Information processing device, information processing program, and information processing method

Abstract

To provide biomarkers for discriminating patients into two states from intestinal flora by machine learning.SOLUTION: An information processing device 1 identifies, from among a plurality of candidate bacteria, an intestinal bacterium that serves as a biomarker to be used for discriminating a patient belonging to either a specific state or a non-specific state. The information processing device includes a storage unit 12, a discriminator generation unit 131 configured to use some sample patients as learning patients and some other sample patients as verification patients, to generate a discriminator to be used for discrimination of verification patient based on intestinal bacteria information of each of the learning and verification patients stored in a memory unit, and a marker identification unit 132 configured to identify an intestinal bacterium serving as a biomarker based on the importance of each candidate bacterium in the discriminator. The discriminator generation unit generates the discriminator for each combination of the learning patient and the verification patient. The marker identification unit identifies the intestinal bacterium serving as the biomarker based on the importance of each candidate bacterium in each discriminator for each combination.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing program, and an information processing method. [Background technology]

[0002] Intestinal bacteria are excreted from the human body as waste. Therefore, the types and proportions of intestinal bacteria present in the human body can be searched for without invasive techniques. In particular, in recent years, advances in technology based on next-generation sequencers have made it more accurate and convenient to search for masses of intestinal bacteria (intestinal microbiota). As a result, research into using intestinal microbiota as a biomarker (a factor that predicts the occurrence of a specific abnormality) has been attracting attention.

[0003] Methods of utilizing the intestinal flora as a biomarker have been proposed (see, for example, Non-Patent Documents 1-3).

[0004] The methods disclosed in Non-Patent Documents 1-3 identify the species and genera of bacteria that are statistically significantly different between one group that belongs to a state where a specific abnormality occurs (e.g., disease recurrence) and another group that belongs to a state where the same abnormality does not occur (e.g., disease recurrence not occurring) based on the amount of each species and genera in the content ratio data of the species and genera of bacteria. However, this method cannot distinguish between the two using the species and genera of enterobacteria when the distributions of the two data overlap, even if there is a statistically significant difference between them. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] B. Routy et.al., “Gut microbiome influences efficacy of PD-1-based immunotherapy against epithelial tumors”, Science. 2018 Jan 5, 359(6371), pp.91-97. doi: 10.1126 / science.aan3706. [Non-Patent Document 2] V. Matson et.al., “The commensal microbiome is associated with anti-PD-1 efficacy in metastatic melanoma patients”, Science. 2018 Jan 5, 359(6371), pp.104-108. doi: 10.1126 / science.aao3290. [Non-Patent Document 3] V. Gopalakrishnan et.al., “Gut microbiome modulates response to anti-PD-1 immunotherapy in melanoma patients”, Science. 2018 Jan 5, 359(6371), pp.97-103. doi: 10.1126 / science.aan4236. Summary of the Invention [Problem to be solved by the invention]

[0006] The present invention aims to provide a technology for searching for (identifying) biomarkers for classifying patients into one of two states from their intestinal flora by machine learning, a nonlinear analysis, without relying on linear analysis, a conventionally used statistical analysis method, in the search for biomarkers using the intestinal flora. [Means for solving the problem]

[0007] The information processing device of the present invention is an information processing device that identifies intestinal bacteria from a plurality of candidate bacteria that serve as biomarkers used to identify patients belonging to either a specific state or a non-specific state, and comprises: a memory unit that stores intestinal bacteria information for each of a plurality of sample patients; a classifier generation unit that generates a classifier used to identify the verification patient based on the intestinal bacteria information of the learning patients and the intestinal bacteria information of the verification patients stored in the memory unit, with some of the sample patients among the plurality of sample patients being learning patients and some of the sample patients being verification patients; and a marker identification unit that identifies the intestinal bacteria that will serve as the biomarker based on the importance of each candidate bacterium in the classifier, wherein the classifier generation unit generates a classifier for each combination of the learning patient and the verification patient, and the marker identification unit identifies the intestinal bacteria that will serve as the biomarker based on the importance of each candidate bacterium in each of the classifiers for each combination. Effect of the Invention

[0008] According to the present invention, biomarkers for classifying a patient into one of two states from the intestinal microbiota can be searched for (identified) by machine learning, which is a nonlinear analysis. [Brief description of the drawings]

[0009] [Figure 1] 1 is a network connection diagram showing an embodiment of an information processing device according to the present invention; [Diagram 2] 2 is a functional block diagram showing an embodiment of the information processing device of FIG. 1; [Diagram 3] FIG. 2 is a schematic diagram showing an example of intestinal bacteria information stored in a storage unit provided in the information processing device of FIG. 1. [Figure 4] 1 is a flowchart illustrating an embodiment of an information processing method according to the present invention. [Diagram 5] 5 is a flowchart of a classifier generation process included in the information processing method of FIG. 4. [Figure 6] 5 is a flowchart of a marker identification process included in the information processing method of FIG. 4. [Figure 7]5 is a flowchart of an evaluation information generation process included in the information processing method of FIG. 4. [Figure 8] 8 is a schematic diagram showing an example of evaluation information generated in the evaluation information generating process of FIG. 7 and displayed on a display unit provided in the information processing device of FIG. 1. [Figure 9] 1 is a graph showing sensitivity, specificity, and correct diagnosis rate calculated in an example of the present invention. [Figure 10] FIG. 2 is a schematic diagram showing an example of evaluation information for each marker candidate bacterium generated in an embodiment of the present invention, where (a) shows evaluation information at the genus level, and (b) shows evaluation information at the species level. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Below, with reference to the drawings, we will explain embodiments of an information processing device (hereinafter referred to as "this device"), an information processing program (hereinafter referred to as "this program"), and an information processing method (hereinafter referred to as "this method") according to the present invention.

[0011] The present invention identifies, from among a plurality of candidate bacteria, intestinal bacteria that serve as biomarkers for use in identifying patients belonging to either a specific state or a non-specific state.

[0012] The "specific state" refers to a state in which a specific abnormality occurs in a patient. In this embodiment, the specific state includes a state in which the same abnormality occurs again after treatment of the specific abnormality (i.e., a state in which the abnormality recurs: a recurrent state).

[0013] The "non-specific state" refers to a state in which a patient does not have a specific abnormality. In this embodiment, the non-specific state includes a state in which a specific abnormality does not occur after treatment for the same abnormality (i.e., a state in which the same abnormality does not recur: a non-recurrence state).

[0014] A "specific abnormality" is a disease that affects the types and proportions of intestinal bacteria contained in the patient's intestinal flora, such as a malignant tumor (e.g., cancer) that occurs in the patient's body.

[0015] A "biomarker" is an index used to objectively measure and evaluate a patient's physical condition.

[0016] "Candidate bacteria" are intestinal bacteria that can be candidates for biomarkers in the present invention, and are intestinal bacteria collected from patients belonging to multiple specific states (hereinafter referred to as "specific patients") and patients belonging to multiple non-specific states (hereinafter referred to as "non-specific patients") who make up the sample patients. In other words, the candidate bacteria are intestinal bacteria contained in the logical OR of the intestinal bacteria of each of the specific patients and non-specific patients who make up the sample patients. In this embodiment, the candidate bacteria are identified in advance, for example, by DNA analysis of the intestinal bacteria of each of the multiple specific patients and the multiple non-specific patients.

[0017] In the present invention, a "sample patient" is a patient carrying enterobacteria used to generate a classifier for classifying a patient into either a specific patient or a non-specific patient. As described above, the sample patient is composed of a plurality of specific patients and a plurality of non-specific patients. In other words, the sample patient is a patient whose condition is known.

[0018] <<Information processing device>> First, an embodiment of the present device will be described.

[0019] FIG. 1 is a network connection diagram showing an embodiment of the present device.

[0020] The device 1 identifies, from among a plurality of candidate bacteria, intestinal bacteria that serve as biomarkers used to identify patients belonging to either a specific state or a non-specific state. The specific configuration and operation of the device 1 will be described later.

[0021] The information storage device 2 stores information on the condition of a plurality of patients, including the sample patient, in association with the intestinal bacteria information of each patient. The information storage device 2 is, for example, an information storage server, a NAS (Network-Attached Storage), or a personal computer. The intestinal bacteria information will be described later.

[0022] The network NW is, for example, a communication network such as the Internet, a mobile communication network, a local area network (LAN), a wide area network (WAN), Wi-Fi (registered trademark), or Bluetooth (registered trademark).

[0023] <Configuration of information processing device> FIG. 2 is a functional block diagram showing an embodiment of the present device 1. As shown in FIG.

[0024] The present device 1 is realized, for example, by a personal computer. The present program runs on the present device 1, and the present program cooperates with the hardware resources of the present device 1 to realize the present method.

[0025] Here, by causing a computer (not shown) to execute this program, the program can cause the computer to function in the same manner as the device 1, and to execute the method.

[0026] The device 1 includes a communication unit 11, a storage unit 12, a control unit 13, an operation unit 14, and a display unit 15.

[0027] The communication unit 11 acquires information on the condition of the sample patient and the intestinal bacteria information of each sample patient from the information storage device 2 via the network NW. The communication unit 11 is configured with, for example, a communication module and a communication interface. The information on the condition of the sample patient is stored in the storage unit 12, for example, in association with the intestinal bacteria information of each sample patient.

[0028] The storage unit 12 stores information (e.g., intestinal bacteria information) necessary for the device 1 to execute the method described below. The storage unit 12 is composed of, for example, a recording device such as a hard disk drive (HDD) or a solid state drive (SSD) provided in the device 1, a volatile memory such as a random access memory (RAM), and / or a portable storage medium such as a flash memory.

[0029] The "intestinal bacteria information" is information on the intestinal bacteria of each sample patient. The intestinal bacteria information includes, for example, information indicating the classification name of the intestinal bacteria of each sample patient (for example, the names of the phylum, class, order, family, genus, and species of the intestinal bacteria) and information indicating the proportion of the intestinal bacteria of each sample patient.

[0030] The intestinal bacteria information in the present invention may include information on intestinal bacteria whose species or genus has not been identified.

[0031] FIG. 3 is a schematic diagram showing an example of intestinal bacteria information stored in the storage unit 12. As shown in FIG. The "sample patient ID" is an identification mark unique to each sample patient corresponding to each sample patient. The "state ID" is an identification mark unique to the state specified for each sample patient (in this embodiment, "1": state in which cancer has recurred, "0": state in which cancer has not recurred). The "name ID" is an identification mark unique to the name (classification name) of each intestinal bacterium (candidate bacterium) of each sample patient. In this embodiment, the name ID is associated with each rank of phylum, class, order, family, genus, and species identified for the corresponding intestinal bacterium. The "proportion" is information indicating the proportion of intestinal bacteria corresponding to the name ID in each sample patient. The figure shows that the sample patient ID, the state ID, the name ID, and the proportion are stored in the storage unit 12 in association with each other. For example, the device 1 can read out the state ID, the name ID, and the proportion stored in the storage unit 12 in association with the sample patient ID by referring to the storage unit 12 using the sample patient ID.

[0032] For example, the figure shows that the condition ID corresponding to the sample patient ID "P001" is "1", the proportion of the name ID "C001" is "x11", and the proportion of the name ID "C003" is "-: undetected".

[0033] Return to Figure 2. The control unit 13 controls the overall operation of the device 1 and executes the method described below. The control unit 13 is composed of, for example, a CPU (Central Processing Unit) provided in the device 1, a volatile memory such as a RAM that functions as a working area for the CPU, and a non-volatile memory such as a ROM (Read Only Memory) that stores various information such as the program. The control unit 13 includes a classifier generation unit 131, a marker identification unit 132, and an evaluation information generation unit 133.

[0034] The classifier generation unit 131 sets some of the sample patients as learning patients and the other sample patients as verification patients among the multiple sample patients, and generates a classifier to be used to identify the verification patients based on the intestinal bacteria information of the learning patients and the intestinal bacteria information of the verification patients stored in the storage unit 12. The specific operation of the classifier generation unit 131 will be described later.

[0035] A "learning patient" is a sample patient who possesses intestinal bacteria information to be used as supervised data (learning data) for machine learning of a classifier in the method described below.

[0036] A "verification patient" is a sample patient to be verified (classified) by a classifier trained based on intestinal bacteria information of a training patient in the method described below.

[0037] In this embodiment, the "classifier" is a model of a machine learning algorithm that classifies a plurality of sample patients into either sample patients belonging to a specific state or sample patients belonging to a non-specific state.

[0038] The marker identifying unit 132 identifies the intestinal bacteria to be the biomarkers based on the importance of each candidate bacterium in the classifier. The specific operation of the marker identifying unit 132 will be described later.

[0039] "Importance" is the degree of influence that each candidate bacterium has on the classification result in the classifier. That is, for example, a candidate bacterium with a high importance has a large influence on the classification result, and a candidate bacterium with a low importance has a small influence (almost no influence) on the classification result. In this embodiment, the importance is specified (calculated) for each candidate bacterium.

[0040] The evaluation information generating unit 133 generates evaluation information of the relative importance of each of the plurality of candidate bacteria. The specific operation of the evaluation information generating unit 133 will be described later.

[0041] The "evaluation information" is an index that relatively evaluates the importance of each candidate bacterium.

[0042] The operation unit 14 is a device that receives operations (such as input of information and selection) from a user of the present device 1 (for example, a doctor; hereinafter simply referred to as "user").

[0043] Display unit 15 is a device that displays information (such as evaluation information) necessary for carrying out the present method described below. In this embodiment, operation unit 14 and display unit 15 are, for example, touch panel displays.

[0044] The operation unit in the present invention may be an input device such as a mouse or a keyboard, and the display unit in the present invention may be a monitor or display connected to the device.

[0045] <<Information processing method>> Next, the operation of the present device 1, ie, the present method executed by the present device 1, will be described.

[0046] In the following description of the method, the sample patients are assumed to be composed of "n" sample patients in a specific state and "m" sample patients in a non-specific state (n and m are integers). The specific abnormality is assumed to be "esophageal cancer", the specific state is assumed to be "a state in which esophageal cancer has recurred", and the non-specific state is assumed to be "a state in which esophageal cancer has not recurred". Furthermore, the intestinal bacteria information of each of the "n+m" sample patients is assumed to be specified in advance and stored in the memory unit 12.

[0047] FIG. 4 is a flow chart illustrating an embodiment of the method.

[0048] The device 1 executes a classifier generation process (S1), a marker identification process (S2), and an evaluation information generation process (S3).

[0049] <Classifier generation process> FIG. 5 is a flowchart of the classifier generation process (S1).

[0050] The "classifier generation process (S1)" in the present invention is a process for generating a classifier for classifying a sample patient into either a sample patient in a specific state or a sample patient in a non-specific state.

[0051] First, the classifier generation unit 131 reads out the intestinal bacteria information of each of the "n+m" sample patients from the storage unit 12 (S101). That is, the classifier generation unit 131 acquires the names and proportions of the intestinal bacteria (i.e., candidate bacteria) of each sample patient.

[0052] Next, the classifier generation unit 131 performs primary learning and validation of a known machine learning algorithm (in this embodiment, random forest) using the intestinal bacteria information of each sample patient as a feature (S102).

[0053] "Primary learning and validation" refers to learning and validation of a machine learning algorithm for identifying an optimal combination of hyperparameters for the machine learning algorithm, which will be described later. In this embodiment, the classifier generation unit 131 executes leave-one-out cross-validation on "n+m" sample patients to execute the primary learning and validation of the machine learning algorithm. Here, in the initial learning and validation of the machine learning algorithm, for example, a default combination designated by the user of the device 1 is used as the combination of hyperparameters, which will be described later.

[0054] Specifically, the classifier generating unit 131 extracts "1" sample patient from "n+m" sample patients as a validation patient, and extracts the remaining "n+m-1" patients as training patients. Next, the classifier generating unit 131 performs primary learning of a machine learning algorithm based on the intestinal bacteria information of the "n+m-1" training patients, and generates a provisional classifier (primary classifier) ​​to be used for primary identification of the validation patient. Next, the classifier generating unit 131 performs primary identification of the validation patient using the generated primary classifier. Next, the classifier generating unit 131 repeats the generation of the primary classifier and the primary identification of the validation patient until all sample patients are extracted as validation patients. That is, the classifier generating unit 131 repeats the generation of the primary classifier and the primary identification of the validation patient "n+m" times. Here, the primary classifier is a provisional classifier in the present invention, and the primary identification is a provisional identification in the present invention.

[0055] Next, the classifier generation unit 131 calculates the sensitivity, specificity, and accuracy rate based on the result of the primary classification for each primary classifier (S103).

[0056] "Sensitivity" is the rate at which sample patients in a specific state are correctly identified in "n+m" times of identification (primary identification). In this embodiment, sensitivity is the rate at which sample patients with recurrent esophageal cancer are identified as sample patients in a specific state.

[0057] "Specificity" is the rate at which sample patients in a non-identified state are correctly identified in "n+m" instances of identification (primary identification). In this embodiment, specificity is the rate at which sample patients with no recurrence of esophageal cancer are identified as non-identified patients.

[0058] The "correct diagnosis rate" is the proportion of cases in which the condition of a sample patient was correctly identified in "n+m" instances of identification (primary identification).

[0059] Next, the discriminator generation unit 131 determines whether the first learning and validation (process S102) of the machine learning algorithm has been executed a predetermined number of times (e.g., 1000 times) (S104). The predetermined number of times is set in advance by the user of the device 1, etc.

[0060] When the primary learning and validation of the machine learning algorithm has not been performed the predetermined number of times ("N" in S104), the classifier generation unit 131 adjusts the hyperparameters of the machine learning algorithm constituting the primary classifier based on at least one of the sensitivity, specificity, and accuracy rate (S105). In this embodiment, the classifier generation unit 131 adjusts the hyperparameters so that the sensitivity and specificity are increased (increased). In this embodiment, the hyperparameters are automatically adjusted using a known adjustment method (in this embodiment, Optuna (registered trademark)).

[0061] A "hyperparameter" is a parameter that sets the behavior (operation) of a machine learning algorithm (e.g., the number of layers of a decision tree, the number of samples in a decision tree, etc.). Since the hyperparameters of known machine learning algorithms are known, detailed description thereof will be omitted.

[0062] Next, the discriminator generation unit 131 repeatedly executes steps S102 to S105 until the primary learning and verification of the machine learning algorithm reaches a predetermined number of times.

[0063] When the primary learning and validation of the machine learning algorithm has been executed a predetermined number of times ("Y" in S104), the classifier generation unit 131 identifies the combination of hyperparameters having the highest sensitivity and specificity as the optimal combination of hyperparameters (S106).

[0064] Next, the classifier generation unit 131 determines the machine learning algorithm set to the optimal combination of hyperparameters as a set classifier to be used in the marker identification process (S2) (S107). The set classifier is stored in the storage unit 12.

[0065] In the classifier generation process, the classifier generation unit of the present invention may repeatedly execute the primary learning and validation of the machine learning algorithm and the adjustment of the hyperparameters until at least one of the sensitivity, specificity, and accuracy rate exceeds a predetermined threshold, instead of a predetermined number of times.

[0066] Furthermore, in the classifier generation process, the classifier generation unit of the present invention may adjust hyperparameters, in addition to or instead of the adjustments to increase the sensitivity and specificity, so as to increase the area under the curve (AUC) in a receiver operating characteristic (ROC) curve using the sensitivity and specificity.

[0067] <Marker specific processing> FIG. 6 is a flowchart of the marker identification process (S2).

[0068] The “marker identification process (S2)” is a process that uses the classifier generated in the classifier generation process (S1) to identify intestinal bacteria that are significant for distinguishing between a specific state and a non-specific state as biomarkers from among the candidate bacteria.

[0069] First, the classifier generation unit 131 reads out the intestinal bacteria information of each of the "n+m" sample patients from the storage unit 12 (S201). That is, the classifier generation unit 131 acquires the names and proportions of the intestinal bacteria (i.e., candidate bacteria) of each sample patient.

[0070] Next, the classifier generation unit 131 reads out the set classifier (that is, the machine learning algorithm set to the optimal combination of hyperparameters) generated in the classifier generation process (S1) from the storage unit 12 (S202).

[0071] Next, the classifier generation unit 131 performs secondary learning and verification of the set classifier (machine learning algorithm) read out in step S202, using the intestinal bacteria information of each sample patient as a feature (S203). "Secondary learning and validation" is learning and validation of a machine learning algorithm for identifying sample patients and acquiring the importance of each candidate bacterium using a preset classifier. In this embodiment, the classifier generation unit 131 executes leave-one-out cross validation for "n+m" sample patients.

[0072] Specifically, the classifier generation unit 131 extracts "1" sample patient from the "n+m" sample patients as a validation patient, and extracts the remaining "n+m-1" patients as training patients. Next, the classifier generation unit 131 performs machine learning algorithm learning for the "n+m-1" patients based on the intestinal bacteria information of the training patients, and generates a provisional classifier (secondary classifier) ​​to be used for secondary classification of the validation patient. Next, the classifier generation unit 131 uses the generated secondary classifier to classify the validation patient. Next, the classifier generation unit 131 repeats the generation of the secondary classifier and the secondary classification of the validation patient until all sample patients are extracted as validation patients. That is, the classifier generation unit 131 repeats the generation of the secondary classifier and the secondary classification of the validation patient "n+m" times. In other words, the classifier generation unit 131 generates a classifier (secondary classifier) ​​for each combination of a training patient and a validation patient, and classifies the validation patient for each classifier (secondary classification).

[0073] Next, the classifier generation unit 131 acquires the importance of each candidate bacterium in each of the secondary classifiers for each combination of the training patient and the validation patient (i.e., "n+m" secondary classifiers) (S204). In this embodiment, the machine learning algorithm is Random Forest. Therefore, the classifier generation unit 131 acquires the importance of each candidate bacterium, which is a feature, by referring to a variable (importance) that means importance in the Random Forest, for example.

[0074] Note that the classifier generating unit in the present invention may obtain the importance of each candidate bacterium by a method of shuffling the feature amounts of each classifier (permutation importance) instead of using variables.

[0075] Next, the classifier generation unit 131 selects "N (N is an integer)" types of candidate bacteria as marker candidate bacteria in order of importance in each secondary classifier for each combination of learning patients and verification patients based on the importance of each candidate bacterium (S205). That is, the classifier generation unit 131 selects the top "N" types of marker candidate bacteria for each "n+m" number of secondary learning and verifications. Here, in this embodiment, the classifier generation unit 131 selects "10" types of marker candidate bacteria. The selected marker candidate bacteria are stored in the memory unit 12 in association with their respective importance.

[0076] "Candidate marker bacteria" are candidate bacteria that can ultimately become biomarkers among the candidate bacteria. The selected candidate marker bacteria may be the same for each secondary classifier, or may be partially different for each secondary classifier.

[0077] The classifier generation unit in the present invention may select marker candidate bacteria for each generation of a secondary classifier and secondary identification of a verification patient. That is, for example, the classifier generation unit in the present invention may repeat "n+m" times a set of generation of a secondary classifier, secondary identification of a verification patient, and selection of marker candidate bacteria.

[0078] Next, the discriminator generation unit 131 determines whether or not the secondary learning and validation (process S203) of the machine learning algorithm has been executed a predetermined number of times (e.g., 500 times) “n+m” times (S206). The predetermined number of times is set in advance by the user of the present device 1, etc.

[0079] When the secondary learning and validation of the machine learning algorithm has not been performed the predetermined number of times ("N" in S206), the classifier generation unit 131 executes secondary learning and validation of the machine learning algorithm (S203), obtains the importance (S204), and selects marker candidate bacteria (S205).

[0080] On the other hand, when the secondary learning and verification of the machine learning algorithm has been performed a predetermined number of times ("Y" in S206), the marker identification unit 132 generates a frequency distribution of all (in this embodiment, "n+m" x "N" x a predetermined number of times) marker candidate bacteria (S207).

[0081] Next, the marker identifying unit 132 identifies "M (M is an integer)" types of marker candidate bacteria (i.e., intestinal bacteria) in descending order of frequency as intestinal bacteria that can serve as biomarkers (S208). Here, in this embodiment, the marker identifying unit 132 identifies "10" types of marker candidate bacteria as intestinal bacteria that can serve as biomarkers. The identified intestinal bacteria are stored in the memory unit 12.

[0082] Generally, since a machine learning algorithm has parameters that are random numbers, even if a pre-configured classifier has optimal hyperparameters set, there may be fluctuations in the classification results and importance for each classification. Here, in this embodiment, the secondary learning and validation are repeated a predetermined number of times (500 times), and the intestinal bacteria that serve as biomarkers are identified based on the frequency distribution of the importance for each time. As a result, the present invention can eliminate the influence of random numbers as much as possible.

[0083] <Evaluation information generation process> FIG. 7 is a flowchart of the evaluation information generation process (S3).

[0084] The "evaluation information generation process (S3)" is a process for generating evaluation information for each marker candidate bacterium.

[0085] First, the evaluation information generating unit 133 reads out the marker candidate bacteria and their importance for each of all secondary classifiers from the storage unit 12 (S301).

[0086] Next, the evaluation information generating unit 133 counts the number of times each candidate bacterium selected as a marker candidate bacterium has been selected as a marker candidate bacterium (S302).

[0087] Next, the evaluation information generating unit 133 ranks the candidate bacteria in descending order of frequency and generates information (evaluation information) associating the candidate bacteria with their ranks (S303). The evaluation information is associated with, for example, the name of the corresponding enterobacteria and stored in the storage unit 12. The evaluation information is displayed, for example, on the display unit 15 and viewed by a user of the device 1. In this case, the display unit 15 reads out the evaluation information from the storage unit 12 and displays the evaluation information of the enterobacteria identified as a biomarker.

[0088] FIG. 8 is a schematic diagram showing an example of evaluation information displayed on the display unit 15. As shown in FIG.

[0089] The figure shows a list of the names of intestinal bacteria that serve as biomarkers for the recurrence of esophageal cancer, and the evaluation information corresponding to the names.

[0090] The evaluation information generated by the evaluation information generation process of the present invention is not limited to the present embodiment. That is, for example, the evaluation information generation unit of the present invention may weight each classifier in order of importance, and use the sum of the weights for each candidate bacterium as the evaluation information.

[0091] <<Example>> Next, an embodiment of the present invention will be described. In the following description of the present method, the sample patients are assumed to be composed of "18" sample patients in a specific state and "33" sample patients in a non-specific state. In addition, the specific abnormality is "esophageal cancer", the specific state is "a state in which esophageal cancer has recurred", and the non-specific state is "a state in which esophageal cancer has not recurred". In this embodiment, there were 220 types of candidate bacteria at the genus level and 284 types at the species level.

[0092] 9 is a graph showing the sensitivity, specificity, and accuracy rate calculated in an embodiment of the present invention. The figure shows the sensitivity, specificity, and accuracy rate when the first learning and validation of the machine learning algorithm and the adjustment of the hyperparameters are repeatedly performed at the genus level of the candidate bacteria.

[0093] The figure shows that as the number of iterations increases from before hyperparameter adjustment to 100 times and then to 1000 times, the sensitivity increases, the specificity is maintained at a high value, and the accuracy rate increases. In this way, the sensitivity increases by repeating hyperparameter adjustment based on the sensitivity. Furthermore, this method can identify a combination of hyperparameters with a sensitivity of approximately 50% even when the number of sample patients is small at 51 and there is a relatively large difference between the number of sample patients in a specific state and the number of sample patients in a non-specific state.

[0094] FIG. 10 is a schematic diagram showing an example of evaluation information for each marker candidate bacterium generated in an embodiment of the present invention, where (a) shows evaluation information at the genus level and (b) shows evaluation information at the species level. The figure shows an example of evaluation information for each marker candidate bacterium when one cycle (51 times) of secondary learning and validation of the machine learning algorithm is performed at the genus level and the species level. Here, the "species level" means that the intestinal bacteria information that is the feature of the machine learning algorithm is composed of the species of the candidate bacteria and their proportions. The "genus level" means that the intestinal bacteria information that is the feature of the machine learning algorithm is composed of the genus of the candidate bacteria and their proportions.

[0095] The figure shows that at the genus level, the marker candidate bacteria of the genus "Butyricimonas" were selected 51 / 51 times, the marker candidate bacteria of the genus "holdemania" were selected 49 / 51 times, and the marker candidate bacteria of the genus "Coprobacillus" were selected 22 / 51 times. The figure also shows that at the species level, most of the top-ranked species are unidentified, and that even unidentified species can provide information on whether they can be biomarkers. Furthermore, the figure shows that "holdemania" was selected as a marker candidate bacterium at both the genus and species levels, and that the results differ between the species and genus levels.

[0096] <<Summary>> According to the embodiment described above, the present device 1 includes a classifier generating unit 131 that generates a classifier used to identify the validation patient based on the intestinal bacteria information of the validation patient and the intestinal bacteria information of the learning patient stored in the storage unit 12, with some of the sample patients being learning patients and the other sample patients being validation patients, and a marker identifying unit 132 that identifies the intestinal bacteria that will be the biomarker based on the importance of each candidate bacterium in the classifier. The classifier generating unit 131 generates a classifier for each combination of the learning patient and the validation patient. The marker identifying unit 132 identifies the intestinal bacteria that will be the biomarker based on the importance of each candidate bacterium in each classifier for each combination. According to this configuration, the present device 1 can learn a machine learning algorithm based on the intestinal bacteria information of the sample patient whose state is known, and generate a classifier for each combination of the learning patient and the validation patient. Then, the present device 1 can obtain the importance, which is the degree to which the candidate bacterium affects the identification result, for each candidate bacterium in each classifier. As a result, the device 1 can identify intestinal bacteria that serve as biomarkers using the importance as an objective indicator. That is, the device 1 can search for (identify) biomarkers for classifying patients into two states from the intestinal flora by machine learning, which is a nonlinear analysis.

[0097] Furthermore, according to the embodiment described above, the marker identification unit 132 identifies intestinal bacteria that serve as biomarkers based on the frequency of each candidate bacterium in the frequency distribution of the importance of each candidate bacterium in each classifier. With this configuration, the present device 1 can eliminate, as much as possible, deviations in the importance of each candidate bacterium in each classifier that is generated multiple times due to the influence of random numbers in the machine learning algorithm. As a result, the present device 1 can accurately identify intestinal bacteria that serve as biomarkers through machine learning.

[0098] Furthermore, according to the embodiment described above, the classifier generation unit 131 generates a plurality of provisional classifiers (primary classifiers) used for provisional classification (primary classification) of a verification patient, and adjusts hyperparameters of a machine learning algorithm constituting the provisional classifier based on any one of sensitivity, specificity, and accuracy obtained by provisional classification for each provisional classifier. According to this configuration, the present device 1 can obtain a machine learning algorithm (i.e., a classifier) ​​having any one of desired sensitivity, specificity, and accuracy. As a result, the accuracy of identifying intestinal bacteria as biomarkers by machine learning is improved.

[0099] Furthermore, according to the embodiment described above, the classifier generation unit 131 repeats the primary learning and validation and the adjustment of the hyperparameters multiple times. With this configuration, the device 1 can increase any one of the classification sensitivity, specificity, and accuracy rate to a desired value depending on the number of times the hyperparameters are adjusted. As a result, the accuracy of identifying intestinal bacteria that serve as biomarkers by machine learning is improved.

[0100] Furthermore, according to the embodiment described above, the present device 1 comprises an evaluation information generating unit 133 that generates relative evaluation information on the importance of each of a plurality of candidate bacteria, and a display unit 15 that reads out and outputs evaluation information on the intestinal bacteria identified as a biomarker from the storage unit 12. With this configuration, the user of the present device 1 can easily obtain an index showing how superior the intestinal bacteria selected as a biomarker is compared to other intestinal bacteria.

[0101] In addition, the machine learning algorithm used by the classifier generating unit in the present invention is not limited to random forest. That is, for example, the machine learning algorithm may be other machine learning algorithms such as Naive Bayes or Support Vector Machine. In this case, the importance can be obtained by, for example, permutation importance.

[0102] Furthermore, the primary learning and validation of the classifier generation process and the secondary learning and validation of the marker identification process are not limited to leave-one-out cross-validation. That is, for example, the classifier generation unit in the present invention may perform the primary learning and validation and the secondary learning and validation using other cross-validation (e.g., k-fold cross-validation, etc.).

[0103] Furthermore, the present device only needs to identify intestinal bacteria that serve as biomarkers, and does not need to execute evaluation information generation processing.

[0104] Furthermore, in the marker identification process, the present device may identify the intestinal bacteria that will be the final biomarker by taking into consideration the identification result using another classifier different from the set classifier. That is, for example, the present device may identify the intestinal bacteria that will be the biomarker using a classifier whose hyperparameters are not adjusted (an unadjusted classifier), and identify the intestinal bacteria that are common to the intestinal bacteria that will be the biomarker identified in the set classifier and the intestinal bacteria that will be the biomarker identified in the unadjusted classifier as the intestinal bacteria that will be the final biomarker. In this configuration, since the result using the unadjusted classifier is taken into consideration, the influence of overlearning of the machine learning algorithm can be suppressed. Here, the present device may use, for example, a classifier in which a random number is not specified in the set classifier as the other classifier, instead of the unadjusted classifier. Here, the unadjusted classifier and the classifier in which a random number is not specified are examples of classifiers before adjustment in the present invention is completed.

[0105] Furthermore, the present device may identify the intestinal bacteria to be biomarkers based on the importance acquired at the species level and the importance acquired at the genus level. That is, for example, the present device may identify the top "M" types of marker candidate bacteria with a high frequency of the importance at the species level and a high frequency of the importance at the genus level as the intestinal bacteria to be biomarkers. That is, the present device may identify common intestinal bacteria among the intestinal bacteria identified as the intestinal bacteria to be biomarkers at the genus level and the species level as the intestinal bacteria to be the final biomarker. In this configuration, the accuracy of identifying the intestinal bacteria to be biomarkers by machine learning may be improved compared to the species level and the genus level alone. Also, for example, the present device may narrow down the feature amount at the species level (type of candidate bacteria) based on the importance acquired at the genus level to obtain the importance at the species level, and identify the intestinal bacteria to be biomarkers based on the importance at the species level. In this configuration, when the candidate bacteria cover a huge number of species, the number of species of candidate bacteria can be reduced in advance.

[0106] Furthermore, in the classifier generation process, the present device may execute a dimensionality reduction process of the feature amount (types of bacteria and their ratios) using a known dimensionality reduction technique (e.g., principal component analysis, t-SNE, Isomap, etc.) as a preprocessing of step S102. In this configuration, the efficiency of the first learning and validation in the classifier generation process is improved.

[0107] Furthermore, in the embodiment described above, the device 1 includes the display unit 15 as an example of the output unit in the present invention. Instead of this, the device may include a printer as an example of the output unit in the present invention.

[0108] Furthermore, in the above-described embodiment, the present device 1 is configured by one computer. Alternatively, the present device may be configured by multiple computers. That is, for example, the present device may be configured by a group of multiple computers that function as the present device. Specifically, for example, the present device (group of computers) may be configured by a computer having a storage unit and a computer having a control unit that executes the present method. Also, for example, the multiple computers may have the respective functions of the classifier generation unit, the marker identification unit, and the evaluation information generation unit in a distributed manner. In this case, the multiple computers that make up the group of computers may send and receive information through a communication line, or may transfer information using a portable storage medium. [Explanation of symbols]

[0109] 1. Information processing device 12 Storage section 131 Discriminator generator 132 Marker Identification Unit 133 Evaluation information generation unit 15 Display section (output section)

Claims

1. An information processing device that identifies an intestinal bacterium that serves as a biomarker for identifying a patient belonging to either a specific state or a non-specific state from among a plurality of candidate bacteria, A memory unit that stores intestinal bacteria information of each of a plurality of sample patients; a classifier generating unit that generates a classifier, which is a model of a machine learning algorithm, for classifying the verification patients into either a patient belonging to the specific state or a patient belonging to the non-specific state, based on the intestinal bacteria information of the learning patients and the intestinal bacteria information of the verification patients stored in the storage unit, by using some of the sample patients as learning patients and the other sample patients as verification patients, among the plurality of sample patients; a marker identification unit that identifies the intestinal bacteria to be the biomarker based on the importance of each of the candidate bacteria in the classifier; and the classifier generation unit generates the classifier for each combination of the training patient and the validation patient; the marker identifying unit identifies the intestinal bacteria to be the biomarkers based on the importance of each of the candidate bacteria in each of the classifiers for each of the combinations.

23. An information processing apparatus comprising:

2. the marker identification unit identifies the intestinal bacteria to be the biomarker based on a frequency of the importance of each of the candidate bacteria in a frequency distribution of the importance of each of the candidate bacteria in each of the classifiers.

2. The information processing device according to claim 1.

3. The discriminator generation unit generating a plurality of provisional classifiers that are models of the machine learning algorithm for provisionally classifying the verification patient into either a patient belonging to the specific state or a patient belonging to the non-specific state; adjusting a hyperparameter of the machine learning algorithm constituting the provisional classifier based on any one of sensitivity, specificity, and accuracy obtained by the provisional classification for each of the provisional classifiers; 2. The information processing device according to claim 1.

4. the classifier generation unit repeats the adjustment of the hyperparameters and the generation of the temporary classifier a plurality of times.

4. The information processing device according to claim 3.

5. the marker identification unit identifies the intestinal bacteria to be the biomarker based on the importance of each of the classifiers after the adjustment of the hyperparameters is completed and the importance of each of the classifiers before the adjustment of the hyperparameters is completed.

5. The information processing device according to claim 4.

6. The intestinal bacteria information is The intestinal bacteria information at the genus level of the intestinal bacteria; The gut bacteria information at the species level of the gut bacteria; Including, The marker identification unit identifies the intestinal bacteria to be the biomarker based on the importance acquired at the species level and the importance acquired at the genus level.

3. The information processing device according to claim 1 or 2.

7. an evaluation information generating unit that generates relative evaluation information of the importance of each of the plurality of candidate bacteria; It is made of 2. The information processing device according to claim 1.

8. The evaluation information generation unit stores the evaluation information for each of the candidate bacteria in the storage unit, an output unit that reads out from the storage unit the evaluation information of the intestinal bacteria identified as the biomarker and outputs it; having 8. The information processing device according to claim 7.

9. A computer is caused to function as the information processing device according to claim 1.

2. An information processing program comprising:

10. An information processing method executed by an information processing device for identifying an intestinal bacterium that serves as a biomarker for identifying a patient belonging to either a specific state or a non-specific state from among a plurality of candidate bacteria, comprising: The information processing device includes: A memory unit that stores intestinal bacteria information of each of a plurality of sample patients; Equipped with The information processing device, a classifier generating step of generating a classifier that is a model of a machine learning algorithm that classifies the verification patients into either a patient belonging to the specific state or a patient belonging to the non-specific state based on the intestinal bacteria information of the learning patients and the intestinal bacteria information of the verification patients stored in the storage unit, with some of the sample patients being used as learning patients and the other sample patients being used as verification patients; a marker identification step of identifying the intestinal bacteria to be the biomarker based on the importance of each of the candidate bacteria in the classifier; and The information processing device includes: In the classifier generating step, the classifier is generated for each combination of the training patient and the validation patient; In the marker identification step, the enterobacteria to be the biomarkers are identified based on the importance of each of the candidate bacteria in each of the classifiers for each of the combinations.

23. An information processing method comprising:

Images