Method and system for learning a cytotoxicity prediction model utilizing drug response data, and method and system for predicting cytotoxicity using the same

Abstract

The present invention provides a method for predicting cytotoxicity, comprising receiving information related to the molecular structure of a specific drug and the genes of a specific cell line, generating first feature data corresponding to the molecular structure of the specific drug and second feature data corresponding to the genes of the specific cell line based on the received information, generating input data by concatenating the first feature data and the second feature data, and inputting the input data into a pre-trained cytotoxicity prediction model to predict cytotoxicity data indicating the toxicity of the specific drug according to the gene expression of the specific cell line.

Description

Technology Field

[0001] The present invention relates to a method and system for learning a cytotoxicity prediction model utilizing drug response data, and a method and system for predicting cytotoxicity using the same. Background Technology

[0002] Drug toxicity information is essential for selecting drug candidates in the early stages of development, and is particularly crucial for developing effective drugs that are not harmful to the human body.

[0003] Conventionally, methods for evaluating drug toxicity based on specific biomarkers or data related to drug toxicity have been primarily used. For example, hERG (the Human Ether-a-go-go-Related Gene) is a method for evaluating drug toxicity by assessing changes in cardiac hERG channels caused by the drug, LD50 (Lethal Dose 50%) is a method for evaluating drug toxicity by assessing the dose at which 50% of animals die upon drug administration, and DILI (Drug-Induced Liver Injury) is known as a method for evaluating the effect of a drug on the liver through liver enzyme levels, pharmacological mechanisms, and pathological examinations.

[0004] Meanwhile, cell viability indicates the effect of a drug on cells, and the toxicity of a drug is determined by measuring the cell survival rate or cell death rate. In this regard, the drug response value (IC50, The Half Maximal Inhibitory Concern) refers to the concentration of a drug that reduces cell survival rate by 50% and is used to evaluate how potent the drug's toxicity is to cells.

[0005] Cytotoxicity provides a basis for selecting drugs suitable for specific patient groups, and for this purpose, predicting cytotoxicity is known to be important in anticancer drugs and personalized therapies. In particular, since anticancer drugs can affect both normal and cancer cells, there is a need for methods to predict drug toxicity at the cellular level. The problem to be solved

[0006] The present invention relates to a method and system for learning a cytotoxicity prediction model utilizing drug response data, which can effectively identify a toxic response induced by a drug in cells having a specific gene expression state, and a method and system for predicting cytotoxicity using the same.

[0007] In addition, the present invention relates to a method and system for learning a cytotoxicity prediction model utilizing drug response data, which can effectively predict the toxic response of a drug to cells having various genes, and a method and system for predicting cytotoxicity using the same. means of solving the problem

[0008] To solve the problem described above, a cytotoxicity prediction method according to the present invention may include: receiving information related to the molecular structure of a specific drug and the genes of a specific cell line; generating, based on the received information, first feature data corresponding to the molecular structure of the specific drug and second feature data corresponding to the genes of the specific cell line; generating input data by concatenating the first feature data and the second feature data; and inputting the input data into a pre-trained cytotoxicity prediction model to predict cytotoxicity data indicating the toxicity of the specific drug according to the gene expression of the specific cell line.

[0009] In addition, the cytotoxicity prediction system according to the present invention includes: a storage unit in which information related to the molecular structure of a specific drug and the gene of a specific cell line is stored; and a control unit that predicts cytotoxicity data indicating the toxicity of the specific drug according to the gene expression of the specific cell line using a pre-trained cytotoxicity prediction model based on the stored information. The control unit generates first feature data corresponding to the molecular structure of the specific drug and second feature data corresponding to the gene of the specific cell line based on the stored information, generates input data by connecting the first feature data and the second feature data, and predicts the cytotoxicity data by inputting the input data into the pre-trained cytotoxicity prediction model.

[0010] Additionally, the method for training a cytotoxicity prediction model according to the present invention may include: receiving a dataset containing learning feature data related to the molecular structure of a specific drug, learning gene data related to the gene of a specific cell line, and drug response data related to gene expression of the cell line by the drug; generating a first learning feature data for the molecular structure of the drug and a second learning feature data for the gene of the cell line using the dataset; generating ground truth cytotoxicity data representing gene expression of the cell line by the drug using the dataset; and training a cytotoxicity prediction model using the first learning feature data, the second learning feature data, and the ground truth cytotoxicity data to predict the toxicity of the specific drug according to gene expression of the specific cell line in correspondence with information related to the molecular structure of the specific drug and the gene of the specific cell line.

[0011] In addition, the cytotoxicity prediction model learning system according to the present invention includes: a storage unit storing a data set comprising learning characteristic data related to the molecular structure of a specific drug, learning gene data related to the gene of a specific cell line, and drug response data related to gene expression of the cell line by the drug; and a control unit for learning a cytotoxicity prediction model based on the data set. The control unit generates learning first characteristic data regarding the molecular structure of the drug and learning second characteristic data regarding the gene of the cell line using the data set, generates ground truth cytotoxicity data representing gene expression of the cell line by the drug using the data set, and can learn a cytotoxicity prediction model using the learning first characteristic data, learning second characteristic data, and ground truth cytotoxicity data to predict the toxicity of the specific drug according to gene expression of the specific cell line in correspondence with information related to the molecular structure of the specific drug and the gene of the specific cell line. Effects of the invention

[0012] According to various embodiments of the present invention, a method and system for learning a cytotoxicity prediction model utilizing drug response data, and a method and system for predicting cytotoxicity using the same, can effectively identify the toxic response induced by a drug in cells having a specific gene expression state by predicting the toxicity of a drug according to the characteristics of the drug and the gene expression state of the cell using a pre-trained cytotoxicity prediction model.

[0013] In addition, according to various embodiments of the present invention, a method and system for learning a cytotoxicity prediction model utilizing drug response data, and a method and system for predicting cytotoxicity using the same, can effectively predict the toxic response of a drug to cells having various genes by training a cytotoxicity prediction model based on the degree of drug response according to the characteristics of the drug and the gene expression state of the cell. Brief explanation of the drawing

[0014] FIG. 1 illustrates a cytotoxicity prediction system according to the present invention. FIG. 2 illustrates a cytotoxicity prediction model learning system according to the present invention. FIGS. 3 to 5 illustrate an example of a cytotoxicity prediction model according to the present invention. Figure 6 is a flowchart illustrating a method for predicting cytotoxicity according to the present invention. FIG. 7 illustrates an embodiment of generating first feature data. FIG. 8 illustrates an embodiment of generating second feature data. Figure 9 illustrates an example of generating cytotoxicity data. FIG. 10 is a flowchart illustrating a cytotoxicity prediction model learning method according to the present invention. FIG. 11 illustrates an embodiment of generating first feature data for training. FIG. 12 illustrates an example of training a discriminator. FIG. 13 illustrates an example of training a classification unit. Specific details for implementing the invention

[0015] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components are assigned the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not have distinct meanings or roles in themselves. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the present invention.

[0016] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0017] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0018] Singular expressions include plural expressions unless the context clearly indicates otherwise.

[0019] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0020] FIG. 1 illustrates a cytotoxicity prediction system according to the present invention. FIG. 2 illustrates a cytotoxicity prediction model learning system according to the present invention. FIG. 3 to 5 illustrate an embodiment of a cytotoxicity prediction model according to the present invention.

[0021] Referring to FIG. 1, the cytotoxicity prediction system (100) according to the present invention can generate cytotoxicity data (10) based on information related to the molecular structure of a specific drug and the genes of a specific cell line by using a pre-learned cytotoxicity prediction model (1).

[0022] Additionally, with reference to FIG. 2, the cytotoxicity prediction model learning system (200) according to the present invention can learn a cytotoxicity prediction model (1) based on a data set (20) that includes information related to the molecular structure of a predetermined drug and the genes of a predetermined cell line, and information related to the gene expression of the cell line by the drug.

[0023] That is, the cytotoxicity prediction model learning system (200) can train the cytotoxicity prediction model (1) to predict the toxicity of a specific drug according to the gene expression of a specific cell line in response to information related to the molecular structure of a specific drug and a specific cell line.

[0024] Here, a drug is a chemical substance that affects biological functions and may be a chemical substance configured to treat or prevent a disease, or to delay the exacerbation of a disease. Therefore, the drug may have a chemical molecular structure.

[0025] In this regard, the cytotoxicity prediction system (100) and the cytotoxicity prediction model learning system (200) each can obtain characteristic data (11) (or learning characteristic data (21)) which represents the molecular structure of the drug as a string, as information related to the molecular structure of the drug.

[0026] For example, information related to the molecular structure of a drug may include characteristic data (11) (or learning characteristic data (21)) that represents the molecular structure of the drug as a string based on SMILES (Simplified Molecular Input Line Entry System).

[0027] Meanwhile, genes are factors that constitute physical traits and can refer to units of genetic information contained in DNA (Deoxyribonucleic Acid), which makes up the chromosomes of cells. Therefore, genes can appear in different forms depending on the body part.

[0028] In this regard, the cytotoxicity prediction system (100) and the cytotoxicity prediction model learning system (200) can each obtain gene data (13) (or learning gene data (23)) representing the expression state of a gene appearing from a specific cell line as information related to genes.

[0029] Here, a cell line is a collection of cells having the same genetic composition and may include cultured cell lines and somatic cell lines. In this case, a cultured cell line may refer to a cell line artificially cultured in a laboratory, etc., and a somatic cell line may refer to a cell line extracted from the body of a normal person or a patient suffering from a different disease.

[0030] In this regard, cultured cell lines may include cell lines based on GDSC (Genomics of Drug Sensitivity in Cancer), and somatic cell lines may include cell lines based on TCGA (The Cancer Genome Atlas).

[0031] In addition, the cell line may include multiple cell lines corresponding to different body parts. Accordingly, the gene-related information may include information regarding the gene expression status appearing from cultured cell lines or body cell lines, and in this case, the gene-related information may include information regarding the gene expression status appearing from each of the multiple cell lines corresponding to a predetermined body part.

[0032] Meanwhile, gene expression induced by drugs can manifest as changes in the active state of a cell line through the drug's target pathway. In other words, drug-induced gene expression can appear in forms that alter the active state of the cell line, such as by causing cell destruction or regulating the cell division cycle.

[0033] Accordingly, the cytotoxicity prediction model learning system (200) can obtain drug response data (27) indicating the active state of a cell line corresponding to the drug's target pathway as information related to gene expression of a cell line by a drug, or obtain drug response data (27) indicating the active state of cell lines according to the drug's target pathway for a predetermined plurality of cell lines (e.g., a plurality of cell lines corresponding to gene data).

[0034] In this regard, the cytotoxicity data (10) may classify the toxicity of a drug according to the degree to which the gene of a cell line is expressed by the drug. For example, the cytotoxicity data may be implemented to indicate strong toxicity, weak toxicity, and no toxicity, and in one embodiment, the cytotoxicity data (10) may indicate the toxicity of the drug by classifying the response value by a specific drug according to predetermined reference values ​​(e.g., 10 micromol and 30 micromol).

[0035] Meanwhile, the cytotoxicity prediction model (1) may be trained such that when characteristic data (11) related to the molecular structure of a specific drug and gene data (13) indicating the active state of a gene corresponding to a specific cell line are input, cytotoxicity data corresponding to the characteristic data (11) and gene data (13) is output.

[0036] That is, the cytotoxicity prediction model (1) may be trained to predict the degree of toxicity (or effect) that the molecular structure of a specific drug has on the genes of a specific cell line.

[0037] To this end, a cytotoxicity prediction model learning system (200) can train a cytotoxicity prediction model (1) using a data set (20) that includes learning characteristic data (21) and learning gene data (23). At this time, the learning gene data (23) may be labeled with cell line data (25) indicating whether the learning gene data (23) corresponds to a cultured cell line or a body cell line, and the learning characteristic data (21) and the learning gene data (23) may be labeled with drug response data (27) indicating gene expression of the cell line by a predetermined drug.

[0038] Here, a predetermined drug may mean a drug corresponding to the learning characteristic data (21), and the gene expression of the cell line may mean the gene expression of the cell line corresponding to the learning gene data (25).

[0039] In this regard, referring to FIG. 3, the cytotoxicity prediction model (1) may include a first feature extraction unit (3), a second feature extraction unit (5), and a classification unit (7).

[0040] The first feature extraction unit (3) may be trained to generate first feature data (31) corresponding to feature data (12), and this first feature extraction unit (3) may be composed of a foundation model (4) that has been trained in advance based on a large-scale training data set.

[0041] The second feature extraction unit (5) may be trained to generate second feature data (33) corresponding to the genetic data (14), and the second feature extraction unit (5) may include an encoder (6) that generates second feature data (33) from the genetic data (14) and a discriminator that determines whether the genetic data (14) corresponds to a cultured cell line or a body cell line based on the second feature data (33).

[0042] At this time, referring to FIG. 4, the encoder (6a) and discriminator (9a) of the second feature extraction unit (5a) can be learned through adversarial learning. To this end, the second feature extraction unit (6a) can be learned based on learning gene data (24) and cell line data (26) labeled on the learning gene data (24). That is, the cytotoxicity prediction model learning system (200) generates second feature data (33a) for learning from genetic data (24) through the encoder (6a) of the second feature extraction unit (5a), generates discrimination result data (34) corresponding to the second feature data (33a) through the discriminator (9) of the second feature extraction unit (5a), and can train the discriminator (9) based on the loss (37) for the second feature extraction unit (5a) calculated by comparing the cell line data (26) and the discrimination result data (34).

[0043] At this time, referring to FIG. 5, the cytotoxicity prediction model learning system (200) can perform primary learning on the discriminator (9) of the second feature extraction unit (5a), and then perform secondary learning on the encoder (6a) and classification unit (7a) of the second feature extraction unit (5a). That is, when the learning on the discriminator (9) of the second feature extraction unit (5a) is finished (or completed), the cytotoxicity prediction model learning system (200) can perform learning on the encoder (6a) and classification unit (7a) of the second feature extraction unit (5a). The learning on the encoder (6a) and classification unit (7a) of the second feature extraction unit (5a) will be explained in detail below.

[0044] Meanwhile, referring again to FIG. 3, the classification unit (7) may be trained to generate cytotoxicity data (35) based on the first feature data (31) and the second feature data (33). Specifically, the classification unit (7) may be trained to generate cytotoxicity data (35) in response to input data in which the first feature data (31) and the second feature data (33) are combined. For example, the classification unit (7) may be implemented as a Multi Layer Perceptron (MLP) (8). Thus, the classification unit (7) may be trained to output a probability of drug toxicity corresponding to the input data when input data is input. At this time, when the cytotoxicity prediction system (100) (or the cytotoxicity prediction model learning system (200)) outputs a probability for each toxicity corresponding to the cytotoxicity data (35) in response to the input data (or learning input data) from the classification unit (7), the toxicity having the highest probability can be identified as the cytotoxicity data (35).

[0045] To this end, with reference to FIG. 5, when the learning of the discriminator of the second feature extraction unit (5a) is finished (or completed), the cytotoxicity prediction model learning system (200) can generate a first feature data (31a) for learning corresponding to the feature data (22) using the first feature extraction unit (3a), generate a second feature data (33a) for learning corresponding to the gene data (24) and a discrimination result data using the second feature extraction unit (5a), generate a learning input data by connecting the first feature data (31a) and the second feature data (33a) for learning, and generate a learning cytotoxicity data (35a) for learning corresponding to the learning input data using the classification unit (7a).

[0046] Accordingly, the cytotoxicity prediction model learning system (200) can train the encoder (6a) and the classification unit (7a) (or multi-layer perceptron) based on the loss (37a) for the second feature extraction unit calculated based on the comparison of cell line data and discrimination result data, and the loss (38) for the classification unit (7a) calculated based on the comparison of the training cytotoxicity data (35a) and the correct cytotoxicity data (28).

[0047] Meanwhile, referring again to FIG. 1, the cytotoxicity prediction system (100) may include an input unit (110), a storage unit (120), a control unit (130), and an output unit (140).

[0048] The input unit (110) may receive information necessary for the operation of the cytotoxicity prediction system (100) according to the present invention. To this end, the input unit (110) may be connected to a separate input device, server, or external storage device via a wireless or wired network.

[0049] Accordingly, the input unit (110) can receive characteristic data (11) and gene data (13) from a separate input device, server, or external storage device. Additionally, the input unit (110) can receive user input required during the process of predicting cytotoxicity data (10) using the cytotoxicity prediction model (1).

[0050] Additionally, the storage unit (120) may store commands and information necessary for the operation of the cytotoxicity prediction system (100) according to the present invention. For example, the storage unit (120) may store characteristic data (11) and gene data (13) input through the input unit (110).

[0051] Additionally, the storage unit (120) may store a cytotoxicity prediction model (1) learned through the cytotoxicity prediction model learning system (200), and the storage unit (120) may store various information generated during the process of predicting cytotoxicity data (10) from the cytotoxicity prediction model (1) in correspondence with characteristic data (11) and genetic data (13). That is, the storage unit (120) may store first characteristic data, second characteristic data, input data, and cytotoxicity data (10) generated from the cytotoxicity prediction model (1).

[0052] The control unit (130) can control the overall operation of the cytotoxicity prediction system (100) according to the present invention. That is, the control unit (130) can generate cytotoxicity data (10) based on information related to the molecular structure of a specific drug and the genes of a specific cell line by using a pre-learned cytotoxicity prediction model (1).

[0053] Specifically, the control unit (130) can receive information related to the molecular structure of a specific drug and the genes of a specific cell line. That is, the control unit (130) can receive characteristic data (11) related to the molecular structure of a specific drug and gene data (13) indicating the active state of the genes corresponding to the specific cell line.

[0054] Accordingly, the control unit (130) can generate first feature data corresponding to the molecular structure of a specific drug and second feature data corresponding to the gene of a specific cell line based on previously received information.

[0055] That is, the control unit (130) can obtain first feature data by inputting characteristic data (11) related to the molecular structure of a specific drug into the first feature extraction unit implemented in the cytotoxicity prediction model (1).

[0056] Additionally, the control unit (130) can obtain second feature data by inputting gene data (13) representing the active state of a specific cell line's gene into the second feature extraction unit implemented in the cytotoxicity prediction model (1).

[0057] Accordingly, the control unit (130) can generate input data by connecting the first feature data and the second feature data. To this end, the control unit (130) can generate input data representing the molecular structure of a specific drug and the gene activity state of a specific cell line by connecting the first feature data and the second feature data in parallel.

[0058] Furthermore, the control unit (130) inputs input data into a pre-learned cytotoxicity prediction model (1) to predict cytotoxicity data (10) indicating the toxicity of a specific drug according to the gene expression of the cell line.

[0059] That is, the control unit (130) inputs the input data generated prior to the classification unit implemented in the cytotoxicity prediction model (1) to predict cytotoxicity data (10) that distinguishes the degree of gene expression of the cell line by a specific drug.

[0060] The output unit (140) may output information generated by the operation of the cytotoxicity prediction system (100) according to the present invention. To this end, the output unit (140) may be connected to a separate visual output device, server, or external storage device via a wireless or wired network.

[0061] Accordingly, the output unit (140) can output characteristic data (11), gene data (13) and cytotoxicity data (10), etc., so that the user can visually check them through a separate output device, server, or external storage device, etc., and according to the embodiment, the characteristic data, gene data (10) and cytotoxicity data (10), etc. may also be transmitted to another device.

[0062] Additionally, the output unit (140) may output various information generated during the process of predicting cytotoxicity data (10) from the cytotoxicity prediction model (1) in correspondence with characteristic data (11) and gene data (13).

[0063] Meanwhile, referring again to FIG. 2, the cytotoxicity prediction model learning system (200) may include an input unit (210), a storage unit (220), and a control unit (230).

[0064] The input unit (210) may receive information necessary for the operation of the cytotoxicity prediction model learning system (200) according to the present invention. To this end, the input unit (210) may be connected to a separate input device, server, or external storage device via a wireless or wired network.

[0065] Accordingly, the input unit (210) can receive a data set (20) containing learning characteristic data (21), learning gene data (23), cell line data (25), and learning cytotoxicity data (27) from a separate input device, server, or external storage device, etc.

[0066] Additionally, the storage unit (220) may store commands and information necessary for the operation of the cytotoxicity prediction model learning system (200) according to the present invention. For example, the storage unit (220) may store a data set (20) containing learning characteristic data (21), learning gene data (23), cell line data (25), and learning cytotoxicity data (27) input through the input unit (210).

[0067] Additionally, the storage unit (220) may store the cytotoxicity prediction model (1a) and may store various information generated during the process of training the cytotoxicity prediction model (1a) using a data set. That is, the storage unit (220) may store the first feature data for training, the second feature data for training, the discrimination result data, the input data for training, and the cytotoxicity data for training generated from the cytotoxicity prediction model (1a). Additionally, the storage unit (220) may store the loss for the second feature extraction unit and the loss for the classification unit, etc.

[0068] The control unit (230) can control the overall operation of the cytotoxicity prediction model learning system (200) according to the present invention. That is, the control unit (230) can train the cytotoxicity prediction model (1a) based on a data set that includes information related to the molecular structure of a predetermined drug and the genes of a predetermined cell line, and information related to the gene expression of the cell line by the drug.

[0069] In other words, the control unit (230) can train a cytotoxicity prediction model (1a) to predict the toxicity of a specific drug according to the gene expression of a specific cell line in response to information related to the molecular structure of a specific drug and a specific cell line.

[0070] Specifically, the control unit (230) can receive a data set containing information related to the molecular structure of a specific drug and the gene of a specific cell line, and information related to the gene expression of the cell line by the drug.

[0071] That is, the control unit (230) can receive and acquire, as a data set, learning characteristic data (21) related to the molecular structure of a predetermined drug and learning gene data (23) indicating the active state of a gene corresponding to a predetermined cell line, and cell line data (25) indicating whether the learning gene data (23) corresponds to a cultured cell line or a body cell line as labeling data for the learning gene data (23).

[0072] At this time, the control unit (230) can acquire drug response data (27) related to gene expression of a cell line by a drug as a data set, corresponding to the previously acquired learning characteristic data (21) and learning gene data (23).

[0073] Accordingly, the control unit (230) can generate first feature data for learning about the molecular structure of the drug and second feature data for learning about the gene of the cell line using the data set.

[0074] To this end, the control unit (230) can obtain learning feature data (21) related to the molecular structure of a specific drug by inputting the learning feature data (21) into the first feature extraction unit implemented in the cytotoxicity prediction model (1a).

[0075] Additionally, the control unit (230) can acquire learning gene data (23) indicating the active state of a specific cell line's gene by inputting it into the encoder of the second feature extraction unit implemented in the cytotoxicity prediction model (1a), and acquire learning second feature data by inputting the learning second feature data into the discriminator of the second feature extraction unit to acquire discrimination result data indicating whether the learning second feature data is a cultured cell line or a somatic cell line.

[0076] Furthermore, the control unit (230) can calculate a loss for the second feature extraction unit by comparing the cell line data (25) labeled on the learning gene data (23) with the discrimination result data obtained from the encoder of the second feature extraction unit, and can train the discriminator of the second feature extraction unit based on the calculated loss.

[0077] Meanwhile, the control unit (230) can generate correct drug toxicity data representing the gene expression of a cell line by a drug using a data set. That is, the control unit (230) can generate correct drug toxicity data by classifying the drug response data (27) included in the data set according to the toxicity intensity of the drug based on a predetermined reference value.

[0078] Accordingly, the control unit (230) can train a cytotoxicity prediction model (1a) using first feature data for training, second feature data for training, and correct drug toxicity data to predict the toxicity of a specific drug according to gene expression of a specific cell line in response to information related to the molecular structure of a specific drug and the genes of a specific cell line.

[0079] That is, the control unit (230) can generate learning input data by connecting the first feature data for learning and the second feature data for learning, and generate learning cytotoxicity data corresponding to the learning input data by inputting the learning input data into the cytotoxicity prediction model (1a).

[0080] Accordingly, the control unit (230) can derive a loss for the cytotoxicity prediction model (1a) based on a comparison of the previously generated correct cytotoxicity data and the learning cytotoxicity data obtained from the classification unit, and can train the cytotoxicity prediction model (1a) based on the derived loss.

[0081] At this time, the control unit (230) can perform primary learning on the discriminator of the second feature extraction unit, and when the learning is finished (or completed), it can perform secondary learning on the cytotoxicity prediction model (1a).

[0082] That is, the control unit (230) can calculate the loss for the cytotoxicity prediction model (1a) by considering the loss for the second feature extraction unit and the loss for the classification unit together, and can train the encoder and classification unit of the second feature extraction unit based on the calculated loss for the cytotoxicity prediction model (1a).

[0083] Based on the configuration of the cytotoxicity prediction system (100) and the cytotoxicity prediction model learning system (200) examined above, the cytotoxicity prediction method and the cytotoxicity prediction model learning method will be explained in more detail below.

[0084] FIG. 6 is a flowchart illustrating a method for predicting cytotoxicity according to the present invention. FIG. 7 illustrates an embodiment for generating first feature data. FIG. 8 illustrates an embodiment for generating second feature data. FIG. 9 illustrates an embodiment for generating cytotoxicity data. FIG. 10 is a flowchart illustrating a method for training a cytotoxicity prediction model according to the present invention. FIG. 11 illustrates an embodiment for generating first feature data for training. FIG. 12 illustrates an embodiment for training a discriminator. FIG. 13 illustrates an embodiment for training a classification unit.

[0085] Referring to FIG. 6, the cytotoxicity prediction system (100) according to the present invention can receive information related to the molecular structure of a specific drug and the genes of a specific cell line (S100).

[0086] Specifically, the cytotoxicity prediction system (100) can receive characteristic data related to the molecular structure of a specific drug and gene data indicating the active state of a gene corresponding to a specific cell line.

[0087] For example, the cytotoxicity prediction system (100) can obtain characteristic data representing the molecular structure of a specific drug as a string. In this case, the characteristic data may be a string representing the molecular structure of a specific drug based on SMILES.

[0088] Additionally, the cytotoxicity prediction system (100) can acquire genetic data for multiple cell lines corresponding to each of a plurality of predetermined body parts. At this time, the plurality of predetermined body parts may include a body part corresponding to a cell line corresponding to the target pathway of a specific drug, or a body part requiring follow-up observation for a specific disease.

[0089] In one embodiment, the cytotoxicity prediction system (100) can obtain genetic data for cell lines corresponding to at least one body part among the prostate, stomach, brain, skin, bladder, kidney, thyroid, blood, lung, breast, pancreas, head, neck, neuroblastoma, cervical, colon, liver, bone marrow, esophagus, ovary, bone, lymph, and adrenal cortex.

[0090] The cytotoxicity prediction system (100) according to the present invention can generate first feature data corresponding to the molecular structure of a specific drug and second feature data corresponding to the gene of a specific cell line based on previously received information (S200).

[0091] Specifically, the cytotoxicity prediction system (100) can obtain first feature data by inputting characteristic data related to the molecular structure of a specific drug into a first feature extraction unit implemented in a cytotoxicity prediction model.

[0092] For example, with reference to FIG. 7, the cytotoxicity prediction system (100) can obtain first feature data (51) for the molecular structure of a specific drug by inputting feature data (11) representing the molecular structure of a specific drug as a string based on SMILES into the first feature extraction unit (41).

[0093] Here, the first feature extraction unit (41) may consist of a basic model that has been pre-trained based on a large-scale learning data set. Accordingly, the cytotoxicity prediction system (100) can obtain first feature data (51) for the molecular structure of a specific drug by inputting feature data (11) into the basic model.

[0094] Furthermore, the cytotoxicity prediction system (100) can obtain second feature data by inputting gene data representing the active state of a specific cell line's gene into a second feature extraction unit implemented in a cytotoxicity prediction model.

[0095] For example, with reference to FIG. 8, the cytotoxicity prediction system (100) can obtain second feature data (53) for genes by inputting gene data (13) for multiple cell lines corresponding to each of the multiple body parts that are predetermined into the second feature extraction unit (42).

[0096] Here, the second feature extraction unit (42) may be trained to extract second feature data (53) corresponding to the gene data (13) and to determine whether the gene data (13) is a cultured cell line or a body cell line based on the extracted second feature data (53).

[0097] Accordingly, the second feature extraction unit (42) may include an encoder (43) that generates second feature data (53) corresponding to the gene data (13) when the gene data (13) is input, and a discriminator (not shown) that determines whether the gene data (13) corresponding to the second feature data (53) corresponds to a cultured cell line or a body cell line when the second feature data (53) is input.

[0098] Here, the discriminator (not shown) of the second feature extraction unit (42) may be used only during the process of learning the cytotoxicity prediction model, and therefore, the discriminator (not shown) of the second feature extraction unit (42) may be omitted in the cytotoxicity prediction model used in the cytotoxicity prediction system (100).

[0099] Meanwhile, referring again to FIG. 6, the cytotoxicity prediction system (100) according to the present invention can generate input data by connecting first feature data and second feature data (S300).

[0100] Specifically, the cytotoxicity prediction system (100) can generate input data representing the molecular structure of a specific drug and the gene activity status of a specific cell line by connecting the first feature data and the second feature data in parallel.

[0101] For example, the cytotoxicity prediction system (100) can generate input data by connecting second feature data in a row direction (or column direction) to first feature data. That is, the size of the input data may be the sum of the size of the first feature data and the size of the second feature data.

[0102] In one embodiment, the first feature data can be implemented as a 128-dimensional embedding vector, and the second feature data can be implemented as a 128-dimensional embedding vector. In this case, the input data can be implemented as a 256-dimensional embedding vector by concatenating the first feature data and the second feature data.

[0103] The cytotoxicity prediction system (100) according to the present invention can predict cytotoxicity data indicating the toxicity of a specific drug according to gene expression of a cell line by inputting input data into a pre-trained cytotoxicity prediction model (S400).

[0104] Specifically, the cytotoxicity prediction system (100) can predict cytotoxicity data that distinguishes the degree to which the gene of a cell line is expressed by a specific drug by inputting input data generated prior to the classification unit implemented in the cytotoxicity prediction model.

[0105] For example, with reference to FIG. 9, the cytotoxicity prediction system (100) can predict cytotoxicity data (50) by inputting input data (54) in which the first feature data (51) and the second feature data (53) are connected to the classification unit (45). Here, the classification unit (45) may be trained to output a value corresponding to the degree of gene expression of a cell line by a specific drug as cytotoxicity data (50) when the input data (54) is input.

[0106] That is, the cytotoxicity data (50) may include values ​​indicating strong toxicity, weak toxicity, and no toxicity depending on the degree to which the gene of the cell line is expressed by a specific drug.

[0107] Accordingly, when cytotoxicity data (50) is output from the classification unit (45), the cytotoxicity prediction system (100) can determine whether the toxicity of a specific drug is strong toxicity, weak toxicity, or non-toxicity based on the value of the output cytotoxicity data (50), and provide the determined result.

[0108] Meanwhile, referring to FIG. 10, the cytotoxicity prediction model learning system (200) according to the present invention may receive a data set including learning characteristic data related to the molecular structure of a predetermined drug, learning gene data related to the gene of a predetermined cell line, and drug response data related to gene expression of the cell line by said drug (S500).

[0109] Specifically, the cytotoxicity prediction model learning system (200) can receive and acquire, as a data set, learning characteristic data related to the molecular structure of a predetermined drug and learning gene data indicating the active state of a gene corresponding to a predetermined cell line, along with cell line data indicating whether the learning gene data corresponds to a cultured cell line or a somatic cell line as labeling data for the learning gene data.

[0110] For example, a cytotoxicity prediction model learning system (200) can acquire training characteristic data representing the molecular structure of a specific drug as a string. In this case, the training characteristic data may represent the molecular structure of a specific drug as a string based on SMILES.

[0111] Additionally, the cytotoxicity prediction system (100) can acquire training gene data for multiple cell lines corresponding to each of a plurality of predetermined body parts. At this time, the plurality of predetermined body parts may include a body part corresponding to a cell line corresponding to the target pathway of a specific drug, or a body part requiring follow-up observation for a specific disease.

[0112] In one embodiment, the cytotoxicity prediction system (100) can acquire training gene data for cell lines corresponding to at least one body part among the prostate, stomach, brain, skin, bladder, kidney, thyroid, blood, lung, breast, pancreas, head, neck, neuroblastoma, cervical, colon, liver, bone marrow, esophagus, ovary, bone, lymph, and adrenal cortex.

[0113] At this time, the cytotoxicity prediction model learning system (200) can acquire learning gene data for each of the cultured cell line and the somatic cell line, and in this case, the cytotoxicity prediction model learning system (200) can acquire cell line data that indicates whether a specific learning gene data is a cultured cell line or a somatic cell line, as data labeled on the learning gene data.

[0114] In one embodiment, the cytotoxicity prediction model learning system (200) can acquire first learning gene data corresponding to a somatic cell line based on TCGA and second learning gene data corresponding to a cultured cell line based on GDSC. In this case, the cytotoxicity prediction model learning system (200) can acquire first cell line data as labeling data for the first learning gene data and acquire second cell line data as labeling data for the second learning gene data.

[0115] Meanwhile, the cytotoxicity prediction model learning system (200) can acquire characteristic data for learning for multiple different drugs and can also acquire genetic data for learning for each of multiple cell lines having different diseases.

[0116] Furthermore, the cytotoxicity prediction model learning system (200) can acquire drug response data related to gene expression of cell lines by a drug as a data set in correspondence with the previously acquired training characteristic data and training gene data.

[0117] For example, the cytotoxicity prediction model learning system (200) can obtain drug response data showing gene expression of a cell line according to learning gene data by means of a predetermined drug according to learning characteristic data.

[0118] That is, the cytotoxicity prediction model learning system (200) can acquire drug response data indicating a change in the active state of a cell line corresponding to the learning gene data according to the target pathway of the drug corresponding to the learning characteristic data.

[0119] The cytotoxicity prediction model learning system (200) according to the present invention can generate first feature data for learning about the molecular structure of the drug and second feature data for learning about the gene of the cell line using a data set (S600).

[0120] Specifically, the cytotoxicity prediction model learning system (200) can obtain first feature data for learning by inputting learning feature data related to the molecular structure of a specific drug into a first feature extraction unit implemented in the cytotoxicity prediction model.

[0121] For example, with reference to FIG. 11, a cytotoxicity prediction model learning system (200) can obtain first feature data (71) for learning the molecular structure of a specific drug by inputting learning feature data (21) representing the molecular structure of a specific drug as a string based on SMILES into a first feature extraction unit (61).

[0122] Here, the first feature extraction unit (61) may consist of a basic model that has been pre-trained based on a large-scale learning data set. Accordingly, the cytotoxicity prediction model learning system (200) can obtain the first feature data (71) for learning the molecular structure of a specific drug by inputting the learning feature data (21) into the basic model.

[0123] Furthermore, the cytotoxicity prediction model learning system (200) can acquire learning gene data representing the active state of a specific cell line's gene by inputting it into the encoder of the second feature extraction unit implemented in the cytotoxicity prediction model, and acquire learning second feature data by inputting the learning second feature data into the discriminator of the second feature extraction unit to obtain discrimination result data indicating whether the learning second feature data is a cultured cell line or a somatic cell line.

[0124] For example, with reference to FIG. 12, the second feature extraction unit (62) may be implemented to extract a second feature data (73) for learning that corresponds to the genetic data (23) for learning, and to determine whether the genetic data (23) for learning corresponds to a cultured cell line or a body cell line based on the extracted second feature data (73).

[0125] Accordingly, the second feature extraction unit (62) may include an encoder (63) that generates second feature data (73) for learning corresponding to the second feature data (23) when the second feature data (73) for learning is input, and a discriminator (64) that determines whether the second feature data (23) for learning corresponding to the second feature data (73) corresponds to a cultured cell line or a body cell line when the second feature data (73) for learning is input.

[0126] Through this, the cytotoxicity prediction model learning system (200) can input learning gene data (23) for multiple cell lines corresponding to each of the predetermined multiple body parts into the encoder (63) of the second feature extraction unit (62) to obtain learning second feature data (73) for the corresponding learning gene data (23).

[0127] Additionally, the cytotoxicity prediction model learning system (200) inputs the second feature data (73) for learning into the discriminator (64) of the second feature extraction unit (62), and can obtain discrimination result data (74) indicating whether the learning gene data (23) input into the encoder (63) of the second feature extraction unit (62) corresponds to a cultured cell line or a body cell line.

[0128] Furthermore, the cytotoxicity prediction model learning system (200) can calculate a loss (81) for the second feature extraction unit (62) by comparing the cell line data (25) labeled in the learning gene data (23) with the discrimination result data (74) obtained from the encoder (63) of the second feature extraction unit (62), and can train the discriminator (64) of the second feature extraction unit (62) based on the calculated loss (81).

[0129] For example, the cytotoxicity prediction model learning system (200) can calculate a loss (81) for the second feature extraction unit (62) based on the difference between the cell line data (25) labeled in the learning gene data (23) and the discrimination result data (74) output from the discriminator (64), and can calculate the slope of the previously calculated loss with respect to the weights provided in the discriminator (64).

[0130] Accordingly, the cytotoxicity prediction model learning system (200) can train the discriminator (64) by correcting the weights provided in the discriminator (64) so ​​that the calculated slope becomes a minimum value. Through this, the cytotoxicity prediction model learning system (200) can terminate the training of the discriminator (64) when the accuracy of the discrimination result data (74) output from the discriminator (64) satisfies a predetermined threshold based on the cell line data (25) labeled in the training gene data (23), or when the number of training cycles (e.g., epoch) for the discriminator (64) satisfies a predetermined standard number.

[0131] Referring again to FIG. 10, the cytotoxicity prediction model learning system (200) according to the present invention can generate ground truth drug toxicity data representing gene expression of a cell line by a drug using a dataset (S700).

[0132] Specifically, the cytotoxicity prediction model learning system (200) can generate correct drug toxicity data by classifying drug response data included in the dataset according to the toxicity intensity of the drug based on a predetermined reference value.

[0133] For example, ground truth cytotoxicity data may include values ​​representing strong toxicity, weak toxicity, and non-toxicity, respectively, depending on the toxicity intensity of the drug. In this case, the cytotoxicity prediction model learning system (200) can generate ground truth cytotoxicity data by classifying and adjusting drug response data into values ​​representing strong toxicity, weak toxicity, and non-toxicity, respectively, according to predetermined reference values.

[0134] In one embodiment, a predetermined reference value may be set to 10 micromol and 30 micromol. In this case, the cytotoxicity prediction model learning system (200) can generate correct cytotoxicity data by adjusting the value of the drug response data to a value corresponding to strong toxicity when the value of the gene expression level of the cell line appearing from the drug response data is less than 10 micromol, adjusting the value of the drug response data to a value corresponding to weak toxicity when the value of the gene expression level of the cell line appearing from the drug response data falls between 10 micromol and 30 micromol, and adjusting the value of the drug response data to a value corresponding to non-toxicity when the value of the gene expression level of the cell line appearing from the drug response data is 30 micromol or more.

[0135] The cytotoxicity prediction model learning system (200) according to the present invention can train a cytotoxicity prediction model using first feature data for training, second feature data for training, and correct drug toxicity data to predict the toxicity of a specific drug according to gene expression of a specific cell line in response to information related to the molecular structure of a specific drug and the genes of a specific cell line (S800).

[0136] Specifically, the cytotoxicity prediction model learning system (200) can generate learning input data by connecting learning first feature data and learning second feature data, and generate learning cytotoxicity data corresponding to the learning input data by inputting the learning input data into the cytotoxicity prediction model.

[0137] For example, with reference to FIG. 13, a cytotoxicity prediction model learning system (200) can generate learning input data (75) representing the molecular structure of a drug and the gene activity state of a cell line by connecting the first feature data (71) and the second feature data (73) for learning in parallel.

[0138] Accordingly, the cytotoxicity prediction model learning system (200) can generate learning cytotoxicity data (77) by inputting the learning input data (75) generated prior to the classification unit (65) implemented in the cytotoxicity prediction model. At this time, the learning cytotoxicity data (77) can be output as a value belonging to a predetermined range (e.g., 0 to 1).

[0139] Furthermore, the cytotoxicity prediction model learning system (200) can derive a loss (85) for the cytotoxicity prediction model based on a comparison of the previously generated correct cytotoxicity data (29) and the learning cytotoxicity data (77) obtained from the classification unit, and can train the cytotoxicity prediction model based on the derived loss (85).

[0140] For example, the cytotoxicity prediction model learning system (200) can calculate a loss (83) for the classification unit (65) based on the difference between the correct cytotoxicity data (29) and the learning cytotoxicity data (77). At this time, the cytotoxicity prediction model learning system (200) can perform primary learning on the discriminator of the second feature extraction unit (62), and perform secondary learning on the cytotoxicity prediction model when the learning is finished (or completed).

[0141] To this end, the cytotoxicity prediction model learning system (200) can calculate a loss (85) for the cytotoxicity prediction model by considering the loss (81) for the second feature extraction unit (62) and the loss (83) for the classification unit (65) together, and can train the encoder of the second feature extraction unit (62) and the classification unit (65) based on the calculated loss (85) for the cytotoxicity prediction model.

[0142] That is, the cytotoxicity prediction model learning system (200) can generate second feature data (73) for learning from genetic data for learning through the encoder of the second feature extraction unit (62) when the learning of the discriminator of the second feature extraction unit (62) is terminated (or completed), generate discrimination result data corresponding to the second feature data (73) for learning through the discriminator of the second feature extraction unit (62), and calculate a loss (81) for the second feature extraction unit (62) by comparing the cell line data labeled on the genetic data for learning with the discrimination result data.

[0143] Additionally, the cytotoxicity prediction model learning system (200) can generate first learning feature data (71) from learning feature data through a first feature extraction unit, generate learning input data (75) by connecting the previously generated second learning feature data (73) and the first learning feature data (71), and generate learning cytotoxicity data (77) corresponding to the learning input data (75) through a classification unit (65).

[0144] Accordingly, the cytotoxicity prediction model learning system (200) calculates a loss (83) for the classification unit (65) based on the difference between the correct cytotoxicity data (29) and the training cytotoxicity data (77), calculates a loss (85) for the cytotoxicity prediction model by combining the previously calculated loss (81) for the second feature extraction unit (62) and the loss (83) for the classification unit (65), and can train the encoder of the second feature extraction unit (62) and the classification unit (65) based on the loss (85) for the cytotoxicity prediction model.

[0145] According to an embodiment, the cytotoxicity prediction model learning system (200) can calculate the loss (85) for the cytotoxicity prediction model by summing the loss (81) for the second feature extraction unit (62) and the loss (83) for the classification unit (65).

[0146] Accordingly, the cytotoxicity prediction model learning system (200) can calculate the slope of the loss (85) for the cytotoxicity prediction model with respect to the weights provided in the encoder of the second feature extraction unit (62), and train the encoder by correcting the weights provided in the encoder so that the calculated slope becomes a minimum value, and can train the classification unit (65) by calculating the slope of the loss (85) for the cytotoxicity prediction model with respect to the weights provided in the classification unit (65), and train the classification unit (65) by correcting the weights provided in the encoder so that the calculated slope becomes a minimum value.

[0147] Through this, the cytotoxicity prediction model learning system (200) can terminate the learning of the cytotoxicity prediction model when, based on the data set, the accuracy of the cytotoxicity data output from the cytotoxicity prediction model satisfies a predetermined threshold, or when the number of learning cycles (e.g., epochs) of the cytotoxicity prediction model satisfies a predetermined standard number.

[0148] Through the above configurations, the cytotoxicity prediction system (100) and the cytotoxicity prediction model learning system (200) according to the present invention can effectively identify the toxic response induced by a drug in cells having a specific gene expression state by predicting the toxicity of the drug according to the characteristics of the drug and the gene expression state of the cell using a pre-learned cytotoxicity prediction model.

[0149] In addition, the cytotoxicity prediction system (100) and the cytotoxicity prediction model learning system (200) according to the present invention can effectively predict the toxic response of a drug to cells having various genes by learning a cytotoxicity prediction model based on the characteristics of the drug and the degree of response of the drug according to the gene expression state of the cell.

[0150] Furthermore, the present invention described above can be implemented as a program that is executed by one or more processes in an electronic device and stored on a computer-readable recording medium.

[0151] Accordingly, the present invention can be implemented as computer-readable code or instructions on a medium on which a program is recorded. That is, various control methods according to the present invention can be provided in the form of integrated or individual programs.

[0152] Meanwhile, computer-readable media include all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0153] Furthermore, the computer-readable medium may be a server or cloud storage that includes a storage and is accessible to an electronic device via communication. In this case, the computer may download the program according to the present invention from the server or cloud storage via wired or wireless communication.

[0154] Furthermore, in the present invention, the computer described above is an electronic device equipped with a processor, namely a CPU (Central Processing Unit), and no special limitations are placed on its type.

[0155] Meanwhile, the above detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

Claim 1 A method for predicting cytotoxicity, comprising: receiving information related to the molecular structure of a specific drug and the genes of a specific cell line; generating, based on the received information, first feature data corresponding to the molecular structure of the specific drug and second feature data corresponding to the genes of the specific cell line; generating input data by connecting the first feature data and the second feature data; and inputting the input data into a pre-trained cytotoxicity prediction model to predict cytotoxicity data indicating the toxicity of the specific drug according to the gene expression of the specific cell line. Claim 2 A cytotoxicity prediction method according to claim 1, wherein the cytotoxicity prediction model is trained according to a cytotoxicity prediction model training method, and the cytotoxicity prediction model training method comprises: receiving a data set including training feature data related to the molecular structure of a specific drug, training gene data related to the gene of a specific cell line, and drug response data related to gene expression of the cell line by the drug; generating a first training feature data for the molecular structure of the drug and a second training feature data for the gene of the cell line using the data set; generating ground truth cytotoxicity data representing gene expression of the cell line by the drug using the data set; and training a cytotoxicity prediction model using the first training feature data, the second training feature data, and the ground truth cytotoxicity data to predict the toxicity of the specific drug according to gene expression of the specific cell line in correspondence with information related to the molecular structure of the specific drug and the gene of the specific cell line. Claim 3 A cytotoxicity prediction system comprising: a storage unit storing information related to the molecular structure of a specific drug and the genes of a specific cell line; and a control unit that predicts cytotoxicity data indicating the toxicity of the specific drug according to the gene expression of the specific cell line using a pre-trained cytotoxicity prediction model based on the stored information, wherein the control unit generates first feature data corresponding to the molecular structure of the specific drug and second feature data corresponding to the genes of the specific cell line based on the stored information, generates input data by connecting the first feature data and the second feature data, and predicts the cytotoxicity data by inputting the input data into the pre-trained cytotoxicity prediction model. Claim 4 A method for training a cytotoxicity prediction model, comprising: receiving a dataset containing learning feature data related to the molecular structure of a specific drug, learning gene data related to the gene of a specific cell line, and drug response data related to gene expression of the cell line by the drug; generating a first learning feature data for the molecular structure of the drug and a second learning feature data for the gene of the cell line using the dataset; generating ground truth cytotoxicity data representing gene expression of the cell line by the drug using the dataset; and training a cytotoxicity prediction model using the first learning feature data, the second learning feature data, and the ground truth cytotoxicity data to predict the toxicity of the specific drug according to gene expression of the specific cell line in correspondence with information related to the molecular structure of the specific drug and the gene of the specific cell line. Claim 5 In claim 4, the cytotoxicity prediction model comprises: a first feature extraction unit that generates the first learning feature data corresponding to the learning feature data; a second feature extraction unit including an encoder that generates the second learning feature data from the learning gene data and a discriminator that determines from the second learning feature data whether the learning gene data corresponds to a cultured cell line or a somatic cell line; and a classification unit that generates cytotoxicity data based on the first learning feature data and the second learning feature data. Claim 6 A method for training a cytotoxicity prediction model according to claim 5, wherein the step of training the cytotoxicity prediction model comprises: a step of performing primary training on the discriminator of the second feature extraction unit; and a step of performing secondary training on the cytotoxicity prediction model after the primary training is completed. Claim 7 In claim 6, the step of performing the primary learning comprises: a step of inputting the learning gene data into the encoder of the second feature extraction unit to obtain the second feature data for learning; a step of inputting the second feature data for learning into the discriminator of the second feature extraction unit to obtain discrimination result data indicating whether the second feature data for learning is a cultured cell line or a somatic cell line; a step of calculating a loss for the second feature extraction unit by comparing the cell line data labeled on the learning gene data with the discrimination result data obtained from the encoder of the second feature extraction unit; and a step of training the discriminator of the second feature extraction unit based on the calculated loss. Claim 8 In claim 6, the step of performing the secondary learning comprises: a step of generating the learning input data by connecting the learning first feature data and the learning second feature data; a step of generating the learning cytotoxicity data corresponding to the learning input data by inputting the learning input data into the cytotoxicity prediction model; a step of deriving a loss for the cytotoxicity prediction model based on a comparison of the ground truth cytotoxicity data and the learning cytotoxicity data obtained from the classification unit; and a step of training the cytotoxicity prediction model based on the derived loss. Claim 9 In claim 8, the step of deriving a loss for the cytotoxicity prediction model comprises: a step of calculating a loss for the classification unit based on the difference between the ground truth cytotoxicity data and the training cytotoxicity data; and a step of calculating a loss for the cytotoxicity prediction model by combining the calculated loss for the second feature extraction unit and the loss for the classification unit; and the step of training the cytotoxicity prediction model based on the derived loss comprises a step of training the encoder of the second feature extraction unit and the classification unit based on the loss for the cytotoxicity prediction model. Claim 10 A cytotoxicity prediction model learning system comprising: a storage unit storing a data set including learning feature data related to the molecular structure of a specific drug, learning gene data related to the gene of a specific cell line, and drug response data related to gene expression of the cell line by the drug; and a control unit for learning a cytotoxicity prediction model based on the data set, wherein the control unit generates learning first feature data regarding the molecular structure of the drug and learning second feature data regarding the gene of the cell line using the data set, generates ground truth cytotoxicity data representing gene expression of the cell line by the drug using the data set, and learns a cytotoxicity prediction model using the learning first feature data, learning second feature data, and ground truth cytotoxicity data to predict the toxicity of the specific drug according to gene expression of the specific cell line corresponding to information related to the molecular structure of the specific drug and the gene of the specific cell line.

Images