Method for estimating drug efficacy, method for generating drug efficacy estimation model, drug efficacy estimation system, control program, recording medium, and trained model

Abstract

To precisely estimate the treatment effect of a drug by using a medical image and information of genetic expression.SOLUTION: The method for estimating a drug efficiency includes: an acquisition step of acquiring a first medical image of a lesion area of a first patient and first expression information regarding a genetic expression in the lesion area of the first patient; and an estimation step of inputting the first medical image and the first expression information into a trained model trained by using second drug information of a second drug input into a second patient and second effect information showing the effect of the treatment with the second drug as purpose variables, by using a second medical image and second expression information of the second patient as explanatory variables, and outputting a first drug which can be input into the first patient and the estimated effect of the treatment.SELECTED DRAWING: Figure 3

Description

Technical Field

[0001] The present invention relates to a method for estimating drug efficacy, a method for generating a drug efficacy estimation model, a drug efficacy estimation system, a control program, a recording medium, and a learned model.

Background Art

[0002] In the field of medicine, Radiomics technology, which estimates the presence or absence of a disease or the prognosis of treatment using a large number of feature quantities extracted from medical images such as CT (Computed Tomography) images, has attracted attention. For example, Non-Patent Document 1 discloses that the Radiomics technology estimates the genetic properties of cancer from medical images and predicts the effect of radiotherapy for the cancer.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] As shown in Non-Patent Document 1, the relationship between the feature quantities included in medical images and gene mutations has been discussed. However, the correlation between the gene expression pattern, which is considered to be more directly related to the biological phenotype of the lesion site than gene mutations, and the feature quantities included in medical images has not been conventionally studied. Therefore, it has not been done to estimate the treatment effect of a drug by associating the feature quantities of a medical image with both the gene expression pattern of the lesion site and the treatment effect data of the drug.

[0005] One aspect of the present invention aims to realize a drug efficacy estimation method or the like that can accurately estimate the treatment effect of a drug using medical images and gene expression information.

Means for Solving the Problem

[0006] To solve the above problems, a method for estimating drug efficacy according to one aspect of the present invention includes an acquisition step of acquiring a first medical image obtained by imaging a lesion site related to a target disease of a first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image; and inputting the first medical image and the first expression information into a learned model that has been learned using learning data including, as explanatory variables, a second medical image obtained by imaging the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and including, as objective variables, second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease, and causing the learned model to output first drug information indicating a first drug administrable to the first patient and first effect information indicating the estimated therapeutic effect of the first drug on the target disease.

[0007] To solve the above problems, a method for generating a drug efficacy estimation model according to one aspect of the present invention generates a learned model by machine learning using learning data including, as explanatory variables, a second medical image obtained by imaging a lesion site related to a target disease of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and including, as objective variables, second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease.

[0008] To solve the above problems, a drug efficacy estimation system according to an aspect of the present invention includes an acquisition unit that acquires a first medical image of a lesion site related to a target disease of a first patient and first expression information related to gene expression in the lesion site corresponding to the first medical image, and a second medical image of the lesion site of a second patient, and second expression information related to gene expression in the lesion site corresponding to the second medical image are used as explanatory variables, and a learned model that has been learned using learning data including second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease as target variables, inputs the first medical image and the first expression information, and outputs, to the learned model, first drug information indicating a first drug that can be administered to the first patient and first effect information indicating the estimated therapeutic effect of the first drug on the target disease.

[0009] To solve the above problems, a learned model according to an aspect of the present invention is a learned model for estimating a first drug that can be administered to a first patient and the estimated therapeutic effect of the first drug on the target disease based on a first medical image of a lesion site related to a target disease of the first patient and first expression information related to gene expression in the lesion site corresponding to the first medical image, and is obtained by machine learning using learning data including a second medical image of the lesion site of a second patient and second expression information related to gene expression in the lesion site corresponding to the second medical image as explanatory variables, and second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease as target variables, and is for causing a computer to function so as to output first drug information indicating the first drug and first effect information indicating the estimated therapeutic effect of the first drug from the first medical image and the first expression information.

[0010] The drug efficacy estimation system according to each aspect of the present invention may be implemented by a computer. In this case, a computer control program for causing the computer to operate as each part (software element) included in the drug efficacy estimation system to implement the drug efficacy estimation system on the computer, and a computer-readable recording medium on which the program is recorded also fall within the scope of the present invention.

Effect of the Invention

[0011] According to one aspect of the present invention, it is possible to realize a drug efficacy estimation method or the like capable of accurately estimating the therapeutic effect of a drug using medical image and gene expression information.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Modes for Carrying Out the Invention

[0013] 〔Embodiment 1〕 <Schematic Configuration of Drug Efficacy Estimation System> Hereinafter, the schematic configuration of the drug efficacy estimation system 100 will be described with reference to FIG. 1. FIG. 1 is a block diagram showing an example of the schematic configuration of the drug efficacy estimation system 100.

[0014] As shown in FIG. 1, the drug efficacy estimation system 100 may include a drug efficacy estimation device 1 and a display device 4. FIG. 1 shows a drug efficacy estimation system 100 including one drug efficacy estimation device 1 and one display device 4 each. However, the configuration of the drug efficacy estimation system 100 is not limited to this. For example, the drug efficacy estimation system 100 may not include the display device 4, or on the other hand, may include a plurality of display devices 4.

[0015] In the drug efficacy estimation system 100, the drug efficacy estimation device 1 and the display device 4 are communicably connected to each other. The drug efficacy estimation device 1 and the display device 4 may be directly connected by wire or wirelessly, or may be connected via a communication network. The mode of the communication network is not limited, and it may be a local area network (LAN) or the Internet.

[0016] The drug efficacy estimation device 1 is a device that estimates the therapeutic effect of a drug on a target disease using a medical image of a lesion site and expression information regarding the gene expression of the lesion site in a first patient who is the patient to be estimated. The estimation result by the drug efficacy estimation device 1 may be transmitted from the drug efficacy estimation device 1 to the display device 4.

[0017] The display device 4 may be a computer, a smartphone, a tablet terminal, etc. used by a user who uses the drug efficacy estimation system 100. Note that FIG. 1 shows a drug efficacy estimation system 100 in which the display device 4 is separate from the drug efficacy estimation device 1. However, the configuration of the drug efficacy estimation system 100 is not limited to this. For example, the display device 4 may be a device integrated with the drug efficacy estimation device 1, and in this case, the display device 4 may be a display unit (such as a display) included in the drug efficacy estimation device 1.

[0018] <Configuration of the drug efficacy estimation device 1> Next, the configuration of the drug efficacy estimation device 1 will be described. The drug efficacy estimation device 1 includes a control unit 10, a storage unit 20, and an input unit 30.

[0019] The storage unit 20 may be, for example, an HDD (Hard Disk Drive) or an SSD (Solid State Drive). The storage unit 20 stores the information transmitted from the control unit 10 and may also read the stored information by the control unit 10.

[0020] The input unit 30 is configured for the user of the drug efficacy estimation device 1 to input information. The input unit 30 may be, for example, at least any one of a keyboard, a mouse, or a touch pad. Also, the input unit 30 may be an interface that accepts input of information from an external storage device such as a USB memory.

[0021] The control unit 10 is a control device that comprehensively controls each part of the drug efficacy estimation device 1. The control unit 10 may be, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The control unit 10 includes an acquisition unit 11, a model generation unit 12, and an estimation unit 13.

[0022] The control unit 10 may read the control program, which is software stored in the storage unit 20, and expand it in a memory such as a RAM (Random Access Memory) to execute the functions of each part. In the storage unit 20 shown in FIG. 1, for simplicity of explanation, the illustration of storage data such as the control program is omitted.

[0023] (Acquisition Unit 11) The acquisition unit 11 acquires a first medical image that images a lesion site related to the target disease of the first patient, and first expression information regarding gene expression in the lesion site corresponding to the first medical image. In this specification, the first patient refers to a patient suffering from the target disease and is the patient to be the subject of drug efficacy estimation. Also, the second patient described later refers to a patient suffering from the target disease and is a patient with known information regarding the treatment process.

[0024] In this specification, various types of information such as medical images related to the first patient are labeled with "first", and various types of information related to the second patient are labeled with "second". Also, for the description of general matters that do not specify the type of patient, the names of these various types of information may not be labeled with "first" and "second".

[0025] The target disease is not particularly limited as long as a lesion site occurs due to the onset or progression of the disease and a medical image of the lesion site is obtained. The lesion site does not need to be visually confirmable. For example, it may be a feature amount in a medical image that can be extracted by image analysis using a computer or the like and that changes compared to the normal state.

[0026] Examples of the target disease include cancer and stroke. The cancer may be a solid cancer, and examples include brain tumor, head and neck cancer, lung cancer, kidney cancer, liver cancer, digestive tract cancer, skin cancer, bladder cancer, breast cancer, uterine cancer, ovarian cancer, prostate cancer, and sarcoma.

[0027] The lesion site is appropriately set according to the type of the target disease. For example, if the target disease is lung cancer, it may be the whole lung or a specific part of the lung. Examples of the medical image of the lesion site include CT (Computed Tomography) image, MRI (Magnetic Resonance Imaging) image, X-ray image, endoscopic image, and tissue staining image.

[0028] The first expression information is information related to gene expression in the lesion site corresponding to the first medical image. The information related to gene expression may be, for example, information indicating the RNA expression level. Such expression information may be obtained by a comprehensive RNA expression analysis method such as RNA sequencing (RNA-seq).

[0029] The first medical image and the first expression information of the first patient obtained by the acquisition unit 11 are used to estimate the therapeutic effect of the first drug administrable to the first patient in the first patient.

[0030] The acquisition unit 11 may acquire various information from the storage unit 20. Further, the acquisition unit 11 may acquire at least a part of the various information as input information from the input unit 30.

[0031] (Model generation unit 12) The model generation unit 12 generates a learned model used to estimate the treatment effect of the first drug in the first patient. Specifically, the model generation unit 12 generates a learned model by machine learning using learning data including the second medical image and the second expression information as explanatory variables and including the second drug information and the second effect information as objective variables.

[0032] The second medical image and the second expression information are the medical image and the expression information in the second patient, which are the same as those in the first patient. The second medical image and the second expression information are preferably information obtained before the start of treatment with the drug in the second patient. Note that the model generation unit 12 preferably generates a learned model using various information regarding a plurality of second patients, and the larger the number of second patients with available information, the better.

[0033] Each piece of information of the first medical image and the second medical image may include information indicating the type of prognostic-related feature amount, which is a feature amount related to the prognosis of the target disease. Further, each of the first expression information and the second expression information may include functional information indicating the function of a gene whose expression level changes as the prognostic-related feature amount changes.

[0034] The feature amount may be an index that quantitatively indicates the state of each feature included in the medical image. The feature amount may be represented as a one-dimensional scalar amount or as a two-dimensional or higher-dimensional vector amount. From the viewpoint of reducing the processing load of the drug effect estimation device 1, dimensionality reduction processing may be performed on a feature amount represented as a two-dimensional or higher-dimensional vector amount.

[0035] Examples of the types of feature quantities include the size, shape, density, and texture of the lesion site. The types of feature quantities may be one or more, and may be several to tens of thousands of types, or more. The feature quantities included in the medical image may be extracted by known methods. For example, they can be extracted by performing image processing such as wavelet transform on the medical image.

[0036] The prognosis-related feature quantity is a feature quantity among the feature quantities included in the medical image that is related to the prognosis of the target disease. Information indicating the relationship with the prognosis is, for example, information including the overall survival period after treatment in patients with the target disease. The correlation between each feature quantity of the medical image and the prognosis of the patient with the target disease can be estimated using a known regression model or the like. Examples of such a regression model include the Cox regression model when the information regarding the prognosis is information regarding the survival period such as the overall survival period.

[0037] The "feature quantity related to the prognosis" indicates a feature quantity related to the course or prediction of the course after treatment of the target disease, and may be, for example, a feature quantity that changes in correlation with the length of the overall survival period before and after treatment. The model generation unit 12 may extract and use only the prognosis-related feature quantities among the respective feature quantities included in the second medical image based on the information indicating the types of such prognosis-related feature quantities.

[0038] Regarding the functional information indicating the function of a gene whose expression level changes with the change in the prognosis-related feature quantity, the function of the gene may be, for example, a function classified by GO (Gene Ontology) terms, or a function classified by other methods. When a plurality of genes whose expression levels change with the change in the prognosis-related feature quantity are found, common functions or representative functions, etc. in the gene group can be extracted, for example, by enrichment analysis.

[0039] Regarding the information indicating the types of the prognostic-related feature quantities and the functional information described above, those related to the second patient may be generated by the model generation unit 12, and those related to the first patient may be generated by the estimation unit 13. Hereinafter, an example of the process when the model generation unit 12 generates the information indicating the types of the prognostic-related feature quantities and the functional information related to the second patient will be shown. The estimation unit 13 may also generate these pieces of information related to the first patient by a method according to this.

[0040] The model generation unit 12 generates the information indicating the types of the prognostic-related feature quantities of the second patient by using the information of the second medical image and the second effect information described later, which includes information such as the overall survival period of the second patient after treatment, regarding the types of the prognostic-related feature quantities. In addition to the information indicating the types of the prognostic-related feature quantities of the second patient, the model generation unit 12 also acquires information such as the magnitude of the prognostic-related feature quantities.

[0041] The model generation unit 12 may estimate, for example, the correlation between each feature quantity included in the second medical image and the prognosis such as the overall survival period of the second patient by using a known regression model or the like. Then, the model generation unit 12 may extract, as the prognostic-related feature quantities, the feature quantities that are estimated to be statistically significantly correlated with the prognosis of the second patient among the feature quantities included in the second medical image.

[0042] In addition, the model generation unit 12 generates the functional information indicating the functions of the genes whose expression levels change with the change of the prognostic-related feature quantities by using the information of the extracted prognostic-related feature quantities and the second expression information. The model generation unit 12 may estimate, for example, the correlation between the comprehensive gene expression information such as RNA sequencing included in the second expression information and the prognostic-related feature quantities by using a known regression model or the like. Also, the functions of the gene groups estimated to be correlated with the prognostic-related feature quantities may be estimated by enrichment analysis.

[0043] Further, the model generation unit 12 may use an ensemble learning method that combines a plurality of learning algorithms for estimating genes correlated with prognosis-related features and generating functional information thereof. For example, the model generation unit 12 may use an algorithm that combines processing by a learning model suitable for natural language processing and correlation analysis by other regression models or the like. Examples of the learning model suitable for natural language processing include, for example, the Transformer model.

[0044] When the model generation unit 12 uses the Transformer model, the Transformer model may be adjusted to generate a connection between the feature information included in the second medical image and the gene expression information included in the second expression information. For example, in the Transformer model, it is possible to read the relationships such as the order and relevance of each prognosis-related feature amount, including the position information of the prognosis-related feature amount in the second medical image, and include the relationships in learning. Further, the information on the relationships used for learning preferably includes the relevance between the prognosis-related feature amount included in the second medical image and the gene expression information. The Transformer model may perform learning using such relationship information so as to be able to estimate a combination with a high relevance between the prognosis-related feature amount and the gene expression information.

[0045] Further, the model generation unit 12 may use a reinforcement learning algorithm as the learning model used for generating functional information. By using such a reinforcement learning algorithm, the model generation unit 12 can perform learning weighting by reward setting and accurately estimate a gene group correlated with prognosis-related features and the functions of the gene group.

[0046] The information used for reward setting may be, for example, a correlation value obtained by correlation analysis between the prognosis-related feature amount extracted from the second medical image and the second expression information. Further, the information used for reward setting may be an enrichment score obtained by enrichment analysis of a gene group estimated to be correlated with the prognosis-related feature amount. Further, it may be information that combines these correlation values and enrichment scores.

[0047] The model generation unit 12 may perform weighting and learning on the correlation between the prognosis-related feature amount and the second expression information using the enrichment score and the correlation value as information (values) used for such reward setting.

[0048] The second drug information is information regarding a second drug that is a drug administered to a second patient. The second drug information may be, for example, identification information of the second drug such as the generic name or compound name of the second drug. Further, when the second drug is a molecular target therapeutic agent, the second drug information preferably further includes information on the target molecule.

[0049] The second effect information is information indicating the therapeutic effect of the second drug on the target disease. The second effect information may be information such as the overall survival period (OS, number of days from the start of treatment to death or discontinuation) and life and death of the second patient to whom the drug has been administered.

[0050] Further, the second effect information may further include information on a drug response value indicating the influence exerted by the second drug on the cells of the second patient as information indicating a more specific therapeutic effect. The second drug may selectively damage the cells (abnormal cells) of the lesion site as a therapeutic effect of reducing the lesion site. In addition, it is also conceivable that the second drug may cause some damage to cells (normal cells) other than the lesion site.

[0051] The information on the drug response value of the second patient may be, as described above, information obtained by quantifying the positive or negative reaction exerted by the second drug on the cells of the second patient from the perspective of the treatment of the target disease. In other words, the second effect information may include, for example, information indicating the degree to which the second drug causes cell damage to the normal cells and abnormal cells of the second patient.

[0052] The second medical image, second expression information, second drug information, and second effect information used by the model generation unit 12 may be acquired by the acquisition unit 11. When the acquisition unit 11 acquires these pieces of information from the storage unit 20, the storage unit 20 may store a database in which various pieces of information indicating the treatment process of the target disease, such as the second medical image, second expression information, second drug information, and second effect information of the second patient, are stored.

[0053] The model generation unit 12 generates a learned model by machine learning using learning data that includes the second medical image and second expression information as explanatory variables and the second drug information and second effect information as target variables. Specifically, the model generation unit 12 performs machine learning by associating the correlation between the second medical image and second expression information and the second effect information with the second drug information for each second drug, and generates a learned model.

[0054] When such a learned model is input with the first medical image and first expression information, it can output the first effect information, which is the estimated treatment effect, for each first drug corresponding to each second drug used in the learning. Further, the learned model may extract the first drugs for which the estimated treatment effect is equal to or greater than a predetermined threshold and output only the corresponding first drug information and first effect information.

[0055] The model generation unit 12 is not limited to the second medical image and second expression information themselves, and may perform machine learning using, as explanatory variables, prognostic-related feature quantities including functional information of related genes that can be generated from the second medical image and second expression information. According to the learned model learned by such a method, it is possible to output the first effect information that accurately estimates the influence on the prognosis of the first patient administered the first drug based on the first medical image and first expression information.

[0056] Thus, the model generation unit 12 may generate a learned model using information that can be generated from the second medical image and the second expression information, not limited to the second medical image and the second expression information themselves as explanatory variables included in the learning data. This method is included as one of the modes in which the model generation unit 12 learns a learned model using the second medical image and the second expression information as explanatory variables.

[0057] The machine learning algorithm for learning the learned model used by the model generation unit 12 is not particularly limited. For example, it may be a reinforcement learning algorithm or a supervised learning algorithm. These algorithms are preferably deep learning algorithms that use a neural network at least in part.

[0058] The model generation unit 12 preferably uses a reinforcement learning algorithm. However, it is difficult to say that the reinforcement learning algorithm has been widely spread, especially in the medical field, compared with the supervised learning algorithm. This is presumably because, for example, the absolute number of AI (Artificial Intelligence) developers among medical workers such as doctors is small, and the absolute number of those involved in clinical sites among AI developers is small. Therefore, while supervised learning algorithms with abundant general-purpose development tools have been utilized in the medical field, other algorithms such as reinforcement learning algorithms have not been actively utilized.

[0059] In the reinforcement learning algorithm, it is important to appropriately set the state, policy, and reward of the agent in order to obtain an accurate estimation result. The inventors of the present invention have intensively studied, particularly on reward setting, and have found a method for accurately obtaining the estimated therapeutic effect of a drug, particularly the estimated therapeutic effect by the combined use of multiple drugs.

[0060] When the model generation unit 12 generates a learned model using a reinforcement learning algorithm, the learned model may be learned by a reinforcement learning algorithm that sets the state, policy, and reward of the agent as follows. State: the second medical image and the second expression information, Policy: selection of one or more second drugs, Reward: integrated information of a plurality of corresponding second effect information when a plurality of second drugs are selected.

[0061] The model generation unit 12 sequentially causes the agent to select one or more selectable second drugs in each state set by the second medical image and the second expression information, and for each selected second drug or its combination, learns the correlation between each state and the second effect information. Further, the model generation unit 12 performs learning by weighting the correlation between each state and the second effect information based on the reward setting.

[0062] The model generation unit 12 may cause the agent to execute a policy for each type of prognostic-related feature amount in the second medical image. Information indicating the type of prognostic-related feature amount at this time may include functional information of related genes. Thereby, the model generation unit 12 can generate a learned model for estimating the therapeutic effect of a drug for each type of prognostic-related feature amount.

[0063] In this way, the state of the agent is not limited to the second medical image and the second expression information themselves, and may be set based on information that can be generated based on the second medical image and the second expression information. Such a method is included as one of the modes in which the model generation unit 12 sets the second medical image and the second expression information as the state of the agent and learns the learned model.

[0064] The second drug that the model generation unit 12 causes the agent to select may be one or a plurality. As a combination in which a plurality of second drugs are selected, when the second effect information of a certain second patient is information indicating an effect obtained by the combined use of a plurality of second drugs, it is preferably the combination of the combined second drugs. Note that the model generation unit 12 may also select, as a combination of a plurality of second drugs, a combination in which there is no second effect information by combined use.

[0065] When the agent selects a plurality of second drugs, the model generation unit 12 learns by weighting the correlation between the information indicating the biological phenotype related to the prognosis-related feature amount and the second effect information based on the integrated information of the corresponding plurality of second effect information.

[0066] The integrated information of the second effect information may be, for example, the integrated value, average value, and / or correlation value of the drug response values in each second drug when the second effect information includes information on drug response values. Further, when the second effect information only includes information on the overall survival period, it may be the average value and / or correlation value of the overall survival period after administration of each second drug.

[0067] The correlation values of a plurality of drug response values or the overall survival period for a plurality of second drugs can be calculated, for example, from Pearson's correlation analysis, Spearman's rank correlation analysis, or mutual information amount.

[0068] The model generation unit 12 may learn, for example, so that the larger any one of the integrated value, average value, and correlation value of the drug response values is, the larger (smaller) the gene expression level related to the prognosis-related feature amount is. Note that when the second drug selected by the agent is one, the model generation unit 12 may learn the correlation between the information indicating the biological phenotype related to the prognosis-related feature amount and the second effect information as it is without performing weighting.

[0069] The trained model generated in this way is a drug efficacy estimation model that can output first drug information indicating a first drug administrable to a first patient and an estimated therapeutic effect of the first drug on the target disease when the first medical image and the first expression information of the first patient are input.

[0070] Note that the control unit 10 does not necessarily have to include the model generation unit 12. In other words, the drug efficacy estimation device 1 may be a device that performs a process of estimating the therapeutic effect of a drug administrable to a first patient using a pre-generated trained model. In this case, the drug efficacy estimation device 1 may store the above-mentioned trained model in the storage unit 20 in advance.

[0071] (Estimation unit 13) The estimation unit 13 inputs the first medical image and the first expression information into the trained model, and causes the trained model to output the first drug information and the first effect information.

[0072] The first drug information is information indicating a first drug administrable to a first patient. Similar to the second drug information, the first drug information may be the identification information of the first drug.

[0073] The information of the first medical image may include information indicating the types of prognosis-related feature amounts, similar to the information of the second medical image. Also, the first expression information may include functional information indicating the functions of genes whose expression levels change as the prognosis-related feature amounts change, similar to the second expression information. Further, the estimation unit 13 may generate information indicating the types of prognosis-related feature amounts and functional information based on the first medical image and the first expression information, similar to the model generation unit 12.

[0074] The trained model may output an estimated therapeutic effect for each first drug. For example, the trained model may use each second drug included in the learning data as a first drug administrable to the first patient, and for each first drug, associate and output the estimated therapeutic effect on the first patient with the first drug information. That is, the estimation unit 13 may cause the trained model to output information of a list including combinations of a plurality of first drugs and estimated therapeutic effects.

[0075] The estimation unit 13 may include, in the list, correspondence information between a combination of a plurality of first drugs and an estimated therapeutic effect by the combination, and output it to the learned model. A user of the drug efficacy estimation device 1 can use such a list to refer to the estimated therapeutic effects of the respective first drugs from among the plurality of first drugs and determine one or more types of drugs to be administered to the first patient.

[0076] The estimation unit 13 may extract a first drug or a combination thereof whose estimated therapeutic effect is equal to or higher than a predetermined threshold from the output of the learned model, and output only the corresponding first drug information and first effect information.

[0077] Further, when the learned model is learned by the reinforcement learning algorithm as described above, the estimation unit 13 can output the estimated therapeutic effect of the first drug for each type of prognostic-related feature amount. For example, when a favorable estimated therapeutic effect is output for the same first drug or a combination thereof with respect to a plurality of types of prognostic-related feature amounts, the user can estimate that the first drug or the combination thereof has a high possibility of exhibiting a good therapeutic effect.

[0078] Note that the first patient and the second patient may each be a patient with poor prognosis or poor course. In other words, the estimation unit 13 may cause the learned model learned using the second medical image and the second manifestation information of the second patient diagnosed with poor prognosis or poor course to output the estimated therapeutic effect of the first drug that can be administered to the first patient diagnosed with poor prognosis or poor course.

[0079] According to such a configuration, the estimation unit 13 can cause the learned model to output the estimated therapeutic effect of the first drug for the first patient diagnosed with poor prognosis or poor course, who is considered to be relatively difficult to treat. Therefore, the user of the drug efficacy estimation device 1 can easily select an effective first drug for the first patient who is considered to be relatively difficult to treat using the output result.

[0080] As described above, according to the drug efficacy estimation device 1, an effective first drug can be estimated with high accuracy based on the first medical image and the first expression information. In addition, the drug efficacy estimation device 1 can also be applied to the estimation of drugs effective for a first patient with poor prognosis or poor course, and can also be applied to the promotion of treatment for patients who have been difficult to treat conventionally. Such an effect also contributes to the achievement of, for example, Goal 3, "Good health and well-being for all," of the Sustainable Development Goals (SDGs) proposed by the United Nations.

[0081] <Drug Efficacy Estimation Method> Regarding the drug efficacy estimation method according to an embodiment of the present invention, the flow of processing executed by the drug efficacy estimation device 1 will be described using FIGS. 2 and 3 as examples. FIG. 2 is a flowchart showing an example of the flow of processing in the model generation step executed by the drug efficacy estimation device 1. FIG. 3 is a flowchart showing an example of the flow of processing in the estimation step executed by the drug efficacy estimation device 1. FIGS. 2 and 3 are also the flow of processing executed by the drug efficacy estimation system 100 including the drug efficacy estimation device 1. Note that the content already described in the items of the drug efficacy estimation system 100 described above will be omitted here.

[0082] First, the model generation step will be described. As shown in FIG. 2, the acquisition unit 11 acquires a second medical image, second expression information, second drug information, and second effect information for each second patient (S1). The various types of information acquired here are used as learning data by the model generation unit 12.

[0083] Next, the model generation unit 12 generates information indicating the types of prognosis-related feature amounts based on the second medical image and the second effect information (S2). Note that the information indicating the types of prognosis-related feature amounts may be included in the information of the second medical image acquired by the acquisition unit 11. In this case, the model generation unit 12 omits the processing of S2.

[0084] Further, based on the information indicating the types of prognosis-related feature quantities and the second expression information, the model generation unit 12 estimates a gene group whose expression level changes as the prognosis-related feature quantity changes. Then, it generates function information indicating the biological function of the estimated gene group (S3). Note that the gene group whose expression level changes as the prognosis-related feature quantity changes and the function information indicating its biological function may be included in the second expression information acquired by the acquisition unit 11. In this case, the model generation unit 12 omits the process of S3.

[0085] Next, the model generation unit 12 generates learning data using, as explanatory variables, the information of the second medical image including the information indicating the types of prognosis-related feature quantities and the function information, and using, as objective variables, the second drug information and the second effect information (S4). Note that the model generation unit 12 may generate the learning data using the second medical image and the second expression information themselves as explanatory variables.

[0086] Next, the model generation unit 12 generates a learned model by machine learning using the generated learning data (S5). The model generation unit 12 preferably uses a reinforcement learning algorithm as the machine learning algorithm for generating the learned model. Note that the drug efficacy estimation device 1 may acquire the learning data generated by a computer other than the drug efficacy estimation device 1 executing the processes of S1 to S4, and execute the process of generating the learned model of S5.

[0087] Next, the estimation process will be described. As shown in FIG. 3, the acquisition unit 11 acquires the first medical image and the first expression information (S6, acquisition process). The information of the first medical image may include information indicating the types of prognosis-related feature quantities, similar to the information of the second medical image, and this information may be generated by the estimation unit 13. Further, the first expression information may include, similar to the second expression information, function information indicating the function of a gene related to the prognosis-related feature quantity, and this information may be generated by the estimation unit 13.

[0088] Next, the estimation unit 13 inputs the first medical image and the first expression information into the learned model generated by the model generation unit 12 (S7). Then, the estimation unit 13 causes the learned model to output the first drug information and the first effect information (S8). The output of the learned model includes the first drug information and the first effect information for each first drug that can be administered to the first patient.

[0089] <learned model> In the present invention, the above-described learned model is also included in the scope. That is, the learned model according to one aspect of the present invention is a learned model for estimating the first drug that can be administered to the first patient and the estimated therapeutic effect of the first drug on the target disease based on the first medical image and the first expression information of the first patient. The learned model includes the second medical image and the second expression information of the second patient as explanatory variables, and is obtained by machine learning using learning data including the second drug information and the second effect information of the second drug administered to the second patient as objective variables. The learned model thus obtained is for causing a computer to function so as to output the first drug information indicating the first drug and the first effect information indicating the estimated therapeutic effect of the first drug from the first medical image and the first expression information.

[0090] 〔Embodiment 2〕 Other embodiments of the present invention will be described below.

[0091] The drug effect estimation system 100 shown in FIG. 1 includes an input unit 30 that receives input of the first medical image, the first expression information, etc. by the user, and is configured to output the drug effect estimation result to the display device 4, but is not limited thereto. For example, as shown in FIG. 4, the drug effect estimation system 100a may include a drug effect estimation device 1a that is communicably connected to communication terminals 5a and 5b used by each user via a communication network 9.

[0092] In the drug efficacy estimation system 100a shown in FIG. 4, the drug efficacy estimation device 1a receives information such as the first medical image and the first manifestation information from each of the communication terminals 5a and 5b. Then, the drug efficacy estimation device 1a transmits the estimation result corresponding to the information received from the communication terminal 5a to the communication terminal 5a, and transmits the estimation result corresponding to the information received from the communication terminal 5b to the communication terminal 5b. The estimation result is the first drug information and the first effect information output by the estimation unit 13 to the learned model.

[0093] Note that FIG. 4 shows the drug efficacy estimation system 100a including the communication terminals 5a and 5b and the drug efficacy estimation device 1a, but is not limited thereto. In the drug efficacy estimation system 100a, the drug efficacy estimation device 1a may be communicable with, for example, three or more communication terminals.

[0094] (Configuration of the drug efficacy estimation device 1a) The configuration of the drug efficacy estimation device 1a will be described with reference to FIG. 5. FIG. 5 is a functional block diagram showing a configuration example of the drug efficacy estimation system 100a according to an embodiment of the present invention. For the sake of convenience of explanation, members having the same functions as the members described in the above embodiment are denoted by the same reference numerals, and the description thereof will not be repeated.

[0095] As shown in FIG. 5, the drug efficacy estimation device 1a includes a communication unit 60 that functions as a communication interface with the communication terminals 5a and 5b. The acquisition unit 11 receives information such as the first medical image and the first manifestation information via the communication unit 60.

[0096] The estimation unit 13 transmits the estimation result to each of the communication terminals 5a and 5b via the communication unit 60. Note that the drug efficacy estimation device 1a may generate a web page showing the estimation result regarding the received information, and provide information for the user who is the transmission source of the information to access the web page.

[0097] 〔Example of implementation by software〕 The function of the drug efficacy estimation device 1 (hereinafter referred to as the "device") is a program for causing a computer to function as the device, and can be realized by a program for causing a computer to function as each control block of the device (especially each part included in the control unit 10).

[0098] In this case, the above device includes, as hardware for executing the above program, a computer having at least one control device (for example, a processor) and at least one storage device (for example, a memory). By executing the above program by this control device and storage device, each function described in the above embodiments is realized.

[0099] The above program may be recorded on one or more computer-readable recording media, not temporarily. This recording medium may or may not be provided in the above device. In the latter case, the above program may be supplied to the above device via any wired or wireless transmission medium.

[0100] Also, part or all of the functions of each of the above control blocks can also be realized by a logic circuit. For example, an integrated circuit in which a logic circuit functioning as each of the above control blocks is formed is also included in the scope of the present invention. In addition to this, for example, it is also possible to realize the functions of each of the above control blocks by a quantum computer.

[0101] 〔Summary〕 The method for estimating drug efficacy according to Aspect 1 of the present invention includes an acquisition step of acquiring a first medical image obtained by imaging a lesion site related to a target disease of a first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image, and a second medical image obtained by imaging the lesion site of a second patient, and second expression information regarding gene expression in the lesion site corresponding to the second medical image as explanatory variables, and using learning data including second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease as objective variables, inputting the first medical image and the first expression information into a learned model learned using the learning data, and outputting, to the learned model, first drug information indicating a first drug administrable to the first patient and first effect information indicating the estimated therapeutic effect of the first drug on the target disease.

[0102] In the method for estimating drug efficacy according to Aspect 2 of the present invention, in the above Aspect 1, each piece of information of the first medical image and the second medical image includes information indicating the type of prognostic-related feature amount that is a feature amount related to the prognosis of the target disease, and the first expression information and the second expression information each include function information indicating the function of a gene whose expression level changes as the prognostic-related feature amount changes. The first patient and the second patient may each be a patient with poor prognosis or poor course.

[0103] In the method for estimating drug efficacy according to Aspect 3 of the present invention, in the above Aspect 1 or 2, the learned model may be learned by a reinforcement learning algorithm in which the second medical image and the second expression information are used as the state of an agent, the selection of the second drug is used as the policy of the agent, and integrated information of a plurality of corresponding second effect information is set as the reward for the policy when a plurality of the second drugs are selected.

[0104] The method for generating a drug efficacy estimation model according to Aspect 4 of the present invention includes, as explanatory variables, a second medical image obtained by imaging a lesion site related to a target disease of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and uses learning data including, as objective variables, second drug information indicating a second drug administered to the second patient and second efficacy information indicating the therapeutic effect of the second drug on the target disease to generate a learned model by machine learning.

[0105] The drug efficacy estimation system according to Aspect 5 of the present invention includes an acquisition unit that acquires a first medical image obtained by imaging a lesion site related to a target disease of a first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image, and inputs the first medical image and the first expression information into a learned model learned using learning data including, as explanatory variables, a second medical image obtained by imaging the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and including, as objective variables, second drug information indicating a second drug administered to the second patient and second efficacy information indicating the therapeutic effect of the second drug on the target disease, and an estimation unit that causes the learned model to output first drug information indicating a first drug administrable to the first patient and first efficacy information indicating the estimated therapeutic effect of the first drug on the target disease.

[0106] The control program according to Aspect 6 of the present invention is a control program for controlling a computer, and causes the computer to perform an acquisition step of acquiring a first medical image of a lesion site related to a target disease of a first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image, and a learning model that includes, as explanatory variables, a second medical image of the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and includes, as target variables, second drug information indicating a second drug administered to the second patient and second effect information indicating a therapeutic effect of the second drug on the target disease. The learning model is trained using learning data. The method further includes an estimation step of inputting the first medical image and the first expression information into the learning model and causing the learning model to output first drug information indicating a first drug administrable to the first patient and first effect information indicating an estimated therapeutic effect of the first drug on the target disease.

[0107] The recording medium according to Aspect 7 of the present invention is a computer-readable recording medium that records the control program according to Aspect 6.

[0108] The learned model according to Aspect 8 of the present invention is a learned model for estimating a first drug administrable to a first patient and an estimated therapeutic effect of the first drug on a target disease based on a first medical image of a lesion site related to the target disease of the first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image. The learned model is obtained by machine learning using learning data that includes, as explanatory variables, a second medical image of the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and includes, as target variables, second drug information indicating a second drug administered to the second patient and second effect information indicating a therapeutic effect of the second drug on the target disease. The learned model causes a computer to function so as to output first drug information indicating the first drug and first effect information indicating the estimated therapeutic effect of the first drug from the first medical image and the first expression information.

[0109] 〔Supplementary Notes〕 The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope indicated in the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

Example

[0110] An example of the present invention will be described below. However, the present invention is not limited to the scope of each example shown below.

[0111] (Example 1. Prognosis prediction using medical images and estimation of prognosis-related features) Prognosis prediction was attempted using medical images of the lungs of non-small cell lung cancer patients, and the accuracy was evaluated. As a control group, the prognosis prediction accuracy was similarly estimated from the gene expression analysis data of non-small cell lung cancer patients and compared with the case using medical images.

[0112] Non-small cell lung cancer patients satisfying each of the following conditions (1) to (3) were selected, and CT image data and postoperative follow-up data were obtained. (1) Histopathologically diagnosed as non-small cell lung cancer, (2) Preoperative lung CT image data exists, (3) As postoperative follow-up data, the overall survival period (OS, number of days from the start of irradiation to death or discontinuation) and information on life and death exist.

[0113] These data were obtained from The Cancer Imaging Archive portal and The Cancer Genome Atlas database. Data of 150 cases were obtained, 100 cases were used as a learning dataset, and 50 cases were used as a validation dataset.

[0114] As a medical image, to reduce the variation between facilities, non-contrast simple CT images of the lungs in non-small cell lung cancer patients were used. The simple CT images were visually classified into tumor parts and non-tumor parts. Feature quantities in each of the tumor part and the non-tumor part were extracted using the wavelet imaging filter function of PyRadiomics software. The extracted feature quantities were each subjected to dimensionality reduction by Lasso regression.

[0115] Next, among the extracted feature quantities, prognostic-related feature quantities related to prognosis were estimated. Using the Lasso-Cox regression model, for each feature quantity in the medical image, the values of the tumor part and / or the non-tumor part were set as covariates, and the influence on the overall survival period of the patients was evaluated. Feature quantities evaluated to have a significant (p-value < 0.05) influence on the overall survival period of the patients by the model were estimated as prognostic-related feature quantities.

[0116] The types of the estimated prognostic-related feature quantities are as shown in Table 1 below. The names of the types of prognostic-related feature quantities are based on the classification of PyRadiomics software.

[0117]

Table 1

[0118] A Radiomics-score (Rad-score) was calculated to extract the importance for each prognostic-related feature quantity obtained by Lasso regression using data associating the prognostic-related feature quantities and the course data in the learning dataset. A learned model was created by an ensemble learning model combining a one-dimensional machine learning model with the Rad-Score as input data and a multi-dimensional machine learning model with all feature quantities in the medical image obtained by Lasso regression as input. The medical images included in the validation dataset were input to the obtained learned model to perform prognosis prediction.

[0119] As a control, a model trained with gene expression data instead of medical images was also created. RNA sequencing expression analysis data of lung tissue obtained from the The Cancer Genome Atlas database in the above patients was associated with the longitudinal data as input data and input into the above ensemble learning model to create a trained model. The RNA sequencing expression analysis data included in the validation dataset was input into the obtained trained model to perform prognosis prediction.

[0120] The prediction accuracy of each trained model was evaluated by the AUC (Area Under Curve) after ROC (Receiver Operating Characteristic) analysis. The AUC of the trained model trained using medical images was 0.90, and the AUC of the trained model trained using gene expression data was 0.71. That is, the trained model trained using medical images showed better prognosis prediction accuracy compared to the trained model trained using gene expression data. From this result, it was shown that information regarding prognosis can be obtained with high accuracy by using medical images.

[0121] (Example 2. Association between prognosis-related features in medical images and biological phenotypes) An attempt was made to associate prognosis-related features with biological phenotypes based on gene expression data.

[0122] Regarding the prognosis-related features estimated in Example 1, the association with biological phenotypes was analyzed. For gene groups whose expression levels change in conjunction with changes in prognosis-related features, the Transformer model that has been utilized in language models was used. The Transformer model is a model improved for medical use so as to generate a connection between features in medical images and gene expression data. Also, enrichment analysis was combined with the estimation to estimate the functional classification of the estimated gene groups.

[0123] As gene expression data, the above-described RNA sequencing expression analysis data was used. As gene functional classification data, the MSigDB C2 and C5 collections registered in the Molecular Signature Database (MSigDB, https: / / www.gsea-msigdb.org / gsea / msigdb / ) were used. The C2 collection mainly contains information on gene annotation from literature and the like, and the C5 collection mainly contains classification information by GO (Gene Ontology) terms.

[0124] Show a biological phenotype that is presumed to be related to prognostic-related features in the medical images of non-small cell lung cancer patients; ·RNA binding: A function that plays an important role in intracellular information transmission and protein synthesis. It may be related to the growth of cancer cells and may affect the prognosis. ·Nucleolus: A specific region within the cell nucleus, involved in rRNA synthesis and ribosome assembly, etc. It may be related to the growth of cancer cells and may affect the prognosis. ·Macroautophagy (macrophagy), Regulation of autophagy: Autophagy is a function that decomposes intracellular waste products and maintains energy supply and survival. Cancer cells may utilize autophagy to maintain survival and may affect the prognosis.

[0125] All of the presumed phenotypes were suggested to be related to the treatment prognosis of non-small cell lung cancer. That is, it was suggested that the prognostic-related features extracted from medical images may be effective indicators showing biological phenotypes related to the prognosis.

[0126] (Example 3. Drug Discovery Method) Drug discovery was attempted to estimate drugs suitable for patients with poor prognosis using medical images. A reinforcement learning model was used for drug discovery. As learning data for the reinforcement learning model, data including information showing the relationship between prognostic-related features and biological phenotypes based on gene expression data, which were medical images of non-small cell lung cancer patients, were used.

[0127] In addition, drug treatment response data showing the effects on each gene in cells (drug response values) after drug administration, which are registered in the LINCS1000 database, were obtained and input into the reinforcement learning model. The effects on cells indicate cell repair (proliferation) or damage (decrease), etc. When evaluating these effects on tumor cells (abnormal cells) as treatment effects, it can be said that cell repair is a negative treatment effect and cell damage is a positive treatment effect.

[0128] In the learning of the reinforcement learning model, various conditions were set as follows. · Environment, Agent: A simulation model by a reinforcement learning algorithm. Learn the correlation between information showing a biological phenotype related to prognostic-related features and drug response values. · State: Prognostic-related features including the information showing the relationship with the biological phenotype based on the gene expression data described above · Policy: Selection of one or more drugs · Reward: Integrated value, average value, and correlation value of drug response values when multiple drugs are selected Data including information showing the relationship between prognostic-related features and biological phenotypes based on gene expression data, which were medical images of the patient to be estimated, were input into the learned model. Then, for each prognostic-related feature, the estimation result of the treatment response of each drug was output from the learned model, and the drug estimated to show the optimal treatment response for the patient to be estimated was selected from among them.

[0129] Drugs estimated to be suitable for patients with poor prognosis or patients with poor disease progression are shown below. · GSK650394: SGK1 inhibitor · SB.366791: Ligand of TRPV1 · VE-922 (Berzosertib): ATR inhibitor All of these drugs have been used for some lung cancer patients. Since these drugs were estimated in the above-mentioned reinforcement learning model, it is considered that the therapeutic effect can be improved by selectively applying them not to conventional comprehensive options for small cell lung cancer patients but to patients with poor prognosis or poor course.

[0130] By using the reinforcement learning model as described above, one drug estimated to have an optimal therapeutic effect can be selected from one type of prognosis-related feature amount. Therefore, by using a plurality of prognosis-related feature amounts, it is possible to select a plurality of types of drugs, and it is also possible to find a drug estimated to be optimal in common for two or more prognosis-related feature amounts.

[0131] According to the drug discovery method according to one aspect of the present invention, compared with conventional methods such as correlation analysis, drug discovery considering a plurality of types of biological phenotypes can be easily performed using a plurality of types of prognosis-related feature amounts. That is, compared with the conventional method, it is possible to more effectively search for drugs having an excellent therapeutic effect.

Explanation of reference numerals

[0132] 11 Acquisition unit 12 Model generation unit 13 Estimation unit 100 Drug efficacy estimation system

Claims

1. An acquisition step of acquiring a first medical image obtained by imaging a lesion site related to a target disease of a first patient, and first expression information regarding gene expression in the lesion site corresponding to the first medical image; An estimation step of inputting the first medical image and the first expression information into a learned model that is learned using learning data including, as explanatory variables, a second medical image obtained by imaging the lesion site of a second patient, and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and including, as objective variables, second drug information indicating a second drug administered to the second patient and second effect information indicating a therapeutic effect of the second drug on the target disease, and causing the learned model to output first drug information indicating a first drug administrable to the first patient and first effect information indicating an estimated therapeutic effect of the first drug on the target disease. A drug efficacy estimation method comprising:

2. Each piece of information of the first medical image and the second medical image includes information indicating a type of prognostic-related feature amount that is a feature amount related to the prognosis of the target disease; The first expression information and the second expression information each include function information indicating the function of a gene whose expression level changes as the prognostic-related feature amount changes; The method for estimating drug efficacy according to claim 1, wherein the first patient and the second patient are each a patient with poor prognosis or poor course.

3. The learned model is learned by a reinforcement learning algorithm in which the second medical image and the second expression information are used as the state of an agent, the selection of one or more of the second drugs is used as the policy of the agent, and when a plurality of the second drugs are selected, integrated information of the corresponding plurality of the second effect information is set as the reward for the policy, respectively. The method for estimating drug efficacy according to claim 1 or 2.

4. A method for generating a drug efficacy estimation model, comprising generating a learned model by machine learning using learning data including, as explanatory variables, a second medical image obtained by imaging a lesion site related to a target disease of a second patient, and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and including, as objective variables, second drug information indicating a second drug administered to the second patient and second effect information indicating a therapeutic effect of the second drug on the target disease.

5. An acquisition unit that acquires a first medical image of a lesion site related to a target disease of a first patient, and first expression information regarding gene expression in the lesion site corresponding to the first medical image; An estimation unit that includes, as explanatory variables, a second medical image of the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and inputs the first medical image and the first expression information into a learned model learned using learning data including second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease, and causes the learned model to output first drug information indicating a first drug administrable to the first patient and first effect information indicating the estimated therapeutic effect of the first drug on the target disease. A drug efficacy estimation system.

6. A control program for controlling a computer, wherein the computer executes an acquisition step of acquiring a first medical image of a lesion site related to a target disease of a first patient, and first expression information regarding gene expression in the lesion site corresponding to the first medical image; and an estimation step of inputting the first medical image and the first expression information into a learned model learned using learning data including a second medical image of the lesion site of a second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image as explanatory variables, and including second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease as objective variables, and causing the learned model to output first drug information indicating a first drug administrable to the first patient and first effect information indicating the estimated therapeutic effect of the first drug on the target disease. A control program for this purpose.

7. A computer-readable recording medium recording the control program according to Claim 6.

8. A learned model for estimating a first drug administrable to the first patient and the estimated therapeutic effect of the first drug on the target disease based on a first medical image of a lesion site related to the target disease of the first patient and first expression information regarding gene expression in the lesion site corresponding to the first medical image. Obtained by machine learning using learning data that includes, as explanatory variables, a second medical image of the lesion site of the second patient and second expression information regarding gene expression in the lesion site corresponding to the second medical image, and includes, as target variables, second drug information indicating a second drug administered to the second patient and second effect information indicating the therapeutic effect of the second drug on the target disease. A learned model for causing a computer to function so as to output first drug information indicating the first drug and first effect information indicating the estimated therapeutic effect of the first drug from the first medical image and the first expression information.

Images