Drug interaction prediction method

By establishing a PK interaction network and using a multi-layer perceptron model, the molecular fingerprint of drugs is used to calculate structural similarity scores, which solves the problem that the existing technology cannot discover new drug interactions, and achieves high-accurate drug interaction predictions. Simplified operation process.

CN118983023APending Publication Date: 2024-11-19SHANGHAI FIRST PEOPLES HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410968428.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing drug interaction prediction methods cannot detect new drug interactions outside the text and cannot provide recommendations before combination therapy, and can only show that drug similarity characteristics in some aspects are independent of the DDI prediction task.

Method used

By distinguishing and defining three types of drugs, establishing a PK interaction network, calculating structural similarity scores based on the molecular fingerprint of the drug, defining five similarity scores, and calculating the maximum similarity between drugs to obtain predicted results, only The structure information of the drug is used as input and the compound characteristics are automatically obtained from the dataset using a multi-layer perceptron model.

Benefits of technology

The drastically reduced calculation improves prediction accuracy, simplifies operation, requires only drug structure information, has better performance, and reduces calculation while ensuring the accuracy of prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118983023A_ABST
    Figure CN118983023A_ABST
Patent Text Reader

Abstract

The invention discloses a drug interaction prediction method which is characterized by comprising the following steps: distinguishing and defining three types of drugs; establishing a PK interaction network; calculating a structural similarity score based on the molecular fingerprint of the drug; defining and obtaining five similarity scores; and calculating the maximum similarity among the drugs, and obtaining a prediction result according to the scores of the five similarities. According to the drug interaction prediction method provided by the invention, the result can be obtained only by taking the structural information of the drug as input, the method is used for small molecule drugs, the prediction accuracy is improved, drug research and development are facilitated, and the development cost and time are reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medicine, and in particular to a method for predicting drug interaction. Background Art

[0002] Drug-drug interactions (DDIs) can lead to adverse reactions and may result in the withdrawal of drugs from the market. Predicting DDIs during drug development will help developers reduce development costs and time. Many studies have been reported on predicting DDIs based on various methods, such as physiologically based pharmacokinetic (PBPK) modeling, molecular structure similarity analysis, ontology and annotation-based analysis, network modeling, QSAR modeling, and data mining from clinical data. Text mining-based methods cannot discover new DDIs outside of the text and cannot give advice to doctors before combined treatment. Machine learning-based methods mainly include logistic regression, etc., and are based on one or several ready-made drug feature similarities, such as chemical structure and side effects. However, the characteristics of similarity shown by drugs in a certain aspect may be irrelevant to the DDI prediction task we are concerned with. Summary of the invention

[0003] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that the existing methods for predicting DDI cannot discover new DDIs outside the text, cannot provide suggestions before combined treatment, or can only show the similar characteristics of drugs in certain aspects, which is irrelevant to the DDI prediction task. The present invention provides a drug interaction prediction method, which only needs the structural information of the drug as input to obtain the result, is used for small molecule drugs, improves the prediction accuracy, and helps to reduce the development cost and time of drug research and development.

[0004] To achieve the above object, the present invention provides a method for predicting drug interactions, comprising the following steps:

[0005] Distinguish and define three types of drugs;

[0006] Establish PK interaction network;

[0007] Calculate structural similarity scores based on the molecular fingerprints of drugs;

[0008] Five similarity scores were obtained by definition;

[0009] The maximum similarity between drugs was calculated, and prediction results were obtained based on five similarity scores.

[0010] Furthermore, the three types of drugs include all clinically approved drugs DA, drugs for drug interactions to be tested in a certain type of disease De, and query drugs Dq, wherein all Dq drugs need to be excluded from DA drugs.

[0011] Furthermore, the PK interaction network includes five parts, namely, direct interaction between De and DA drugs, drug-metabolizing enzyme-related drug interaction, drug-transporter-related drug interaction, drug-genetic metabolic enzyme-related drug interaction, and drug-genetic transporter-related drug interaction. Drug interactions are predicted through the above five interaction networks.

[0012] Furthermore, a PK interaction network was constructed through the five drug interactions.

[0013] Furthermore, the relevant data of De drugs were obtained using the DrugBank database.

[0014] Furthermore, first, all drugs used in a certain disease are searched in the DrugBank database according to the type of the disease;

[0015] And search the Enzymes, Carriers, and Transporters information of all De drugs in the DrugBank database;

[0016] Then, the pharmacogenetic metabolizing enzymes and pharmacogenetic transporter information of all De drugs were retrieved in the PharmGKB database;

[0017] Obtain all these drugs interacting with enzymes, carriers and transporters in DrugBank database, and obtain all these drugs interacting with pharmacogenetic metabolizing enzymes and drugs interacting with pharmacogenetic transporters in PharmGKB;

[0018] Drugs metabolized by the same enzyme as De were denoted as De-Dme, drugs metabolized by the same transporter as De as De as De-Dtr, drugs metabolized by the same pharmacogenetic enzyme as De-Dmeg, and drugs metabolized by the same pharmacogenetic transporter as De-Dtrg.

[0019] Furthermore, the drug interaction information list of all clinically approved small molecule drugs was crawled from the DrugBank database, and the SMILES numbers of all clinically approved small molecule drugs were obtained;

[0020] The acquired small molecule drugs are treated as DA drugs after excluding Dq drugs;

[0021] All DA drugs that interact with De form DDI pairs with De drugs, and all other DA drugs that do not interact with De form non-DDI pairs with De drugs.

[0022] Furthermore, the structural similarity score was calculated based on the molecular fingerprint of the drug, specifically using Python software to calculate the topological fingerprint RDkit; the calculation formula is as follows:

[0023] It involves converting the drug's SMILES number into a topological fingerprint and then calculating the molecular structure similarity.

[0024] Technical Effects

[0025] The present invention provides a drug interaction prediction method, which uses the maximum structural similarity score to greatly reduce calculations, and only uses drug metabolism kinetics to construct a drug interaction network, thereby reducing calculations while ensuring the accuracy of prediction results. It only requires the structural information of the drug as input, is simple and convenient to operate, and uses a multi-layer perceptron model to automatically obtain compound features from a data set, thereby having better performance.

[0026] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a box plot of a drug interaction prediction method according to a preferred embodiment of the present invention;

[0028] Figure 2 It is a schematic diagram of a quadruple cross-validation of a drug interaction prediction method according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] The present invention provides a method for predicting drug interactions, comprising the following steps:

[0031] Step 100, distinguishing and defining three types of drugs;

[0032] Step 200, establishing a PK interaction network;

[0033] Step 300, calculating a structural similarity score based on the molecular fingerprint of the drug;

[0034] Step 400, defining and obtaining five similarity scores;

[0035] Step 500, calculating the maximum similarity between drugs, and obtaining prediction results according to five similarity scores.

[0036] In step 100, the three types of drugs include all clinically approved drugs DA, drugs De for drug interactions to be tested in a certain type of disease, and query drugs Dq, wherein all Dq drugs need to be excluded from DA drugs.

[0037] Further, the PK interaction network includes five parts, namely, direct interaction between De and DA drugs, drug interactions related to drug metabolizing enzymes, drug interactions related to drug transporters, drug interactions related to drug genetics metabolizing enzymes, and drug interactions related to drug genetics transporters. Drug interactions are predicted through the above five interaction networks. Dq is the drug to be queried, which is a drug specifically used to treat a certain disease. Dq may be a newly developed drug or a drug already on the market; De is all the common drugs used in the above disease. In the embodiment of the present invention, it can be predicted whether the Dq drug will interact with the existing drugs used for the disease, i.e., De drugs, in the process of treating the disease. For example, voriconazole is studied for antifungal prevention of leukemia. In this case, voriconazole is the Dq drug, and all other common drugs for leukemia, including cyclophosphamide, paclitaxel, cisplatin, carboplatin, etc., are De drugs, which are used to predict whether voriconazole will interact with other common drugs for leukemia. In addition, it can also be used in the research stage of new drugs to predict whether new drugs will interact with other drugs, and only the structure of the new drug needs to be known. As another preferred embodiment: Cerivastatin sodium is a statin lipid-regulating drug with good effects, but it causes rhabdomyolysis due to drug interactions with gemfibrozil. For another example, when cerivastatin sodium was first developed, before it was put on the market, the drug's interactions with other drugs, such as cardiovascular drugs, were predicted to foresee possible risk factors. DA is a small molecule drug approved for clinical use on the market. The interaction information between DA drugs and De drugs is needed to construct an interaction network. Therefore, DA drugs are used to help us build models. We need to know the structures of Dq, De, and DA drugs, as well as the interaction information between DA drugs and De drugs.

[0038] After determining the Dq drug and the disease for which it is used, search the clinical guidelines and drugbank database according to the disease, and select all the commonly used drugs for the disease as De drugs. Then search the Enzymes, Carriers, and Transporters information of all De drugs in the drugbank database; then search the pharmacogenetic metabolic enzymes and pharmacogenetic transporter information of all De drugs in the PharmGKB database; then obtain all the drugs that interact with all these Enzymes, Carriers, and Transporters in the DrugBank database, and obtain all the drugs that interact with all these pharmacogenetic metabolic enzymes and pharmacogenetic transporters in PharmGKB. Drugs that are metabolized by the same enzyme as De are recorded as De-Dme, drugs that pass through the same transporter are recorded as De-Dtr, drugs that pass through the same pharmacogenetic metabolic enzymes are recorded as De-Dmeg, and drugs that pass through the same pharmacogenetic transporter are recorded as De-Dtrg.

[0039] All clinically approved small molecule drugs, as well as the interaction information and SMILES numbers of all these small molecule drugs, are crawled from the DrugBank database. After excluding Dq and De from these small molecule drugs, the remaining drugs are regarded as DA drugs. From these interaction information, we can know the interaction between DA drugs and De drugs. All DA drugs that interact with De form a DDI pair with De drugs, and all other DA drugs that do not interact with De form a non-DDI pair with De drugs.

[0040] Step 200, establish a PK interaction network; the PK interaction network includes five parts, namely, direct interaction between De and DA drugs, drug interaction related to drug metabolizing enzymes, drug interaction related to drug transporters, drug interaction related to pharmacogenetic metabolizing enzymes, and drug interaction related to pharmacogenetic transporters. The drug interactions are predicted through the above five interaction networks. A PK interaction network is constructed through 5 drug interactions. Specifically, the first is direct interaction, which is the drug interaction between DA drugs and De drugs based on the degree of structural similarity; the second is drug interaction related to drug metabolizing enzymes (Enzymes), which means that De drugs share the same drug metabolizing enzymes with other drugs, thus forming drug interactions. Specifically, De drugs and multiple drug metabolizing enzymes through which De drugs are metabolized, and all other drugs metabolized by these drug metabolizing enzymes (not necessarily DA drugs), which are called drug interactions related to drug metabolizing enzymes. Similarly, De drugs and multiple drug transporters (Carriers and Transporters) through which De drug transporters act, and all other drugs transported by these drug transporters (Carriers and Transporters) (not necessarily DA drugs).

[0041] [8] In this step, a multi-layer perceptron is used to construct a drug interaction prediction model. The main idea of ​​building the model is that it has been found that the five maximum structural similarity scores between DDI pairs and non-DDI pairs are not the same. The average and combined values ​​of the maximum similarity scores of each type in the combination with drug interaction are greater than those in the combination without drug interaction. This difference shows that the maximum similarity score can be used to predict DDI. We first input the five maximum structural similarity scores of the DDI pair and set the dependent variable to 1, that is, there is a drug interaction. Then, we input the five maximum structural similarity scores of the non-DDI pair and set the dependent variable to 0, that is, there is no drug interaction. The deep learning model can automatically learn according to the input data, adjust the internal parameters, and achieve the optimal prediction result. Finally, the five maximum structural similarity scores of the drug to be tested are input to obtain the prediction results output by the model. The basic principle of the model is as follows.

[0042] t1=β10+β11SSd+β12SSme+β13SSmeg+β14SStr+β15SStrg

[0043] t2=β20+β21SSd+β22SSme+β23SSmeg+β24SStr+β25SStrg ......

[0045] t20=β200+β201SSd+β202SSme+β203SSmeg+β204SStr+β

[0046] 205SStrg

[0047] hi=F(ti),i=1,2,3...20; F(t)=1 / (1+e -t )

[0048] X=y0+h1y1+h2y2+h3y3+……+h20y20

[0049] Finally, we get F(X), F(X) = 1 / (1+e -X ).

[0050] β10, β11......β205, y0, y1......y20 are all parameters, totaling 5*21+21, a total of 126.

[0051] Step 300, calculating the structural similarity score based on the molecular fingerprint of the drug, specifically using Python software to calculate it as a topological fingerprint RDkit; the calculation formula is as follows:

[0052] It involves converting the drug's SMILES number into a topological fingerprint and then calculating the molecular structure similarity.

[0053] Topological (TOPOLOGICAL) fingerprint is similar to the famous Daylight fingerprint in hash fingerprint, also known as RDKit fingerprint. By facilitating the storage of atoms and bond types corresponding to a certain number of bonds, there is no need to prepare partial structures in advance, so molecular structures can be expressed more flexibly. This is achieved through the RDKit open source cheminformatics toolkit in Python. RDKit can provide a series of powerful functions, including compound descriptor vectorization and partial structure retrieval, so that users can easily extract molecular features.

[0054] The second code is to calculate the molecular structure similarity, which is achieved through the Tanimoto coefficient. The calculation formula is:

[0055]

[0056] Using set comprehension is equivalent to:

[0057]

[0058] The following is a sample Python code:

[0059]

[0060] t=[13,12,20,11]

[0061] m=[11,11,11,11]

[0062] print(getTanimotocoefficient(s,t),getTanimotocoefficient(s,m))

[0063] #0.9845984598459846 0.8312342569269522

[0064] If the elements of two vectors can only be 01, then their Tanimoto similarity coefficient can be simplified to:

[0065] Tst=c÷(a+bc)

[0066] Where a is the number of 1s in vector s, b is the number of 1s in vector t, and c is the number of 1s in the same positions in both vectors.

[0067] The structural similarity score is then calculated using the Tanimoto similarity coefficient. The score is [0, 1], and the higher the score, the greater the degree of structural similarity. The Tanimoto similarity coefficient can be used to calculate the similarity between molecules. The method of representing molecules as a string of 01, i.e., a molecular fingerprint, is the first code mentioned above, which is the step of converting the SMILES number of the drug into a topological fingerprint.

[0068] In step 400, five maximum structural similarity scores are defined, which are explained in detail as follows.

[0069] SSd refers to the direct structural similarity score between DA and De. For example, if a DDI pair of a De drug has 100 DA drugs, then the direct structural similarity score in the DDI pair of the De drug is the structural similarity score between De and these 100 DA drugs.

[0070] SSme refers to the maximum structural similarity score between drugs interacting with De enzymes in DA and De networks, where SSm e =max(SSDe-Dme1, SSDe-Dme2). Similarly, SSmeg is the maximum similarity score between drugs with pharmacogenetic associations with the De enzyme gene in the DA and De networks. eg =max(SSDe-Dmeg1,SSDe-Dmeg2)

[0071] SS tr is the maximum similarity score between DA and drugs transported by the same transporter of De in the De network, where SStr = max(SSDe-Dtr1, SSDe-Dtr2).

[0072] Strg is the maximum similarity score between DA and drugs with pharmacogenetic association with De transporter gene in De network (De-Dtrg1, De-Dtrg2). Where, SStrg = max(SSDe-Dtrg1, SSDe-Dtrg2).

[0073] Step 500, calculating the maximum similarity between drugs, and obtaining prediction results according to five similarity scores.

[0074] If there were no drugs in the category in SS me, SS meg, SS tr, SS trg, a score of -10 was assigned, which was chosen empirically to facilitate the processing of the scores. In all networks, DA and De were removed when they also appeared as drugs in De's PK network. The reason for using the maximum score when there were multiple drugs in the subnetwork was based on the fact that structurally similar drugs may interact with the same protein, and the drug with the greatest structural similarity may interact with the protein most effectively and therefore have the greatest impact on DDI.

[0075] A specific example is given below to illustrate the method of the embodiment of the present invention.

[0076] Take the commonly used drugs for leukemia as an example: predict the drug interactions of fluconazole in leukemia drugs. Voriconazole is the drug to be tested, that is, Dq drugs, and other leukemia drugs are De drugs, a total of 60 drugs, that is, predict the drug interactions between fluconazole and these 60 drugs. Obtain all clinically approved small molecule drugs in Drugbank, excluding De and fluconazole, and there are 2577 remaining as Dq drugs. Then obtain the drug interaction information of these 60 De drugs and Dq drugs, and divide those with interactions into DDI pairs, and those without interactions into non-DDI pairs. Then obtain the enzymes, transporters, pharmacogenetic enzymes, and pharmacogenetic transporters that interact with De drugs in the database, and then obtain the drugs that all these enzymes, transporters, pharmacogenetic enzymes, and pharmacogenetic transporters interact with. Finally, calculate the maximum similarity. For example Figure 1 As shown in Figure 2, the average and combined values ​​of the maximum similarity scores of each type in DDI pairs are greater than those in non-DDI pairs. This difference indicates that the maximum similarity score can be used to predict DDI.

[0077] Finally, a four-fold cross-validation test was performed; Figure 2As shown, the average AUC = 0.975 was obtained, the model was stable, and the prediction accuracy was 91%.

[0078] The prediction results are as follows

[0079] Posaconazole

[0080]

[0081]

[0082] It can be seen that among the 33 drugs predicted to have drug interactions, 30 are clinically relevant.

[0083] Fluconazole

[0084]

[0085]

[0086]

[0087] It can be seen that among the 39 drugs predicted to have drug interactions, 37 are clinically relevant.

[0088] Voriconazole

[0089]

[0090]

[0091] It can be seen that among the 37 drugs predicted to have drug interactions, 37 are clinically relevant.

[0092] Amphotericin B

[0093]

[0094]

[0095] It can be seen that among the 18 drugs predicted to have drug interactions, 18 are clinically relevant.

[0096] Itraconazole

[0097]

[0098]

[0099]

[0100] It can be seen that among the 46 drugs predicted to have drug interactions, 45 are clinically relevant.

[0101] Drugs with clinical relevance are those that have been clinically confirmed to have drug interactions. For example, among the 46 drugs that we predicted to have possible drug interactions, 45 have been clinically confirmed, indicating that the prediction model of the method of the present invention has a high accuracy rate.

[0102] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting drug interactions, characterized in that: The following steps are involved: Distinguish and define three types of drugs; Establish PK interaction network; Calculate structural similarity scores based on the molecular fingerprints of drugs; Five similarity scores were obtained by definition; The maximum similarity between drugs was calculated, and prediction results were obtained based on five similarity scores.

2. A drug interaction prediction method according to claim 1, characterized in that: The three types of drugs include all clinically approved drugs DA, drugs for drug interactions to be tested in a certain type of disease De, and query drugs Dq. Among them, all Dq drugs need to be excluded from DA drugs.

3. A drug interaction prediction method according to claim 1, characterized in that: The PK interaction network includes five parts, namely, direct interaction between De and DA drugs, drug-metabolizing enzyme-related drug interaction, drug-transporter-related drug interaction, drug-genetic metabolic enzyme-related drug interaction, and drug-genetic transporter-related drug interaction. Drug interactions are predicted through the above five interaction networks.

4. A drug interaction prediction method according to claim 3, characterized in that: A PK interaction network was constructed based on the five drug interactions.

5. A drug interaction prediction method according to claim 2, characterized in that: The relevant data of drugs were obtained using the DrugBank database.

6. A drug interaction prediction method according to claim 4, characterized in that: First, all drugs used in a disease are searched in the DrugBank database according to the type of the disease; And search the Enzymes, Carriers, and Transporters information of all De drugs in the DrugBank database; Then, the pharmacogenetic metabolizing enzymes and pharmacogenetic transporter information of all De drugs were retrieved in the PharmGKB database; Obtain all these drugs interacting with enzymes, carriers and transporters in DrugBank database, and obtain all these drugs interacting with pharmacogenetic metabolizing enzymes and drugs interacting with pharmacogenetic transporters in PharmGKB; Drugs metabolized by the same enzyme as De were denoted as De-Dme, drugs metabolized by the same transporter as De as De as De-Dtr, drugs metabolized by the same pharmacogenetic enzyme as De-Dmeg, and drugs metabolized by the same pharmacogenetic transporter as De-Dtrg.

7. A drug interaction prediction method according to claim 4, characterized in that: Crawl the drug interaction information list of all clinically approved small molecule drugs from the DrugBank database, and obtain the SMILES numbers of all clinically approved small molecule drugs; The acquired small molecule drugs are treated as DA drugs after excluding Dq drugs; All DA drugs that interact with De form DDI pairs with De drugs, and all other DA drugs that do not interact with De form non-DDI pairs with De drugs.

8. A drug interaction prediction method according to claim 7, characterized in that: The structural similarity score is calculated based on the molecular fingerprint of the drug, specifically using Python software to calculate it as a topological fingerprint RDkit; this includes converting the SMILES number of the drug into a topological fingerprint and then calculating the molecular structure similarity.