benchmark dataset. DrugBank contains 191,808 DDI triplets, 1,706 drugs, and 86 interaction types. Each interaction type is represented by a number starting from 1, describing how one drug affects the metabolism of another drug. For example, (DB04571, DB00460, 1) indicates that Trioxsalen may increase the photosensitivity of Verteporfin. By screening a total of 1,691 drugs contained in the DrugBank dataset and the DRKG knowledge graph, there are 191,286 interactions between them (positive samples), so 191,286 strong negative samples are generated according to the above method.

[0132] Step 4: Drug-drug interaction prediction

[0133] Using known positive samples from DrugBank and strong negative samples generated by our method to train drug-drug interaction prediction models can improve prediction accuracy and robustness. We applied uniform and normal perturbations to the warm-start setting of the MSTE method and tested the evaluation metrics after perturbation. As shown in the figure below, it is clear that the performance change of our proposed MDNDDI sampling method after perturbation is smaller than that of random negative sampling. Figure 9 (a) shows the AUROC values ​​of the MSTE method trained with random negative sampling and MDNDDI sampling on 86 interaction relations after adding uniformly distributed noise U(0,0.2) to the head entity, tail entity and relationship representations for perturbation. Figure 9(b) shows the AUROC values ​​of the MSTE method trained with random negative sampling and MDNDDI sampling on 86 interaction relationships after adding normal distribution noise N(0,1) perturbations to the head entity, tail entity, and relationship representations. The results show that the MSTE method trained with MDNDDI sampling has good robustness.

[0134] The above further describes the technical solution provided by the present invention in detail through several specific embodiments in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific embodiments described above are not intended to limit the present invention. Any reasonable modification and improvement of the present invention, combination of embodiments and equivalent replacement based on the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A model construction method for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, characterized by: The method comprises: Steps to learn drug knowledge graphs, drug structures, and drug physicochemical property information representation; The step of obtaining a drug similarity matrix based on the drug knowledge graph, drug structure and drug physicochemical property information representation; The drug similarity matrix includes the drug knowledge graph similarity matrix, the drug structure similarity matrix and the drug physicochemical property similarity matrix; A step of obtaining a drug similarity network according to the drug similarity matrix; The step of fusing the drug similarity network to obtain a drug fusion similarity matrix; According to the drug fusion similarity matrix, for each positive sample triple, the relationship type between drugs remains unchanged, and the replaced drug is selected with a certain probability. The replacement drug ensures that the fusion similarity between the replaced drug and the remaining drugs is minimized, and a strong negative sample with the same relationship type distribution as the positive sample is generated; The step of training a DDI prediction model according to the strong negative samples.

2. The method for constructing a model for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information according to claim 1, characterized in that: The drug knowledge graph embedding representation is learned by using the knowledge graph embedding representation learning model through the entity and relationship interaction information such as drug-target, drug-disease and drug-side effect covered in the biomedical knowledge graph.

3. The method for constructing a model for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information according to claim 1, characterized in that: The drug structure information is represented by a dictionary-based drug molecule fingerprint.

4. The method for constructing a model for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information according to claim 1, characterized in that: The physicochemical properties of the drugs include quantitative estimate of drug likeness (QED), synthetic accessibility (SA), lipid-water partition coefficient (LogP), molecular weight (MW), topological polar surface area (TPSA), number of hydrogen bond acceptors (HBA) and number of hydrogen bond donors (HBD).

5. The method for constructing a model for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information according to claim 1, characterized in that: The drug similarity matrix was calculated using cosine similarity and Jaccard similarity and normalized to construct a drug similarity network.

6. A model construction device for predicting drug pair interactions based on strong negative samples generated by multi-scale drug information, characterized in that: The device comprises: Modules for learning drug knowledge graphs, drug structures, and drug physicochemical property information representation; A module for obtaining a drug similarity matrix based on the drug knowledge graph, drug structure, and drug physicochemical property information representation; The drug similarity matrix includes the drug knowledge graph similarity matrix, the drug structure similarity matrix and the drug physicochemical property similarity matrix; Obtaining a module of a drug similarity network according to the drug similarity matrix; fusing the drug similarity networks to obtain a module of a drug fusion similarity matrix; According to the drug fusion similarity matrix, for each positive sample triple, the relationship type between drugs remains unchanged, and the replaced drug is selected with a certain probability. The replacement drug ensures the minimum fusion similarity with the remaining drugs, and generates a module of strong negative samples that are non-repeating and have the same relationship type distribution as the positive sample; A module for training a DDI prediction model based on the strong negative samples.

7. A method for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information, characterized in that: The method comprises: The step of using the model constructed by the method according to claim 1 to predict the target drug interaction.

8. A device for predicting drug pair interactions based on strong negative samples generated from multi-scale drug information, characterized in that: The device comprises: A module for predicting target drug interactions using the model constructed using the device of claim 6.

9. A computer storage medium for storing a computer program, characterized in that When the computer program is read by a computer, the computer executes the method according to claim 1 .

10. A computer comprising a processor and a storage medium, characterized in that When the processor reads the computer program stored in the storage medium, the computer executes the method according to claim 1 .

Images