Disease affection determination device, disease affection determination method, and disease affection determination program

Patent metadata

Field	Value
Publication number	US-2019267113-A1
Application number	US-201716346017-A
Country	US
Kind code	A1
Filing date	Oct 31, 2017
Priority date	Oct 31, 2016
Publication date	Aug 29, 2019
Grant date	—

Abstract

To enable disease affection determination by using a neural network to perform learning using data of the expression levels of biomarkers, and to enable extraction of a feature biomarker for a disease by the neural network. Sample data in which respective expression levels of a plurality of types of biomarkers are recorded for each individual is acquired, a learned model in which affection of diseases is determinable obtained in advance by performing machine learning using training data is generated, a plurality of sample data to which label information of disease affection is attached is input to the learned model and calculation is performed, the degrees of importance of respective feature of a plurality of biomarkers obtained with the learned model are quantified by affection determination calculation, for each sample data, and a predetermined number of biomarkers are extracted as feature biomarkers regarding the disease on the basis of the quantified degrees of importance of all the sample data for each biomarker.

First claim

We claim:

1 . A disease affection determination device comprising:
a sample data acquisition unit configured to acquire sample data including respective expression levels of biomarkers including a plurality of types of miRNAs in an individual organism-derived sample; a learned model in which affection of diseases is determinable, obtained in advance by performing machine learning using a plurality of training data including sample data each including items for identifying presence or absence of affection of a plurality of diseases and to which label information is attached, the label information indicating whether individuals are affected with any of the diseases; and an affection determination unit configured to perform affection determination as to whether sample data to be determined is affected with a plurality of diseases, using the learned model.

2 . The disease affection determination device according to claim 1 , comprising:
a determination contribution biomarker output unit configured to extract a biomarker that has contributed to a disease affection determination result, of the biomarkers included in the sample data to be determined for disease affection, and output the extracted biomarker.

3 . The disease affection determination device according to claim 2 , wherein the determination contribution biomarker output unit calculates, by a process of calculating a loss function L, using the learned model, for the sample data, and a process of performing error back propagation with a value L of the loss function as a starting point and calculating a gradient g i =∂L/∂x j for a feature x j corresponding to each of a plurality of types of biomarkers, the degree of importance of each feature dimension corresponding to the biomarker as the gradient g i for the feature x j , and extracts a predetermined number of biomarkers as the biomarkers that have contributed to the disease affection determination result on the basis of the magnitude of the degree of importance.

4 . The disease affection determination device according to claim 2 , wherein the determination contribution biomarker output unit learns a linear learner that approximates the learned model in the affection determination unit by LIME, calculates a coefficient of the linear learner, the coefficient corresponding to the feature dimension of each biomarker of when the sample data to be determined for affection is input to the linear learner, as the degree of importance of each biomarker, and extracts a predetermined number of biomarkers as the biomarkers that have contributed to the disease affection determination result on the basis of the magnitude of the degree of importance.

5 . The disease affection determination device according to claim 2 , wherein the determination contribution biomarker output unit performs forward propagation by providing a feature of sample data of a patient to be determined for affection to the learned model in the affection determination unit by LRP, recursively calculates an importance vector R representing the degree of importance in each layer, crossing layers in reverse order from the output unit, calculates the importance vector R as the degree of importance of each feature dimension corresponding to each biomarker, and extracts a predetermined number of biomarkers as the biomarkers that have contributed to the disease affection determination result on the basis of the magnitude of the degree of importance.

6 . A disease affection determination device comprising:
a sample data acquisition unit configured to acquire sample data including respective expression levels of biomarkers including a plurality of types of miRNAs in an individual organism-derived sample; at least two or more machine learners configured to perform machine learning commonly using a plurality of training data including sample data each including items for identifying presence or absence of affection of a plurality of diseases and to which label information is attached, the label information indicating whether individuals are affected with any of the diseases, the machine learners respectively including different types of learned models that have learned in advance to determine affection of the same disease, the machine learners configured to output a prediction result as to whether sample data to be determined for disease affection has affected a disease; and a stacking machine learner that has learned in advance to output a final determination result, using the prediction results from the plurality of machine learners as inputs, and configured to output a determination result as to whether the sample data to be determined for affection is affected with a disease on the basis of the prediction results from the plurality of machine learners.

7 . The disease affection determination device according to any one of claims 1 to 6 , wherein the plurality of diseases includes at least two types of breast cancer, breast benign disease, prostate cancer, benign prostate disease, pancreatic cancer, biliary tract cancer, colon cancer, gastric cancer, esophageal cancer, liver cancer, and benign pancreatic disease.

8 . A disease affection determination device comprising:
a plurality of sample data respectively acquired from individual organisms and including respective expression levels of a plurality of types of biomarkers including miRNA in individual organism-derived samples; a learned model in which presence or absence of affection of a plurality of diseases is determinable, the plurality of diseases being output as a result of machine learning using, as training data, sample data with label information in which items for identifying whether each individual organism has affected the plurality of diseases are provided as label information, for each of the plurality of sample data; and an affection determination unit configured to determine presence or absence of affection of each of the plurality of diseases, using the learned model, for sample data newly acquired from another organism for which affection determination is to be performed.

9 . A disease affection determination device comprising:
a plurality of sample data respectively acquired from individual organisms and including respective expression levels of a plurality of types of biomarkers including miRNA in an individual organism-derived sample; a learned model in which presence or absence of affection of a predetermined disease is determinable, the predetermined disease being output as a result of machine learning using, as training data, sample data with label information in which items for identifying whether each individual organism is affected with any one of a predetermined group of diseases determined in advance or whether each individual organism is not affected with any of the predetermined group of diseases determined in advance, as information regarding the disease when affected with the disease or information indicating that the individual organism is not affected when not affected, as label information for each of the plurality of sample data; and an affection determination unit configured to determine whether affected with any one of the predetermined group of diseases or whether not affected with any of the predetermined group of diseases, using the learned model, for sample data newly acquired from another organism for which affection determination is to be performed.

10 . A disease affection determination method comprising the steps of:
acquiring sample data including respective expression levels of biomarkers including a plurality of types of miRNAs in an individual organism-derived sample; generating a learned model in which whether affected with a plurality of dise

Assignees

• Preferred Networks Inc

Inventors

• Okanohara Daisuke
• Oono Kenta
• Ota Nobuyuki
• Hamzaoui Karim
• Akiba Takuya

Classifications

• C12Q2600/178

  miRNA, siRNA or ncRNA · CPC title

• G16B20/00

  ICT specially adapted for functional genomics or proteomics, e.g. genotype-phenotype associations · CPC title

• C12Q1/6883

  for diseases caused by alterations of genetic material · CPC title

• C12Q1/6886

  for cancer (immunoassay for cancer G01N33/575) · CPC title

• G16H50/70

  for mining of medical data, e.g. analysing previous cases of other patients · CPC title