Method for predicting drug-drug or drug-food interactions using drug structural information

Abstract

The present invention relates to a method for predicting drug-drug interactions and drug-food interactions using drug structural information, and more particularly to a method for predicting the mechanism of action and activity of drug interactions by using interaction prediction results expressed by standardized statements. When the method for predicting drug interactions according to the present invention is used, drug interactions can be predicted quickly and accurately, and in particular, activity information of unknown compounds can also be predicted by expressing prediction results with the aid of statements, and therefore, the method is very useful for developing drugs that exhibit desired activity without causing adverse reactions.

Description

Technical Field

[0001] The present invention relates to a method for predicting drug-drug interaction and drug-food interaction using structural information of drugs, and more particularly to a method for predicting the mechanism of action and drug activity of drug interaction through the prediction results of drug interaction expressed in standardized statements. Background Art

[0002] A drug interaction is a change in the effectiveness of a drug when two or more drugs are given together. Drug interactions include interactions between drugs and ingredients in food.

[0003] Drug interactions are important because they can cause serious adverse drug events (ADEs). It has been reported that drug interactions are responsible for up to 30% of all drug side effects (Tatonetti et al., J. Am. Med. Inf. Assoc. 19:79-85, 2012, Pirmohamed, M. & Orme, M., Chapman & Hall, London, 1998).

[0004] Drug interactions are very important in the pharmaceutical industry because they are known to be the main cause of clinical drug trial failures. Since 1990, 33 drugs have withdrawn from the market due to toxicity issues. Analysis of their respective reasons showed that 8 cases (24%) were withdrawn from the market due to drug interactions (Shah RR et al., Drug Saf. 27: 145-72, 2004). The U.S. FDA suggested in the guidelines updated in 2012 that, in addition to the existing five types of cytochrome P450 (CYP), seven types of cytochrome P450 (CYP) should also be subjected to inhibition studies.

[0005] However, the evaluation of all drug interactions under various conditions is limited due to time and cost issues. To address this problem, methods for predicting drug interactions have been developed, but there are no methods for applying various interactions to drugs and food ingredients.

[0006] Therefore, due to the great investment in developing methods for predicting drug interactions using only the structural information of drugs, the inventors can reveal the mechanism of adverse drug interactions (whose mechanism of action is unknown), and can propose drug combinations that can reduce the side effects caused by drug interactions. In addition, the inventors found that the interaction between food ingredients and drugs can be predicted using only the structural information of drugs, and the prediction results expressed in the form of statements describing the interaction can be used to predict the activity of food ingredients; the inventors have completed the present invention based on this discovery. Summary of the invention

[0007] Technical issues

[0008] An object of the present invention is to provide a method for predicting various types of drug interactions using only the structural similarity information of drugs, and predicting the mechanism of action and drug activity of drug-drug interactions and drug-food interactions by expressing the prediction results through standardized statements.

[0009] Technical Solution

[0010] According to one aspect of the present invention, the above and other objects can be achieved by providing a method for predicting drug interactions using structural information of drugs, the method comprising: calculating the structural similarity features of each of two compounds of interest to predict the interaction between the two by comparing the structural similarity between each of the two compounds and any different predefined compounds; inputting the calculated structural similarity features of each of the two compounds into a training model; predicting the interaction between the two compounds through a prediction model using the input structural similarity features, and outputting the predicted results in the form of a statement describing the interaction; and predicting the mechanism of action of the interaction between the two compounds through the output statement.

[0011] In another aspect of the present invention, a device for predicting drug interactions using drug structural information is provided, the device comprising: a calculation unit for calculating the structural similarity characteristics of two compounds of interest by comparing the structural similarities to predict the interaction between the two; an input unit for combining the calculated structural similarity characteristics of the two compounds and inputting the combined result; an output unit for predicting the interaction between the two compounds using the input structural similarity characteristics and outputting the predicted result in the form of a statement describing the interaction; and a calculation unit for explaining the mechanism of action of the interaction between the two compounds through the output statement.

[0012] In another aspect of the present invention, a method for screening drug combinations with a low probability of adverse interactions using drug structural information is provided, the method comprising: calculating the structural similarity features of each of two compounds of interest to predict the interaction between the two by comparing the structural similarity between each of the two compounds and any different predefined compounds; inputting the calculated structural similarity features of each of the two compounds into a training model; predicting the interaction between the two compounds through a prediction model using the input structural similarity features; and determining the compound combination as having a low probability of adverse interaction when the interaction is not predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a schematic diagram showing the general principles used to predict drug interactions based on structural similarity.

[0014] Figure 2 is a schematic diagram showing the process of expressing the interaction between two drugs based on structural similarity features.

[0015] Figure 3 It is a schematic diagram showing the prediction model (method) proposed by the present invention.

[0016] Figure 4 The optimal structure with the highest accuracy among various deep neural network structures is shown.

[0017] Figure 5 Optimal threshold values ​​for determining the presence or absence of a drug interaction of the present invention are shown.

[0018] Figure 6 The training conditions for deep neural network training of the present invention are shown.

[0019] Figure 7 Shows that the training error decreases as the deep neural network training progresses.

[0020] Figure 8 is a graph showing the accuracy of the prediction method of the present invention.

[0021] Fig. 9 This is a graph for comparing the accuracy between the structural similarity features used in the prediction method of the present invention and the structural similarity features used in other methods.

[0022] Fig.10 It is shown that multiple interactions can be predicted simultaneously based on the predicted results by the prediction method of the present invention.

[0023] Fig.11 It is shown that the prediction method of the present invention reveals the expected mechanism of unknown side effects of drug interactions.

[0024] Fig.12 Schematic diagram of using the prediction method of the present invention to propose a new drug combination that reduces drug side effects caused by drug interactions.

[0025] Fig.13 New drug combinations proposed using the prediction method of the present invention for drug interactions harmful to the human body are shown.

[0026] Fig.14 Examples are shown that may reduce instances of cardiotoxicity caused by drug interactions.

[0027] Fig.15 is a schematic diagram showing a method for exploring the activity of food ingredients using the method for predicting drug interactions of the present invention.

[0028] Fig.16 The results of predicting the activity of food ingredients using the method of predicting drug interactions of the present invention are shown.

[0029] Fig.17 Food ingredients with literature evidence among the food ingredients whose activities were predicted by the drug interaction prediction method are shown. DETAILED DESCRIPTION

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the invention belongs. Generally speaking, the terms used herein are those well known and commonly used in the art.

[0031] In the present invention, a drug interaction prediction model (method) is constructed. When the structural information of two compounds (such as drug-drug and drug-food ingredients) is input through chemical structure expression means, i.e., through chemical structure files such as SMILES (Simplified Molecular Linear Input System), InChI (International Compound Identification), mol and sdf, the drug interaction prediction model (method) can output 86 statements describing the interaction. This drug interaction prediction model can predict drug interactions in the following ways: predict multiple interactions of drug combinations based only on the structural information of the compounds, and output the predicted results in the form of statements. The drug interaction prediction method using the output of standardized (predetermined) statement form can be applied to various compounds including food ingredients, and therefore can also predict the interaction between drugs and food ingredients. In addition, the drug interaction prediction method can use the interaction prediction results to predict the new pharmacological activity of drugs and food ingredients.

[0032] Therefore, in one aspect, the present invention relates to a method for predicting drug interactions using structural information of drugs, the method comprising: calculating the structural similarity features of each of two compounds of interest to predict the interaction between the two by comparing the structural similarity between each of the two compounds and any different predefined compounds; inputting the calculated structural similarity features of each of the two compounds into a training model; predicting the interaction between the two compounds through a prediction model using the input structural similarity features, and outputting the predicted results in the form of a statement describing the interaction; and predicting the mechanism of action of the interaction between the two compounds through the output statement.

[0033] In another aspect, the present invention relates to a device for predicting drug interactions using structural information of drugs, the device comprising: a calculation unit, which is used to calculate the structural similarity characteristics of two compounds of interest by comparing structural similarities to predict the interaction between the two; an input unit, which is used to combine the calculated structural similarity characteristics of the two compounds and input the combined result; an output unit, which is used to predict the interaction between the two compounds using the input structural similarity characteristics, and output the predicted result in the form of a statement describing the interaction; and a calculation unit, which is used to explain the mechanism of action of the interaction between the two compounds through the output statement.

[0034] As used herein, the term "prediction" may refer to the prediction of drug interaction mechanisms whose mechanisms of action are unknown, and may mean the prediction of drug activity or the suggestion of drug combinations with fewer side effects.

[0035] Mechanism of action may refer to the mechanism by which one drug alters the efficacy of another drug.

[0036] The term "prediction model" used in the present invention should be understood to have the same meaning as "prediction method".

[0037] As used herein, the term "predefined compounds" refers to a group of a given number of compounds, and this group of compounds refers to a group of compounds used to calculate structural similarity with an input compound of interest for interaction analysis.

[0038] The compound structure information used as input in the present invention can be input in various forms such as a string (ie, SMILES, InChI) or a file (ie, mol, sdf) that allows a computer to interpret the structure of the compound.

[0039] The "structural similarity" of the present invention can be calculated using various methods capable of quantifying the structural similarity between two compounds. Any structural similarity comparison method can be used, and preferably Tanimoto similarity, cosine similarity, etc. are used.

[0040] The "structural similarity profile" of the present invention means a numerical vector obtained by calculating the structural similarity between the structure of each of two input compounds and the structures of all approved drugs. Therefore, the structural similarity profile of each of the two input compounds is calculated.

[0041] As used herein, the term "structural similarity feature" refers to a numerical vector obtained by comparing the structural similarity (values ​​between 0 and 1) between the structure of a single drug and the structures of all approved drugs. When the total number of approved drugs is 1,000, 1×1000 numerical vectors are obtained from the structural similarity feature of drug a. Since the structures of two compounds are input for drug interaction, a total of 1×1000 plus 1×1000 numerical vectors are created from the two structural similarity features.

[0042] The expression "combining two structural similarity features" means integrating the structural similarity features formed from two compounds. For example, for compounds a and b, a combined structural similarity feature is formed that connects the structural similarity feature of compound a with the structural similarity feature of compound b. In other words, this means that the numerical vector is in the form of 1×2000 (i.e. 1000+1000).

[0043] In one embodiment of the invention, the structural similarity features of all compounds of interest are calculated when determining drug interactions with predefined compounds, and a combination of drug pairs reporting drug interactions is obtained. Then, the statements describing drug interactions are standardized (patterned) and transferred to the classification problem, and the structural similarity features corresponding to the combination of the two drugs are combined to form a new structural similarity feature representing the interaction between the two drugs, and a drug interaction prediction model is constructed using standardized statements describing the interaction.

[0044] In the present invention, the calculation of the structural similarity feature can be represented by the following equation 1 or equation 2.

[0045] [Equation 1]

[0046]

[0047] [Equation 2]

[0048]

[0049] The method for predicting drug interactions of the present invention may be characterized in that the output unit outputs the possibility of each interaction with a probability between 0 and 1, and when the output probability is 0.47 or less, the possibility of determining the drug interaction is low. That is, when the output drug interaction probability is 0.47 or less and therefore no interaction is predicted, it can be interpreted as the combination of the compounds having a lower probability of adverse interactions (fewer side effects).

[0050] In the present invention, the sentences describing the interaction are preferably represented by the 86 sentences listed in Table 1 of Example 1, but are not limited thereto and can be expanded to more than 86 sentences through training.

[0051] Since the number of statements is determined by the data used to train the prediction model, based on the results of the model trained using 86 drug interaction statement data, 86 interactions can be predicted and expressed in statements. When more than 86 interaction statements can be obtained by accumulating more data, more than 86 drug interactions can be predicted and expressed in statements.

[0052] In the present invention, the steps of outputting structural similarity features in the form of statements describing interactions will be described in detail. The training model of the present invention is trained using drug interaction data represented by 86 or more predetermined standardized statements. At this time, when the structural information of two compounds a and b (drugs or food ingredients) that need to be predicted for interaction is input, the value of the possibility of each of the 86 or more interactions is shown to be between 0 and 1. Then, the names of drugs a and b are placed in the statements corresponding to the interactions with one or more specific values ​​to form a new statement.

[0053] That is, in the present invention, the input and output can be characterized as being based on a deep neural network; the input to the deep neural network is a structural similarity curve of two compounds expressing drug interaction; and the output of the deep neural network is a statement describing the interaction between the two compounds.

[0054] The "deep neural network" of the present invention refers to an artificial neural network (ANN) including multiple hidden layers between an input layer and an output layer, and the deep neural network is capable of training various nonlinear relationships including multiple hidden layers.

[0055] The present invention can predict the mechanism of action of drug interactions (whose mechanism of action is still unclear) through the constructed drug interaction prediction model, and can propose the optimal drug combination that exhibits the same drug effect but does not exhibit adverse drug interactions for the interactions that cause adverse drug effects.

[0056] In the present invention, the situation that the interaction is not predicted is a combination of compounds with a low probability of adverse interactions. When the output value of the probability of drug interaction obtained by the output unit is 0.47 or less, it is determined that the possibility of drug interaction is low. That is to say, the output probability of drug interaction is 0.47 or less and therefore the situation that the interaction is not predicted can be considered as a combination of compounds with a low probability of adverse interactions (less side effects).

[0057] The present invention can also predict various drug interactions between various compounds including food ingredients and approved drugs through the constructed drug interaction prediction model; can predict various interactions that may occur when food ingredients and drugs are taken in combination by using the interaction prediction results between food ingredients and drugs of the statement type; and can also predict the activities of various compounds including drugs and food ingredients.

[0058] In the present invention, the compound is preferably a drug or a food ingredient, but is not limited thereto.

[0059] In the present invention, the interaction is preferably action mechanism or activity information, but is not limited thereto.

[0060] The present invention can quickly and accurately predict the interactions occurring in a large number of drugs and drugs, and in a large number of drugs and food ingredients by expressing the results of interaction prediction in sentences, and in particular, it is helpful to understand the mechanism of action of drug interactions, especially unknown mechanisms of action. In addition, the present invention can use the activity information of the drug expressed in sentences to predict the activity of unknown compounds.

[0061] In addition, drugs with fewer side effects can be screened by predicting drug interactions, and drugs that can avoid unexpected side effects can also be designed.

[0062] Therefore, the present invention can detect drug interactions in the early stages of drug development, and thus can quickly identify candidate substances with fewer expected side effects. That is, the present invention can advantageously design drugs with fewer side effects in the structural optimization stage of drug development. The present invention can also avoid drug combinations that are expected to cause serious side effects due to drug interactions. In addition, the present invention can predict the activity information of unknown compounds including drugs, thereby helping to discover candidate drugs with desired activity without causing serious side effects. These advantages are very useful for industrial applications.

[0063] In another aspect, the present invention relates to a method for screening drug combinations with a low probability of adverse interactions using structural information of drugs, the method comprising: calculating the structural similarity features of each of two compounds of interest to predict the interaction between the two by comparing the structural similarity between each of the two compounds and any different predefined compounds; inputting the calculated structural similarity features of each of the two compounds into a training model; predicting the interaction between the two compounds through a prediction model using the input structural similarity features; and determining a compound combination with a low probability of adverse interaction when the interaction is not predicted.

[0064] That is, in a case where it is known that there is an adverse interaction between drug A and drug B, when drug C having the same pharmaceutical effect as drug B is used to replace drug B, the method and device of the present invention predicts that there is no adverse interaction between drug A and drug C.

[0065] In the present invention, the situation that the interaction is not predicted is a combination of compounds with a low probability of adverse interactions. When the output value of the probability of drug interaction obtained by the output unit is 0.47 or less, it is determined that the possibility of drug interaction is low. That is to say, the output probability of drug interaction is 0.47 or less and therefore the situation that the interaction is not predicted can be considered to involve a combination of compounds with a low probability of adverse interactions (less side effects).

[0066] Example

[0067] Hereinafter, the present invention will be described in more detail with reference to examples. However, it will be apparent to those skilled in the art that these examples are provided only to illustrate the present invention and should not be construed as limiting the scope of the present invention.

[0068] Example 1: Development of a drug interaction prediction model

[0069] 1-1: Principles of drug interaction prediction

[0070] The drug interaction prediction method of the present invention is based on the principle of a method using structural similarity. According to this principle, when the interaction between drug a and drug b is known and drug b' has a similar structure to drug b, it is assumed that drug a and drug b' have similar interactions ( Figure 1 ). Since it is difficult to specify a critical (limit) point of a certain structural similarity and the goal is to use various structural information, a structural similarity feature is used.

[0071] 1-2: Calculation of drug structural similarity

[0072] The structural similarity profile of a single drug is an array obtained as a similarity value obtained by comparing the structural similarity between the drug and 2,159 approved drugs ( Figure 2 ). In this embodiment, the structural similarity is calculated using the Tanimoto coefficient described by the following equation 1 to measure the similarity.

[0073] [Equation 1]

[0074]

[0075] Single drug interactions involve two drugs. To represent this, a combination of two structural similarity features obtained from the two drugs is used ( Figure 2 ).

[0076] 1-3: Acquisition of drug interaction training data and expression of data in sentences

[0077] In order to construct a predictive model for drug interactions, drug interaction information was obtained from DrugBank. A total of 191,878 drug combinations were obtained, corresponding to 192,284 drug interactions. Then, 86 statements describing drug interactions were classified (Table 1). The problem of predicting drug interactions can be defined as a classification problem by renumbering the statements describing the interactions.

[0078] [Table 1]

[0079]

[0080]

[0081]

[0082] 1-4: Construction of drug interaction prediction model

[0083] The drug interaction data obtained above were used to construct a drug interaction prediction model. A drug interaction prediction model was established using a mechanical training method based on a deep neural network. The deep neural network was trained using Keras and TensorFlow (Google). As input to the deep neural network, the structural similarity features of the two drugs expressing the drug interaction were used as input, and 86 drug interactions were displayed as results ( Figure 3 ). At this point, since the two drugs may be involved in multiple interactions, a prediction model is implemented using a multiple classification problem.

[0084] The prediction results of the interactions are expressed as statements corresponding to each interaction as defined in Table 1. To optimize the structure of the deep neural network, the structure of the deep neural network with 8 hidden layers with the highest accuracy and 2,048 hidden nodes in each hidden layer was selected. For this purpose, 60% of the data was used for training and 20% of the data was used for evaluation. ( Figure 4 ). In addition, the probability of drug interaction is expressed as a value between 0 and 1, and the optimal threshold that maximizes the accuracy of drug interaction prediction is calculated. When the presence or absence of drug interaction is determined based on a probability of 0.47, the best prediction result is shown ( Figure 5 ). The various parameter conditions used at this time are Figure 6 After the final model was selected, the accuracy of the prediction model was calculated using the remaining 20% ​​of the data.

[0085] The results of confirming the error distribution during the training of the interaction prediction model showed that the training was well performed and there was no overfitting ( Figure 7 The final interaction prediction accuracy was evaluated as 92.4% ( Figure 8 ).

[0086] 1-5: Comparison between structural similarity feature performance

[0087] The methodology based on structural similarity features is a method to represent the drug interactions proposed in the present invention as numerical vectors, which is compared with three methods, namely molecular descriptors, mol2vec (Jaeger S et al., J. Chem. Inf. Model. 58(1): 27-35, 2018), and molecular autoencoders (Gomez-Bombarelli R et al., ACS Cent. Sci. 4(2): 268-276, 2018).

[0088] As a result, the interaction prediction model based on structural similarity features, which requires the most intuitive and simple calculations, showed the most accurate results ( Fig. 9 ).

[0089] Example 2: Prediction of the mechanism of action of unknown interactions using a drug interaction prediction model (method)

[0090] 2-1: Prediction of all possible drug interactions

[0091] Using the drug interaction prediction model (method) constructed in Example 1, 2,329,561 possible drug interactions between all 2,159 approved drugs were predicted. Among them, 487,632 drug combinations had not been reported before, but were predicted to have a high probability of interaction ( Fig.10 ).

[0092] 2-2: Prediction and verification of the mechanism of adverse drug interactions

[0093] In the interaction prediction training data obtained from DrugBank, a total of 60,509 drug combinations have been reported as adverse combinations, but no information on the mechanism of action is available. An analysis was conducted to determine whether all the interaction prediction results obtained in Example 2-1 can explain the mechanism of action of these drug side effects.

[0094] As a result, a total of 12,632 drug combinations were predicted to have one of 86 drug interactions other than side effects ( Fig.11 The predicted drug interactions are mainly related to drug metabolism, that is, the increase or decrease of serum concentration ( Fig.11 ).

[0095] To verify the prediction results, the 12,632 drug combinations were compared with the results from Drugs.com (a database providing drug interaction information). Drugs.com provides information on 182 of the 12,632 drug combinations. Among them, 43 drug interactions were found to be consistent with the interactions predicted in Example 2-1.

[0096] [Table 2]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] Example 3: Using a drug interaction prediction model to propose drug combinations with low probability of adverse interactions

[0108] Using the drug interaction prediction model constructed in Example 1, based on drug interactions that cause known adverse reactions, a drug combination that is expected to exhibit the same drug efficacy without adverse reactions is proposed ( Fig.12 ). Specifically, this was applied to 62,707 drug combinations associated with 14 drug interactions that were found to have significant adverse drug interactions. The 14 drug interactions included increased cardiotoxicity, increased nephrotoxicity, and increased risk of drug side effects ( Fig.13 ).

[0109] As a result, 9,284 of the 62,707 drug combinations were predicted to be substituted with drugs with a lower probability of adverse reactions. Representatively, 168 of the 202 drug combinations reported to have an increased potential for cardiotoxicity were predicted to have alternative drug candidates. In particular, the probability of drug interactions causing cardiotoxicity was predicted to be reduced when the anticancer drug cyclophosphamide was replaced by seven different anticancer drugs (i.e., busulfan, dacarbazine, lomustine, melphalan, pipobroman, temozolomide, or thiotepa). Fig.14 ).

[0110] Example 4: Predicting the interaction between drugs and food ingredients using a drug interaction prediction model

[0111] 4-1: Prediction of drug-food interactions

[0112] The drug interaction prediction model constructed in Example 1 was applied to food ingredients. For this purpose, 1,523 food ingredients were obtained from FooDB (http: / / foodb.ca / ). Using the prediction model, a total of 3,288,157 (2,159 drugs × 1,523 food ingredients) pairs (combinations) of drugs and food ingredients were predicted to have interactions ( Fig.15 ).

[0113] As a result, 690,956 drug interactions were predicted between 319,993 pairs of drugs and food ingredients. The 690,959 drug interactions were all expressed in sentences. Analysis was performed for the cases where the names of food ingredients appeared as topics, i.e., the cases where the food ingredients affected the drug efficacy of the drug. The results showed that 272 food ingredients were predicted to change the drug efficacy of the drug.

[0114] Therefore, in particular, each case where a food ingredient worsens the drug efficacy of a drug is analyzed. There are four types of interactions where a food ingredient reduces the drug efficacy of a drug, namely, reduced absorption of the drug, reduced bioavailability of the drug, reduced serum concentration, and increased metabolism of the drug. A total of 73 food ingredients are predicted to reduce the drug efficacy of at least one of the 430 drugs through one of the four interactions ( Fig.15 ).

[0115] 4-2: Using drug interaction prediction results to predict the activity of food ingredients

[0116] In the prediction results of all drug interactions obtained in Example 4-1 expressed in sentences, the activity information of the food ingredients was analyzed using sentences including information related to the activity of the food ingredients. For example, from the interaction sentence "the drug can reduce the anticoagulant activity of the food ingredient", it can be predicted that the food ingredient has anticoagulant activity. In particular, in order to provide target information of the food ingredients together, food ingredients with structures similar to those of drugs reported to have the same activity (Tanimoto coefficient>0.75) were considered in the analysis.

[0117] As a result, 30 types of activities were predicted in a total of 149 food ingredients ( Fig.16 ). Among the 149 food ingredients, 23 were found to have the same activity information as the results in the existing literature ( Fig.17 ).

[0118] Practicality

[0119] The drug interaction prediction method according to the present invention can quickly and accurately predict drug interactions, and in particular can predict the activity information of unknown compounds by expressing the prediction results in sentences, and is therefore very useful for developing drugs that can exhibit the desired activity without causing side effects.

[0120] Although the specific configuration of the present invention has been described in detail, it will be appreciated by those skilled in the art that this description is provided for illustrative purposes to set forth preferred embodiments and should not be construed as limiting the scope of the present invention. Therefore, the essential scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting drug interactions using drug structural information, the method comprising: The structural similarity profile of each of the two compounds of interest is calculated to predict the interaction between the two compounds by comparing the structural similarity between each of the two compounds of interest and any different predefined compounds using the following Equation 1 or Equation 2: [Equation 1] [Equation 2] combining the calculated structural similarity features of each of the two compounds into a single vectorized data set; Inputting the vectorized data set into a trained deep neural network model, wherein the neural deep network consists of 8 hidden layers and 2048 hidden nodes in each hidden layer; Predicting the interaction between the two compounds using the input vectorized data set through a deep neural network model, and outputting the prediction result in the form of a statement describing the interaction as shown in Table 1 below; [Table 1] as well as The output statement predicts the mechanism of interaction between the two compounds.

2. The method of claim 1, wherein the compound is a pharmaceutical ingredient or a food ingredient.

3. The method of claim 1, wherein the interaction is mechanism of action or activity information. The method of claim 1 , wherein the instances where the interaction is not predicted are combinations of compounds with a low probability of adverse interaction.

5. A device for predicting drug interactions using drug structural information, the device comprising: A calculation unit, which is used to calculate the structural similarity characteristics of two compounds of interest by comparing the structural similarities using the following equation 1 or equation 2 to predict the interaction between the two compounds: [Equation 1] [Equation 2] An input unit, which is used to combine the calculated structural similarity features of the two compounds and input the combined result to a trained deep neural network model, wherein the deep neural network model consists of 8 hidden layers and 2048 hidden nodes in each hidden layer; an output unit, which is used to predict the interaction between the two compounds using the input vectorized data set, and output the predicted result in the form of a statement describing the interaction as shown in the following Table 1; [Table 1] as well as A computing unit is used to interpret the mechanism of action of the interaction between the two compounds by means of the outputted statements.

6. The apparatus of claim 5, wherein the input and the output are based on a deep neural network.

7. The apparatus of claim 6, wherein the input to the deep neural network is a structural similarity feature of two compounds expressing drug interaction, and the output of the deep neural network is a statement describing the interaction between the two compounds.

8. The device of claim 5, wherein the compound is a pharmaceutical ingredient or a food ingredient.

9. The device of claim 5, wherein the interaction is a mechanism of action or activity information.

10. The apparatus of claim 5, wherein the instances where the interaction is not predicted are combinations of compounds with a low probability of adverse interaction.

11. A method for screening drug combinations with low probability of adverse interactions using structural information of drugs, the method comprising: The structural similarity profile of each of the two compounds of interest is calculated to predict the interaction between the two by comparing the structural similarity between each of the two compounds of interest and any different predefined compounds using the following Equation 1 or Equation 2: [Equation 1] [Equation 2] combining the calculated structural similarity features of each of the two compounds into a single vectorized data set; Inputting the vectorized data set into a trained deep neural network model, wherein the deep neural network model consists of 8 hidden layers and 2048 hidden nodes in each hidden layer; Using the vectorized dataset to predict the interaction between the two compounds through the deep neural network model; The prediction results are output in the form of statements describing the interaction as shown in Table 1 below; [Table 1] as well as When such interactions were not predicted, compound combinations were identified as having a low probability of adverse interactions.

12. The method of claim 11, wherein the compound is a pharmaceutical ingredient or a food ingredient.

13. The method of claim 11, wherein the interaction is mechanism of action or activity information.

14. The method of claim 11, wherein a case where the interaction is not predicted has an output value of the structural similarity feature of 0.47 or less.

Images