KR20130004203A - Combined biomarkers information processing method for lung cancer diagnosis - Google Patents

Abstract

PURPOSE: A method for using composite biomarker information for diagnosing lung cancer is provided to enhance diagnosis efficiency. CONSTITUTION: A method for using composite biomarker information for diagnosing lung cancer comprises: a step of acquiring information in which expression level of a first biomarker group of IGF-1 or RANTES and a second biomarker group of A1AT, CYFRA21-1, proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9, or ApoA1 are measured from blood, plasma, serum or other material collected from the body(S41); a step of processing the information and inputting into a preset lung cancer model(S42); and a step of generating lung cancer determining information from the model(S43). [Reference numerals] (S41) Acquiring expression level or rate information for biomarker which forms a biomarker group measured from collected materials separated from the blood, plasma, serum, or other materials of a subject body; (S42) Processing the acquired expression level or rate information with a lung cancer diagnosis predicting module containing a preset lung cancer diagnosis predicting model; (S43) Generating at least one lung cancer diagnosis predicting information from the lung cancer diagnosis predicting module

Description

Combined Biomarkers Information Processing Method for Lung Cancer Diagnosis

The present invention relates to a method for using complex biomarker information for diagnosing lung cancer, and more particularly, to a method for using complex biomarker information for diagnosing lung cancer by using two or more biomarkers specific for lung cancer diagnosis.

Lung cancer is a cancer of the lungs, which is the leading cause of cancer, pollution, smoking, etc. In the 20th century, it started to increase rapidly in Western countries and more than 1.3 million people die from lung cancer worldwide every year, It accounts for the highest percentage of deaths from cancer. In Korea, about 100,000 new cancer cases and 50,000 deaths are reported each year. In addition, the incidence of cancer has increased more recently, and cancer is currently the second largest cause of adult death in Korea. In particular, lung cancer accounts for about 12% of the cancers occurring in Korean adults, followed by stomach cancer and liver cancer. The incidence of lung cancer is significantly higher in men than in women, with a relatively high proportion of younger patients under 45 years of age. Moreover, lung cancer has progressed locally even if it has already metastasized to another organ at the time of diagnosis or no metastasis, and despite the various treatments such as curative resection, chemotherapy, and radiation therapy, the 5-year survival rate is reduced due to recurrence and metastasis after treatment. It is a tumor with a very low cure rate of 5%, which is the leading cancer death rate.

Lung cancer is divided into small cell lung cancer and non-small cell lung cancer. Among them, non-small cell lung cancer is the most representative cancer, which corresponds to about 80% of lung cancer, and is divided into adenocarcinoma, squamous cell carcinoma, and large cell carcinoma. Accurate diagnosis is important because not only the histological characteristics are different according to the type of lung cancer, but also the prognosis and treatment methods. In non-small cell lung cancer, the 10-year survival rate is very low, below 10%, despite recent advances in cancer treatment. This is because most NSCLCs are difficult to diagnose until the advanced stage.

For now, early diagnosis is the best way to increase patient survival. Accordingly, various attempts have been made to diagnose lung cancer using biomarkers.

WO 2009/006323 A2 WO 2007/076439 A2 WO 2006/044946 A2

The first technical problem to be solved by the present invention is to propose a complex biomarker for diagnosing lung cancer.

The second technical problem to be solved by the present invention is to propose a method for using complex biomarker information for lung cancer diagnosis.

The third technical problem to be solved by the present invention is to propose a lung cancer diagnostic kit using a composite biomarker.

In order to achieve the technical problem to be achieved by the present invention, in the complex biomarker group for lung cancer diagnosis, at least one biomarker and individual biomarker A1AT selected from the first biomarker group consisting of individual biomarkers IGF-1 and RANTES, Lung cancer comprising at least one biomarker selected from the second biomarker group consisting of CYFRA21-1, proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9, ApoA1, ApoA1 / proApoA1 A diagnostic composite biomarker is presented.

The biomarkers selected from the first biomarker group are preferably IGF-1 and RANTES.

The selected biomarker is preferably one containing at least one of A1AT and CYFRA21-1.

The biomarker selected from the second biomarker group is preferably at least one of AFP, CA19-9, CYFRA21-1, A1AT, and PAI-1.

The biomarker selected from the second biomarker group is preferably at least two of A1AT, CYFRA21-1, proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9, ApoA1 / proApoA1, ApoA1.

In order to achieve the technical problem to be achieved by the present invention, in the method for using the lung cancer diagnostic complex biomarker information of the lung cancer diagnostic system, the lung cancer diagnostic system, (A) the body of the blood, plasma, serum or other subject of lung cancer diagnosis subject The expression level of each biomarker and the individual biomarkers A1AT, CYFRA21-1, at least one first biomarker group selected from the first biomarker group consisting of individual biomarkers IGF-1 and RANTES measured from the collected material separated from obtaining biomarker-specific expression measurement information of a second biomarker group consisting of proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9 and ApoA1; (B) processing the expression amount information for each biomarker of the first biomarker group and the expression amount information for each biomarker of the second biomarker group and input the biomarker into a predetermined lung cancer determination model; And (C) generating lung cancer determination information from the lung cancer determination model; suggests a method of using the complex biomarker information for lung cancer diagnostics comprising a.

Processing the expression level information for each biomarker in the step (B), when there is expression information of ApoA1 and expression level of proApoA1 in the second biomarker group, generates a ratio value of ApoA1 expression amount and proApoA1 expression amount It is preferable that any one or more of the expression amount of ApoA1, the expression amount of proApoA1, and the ratio of ApoA1 expression amount and proApoA1 expression amount are injected into the lung cancer determination model.

Processing the expression level information for each biomarker may include expressing the expression level information for each biomarker by using partial dependency plot or partial dependency function relationship of an ensemble method using a decision tree. It is preferred to produce.

* The lung cancer judgment model is preferably a logistic regression model.

Preferably, the logistic regression model uses a ridge penalty function.

In order to achieve the technical problem to be achieved by the present invention, in the lung cancer diagnostic kit, at least one protein selected from the first biomarker group consisting of individual biomarkers IGF-1 and RANTES and individual biomarkers A1AT, CYFRA21-1, A kit for diagnosing lung cancer comprising an antibody that specifically binds to at least one protein selected from the second biomarker group consisting of proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9, ApoA1 To present.

The protein selected from the first biomarker group is preferably IGF-1 and RANTES.

Preferably, the protein selected from the second biomarker group includes any one or more of A1AT and CYFRA21-1.

The protein selected from the second biomarker group is preferably one or more of AFP, CA19-9, CYFRA21-1, A1AT, PAI-1. The protein selected from the second biomarker group is preferably at least two of A1AT, CYFRA21-1, proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9, and ApoA1.

The lung cancer diagnostic kit is preferably used for the purpose of lung cancer monitoring, lung cancer screening.

Utilizing the present invention, compared to a single biomarker, it is possible to construct a complex biomarker with high lung cancer diagnosis ability, can increase the efficiency of lung cancer diagnostic method using the lung cancer diagnostic kit and lung cancer diagnostic kit, lung cancer through efficient diagnosis The survival rate of the patient can be improved and it is possible to monitor the patient's response to the treatment and change the treatment accordingly. It can also be used to identify compounds that modulate the expression of one or more biomarkers in vivo or ex vivo of animal models such as mice, rats and the like.

1 is an exemplary flow chart of a method for screening complex biomarkers effective for lung cancer diagnosis in lung cancer diagnostic biomarker candidate groups.
2 is a flowchart of an exemplary method of generating a random forest model for complex biomarker candidates for lung cancer diagnosis.
3 is a conceptual diagram of a method of generating a decision tree using a plurality of biomarkers.
4 is an exemplary diagram of a method of generating a ROC curve as an evaluation index.
5 is an exemplary diagram for a partial dependency plot of RANTES.
6 is an exemplary diagram of a boxplot of cancer patients and normal people with respect to RANTES.
7 is an exemplary diagram for a partial dependency plot of Cyfra21.1.
FIG. 8 is an exemplary diagram of a boxplot of cancer patients and normal people with Cyfra21.1. FIG.
9 is an exemplary diagram for a partial dependency plot of A1AT.
FIG. 10 is an exemplary diagram of a boxplot of cancer patients and normal people with A1AT. FIG.
FIG. 11 is a diagram illustrating an exemplary embodiment of a CP (Coeffiecient Plot) of the present invention. FIG.
12 is a diagram of one embodiment method for selecting a composite biomarker combination consisting of two or more biomarkers.
FIG. 13 is a diagram of another exemplary method of selecting a composite biomarker combination. FIG.
14 is a diagram illustrating an exemplary method for configuring a lung cancer diagnosis system of the present invention and a connection relationship with another information provider.
15 is a diagram illustrating an exemplary method of generating lung cancer diagnostic information in a lung cancer diagnostic system of the present invention.
16 is a view illustrating a method for generating a conversion variable value of a partial dependency plot / function relationship generation unit of a conversion module of a lung cancer diagnosis system of the present invention and a method of using the generated conversion variable value in the lung cancer diagnosis system of the present invention It is a figure concerning.
17 is a diagram illustrating an exemplary method of generating CP information by the CP information generator of the lung cancer diagnosis system of the present invention.
18 is a boxplot of normal samples and cancer samples for each biomarker constituting the complex biomarker group of the present invention.

A detailed explanation follows below with reference to the drawings.

1 is an exemplary flow chart of a method for screening biomarkers effective for lung cancer diagnosis in lung cancer diagnosis biomarker candidate groups. In order to select effective biomarkers for the diagnosis of lung cancer, a variable value for each sample for the lung cancer biomarker candidate group is first generated (S11), and the biomarker group for the lung cancer predictive model among the lung cancer biomarker candidate groups (S12) is selected. The complex biomarker combination generation for the selected lung cancer biomarker group is generated (S13), and the complex biomarker combination selection (S14) having excellent lung cancer diagnosis ability is generated among the generated complex biomarker combinations. This will be described in detail below.

In order to find candidates for complex biomarkers, it is necessary to first select a biomarker that is effective in diagnosing lung cancer. To this end, serum samples of normal human and lung cancer patients are first obtained, and the expression levels of proteins in normal and lung cancer patients are determined using respective protocols using RBM kits, Millipore kits, and kits prepared by the present inventors' population. The measurement result was constructed, and the data of the measurement result was constructed. For the experiment of the present invention, 128 normal (78 male, 50 female) and 121 lung cancer patients (78 male, 43 female) were included. Age distribution of normal subjects ranged from 41 to 65 years old (mean: 50.3, median: 48), and lung cancer patients ranged from 35 to 86 years old (mean: 64.7, median: 66). The distribution of lung cancer patients was stage 1 - 83, stage 2 - 14, stage 3 - stage 21, stage 4 - stage 3. In addition, 37 subjects (16 males and 21 females) and 25 lung cancer patients (10 males and 15 females) who were normal in the blind test were included to verify the classification model. 5 ml of peripheral blood was collected in a Vacutainer SST II tube (Becton Dickinson) from the normal or lung cancer patient and placed at room temperature for 1 hour, followed by centrifugation at 3000 g for 5 minutes, and then serum was obtained from the supernatant. Store at -80 ° C until now.

The inventors of the present invention are A1AT (alpha-1-antitrypsin), A2M (alpha-2 macroglobulin), DD (D-dimer), PAI-1 (total plasminogen activator inhibitor-1), VN (vitronectin), ApoA4 (apolipoprotein-A4) , Hemo (hemoglobin), proApoA1 (proapolipoprotein-A1), VDBP (vitamin D-binding protein), ApoA2 (apolipoprotein-A2), ApoC2 (apolipoprotein-C2), ApoC3 (apolipoprotein-C3), sICAM-1 (soluble intercellular adhesion) molecule-1), soluble vascular cell adhesion molecule-1 (Svcam-1), interleukin-6 (IL-6), regulated upon activation normal T cell expressed and secreted (RANTES), alpha-fetoprotein (AFP), and CA125 (cancer) antigen 125), carbohydrate antigen 19-9 (CA19-9), carcinoembryonic antigen (CEA), prostate specific antigen, free (f-PSA), prostate specific antigen (total), cytokeratin 19 fragment antigen 21 (CYFRA21-1) -1), EGFR (epidermal growth factor receptor), IGF-1 (insulin-like growth factor-1, free), ApoA1 (apolipoprotein-A1), B2M (beta-2 microglobulin), CRP (C-reactive protein), 30 groups including Hp (haptoglobin) and TTR (transthyretin) Buy a kit from different manufacturers or antibodies or produce antibodies were commissioned to analyze the quality. Information such as where to buy an antibody, kit, standard or reagent is shown in Tables 1 to 3 below.

Biomarker Standard Manufacturer Corresponding Antibody Manufacturer1 Corresponding antibody manufacturer 2 A1AT Sigma Acris Biodesign A2M Calbiochem R & D affinity bioreagents DD Abcam Biodesign Biodesign PAI-1 Calbiochem  Abcam USBiological VN Biodesign Biodesign Chemicon ApoA4 BIOINFRA Santa Cruz AB frontier Hemo Sigma Biodesign Bethyl proApoA1 BIOINFRA Biodesign Biodesign or Genscript VDBP Biodesign Abcam Abcam

Biomarker product name manufacturer ApoA2 MILLIPLEX Kit Human Apolipoprotein Millipore ApoC2 MILLIPLEX Kit Human Apolipoprotein Millipore ApoC3 MILLIPLEX Kit Human Apolipoprotein Millipore sICAM-1 MILLIPLEX Kit Human Cardiovascular Disease panel 1 Millipore Svcam-1 MILLIPLEX Kit Human Cardiovascular Disease panel 1 Millipore IL-6 MILLIPLEX Kit Human Cytokine / Chemokine 2 Millipore RANTES MILLIPLEX Kit Human Cytokine / Chemokine 1 Millipore AFP RBM Cancer Antigen Panel 1 RBM CA125 RBM Cancer Antigen Panel 1 RBM CA19-9 RBM Cancer Antigen Panel 1 RBM CEA RBM Cancer Antigen Panel 1 RBM f-PSA RBM Cancer Antigen Panel 1 RBM PSA RBM Cancer Antigen Panel 1 RBM CYFRA21-1 TM-CYFRA21.1 ELISA kit DRG Diagnostics EGFR DuoSet IC ELISA R & D IGF-1 DuoSet IC ELISA R & D

Biomarker Medicine Standard material manufacturer ApoA1 N Antiserum to human Apolipoprotein N Apolipoprotein standard SL Siemens B2M N Latex beta2-microglobulin N Protein standard SL Siemens CRP CardioPhase hsCRP N Rheumatology standard SL Siemens Hp N Antiserum to human Haptoglobin (SMN 10446304) N Protein standard SL Siemens TTR N Antiserum to human PreAlbumin N Protein standard SL Siemens

For standard proteins, ApoA2, ApoC2, ApoC3, sICAM-1, Svcam-1, IL-6, and RANTES proteins are included in the Millipore kit, AFP, CA125, CA19-9, CEA, f-PSA, PSA proteins Is included in RBM's kit, CYFRA21-1 protein is included in DRG Diagnostics' kit, EGFR, IGF-1 protein is used in R &D's kit, ApoA1, B2M, CRP, Hp, TTR protein Was purchased from Siemens, A1AT, Hemo protein was purchased from Sigma, A2M, PAI-1 protein was purchased from Calbiochem, DD protein was purchased from Abcam, VN, VDBP The protein was purchased from Biodesign, and ApoA4 and proApoA1 proteins were prepared and used in Bioinfrastructure (Korea).

If necessary, antibody-bound microspheres were prepared by the following method. First, the microsphere stock solution (Hitachi, Japan) was vortexed and then suspended in a sonication vessel (Sonicor Instrument Corporation, USA) for 20 seconds. 2 × 10 ⁶ microspheres were transferred to a microtube to remove the supernatant by centrifugation, washed with 100 μl of tertiary distilled water, and again in 80 μl of 0.1M sodium phosphate buffer (pH 6.2). Resuspend. Thereafter, 50 mg / ml of N-hydroxy-sulfosuccinimide (Sulfo-NHS) and 1-ethyl-3- (3-dimethylaminopropyl) -carbodiimide hydrochloride (1- 10 µl each of ethyl-3- (3-dimethylaminopropyl) -carbodiimide hydrochloride (Pierce, USA) was treated sequentially and mixed for 20 minutes at room temperature. The supernatant was removed by centrifugation, followed by 50 mM MES, pH 5.0. Wash twice.

Subsequently, the carboxyl-activated microspheres were resuspended with 400 μl of 50 mM MES, and then mixed with 100 μl of 50 mM MES including 25 μg of antibody to be bound, followed by mixing at room temperature for 2 hours. . The reaction was run in the dark. After completion of the antibody binding reaction, the microspheres were washed twice with 500 μl PBS-TBN [PBS, 1% BSA, 0.02% Tween, 20-0.05% sodium azide] using centrifugation. The number was measured by a hemocytometer. Microspheres bound the antibody was stored at 4 ℃ in a dark room with at ^{1 × 10 6/500 ㎕ PBS} -TBN levels.

Subsequently, the antibody-binding microspheres were vortexed and sonicated for 20 seconds to measure the antibody binding efficiency of the antibody-binding microspheres prepared above, and then 2,000 microspheres per well were placed in a filter-bottom 96-well microplate. Depending on the species of antibody bound to the sphere, Phycoerythrin-bound secondary antibody (anti-antibody antibodyPE conjugate, Jackson Immunoresearch, USA) was diluted 1/10 in a 2% BSA / PBS solution. Put into μL / well and mix for 30 minutes at room temperature. The reaction was carried out in the dark so that no light entered. After the reaction was washed twice with PBST and read by Luminex ^TM 200 (Luminex, USA) to confirm that the MFI value is more than 10,000.

Subsequently, biotinylated antibodies were used as detection antibodies. Specifically, the biotinylation reaction was performed according to the manufacturer's method using the EZ-Link Sulfo-NHS-Biotinylation Kit (Pierce, USA), and the degree of biotin binding was determined by HABA (4'-hydroxyazobenzene) included in the kit. -2-carboxylic acid) was carried out according to the kit manufacturer's instructions. As a result, the amount of bound biotin per antibody was measured to 8-12.

Subsequently, the developed assay further optimized the concentration of the detection antibody and the reaction time of the experiment, and the sensitivity was confirmed by analytical measurement values of serially diluted biomarkers. Intra-assay variability is a measure of the CV of nine different concentrations of serum samples in two wells of 12 wells / plate at three different time points. variation) was calculated and averaged 5% to 10%. The developed kit was confirmed to have no cross-reactivity.

Immunoassay of AFP, CA125, CA19-9, CEA, f-PSA, PSA was performed on 96 well V type microplates according to RBM's protocol. At this time, the standard protein provided by the manufacturer was used by serial dilution with serum matrix diluent. Specifically, 20 μl of duplication protein, control serum and patient serum were added to the wells, and 10 μl of the blocking buffer and bead mixture included in the kit were added to the wells. After adding to the mixture and reacted at room temperature for one hour. The detection antibody and streptavidin-PE (Jackson Immunoresearch, USA) were allowed to react sequentially for one hour and 30 minutes, respectively, and the reaction solution was transferred to a filter-bottom 96-well microplate (Millipore, USA), followed by vacuum manifold. Washed twice with (vacuum manifold). The reaction solution treated with 100 μl of the assay buffer included in the kit was transferred to a 96 well microplate and analyzed by Luminex ^™ 200 (Luminex, USA). The results were analyzed by 5-parametric-curve fitting using the beadview software of Upstate, USA.

Immunoassay of ApoA2, ApoC2, ApoC3, sICAM-1, Svcam-1, IL-6, RANTES was performed in a filter-bottomed 96-well microplate (Millipore, USA) according to Millipore's protocol. The filter-bottom 96-well microplates were treated with the assay buffer provided in the kit, blocked for 10 minutes, and then the buffer was removed using a vacuum manifold. The standard protein provided by the manufacturer was used by serial dilution with serum substrate diluent. Specifically, duplicate proteins, duplication serum and patient serum were treated in wells of 25 쨉 l each, and 25 쨉 l of bead mixture was added to each well, followed by reaction at room temperature for one hour. After the reaction plate was washed twice using a vacuum manifold, the detection antibody and streptavidin-PE were sequentially reacted for 1 hour and 30 minutes, respectively. By the analysis buffer solution from the following kits washed the plate the reaction was finished, process 100 ㎕ Luminex ^TM 200 was analyzed. Results were analyzed with 5-parametric curve fitting using Upstate's BeadView software.

Immunoassay of A1AT, A2M, DD, PAI-1, VN, ApoA4, Hemo, proApoA1, VDBP was performed in a filter-bottomed 96-well microplate (Millipore, USA) according to BioInfrasa's protocol. The filter-bottom 96-well microplates were treated with assay buffer (PBS / 2% BSA) for 10 minutes to block and then the buffer was removed using a vacuum manifold. The standard protein provided by the manufacturer was used by serial dilution with serum substrate diluent. Specifically, duplicate proteins, duplication serum and patient serum were treated in wells of 25 쨉 l each, and 25 쨉 l of bead mixture was added to each well, followed by reaction at room temperature for one hour. After the reaction plate was washed twice using a vacuum manifold, the detection antibody and streptavidin-PE were sequentially reacted for 1 hour and 30 minutes, respectively. By the analysis buffer solution from the following kits washed the plate the reaction was finished, process 100 ㎕ Luminex ^TM 200 was analyzed. Results were analyzed with 5-parametric curve fitting using Upstate's BeadView software.

ApoA1, B2M, CRP, Hp and TTR were analyzed by automated method using Behring Nephelometer II (BNII) System according to the manufacturer's instructions.

Cyfra21-1 was analyzed according to the instructions included in the DRG Diagnostics kit, EGFR, and IGF-1 in the R & D DuoSet IC ELISA kit.

Table 4 shows an example of the measurement result data for each biomarker for each sample, thus generating a variable value for each sample for the lung cancer biomarker candidate group (S11). The variable value may be ratio information of the expression level for each biomarker or the expression level for each of the two or more biomarkers.

Sample.ID class Age Sex Stage.S ApoA2 Svcam.1 ............ PAI.1.1 LC01 Lung cancer 53 M One 5.359 2.738 ............ 3.171 LC04 Lung cancer 66 M One 5.617 2.943 ............ 2.950 LC05 Lung cancer 60 M 3 5.385 2.914 ............ 2.770 LC07 Lung cancer 43 F One 5.463 2.752 ............ 2.743 ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ ............ KNF140 normal 51 F 5.600 2.936 ............ 3.116 KNM378 Nor 56 M 5.443 2.923 ............ 3.116 KNF088 Nor 48 F 5.458 2.967 ............ 3.036 KNM151 Nor 55 M 5.542 3.077 ............ 2.986

sample .ID: The sample's unique ID given during the experiment. class: According to the sample classification, Nor is a normal person, Can is a lung cancer patient. Age is age, Sex is sex, Stage.S is stage information of lung cancer (normal: blank, cancer: 1-4), and subsequent columns are biomarkers tested on the biomarker list. The cell value is a list of experimental values of the biomarker candidates, and the experimental values of the input data shown in Table 2 are values obtained by log conversion of the experimental values.

The measurement data we built is based on bioinformatics and statistical analysis, the R Package (R Development Core Team (2007) .R: A language and environment for statistical computing.R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051 -07-0, URL http://www.R-project.org. A random forest algorithm was applied to the input data to determine variable importance, derive a p-value ranking, and perform correlation analysis between biomarkers. In this way, if a random forest ranking includes a biomarker with high correlation with reference to a p-value rank and excludes the lower rank biomarker, 13 lung cancer biomarker candidates will be included in the lung cancer prediction model. Biomarker group was selected (S12).

 Selected biomarkers are A1AT, CYFRA21-1, IGF-1, AFP, proApoA1, EGFR, CEA, RANTES, PAI-1, TTR, CA19-9, ApoA1 / ProApoA1, ApoA1. Table 7 below shows the evaluation index values for the 13 selected biomarkers and each individual biomarker. ApoA1 / proApoA1 is a value obtained by dividing the expression level of ApoA1 by the expression level of proApoA1, which is an example of the expression rate ratio, and shows that the expression rate ratio can be a biomarker. ApoA1 / proApoA1 is a value obtained by dividing the expression level of ApoA1 by the expression level of proApoA1, which is an example of the expression rate ratio, and shows that the expression rate ratio can be a biomarker.

Biomarker accuracy
(accuracy) responsiveness
(ensitivity) Specificity
(specificity) A1AT 0.8326 0.7731 0.8921 CYFRA21-1 0.8525 0.8511 0.8538 IGF-1 0.8035 0.8515 0.7556 RANTES 0.7644 0.7479 0.7809 proApoA1 0.7575 0.6859 0.8291 AFP 0.7347 0.8082 0.6612 EGFR 0.7362 0.6895 0.7829 PAI-1 0.7315 0.6928 0.7703 TTR 0.7156 0.698 0.7332 CEA 0.6869 0.7226 0.6512 CA19-9 0.686 0.7705 0.6015 ApoA1 / proApoA1 0.6583 0.4849 0.8318 ApoA1 0.6679 0.6364 0.6994

Accuracy is cancer, normalized rate, sensitivity is cancer rate, and specificity is normal. An example of how to obtain the above-mentioned evaluation index Sensitivity, Specificity, and Accuracy will be described. It demonstrates that cut-off = 0.5. Assume there is data as shown in Table 6 below.

Actual value (Y axis) Predicted value (X axis) 0 (normal) 1 (female) 0 (normal) 17 3 1 (female) 0 20

If the number of test sets is 40 (normal 20, cancer 20), draw the cross table of the actual value and the predicted value as above. The actual value is 0 (normal), but the number predicted to 0 (normal) is 17, and the actual value is 0 (normal), and the number predicted to 1 (dark) is three. The actual value is 1 and the number predicted by 0 is 0. The actual value is 1 and the number predicted by 1 is 20. Sensitivity is the probability of predicting actual cancer patients as cancer patients. In the table above, 20 out of 20 predicted cancer, so the sensitivity is 100%. Specificity is 85%, because 17 out of 20 people have a probability of predicting normal people. Accuracy is 92.5% because 37 people out of 40 were predicted correctly with the same ratio of actual value and predicted value, that is, the probability of predicting normal to normal and cancer patients as cancer patients.

Meanwhile, in the present invention, sensitivity, specificity, and accuracy are used as evaluation indicators, but various evaluation indicators introduced in the field of statistics or social science may be used. Of course, the present invention naturally introduces such various evaluation indicators. Of course, the biomarkers can be selected through these evaluation indicators. Meanwhile, the ranking of the selected evaluation indicators may be based on any one of the evaluation indicators, but a predetermined function that takes at least one or more evaluation indicators as an input value or a predetermined importance function calculated independently of the evaluation indicators. Of course it is possible. A predetermined function that takes at least one evaluation indicator as an input value or a predetermined importance function calculated independently of the evaluation indicator is called an evaluation function, and a value calculated by the evaluation function is called an evaluation function value.

FIG. 18 is boxplot for normal and cancer samples for the 13 biomarkers. FIG.

Table 7 below summarizes the selected 13 biomarkers, expression patterns and characteristics of each biomarker. The expression pattern is roughly classified into a case where the higher the expression value of each biomarker, the higher the probability of cancer, and the lower, the higher the probability of cancer. In Table 5, Can (high) corresponds to the former, and Can (low) corresponds to the latter.

Biomarker pattern characteristic A1AT Can A1AT is a glycoprotein and is known as an antagonist of serum trypsin. In the body, it protects tissues from various breakdown enzymes (especially elastase) secreted by inflammatory cells and increases in the acute inflammatory phase. Deficiency is associated with birth defects that result in destruction of lung tissue. Hamrita et al. Reported an increase in A1AT in invasive mammary cancer. AFP Can In adults, the highest frequency of germ cell tumors and liver cancer
But also increases in gastic, colon, biliary, pancreatic and lung cancer (~ 20% of patients)
CA19-9 Can Clinically used because it increases in the majority of patients with pancreas, biliary tract, colon, stomach, and breast carcinoma CEA Can Gastrointestinal (GI). Increased in serum of cancer patients of lung, breast, ovary and uterus
CYFRA21.1 Can CYFRA 21-1 (a cytokeratin 19 fragment) is known to be associated with non-small cell lung cancer, and Lai et al. Reported that not only high blood levels but also stage and prognosis, especially in squamous cell carcinoma. EGFR Can Receptor of EGF, involved in cell growth and differentiation IGF-1 Can The expression of IGF-1 is increased in adenocarcinoma of various organs, so Ouban et al. Have endometrial cancer (100%), breast cancer (87.5%), ovarian cancer (100%), gastric cancer (71.1%), and pancreatic cancer (57.1%). , Lung cancer (90.0%), lung cancer (84.6%), such as the tissues are well expressed, but the expression of squamous cell carcinoma of the head and neck is reported to be less. Furstenberger et al. Also reported the association of blood IGF-1 levels with breast cancer, prostate cancer, lung cancer and lung cancer. In other words, IGF-1 is an important mediator in the role of growth hormone, which affects the differentiation and growth of cells and increases their ability to interfere with apoptosis. PAI-1 Can Inhibitor of tissue plasminogen activator (t-PA) and an important enzyme in fibrinolysis. Increasing PAI-1 decreases t-PA activity and impairs fibfinolytic function. increased in deep vein thrombosis, myocardiac imfarction, normal pregnancy, sepsis ApoA1 Can It is a component of HDL (high density lipoprotein) and acts as a cofactor of LCAT (lectin cholesterol acyltransferase) and participates in the transport of cholesterol from tissue to liver. proApoA1 Can Pro form of Apolipoprotein A1 RANTES Can Chemotactic factor for T-cell, eosinophil, basophils
Attracts white blood cells to inflammation sites
asthma, related to allergic rhinitis TTR Can Thyroid hormone-binding protein. Probably transports thyroxine from the bloodstream to the brain.
Defects in TTR are the cause of amyloidosis type 1 (AMYL1). A hereditary generalized amyloidosis due to transthyretin amyloid deposition.

Next, referring to FIG. 2, a method for screening a composite biomarker for lung cancer diagnosis will be described.

First, a composite biomarker combination for the selected lung cancer biomarker group is generated by generating a composite biomarker list that can be combined with 13 biomarkers selected primarily by feature selection (S13). The number of combination biomarker combinations is 13Cr (14> r> 1), which is a total of 8178. For each complex biomarker combination, each cancer / normal predictive statistical model is made, and 8178 statistical models are compared based on the evaluation indexes (Accuracy, Sensitivity, Specificity, etc.) obtained from each model.

The statistical model provides the best model for the data used to create the model. If you create a model with one data set, there is no way to verify that the statistical model works well with normal data. For this reason, create a training set and a test set. For example, if the sample size is 200 (100 arms, 100 normal), randomly 100 (50 arms, 50 normal) can be extracted and used as a training set, and the remaining 100 can be used as a test set. have. (When a sample size is given, how much to use as a training set and how much to use as a test set may vary. Typically, the size of a training set is greater than or equal to the size of a test set.) Create a model using the training set, and then apply the test set to the model (predict the cancer / normal of the test set) to verify how well the given model works by comparing the actual and predicted values. Repeating this multiple times rather than just once with "model creation with training set - model verification with test set" helps to make a more robust model (a more global model that relies less on specific data).

Next, the decision tree will be described. Decision trees are one of the data mining analysis techniques, and they can be used to find decision rules based on the structure of trees. Decision trees are powerful and widely used analytical techniques that tabulate decision rules to classify or predict groups of interest into several subgroups. The general algorithm of decision trees has different formation processes in terms of stopping rules and pruning. The rules used in decision trees are:

1. Separation Criteria: The child nodes are formed by identifying how predictive variables are used to best distinguish the distribution of target variables. It is measured using.

2. Stopping Criteria: A rule that specifies that no further separation takes place and that the current node becomes a terminal node.

3. Pruning: Decision trees with too many nodes can have very high prediction errors when applied to new data. Therefore, it is desirable to select a decision tree having a sub tree structure of a suitable size as a final model by removing inappropriate nodes from the formed decision tree.

If the target variable is discrete (for example, cancer / normal), if the separation occurs based on the frequency belonging to each category of the target variable, a classification tree is formed.

For example, if the biomarker CYFRA21.1 value is greater than 5, the probability of cancer is very high. Of the 100, 50 patients with a CYFRA21.1 value of 5 or more actually had 40 normal cancer patients and 10 normal patients, and CYFRA21. Among 50 patients with a value less than 5, 10 cancer patients and 40 normal patients are summarized in Table 8 below.

Cancer Normal Total CYFRA21.1> 5 40 10 50 CYFRA21.1 <5 10 40 50

Table 8 shows a case where only the CYFRA21.1 value is used. The data is further divided by applying additional criteria to the divided data (CEA <3,> = 3 or CEA <4,> = 4), which is well illustrated in FIG. It demonstrates, referring FIG. For example, if the CYFRA21.1 value of person A is 5 and the CEA value is 4.5, according to the decision tree used as an example, the biomarker combination value corresponds to terminal node 3. According to the principle of majority vote, more than half of Terminal Node 3 is "cancer", so person A is determined to be "cancer". On the other hand, if the biomarker value of person B is CYFRA21.1 = 7.0 and CEA = 2.0, person B enters Terminal Node 4 and is determined as "normal".

Next, the RF algorithm will be described. Random forest (Random forest, RF; Breiman L, Machine Learning 45 (1): 5-32, 2001) is a method of Bagging algorithm that is a combination of CART's decision trees and is proposed by Leo Breiman and Adele Cutler. . The nodes of each tree are organized so that the data with higher dimensions can be broken down into smaller pieces of lower dimensions. Each of these trees completes the final classification by ensemble and voting. Trees generated by random vectors with the same probability distribution are composed independently, and when the number of trees is infinite, the misclassification is generalized and converged. RF is random and out-of It uses the -bag (Random Selection without Replacement) technique to achieve the same accuracy as the Adaboost, shows strong performance on the interface and noise, and helps convergence faster than bagging and boosting.

The RF algorithm generates a decision tree from each of its own (training data set, test data set) (for example 50), the number of which can be optionally adjusted by the user) from the given data. This creates 50 independent decision trees. After creating 50 decision trees, if you put a test set, you get 50 decisions (cancer / normal) per test sample (values from each decision tree) vote) gives the final result. For example, in the case of person A, if 45 decision trees were determined to be cancer and 5 decision trees were judged to be normal, then the score score would be calculated as 45/50 = 0.9. . In this case, when the cut-off value, which is a standard for distinguishing cancer / normal, is assumed to be 0.5, the average score 0.9 of A is greater than 0.5, and thus it is determined as “cancer”.

In this case, a method of synthesizing a decision made from a plurality of statistical models (RF in the case of a decision tree) to a final decision is called an ensemble technique. The present invention is characterized by using such an ensemble technique. In addition to the RF algorithm, there is also a boosting algorithm, which is equivalent in terms of using an ensemble technique. It will be appreciated that a boosting algorithm may be easily employed by those skilled in the art to implement the spirit of the present invention, and it will be obvious that the boosting algorithm is included in the implementation of the present invention.

The basic idea of boosting is to combine multiple weak learners into a strong learner. At this time, weak learner is a better classifier than random guessing, meaning accuracy is 0.5 or more, and can be any statistical classifier such as decision tree and logistic regression. Strong learner means a classifier whose accuracy is much better than random guessing. The algorithm is as follows.

1. When there are N data, all are equally weighted with Wi = 1 / N.

2. Apply weak classifer # 1 to the data using the given weight.

3. Increase the weight of misclassified data with Weak classifer # 1, and decrease the weight of correctly-classified data.

Apply weak classifier # 2 to the data using the re-calculated weights in Section 3.

5. Increase the weight of the data misclassified by Weak classifier # 2 and decrease the weight of the data classified.

like this

Step 1: create weak classifier using given weight

Step 2: Recalculate the weight according to the misclassification / correct classification by the weak classifier.

(Step 1. + Step 2) Repeat the work until the appropriate stop criterion is satisfied. For example, suppose you have 10 weak classifiers. The final result is then synthesized from these 10 weak classifiers.

In this manner, a plurality of cancer / normal prediction statistical models are generated for each of 8178 total biomarker combinations, which are all possible combinations of complex biomarkers, and then an optimal cancer / normal prediction statistical model is selected. If there is a specific complex biomarker combination of complex degree n (X1, X2, ... Xn), the sample among n used biomarkers is divided into a training set, and for a sample belonging to the training set, n A plurality of decision trees as shown in FIG. 3, in which any one or more of the plurality of complex biomarkers participate, are generated, and a plurality of cancer / normal prediction statistical model candidate groups are generated by using the ensemble technique. A test set is constructed of samples that do not participate in the training set for a plurality of cancer / normal predictive statistical model candidate groups, and the prediction performance is verified for the test set. Predictive performance may be an evaluation index or the like. Since there are a great number of methods / combinations for dividing the entire sample into a training set and a test set, it will be obvious that the cancer / normal predictive statistical model candidates will be plural.

Person A's Avg.Score represents the ratio of cancer among n cancer / normal decisions from n decision trees. In the case of a random forest, one prediction model is a collection of multiple decision trees created using specific marker combination information (eg, RANTES + CYFRA21.1).

The generated cancer / normal prediction statistical model candidate group may have a form as shown in Equation 8 below. (X1, X2, ... Xn), the X value specified for the samples utilizing n complex biomarkers is input to the prediction model or the decision tree as shown in Equation 8 below. A plurality of cancers / normals (e.g., expression values of biomarkers called RANTES or expression value ratio information such as ApoA1 / proApoA1) or special treatments (e.g., partial dependency plots / function relationships) When the X values of the samples are input to the predictive statistical model candidate group or each decision tree, each of the model candidate groups or the decision tree has a decision value equal to a value between 0 (normal) and 1 (dark). The average value (Avg.Score) is generated as follows: Of course, since the correct answer value for cancer / normal is known for each sample, any model among the plurality of cancer / normal prediction statistical model candidate groups or each decision tree is the best evaluation index. It has it is possible to determine the map.

Of course, among the cancer / normal prediction statistical model candidate groups, there are many models that ensemble the plurality of decision trees used in the ensemble technique.

Then, through each cancer / normal prediction statistical model, the data shown in Table 9 for each model is obtained.

If Avg.Score exceeds 0.5, it is determined to be cancer, otherwise it is determined to be normal. Of course, the cut-off of 0.5 is only a special example, and may be changed to any number between 0 and 1 according to circumstances. In this way, if Avg.Score is calculated for the composite biomarker and cancer and normal determination are obtained for each composite biomarker, the data shown in Table 9 can be obtained, and the sensitivity of each cancer / normal prediction statistical model is obtained from this data. In addition, diagnostic or predictive performance indicators such as specificity and accuracy can be generated. The cut-off point of the average score is needed to determine whether the cancer is normal with the average score in the RF. In other words, in the above example, if the average score is more than one, cancer is judged to be normal if the average score is over 0.5, otherwise it is determined to be normal. Will be different. The larger the cut-off value, the smaller the ratio determined by the cancer, and the smaller the cut-off value, the larger the ratio determined by the cancer. When the cancer / normal judgment is affected in this way, evaluation index values such as sensitivity and specificity values also vary. Therefore, the cut-off value is varied and the corresponding evaluation index value (sensitivity, 1-specificity) can be plotted. For example, if you use cut-off values of 0.01, 0.02, 0.03, 0.04, .., 0.98, 0.99, 1, you can get the corresponding (sensitivity, 1-specificity) values. These may be displayed in the secondary plane using x and y coordinates, respectively, and an exemplary drawing thereof is shown in FIG. 4. In FIG. 4, the blue line (the line where the sensitivity (Sn) and 1-specificity (Sp) values are specified, and the arc-shaped line in the secondary plane) corresponds to the ROC curve. It will be close to the upper left corner of the box (x = 0.0, y = 1.0 in coordinates). In this case, the area under curve (AUC) near the curve becomes close to 1. ROC curve is a way to compare the performance of the model at the same time in terms of sensitivity and specificity.The closer the area under the curve is to 1, the better the statistical model. You can also use this ROC curve to find cut-off values. .

Subsequently, the composite biomarker combination having excellent lung cancer diagnosis ability is selected among the generated complex biomarker combinations (S14). The following provides an example of how to determine which complex biomarker combination is a more reasonable combination. The combination of the 13Cr complex biomarkers (all individual combinations form one or more cancer / normal predictive statistical models. Of course, the optimal statistical model can be selected for these one or more cancer / normal predictive statistical models ), We calculate the importance of each biomarker in each statistical model.

Importance refers to the magnitude of the association of certain biomarkers to cancer / normal findings in certain statistical models. As shown in FIG. 3, four terminal nodes (nodes at the end of the tree) were generated using two biomarker values of CYFRA21.1 and CEA. It is determined whether it is cancer or normal according to the majority value of the node. A large portion of the sample is divided into cancer / normal by the value of CYFRA21.1 used initially, which means that the CYFRA21.1 biomarker value is significantly associated with cancer / normal. To measure the importance of this CYFRA21.1, we randomly permute the CYFRA21.1 value. That is, since the CYFRA21.1 values are randomly mixed and then assigned to each patient, the correlation between cancer / normal and CYFRA21.1 is almost eliminated. When the randomly permuted data is inserted into the decision tree, the difference between the correct decision ratio (2) at each terminal node and the correct decision ratio (1) at each terminal node is measured when the original data of CYFRA21.1 is used. This value is the importance of CYFRA21.1. If there is a clear pattern of biomarker values by cancer / normal, then the correct decision ratio when using that pattern and the random decision permuted when the biomarker value is ignored becomes large. Conversely, if the biomarker is not related to cancer / normal, there is no significant difference in the correct decision ratio when using the original data or randomly permute.

As described above, when the importance of all the biomarkers participating in each statistical model in each statistical model can be calculated, the importance ranking (sequence) of the biomarkers participating in the statistical model can be given. For example, if there is a statistical model that includes a biomarker combination of IGF.1 + CYFRA21.1 + RANTES, the significance of the biomarkers IGF.1, CYFRA21.1, and RANTES can be seen from this statistical model. For example, the importance ranking could be CYFRA21.1 first, IGF.1 second, RANTES third, and so on. At this time, when all of the 8178 statistical models can know the importance value and importance ranking for each biomarker participating in each model, superior composite biomarkers can be selected using the importance value and importance ranking. Methods of selecting superior biomarkers using the importance value and importance ranking value may be various, but one exemplary method as follows is provided as an example.

From all 8178 statistical models, we can extract the biomarkers of importance ranking 1st and 2nd ranking, and a total of 8178 "ranking top biomarkers + ranking 2nd biomarkers" will be created. With respect to each of the extracted "ranking biomarker + ranking second biomarker", it is possible to calculate the frequency for each "ranking biomarker + ranking second biomarker". At this time, when calculating the frequency of "ranking biomarker + ranking biomarker 2 ranking", even if the sequence of the ranking biomarker and ranking 2nd biomarker different handling method (combination method) and different handling method (Permutation method). In the combination method, "IGF-1 + CYFRA21.1" and "CYFRA21.1 + IGF-1" are the same. In other words, if IGF-1 or CYFRA21.1 is placed first and any one is placed second in all statistical models, "IGF-1 + CYFRA21.1" is equally frequency 1 added. . On the other hand, in the permutation method, the model where IGF-1 ranked first, CYFRA21.1 placed second, and CYFRA21.1 placed first, and IGF-1 placed second, were treated separately. . In other words, "IGF-1 + CYFRA21.1" and "CYFRA21.1 + IGF-1" are handled differently.

Meanwhile, in the combination method, an important biomarker combination may be found based on frequency, including the ranking nth position, such as the third and third rankings, not only the first and second rankings. Alternatively, the permutation method may be used to generate an important biomarker combination by calculating a frequency for the biomarker combination up to the ranking n (n> 1).

In addition, by assigning a weight to each ranking n position (for example, the weight may be the importance value itself, and the weight is given according to a random or statistical basis, such as a weighting method of 1 for a ranking 1 and a weight of 0.5 for a 2nd ranking). In the combination method or the permutation method, an important biomarker combination that reflects both frequency and weight may be found.

Through the above process, relative superiority index value can be calculated for every 13Cr composite biomarkers. Relative superiority is an indication of how much superiority there is with a specific composite biomarker compared to other composite biomarkers.

Meanwhile, performance such as sensitivity, specificity, accuracy, etc. may be calculated for each of the 13Cr complex biomarker combinations, and an optimal complex biomarker may be selected as a performance value for the complex biomarker combination. Sensitivity, specificity, accuracy, etc. are only examples of the performance of each complex biomarker (each complex biomarker equals 1: 1 with the statistical model), and it is obvious that other performance indicators can be calculated, and the ROC curve The area underneath is an example.

The number of single biomarkers participating in the composite biomarker when selecting the composite biomarker is called the composite degree. For example, the composite degree is 2 for IGF-1 + CYFRA21.1 and IGF-1 + CYFRA21.1 + RANTES As the complexity increases, the performance tends to improve as performance indicators such as sensitivity, specificity, accuracy, and the area under the ROC curve can be used. There may be problems such as: 1) increase in manufacturing costs, 2) increase in information processing costs / difficulties, such as data collection and analysis, and 3) increased likelihood of statistical correlation between measurements. In addition, when a complex biomarker of complex degree n, which is a combination of specific biomarkers, provides sufficient and satisfactory performance, the net performance (performance increment) may not be large when additional biomarkers are combined. Therefore, when increasing the complexity, it would be reasonable to increase the complexity by taking into account the net performance and the cost of the complexity increment if it exceeds the lower acceptable limit of performance. In other words, it would be reasonable to have a large value of Benefit variance / Cost variance with increasing complexity. On the other hand, when increasing the complexity, which biomarker is used can be judged by the performance value. For example, if there are enough performances with about 5 biomarker combinations (5 complex biomarkers), if there is no difference in performance even if one or more additional biomarkers are combined, there are about 5 biomarker combinations. Lung cancer diagnostic biomarkers may be made.

Table 10 below shows an example showing the variation (increase) of each evaluation index while adding one biomarker to IGF-1 + Cyfra. As can be seen in Table 10, it can be seen that each evaluation index is saturated as the number of biomarkers is increased. If 93% is sufficient based on accuracy (accuracy 93% is cut off), a model with "IGF-1 + CYFRA21.1 + A1AT + RANTES + CEA + CA19-9" composite biomarker will be sufficient. It may be possible that a model that adds a TTR to this model may be unnecessary.

M_01 M_02 M_03 M_04 M_05 M_06 Accuracy Sensitivity IGF-1 CYFRA21.1 0.8629 0.8213 IGF-1 CYFRA21.1 A1AT 0.8895 0.8708 IGF-1 CYFRA21.1 A1AT RANTES 0.9238 0.9226 IGF-1 CYFRA21.1 A1AT RANTES CEA 0.9266 0.919 IGF-1 CYFRA21.1 A1AT RANTES CEA CA19-9 0.9300 0.9207 IGF-1 CYFRA21.1 A1AT RANTES CEA CA19-9 TTR 0.9315 0.9236

The cell values of the tested biomarkers in Table 2 are a list of experimental values of biomarker candidates, and are values obtained by log conversion. As the experimental values are measured values, errors may occur, and when an outlier exists, the outliers may be a major factor that degrades the evaluation model of the statistical model when used as it is. Therefore, there is a need for a method for effectively removing, minimizing, or correcting outliers. An effective method that can be taken is a technique using a decision tree. The classification tree model ranks given data and partitions the data repeatedly. Each partition is aimed to have all or most of one response value. There are various classification techniques such as Bagging, Boosting, and Random Forest in the ensemble technique using the tree. The ensemble technique uses Decision Tree nodes to create multiple trees and combine them to create more stable and powerful classifiers. Boosting is a technique for creating highly accurate classification models by combining several weak classifiers (typically slightly better performance than random choices). Boosting can also take into account the interaction term of a variable, and the importance of the variable is also observed. Random forest is a method of constructing a large number of classification tree models and combining them randomly instead of building one best classification tree model. The advantages of random forest are good classification accuracy, insensitive to outliers, and quick and simple calculations.

Here, we will explain how to transform the data using partial dependence plots of ensemble techniques (Boosting and Random Forest) to minimize the effects of outliers when constructing cancer / normal prediction models.

In the actual measurement of X (variable), such as the amount of expression of each biomarker, there are various reasons for outliers. When these outliers are used as they are, the outliers included in the sample cause severe distortion of the model even when generating a predictive model. Even when the predictive model is applied, if there is an abnormality in the measured value of the patient or the like, the possibility of significant distortion in the cancer / normal determination increases. This is especially the case when using a complex biomarker combination, when there is an outlier in a particular biomarker included in the combination, the outlier can have a significant impact on the overall judgment model value. There is a need to reduce the impact of direct reflection of these outliers. The decision tree is inherently based on classification, so even if there are outliers, the outliers are not directly reflected, only the relative order, ranking, or correspondence of the outliers, so that the impact of the outliers Greatly reduced.

The logic to eliminate outliers is explained in more detail. Partial dependence plots are intended to show the marginal effect of a particular variable value on a response variable (cancer / normal). In general, the partial dependence plot function relationship is obtained as follows. Random forest is applied first with two biomarker combinations X = (Xs, Xc). For example, suppose that there are 50 decision trees in the random forest. Combining the 50 decision tree results, the following function f (Xs, Xc) can be obtained for each patient's biomarker value X = (Xs, Xc).

f (Xs, Xc) = f (X) = log (p (X) / (1-p (X)))

P (X) corresponds to the rate at which the patient with the marker combination X was selected as cancer in the 50 decision trees, that is, Avg.Score. In this way, the function values f (Xs, Xc) can be calculated for all patients. If you want to determine the partial dependence of the first biomarker (for example, RANTES, the current example is XS), collect patients with the same Xs value and average their f (Xs, Xc) values (g (XS) Let's say).

For example, f (90, Xc) of patients with marker RANTES value Xs = 90 are collected and averaged (g (90)). ,

The f (65, Xc) values of patients with RANTES value Xs = 65 are collected and averaged (g (65)).

In this way, if you find the average (g (Xs)) of f values with the same Xs value,

You can get pair values like (Xs = 90, g (90)), (Xs = 65, g (65)),

If we draw this Xs with the x-axis and g (Xs) with the y-axis,

We can find the marginal effect of Xs on f, and this function is a partial dependence plot.

At this time, since the decision tree used when estimating f (Xs, Xc) from the original data is an algorithm that uses the order rather than the actual value of the data, it can be more insensitive to outliers.

Partial dependence plot or partial dependence function relationships remove the effect of the rest of the variables on a single variable. For example, if an input variable has a joint distribution consisting of two variables, Xs and Xc, and you want to know the effect on the Xs variable, you can average the joint distribution over the Xc variable. A partial dependency function relationship can be created for each X, and the partial dependency function relationship corresponds to a partial dependency plot. Using these partial dependency function relationships or partial dependence plots, X can be transformed. That is, for each of the two or more samples, the original variable values for each biomarker are obtained for each sample (S51), and the partial processing of the partial dependence plot or partial dependence function for each biomarker is performed by performing a predetermined process with the original input variable values for each biomarker. (S52), the biomarker-specific conversion variable values are generated for the original biomarker-specific variable values using the partial dependence plot or partial dependence function-specific biomarker (S53), and the conversion variable values are preset cancer / normal prediction. It may be used (S54) for the generation of statistical models or the execution of cancer / normal prediction statistical models.

Next, the present invention using the partial dependence plot or partial dependence function relation will be described in more detail. A statistical model using a composite biomarker of Complexity 3 consisting of A1AT, CYFRA21.1, and RANTES is described as an example. First, 100 data are extracted from 50 existing samples (50 samples for cancer diagnosis) and 50 samples for normal diagnosis (50 samples for normal diagnosis), and y = 0 for the normal sample and y = 1 for the cancer sample . In this case, data as shown in Table 11 may be prepared.

Sample index A1AT CYFRA21.1 RANTES y 221 3.45308 -2 4.708709 0 223 3.341135 -2 4.958518 0 222 3.568896 -2 4.577357 0 246 3.068592 -2 4.900771 0 207 4.538396 -2 5.014241 0 182 3.674541 -1.94122 4.864592 0 146 3.350815 -2 4.760304 0 197 3.003192 -2 4.741928 0 167 3.36072 -0.5627 4.863431 0 ... ... ... ... ... 120 3.681963 0.072985 4.54931 One 37 3.779961 -2 4.592287 One 6 3.408483 -0.11415 4.698918 One 106 5.341259 0.036621 4.418414 One 121 4.328482 0.550228 5.00923 One 8 3.513981 0.134559 4.732865 One 43 4.122104 -2 4.179332 One 118 5.220087 0.471732 3.972027 One 112 5.117792 0.663135 4.758335 One ... ... ... ... ...

Equation 1 below refers to a sample consisting of 100 pairs of three-dimensional explanatory variables biomarker xi = (A1AT, CYFRA21.1, RANTES) and a categorical response variable yi composed of a specific disease group (lung cancer) and a normal group.

 [Equation 1]

Subsequently, A1AT, l CYFRA21.1 RANTES has a sample consisting of three biomarkers and creates a statistical model using a tree ensemble method.

The decision tree method is expressed by the following equation.

[Equation 2]

Where Rj represents explanatory variable regions that are mutually exclusive at the teminal node. And θ = {Rj, γj} is a parameter to be estimated.

Next, how each partial dependence plot or partial dependence relationship is obtained is explained for each biomarker. Data are transformed using partial dependence plots or partial dependence relations of ensemble techniques (Boosting and Random Forest) to minimize the effects of outliers in lung cancer diagnosis model construction. Partial plot dependence or partial dependence function of the original input variables to remove the influence of other variables for one variable, consider the combined distribution _A1AT consisting of X, X _{Cyfra21 .1,} X 3 _RANTES variable. The joint distribution would like to know the effect on the X _{RANTES A1AT} variable X, X _{Cyfra21. We} can take the mean of ₁ variable. This is the basic idea of partial dependence plot or partial dependence function relationships. The partial dependence function relationship can be expressed by the following expression.

[Equation 3]

This will be described with reference to FIGS. 5 to 10. 5 is a partial dependency plot of RANTES. The function f value from the partial dependency plot is displayed on the vertical axis, and the horizontal axis is the explanatory variable value. 6 is a boxplot of cancer patients and normal people. Boxplot shows that overall RANTES values are lower in cancer patients than in normal patients. That is, the smaller the RANTES value is, the higher the likelihood of being a cancer patient group. This information is reflected in the partial dependency plot. The y-axis value of the partial dependency plot indicates the effect on the RANTES variable. The smaller the horizontal axis value, the larger the y-axis value. The larger the Y value, the more likely it is to be classified as a disease. A partial dependency plot can be drawn for each explanatory variable X. Partial dependency plots and boxplots for Cyfra21.1 are shown in FIGS. 7 and 8, and partial dependency plots and boxplots for A1AT are shown in FIGS. 9 and 10.

Next, we explain how to apply the explanatory variables transformed using partial dependence plot or partial dependence relationship to regression such as logistic regression and ridge regression.

Reflecting this characteristic of partial dependence plot / function relations, we define the value Y converted through partial dependency plot / function relations as a new input variable instead of the original value X, and this new variable is the next step in the logistic model. Become a variable. In FIG. 9, a sample having an A1AT value of 3.0 has a value of -1.5, which is converted through a partial dependency plot / function relationship, and a sample having an A1AT value of 3.5 is converted to 0.5 through a partial dependency plot / function relationship.

The regression model is a method of analyzing the influence of explanatory variables on response variables in general and the results can be used to predict disease diagnosis. There are several regression models such as Lasso regression, Ridge regression, and Logistic regression. The logistic model, one of the classification methods, is a model used when the response variable is a binary variable, and can be estimated easily and easily interpreted. Each regression coefficient can be said to represent the influence (importance) of the variable. If the regression coefficient is greater than 0, the probability of Y becoming 1 (probable disease) increases as the value of X increases. If the regression coefficient is smaller than 0, the probability of Y becoming 1 decreases as the value of X increases. In estimating the regression coefficients in the logistic model, a non-convergence problem can occur, so the probability value is estimated using the ridge function, a regularization method. The regression coefficient using the Ridge function is estimated as shown in Equation 4 below. The Ridge estimator is an estimator that obtains the smallest error when the regression coefficient estimator is limited.

[Equation 4]

 The estimated regression coefficients can be used to obtain predicted probability values of diseases.

Next, the logistic regression model to which the estimated regression coefficients are directly corresponded is shown in Equation 5 below.

[Equation 5]

When applying the value of the regression coefficient in the practical exemplary statistical model of the present invention, Equation 5 is as shown in Equation 6.

[Equation 6]

 In order to predict the probability of being classified as a disease (Yi = 1), the logistic regression model is represented by Equation 7 when the regression coefficient for the marker j of sample xi is βj.

[Equation 7]

Substituting the regression coefficient estimated in the actual exemplary statistical model into Equation 7, it is as Equation 8.

[Equation 8]

In the same manner as described above, the original variable value for each biomarker can be converted for each sample using a partial dependency plot / function relationship, and generation or cancer of a cancer / normal prediction statistical model preset with the converted biomarker variable value. Can be used to run normal predictive statistical models. In this way, a probability function for diagnosing lung cancer as shown in Equation 8 can be obtained by using the biomarker-specific variable values for all statistical models using all the combined biomarker combinations.

On the other hand, when using a composite biomarker, it is difficult to easily determine which biomarker affects how much biomarkers are used. In this case, it is necessary to develop a technique that can easily see the effects of each biomarker and compare it with other biomarkers in the complex biomarker used together with the predicted disease probability as a result of the lung cancer diagnosis model. For this reason, the exploratory data analysis technique, efficient plot (CP), was developed.

11 is an exemplary diagram of a CP. The x-axis shows the biomarkers to be compared and the y-axis shows the degree of impact on the biomarker's disease. In Figure 11 it can be seen at a glance that Cyfra21.1 is an important variable causing the disease.

The degree of impact on the biomarkers of each biomarker used in CP is calculated as follows. g (x) uses the new input variable converted using a partial dependence plot. The discrimination function obtained from the logistic model for a plurality of biomarkers having a complex degree K may be expressed as in Equation 9 below. Standardizing the new input variables and multiplying the beta coefficients with a plot gives a measure of the impact of each biomarker.

[Equation 9]

The method of generating CP includes a method of listing individual biomarkers constituting the composite biomarker on the X axis (S61) and displaying the influence information for each biomarker on the Y axis (S62).

Hereinafter, the spirit of the present invention will be described in more detail with reference to examples. In Table 12, when there are complex biomarker combinations consisting of A1AT, CYFRA21.1, and RANTES, the measured value of each biomarker for each sample and the expression amount of each biomarker converted through the partial dependency plot The measured value is shown.

Sample index A1AT CYFRA21.1 RANTES t (A1AT) t (CYFRA21.1) t (RANTES) 163 3.57 0.21 4.87 1.07 2.88 -0.63 174 2.88 -1.94 4.33 -0.95 -1.50 2.48 205 2.97 0.37 4.98 -0.95 2.88 -0.88 203 3.33 -2.00 4.95 -0.73 -1.50 -0.88 152 3.38 -0.91 4.93 0.13 -1.33 -0.88 130 3.36 -2.00 4.71 -0.47 -1.50 -0.17 229 3.21 -2.00 4.88 -0.90 -1.50 -0.63 156 3.07 -1.26 4.78 -0.95 -1.34 -0.62 168 3.20 -1.83 5.02 -0.90 -1.27 -0.86 228 3.31 -2.00 5.03 -0.73 -1.50 -0.86 ... ... ... ... ... ... ... 23 4.05 0.39 4.56 1.45 2.88 2.31 81 4.10 0.93 4.35 1.45 2.88 2.48 11 3.51 0.29 4.60 0.90 2.88 1.77 49 3.90 -0.38 4.41 1.41 1.60 2.48 8 3.51 0.13 4.73 0.90 2.88 -0.40 104 3.52 0.78 4.86 0.90 2.88 -0.63 45 4.25 -0.69 4.45 1.45 -0.99 2.47 120 3.68 0.07 4.55 1.12 2.88 2.36 5 3.44 -0.80 4.44 0.50 -1.30 2.48 9 4.50 -1.83 4.35 1.45 -1.27 2.48 21 3.65 -0.17 4.82 1.12 1.66 -0.63 ... ... ... ... ... ... ... 74 4.34 -0.27 4.70 1.45 1.66 -0.01

Table 13 shows the actual Y value for each sample (cancer patient or normal person), probability value prob (Y = 1) value predicted through the cancer diagnosis model, predicted value (cancer or normal) through probability value, and for each sample (subject). It shows the result of generating the Coefficient plot value for each biomarker.

Sample index true y Expected probability Estimate coeff_A1AT coeff_Cyfra21.1 coeff_RANTES 163 0 0.95 One 0.59 3.09 -1.29 174 0 0.38 0 -1.67 -2.37 1.90 205 0 0.69 One -1.68 2.28 -1.58 203 0 0.01 0 -1.07 -2.35 -1.59 152 0 0.03 0 -0.33 -2.16 -1.21 130 0 0.03 0 -1.08 -1.78 -0.70 229 0 0.01 0 -1.23 -2.35 -1.28 156 0 0.02 0 -1.67 -2.18 -0.96 168 0 0.01 0 -1.62 -1.57 -1.56 228 0 0.01 0 -1.07 -2.35 -1.57 ... ... ... ... ... ... ... 23 One 1.00 One 1.31 3.09 1.75 81 One 1.00 One 1.31 2.28 2.59 11 One 1.00 One 0.44 3.09 1.71 49 One 1.00 One 1.26 1.50 1.90 8 One 0.96 One 0.62 2.28 -0.99 104 One 0.94 One 0.44 3.09 -1.29 45 One 0.92 One 1.31 -1.73 1.90 120 One 1.00 One 0.90 2.28 2.44 5 One 0.75 One 0.07 -2.11 2.59 9 One 0.89 One 1.31 -2.08 1.90 21 One 0.82 One 0.90 1.15 -1.28 ... ... ... ... ... ... ... 74 One 0.93 One 0.95 1.58 -0.51

As can be seen from Table 13, the actual cancer is not two cancers are diagnosed, it can be seen that the accuracy of the prediction is very high as the cancer is not diagnosed as non-cancer.

Next, a method of selecting a complex biomarker combination with high lung cancer diagnosis ability will be described.

In the present invention, 8178 complex marker combinations having a complexity of 2 in pairs of 13 biomarkers were generated. Cancer diagnosis models corresponding to the generated biomarker combinations are generated, and the normal cancer diagnosis models are generated for 128 normal men (78 males and 50 females) and 121 lung cancer patients (78 males and female 43). Persons) were tested, and the evaluation indicators (accuracy, sensitivity and specificity) for each cancer diagnosis model corresponding to the test results are shown in Tables 14 to 24 below. Each cancer diagnostic model tested corresponds to a separate embodiment from the perspective of the cancer diagnostic model, but all 8178 embodiments should be presented and listed, but in the case of listing, it takes up too much space, and the patent expresses the invention idea. In view of the above, only representative examples are given in the form of a table. Each embodiment presented in the form of a table has a cancer diagnosis model number, and the cancer diagnosis model corresponding to the number corresponds to a biomarker combination that participates in the cancer diagnosis model. The evaluation index, which is the result of the test of 78 males and 50 females and 121 lung cancer patients (78 males and 43 females), is written.

First, in the present invention, a composite marker combination having a composite index of 2 is generated for each of 13 biomarkers. 78 cancer diagnostic models corresponding to the generated biomarker combinations were generated, and an evaluation index was generated for each generated cancer diagnostic model. For each cancer diagnosis model, a test was performed on 128 healthy patients (78 males and 50 females) and 121 lung cancer patients (78 males and 43 females). Some of the evaluation indicators (accuracy, sensitivity and specificity) for each cancer diagnosis model are shown in Table 14 below.

Table 14 below shows the evaluation index of the cancer diagnostic model corresponding to the combination biomarker combination of Complexity 2 corresponding to the top 50% of the accuracy criteria.

Cancer diagnostic model Biomarker Biomarker accuracy responsiveness Specificity 20 A1AT PAI-1 0.8795 0.8505 0.9085 14 A1AT CYFRA21-1 0.8723 0.8541 0.8906 16 A1AT RANTES 0.8702 0.8430 0.8974 31 CYFRA21-1 PAI-1 0.8684 0.8469 0.8900 22 A1AT CEA 0.8663 0.8308 0.9018 27 CYFRA21-1 RANTES 0.8648 0.8708 0.8588 26 CYFRA21-1 IGF-1 0.8629 0.8213 0.9044 32 CYFRA21-1 TTR 0.8626 0.8469 0.8782 21 A1AT TTR 0.8620 0.8197 0.9044 18 A1AT AFP 0.8618 0.8472 0.8765 42 IGF-1 TTR 0.8597 0.8626 0.8568 23 A1AT CA19-9 0.8567 0.8216 0.8918 25 A1AT ApoA1 0.8563 0.8289 0.8838 15 A1AT IGF-1 0.8540 0.8384 0.8697 28 CYFRA21-1 proApoA1 0.8539 0.8331 0.8747 24 A1AT ApoA1 / proApoA1 0.8534 0.7948 0.9121 35 CYFRA21-1 ApoA1 / proApoA1 0.8533 0.8439 0.8626 36 CYFRA21-1 ApoA1 0.8508 0.8413 0.8603 37 IGF-1 RANTES 0.8495 0.8590 0.8400 30 CYFRA21-1 EGFR 0.8494 0.8390 0.8597 17 A1AT proApoA1 0.8492 0.7866 0.9118 40 IGF-1 EGFR 0.8487 0.8397 0.8576 19 A1AT EGFR 0.8484 0.8141 0.8826 29 CYFRA21-1 AFP 0.8472 0.8338 0.8606 34 CYFRA21-1 CA19-9 0.8418 0.8256 0.8579 33 CYFRA21-1 CEA 0.8380 0.8289 0.8471 59 proApoA1 TTR 0.8348 0.8125 0.8571 47 RANTES proApoA1 0.8328 0.8197 0.8459 38 IGF-1 proApoA1 0.8312 0.8151 0.8474 66 AFP TTR 0.8289 0.8272 0.8306 41 IGF-1 PAI-1 0.8278 0.8315 0.8241 43 IGF-1 CEA 0.8203 0.8479 0.7926 46 IGF-1 ApoA1 0.8178 0.8344 0.8012 51 RANTES TTR 0.8168 0.8180 0.8156 39 IGF-1 AFP 0.8097 0.8459 0.7735 57 proApoA1 EGFR 0.8092 0.7698 0.8485 65 AFP PAI-1 0.8089 0.8128 0.8050 54 RANTES ApoA1 / proApoA1 0.8083 0.7780 0.8385 52 RANTES CEA 0.8065 0.8030 0.8100

As can be seen in Table 14, IGF-1, RANTES, A1AT, Cyfra21-1 of the 13 biomarkers it can be seen that significantly more than other biomarkers. On the other hand, in the model of complexity 2, the number of evaluation indicators is more than 85%, and the number is not more than 90%. As described above, the model of complexity 2 has a number of cancer diagnosis models that can be adopted at the 85% level of the evaluation index.

Subsequently, in the present invention, a complex marker combination of complex degree 3 was generated in pairs of 3 for 13 biomarkers. 286 cancer diagnostic models corresponding to the generated biomarker combinations were generated, and an evaluation index was generated for each generated cancer diagnostic model.

Table 15 below shows the evaluation index of the cancer diagnostic model corresponding to the complex biomarker combination corresponding to the top 30 accuracy criteria.

Cancer diagnostic model Biomarker Biomarker Biomarker accuracy responsiveness Specificity 217 IGF-1 RANTES TTR 0.9034 0.9095 0.8974 232 IGF-1 AFP TTR 0.8959 0.9007 0.8912 103 A1AT IGF-1 RANTES 0.8957 0.8944 0.8971 117 A1AT RANTES TTR 0.8925 0.8833 0.9018 131 A1AT AFP PAI-1 0.8924 0.8603 0.9244 145 A1AT PAI-1 CA19-9 0.8899 0.8662 0.9135 261 RANTES proApoA1 TTR 0.8883 0.8793 0.8974 132 A1AT AFP TTR 0.8883 0.8636 0.9129 99 A1AT CYFRA21-1 CEA 0.8879 0.8682 0.9076 114 A1AT RANTES AFP 0.8875 0.8780 0.8971 173 CYFRA21-1 RANTES CEA 0.8872 0.8793 0.8950 107 A1AT IGF-1 PAI-1 0.8862 0.8695 0.9029 116 A1AT RANTES PAI-1 0.8857 0.8679 0.9035 108 A1AT IGF-1 TTR 0.8846 0.8810 0.8882 137 A1AT EGFR PAI-1 0.8819 0.8597 0.9041 225 IGF-1 proApoA1 TTR 0.8811 0.8679 0.8944 150 A1AT TTR ApoA1 / proApoA1 0.8799 0.8485 0.9112 130 A1AT AFP EGFR 0.8796 0.8544 0.9047 119 A1AT RANTES CA19-9 0.8794 0.8689 0.8900 113 A1AT RANTES proApoA1 0.8793 0.8639 0.8947 146 A1AT PAI-1 ApoA1 / proApoA1 0.8789 0.8423 0.9156 147 A1AT PAI-1 ApoA1 0.8782 0.8616 0.8947 120 A1AT RANTES ApoA1 / proApoA1 0.8777 0.8561 0.8994 215 IGF-1 RANTES EGFR 0.8772 0.8711 0.8832 110 A1AT IGF-1 CA19-9 0.8762 0.8607 0.8918 143 A1AT PAI-1 TTR 0.8756 0.8518 0.8994 243 IGF-1 PAI-1 TTR 0.8747 0.8764 0.8729 164 CYFRA21-1 IGF-1 CEA 0.8738 0.8502 0.8974 115 A1AT RANTES EGFR 0.8733 0.8521 0.8944 134 A1AT AFP CA19-9 0.8732 0.8574 0.8891

As can be seen from Table 15, the cancer diagnostic model exceeding 90% of accuracy criteria or extremely close to 90% (when rounding, corresponding to 90%) includes IGF-1 and RANTES among the 13 biomarkers. It can be seen that.

On the other hand, in the cancer diagnostic model showing the high evaluation index, it can be seen that A1AT, Cyfra21-1, TTR is significantly higher than other biomarkers.

Subsequently, in the present invention, a complex marker combination of complex degree 4 was generated in pairs of 4 for 13 biomarkers. 286 cancer diagnostic models corresponding to the generated biomarker combinations were generated, and an evaluation index was generated for each generated cancer diagnostic model.

Table 16 below shows the evaluation indicators of the cancer diagnostic model corresponding to the complex biomarker combination corresponding to the top 30 accuracy criteria, and Table 17 below shows the cancer corresponding to the complex biomarker combination corresponding to the 60th position in the top 31 accuracy criteria. It shows the evaluation index of the diagnostic model.

As can be seen from Table 16, among the 13 biomarkers, IGF-1 and RANTES are included 19 times and 20 times in the top 30 positions in the cancer diagnosis model based on Complexity 4, respectively. It can be seen that. On the other hand, it can be seen that a large number of A1AT and TTR is included.

On the other hand, as shown in Table 17, A1AT, IGF-1, RANTES of the 13 biomarkers 19, 15 and 15 times in the top 31 to 60 in the cancer diagnostic model based on Complexity 4, respectively It can be seen that forms a mode that contains.

As can be seen from Tables 16 to 17, IGF-1 and RANTES are likely to be the major biomarkers in the cancer diagnosis model, and A1AT and TTR are the major biomarkers in the cancer diagnosis model. The probability is high.

On the other hand, as shown in Tables 16 to 17, it can be seen that the evaluation index value of the cancer diagnostic model that falls within the 40th rank shows 90% when the accuracy rounding standard is used.

Subsequently, in the present invention, a complex marker combination of complex degree 5 was generated in pairs of 5 for 13 biomarkers. A cancer diagnostic model corresponding to the generated biomarker combination was generated, and an evaluation index was generated for each generated cancer diagnostic model. Tables 18 to 21 select cancer diagnosis models based on the accuracy evaluation criteria of 90%.

Table 18 shows the evaluation indicators of the cancer diagnostic model corresponding to the complex biomarker combination corresponding to the top 30 accuracy criteria, and Table 19 shows the cancer corresponding to the complex biomarker combination corresponding to the 60th position in the top 31 accuracy criteria. The evaluation index of the diagnostic model is shown, and Table 20 below shows the evaluation index of the cancer diagnostic model corresponding to the complex biomarker combination corresponding to the top 61 to the 90th accuracy criteria. Above, the evaluation index of the cancer diagnostic model corresponding to the combination biomarker of the 117th place is shown.

As can be seen in Table 18, among the thirteen top biomarkers in the cancer diagnostic model based on Complexity 5, IGF-1 and RANTES contained 23 and 27 times, respectively. It can be seen that. On the other hand, it can be seen that A1AT and TTR are also included in a plurality of 15 times and 22 times, respectively.

On the other hand, as shown in Table 19, in the cancer diagnostic model based on Complexity 5, the top 31 to 60, the highest value of the IGF-1, RANTES contained 17 times, 27 times among the 13 biomarkers, respectively It can be seen that the formation, and A1AT and TTR are also included 18 and 16 times, respectively.

On the other hand, as shown in Table 20, the highest 61 to 90 in the cancer diagnostic model based on the complex degree 5, IGF-1, RANTES among the 13 biomarkers, 17 times, 22 times, respectively, the most frequent value It can be seen that it forms, and it can be seen that A1AT is also included 16 times.

On the other hand, as can be seen in Table 21, in the cancer diagnosis model based on the complex degree 5, the top 91 to 117, it is found that a number of A1AT, IGF-1, RANTES, TTR, etc. among the 13 biomarkers are included Can be.

As can be seen from Table 18 to Table 21, in the cancer diagnosis model based on Complexity 5, the highest 1 to 117 are among the 13 biomarkers containing 73 and 88 IGF-1 and RANTES, respectively. It can be seen that the form, and A1AT 65 times, TTR is included 64 times.

Subsequently, in the present invention, a complex marker combination having a complex degree of 6 was formed in pairs of 6 for 13 biomarkers. A cancer diagnostic model corresponding to the generated biomarker combination was generated, and an evaluation index was generated for each generated cancer diagnostic model. Table 22 below shows the evaluation index of the cancer diagnostic model included in the top 30.

As can be seen in Table 22, RANTES is included in all cancer diagnostic models in complex cancer model 6, and A1AT and IGF-1 are included 24 and 24 times, respectively, and Cyfra21-1 and TTR are also included. It can be seen that it is contained.

Subsequently, in the present invention, a compound marker combination of compound degree 7 was generated in pairs of 7 for 13 biomarkers. A cancer diagnostic model corresponding to the generated biomarker combination was generated, and an evaluation index was generated for each generated cancer diagnostic model. Table 23 below shows the evaluation index of the cancer diagnostic model included in the top 30.

As can be seen from Table 23, in the cancer diagnosis model of Complexity 7, IGF-1 and RANTES are included in almost all cancer diagnosis models 29 and 30 times, respectively, and Cyfra21-1 and TTR are included 24 times. Able to know.

On the other hand, as can be seen in Table 23, it can be seen that the degree of saturation of the evaluation indicators increases as the degree of complexity approaches 6 to 7.

The inventors of the present invention have a combination of 8 complex markers in a combination of 8 for 13 biomarkers, a complex marker combination in a complex of 9 in a pair of 9, a complex marker combination of 10 in a complex of 10, a pair of 11 As a result, a cancer diagnostic model including a complex marker combination of complex level 11, a complex marker combination of complex level 12 in pairs of 12, and all 13 biomarkers was generated, and an evaluation index was generated for each generated cancer diagnostic model.

Some of the results for Complexity 12 of the results for Complexities 8-12 are published in Table 24 below. As the complexity increases, the evaluation index tends to improve. However, the higher the complexity, the more saturated the evaluation index or the better the tendency is not necessarily established. Table 24 shows such an example.

Cancer diagnostic model Biomarkers excluded of 13 biomarkers accuracy responsiveness Specificity 8178 ApoA1 0.9047 0.9062 0.9032 8179 ApoA1 / proApoA1 0.9059 0.9023 0.9094 8180 CA19-9 0.9026 0.9023 0.9029 8181 CEA 0.9 0.897 0.9029 8182 TTR 0.9037 0.8997 0.9076 8183 PAI-1 0.8976 0.8941 0.9012 8184 EGFR 0.9074 0.9059 0.9088 8185 AFP 0.9026 0.8987 0.9065 8186 proApoA1 0.9055 0.901 0.91 8187 RANTES 0.8895 0.8852 0.8938 8188 IGF-1 0.8917 0.8928 0.8906 8189 CYFRA21-1 0.8991 0.8967 0.9015 8190 A1AT 0.9002 0.8931 0.9074

As can be seen from Tables 15 to 24, with respect to the biomarker combination candidate constituting the biomarker combination candidate group, lung cancer diagnosis capability is compared with the individual biomarker or configured biomarker combination candidates constituting the biomarker combination candidate (S22). )can do. The comparison may be compared with an evaluation index. Among biomarker combination candidates, a biomarker combination whose lung cancer diagnosis ability is greater than or equal to a predetermined criterion is selected (S23). The predetermined criterion may be different depending on which evaluation indicator is used in the selection. Specificity may be an important evaluation index in the diagnosis of lung cancer, and the area of the ROC curve may be an effective evaluation index.

Meanwhile, at least one biomarker is selected from the first biomarker group (S31), at least one biomarker is selected from the second biomarker group (S32), and at least one biomarker is included. A biomarker combination candidate group including one or more biomarker combinations may be configured (S33), and lung cancer diagnosis capability may be compared with individual biomarkers or configured biomarker combination candidates constituting the biomarker combination candidate (S34).

The present invention provides a kit for lung cancer diagnosis and screening comprising a combination of two or more antibodies that can specifically bind to the 13 biomarkers.

In a specific embodiment of the present invention, 13 proteins with significantly changed expression levels in serum of lung cancer patients were selected as biomarkers for lung cancer diagnosis and screening (see Table 5)), and the classification using a combination of the 13 biomarkers. We confirmed that the model can perform lung cancer classification with higher accuracy. Thus, the kit of the present invention may include an antibody capable of specifically binding to each biomarker constituting the complex biomarker for use in quantifying a complex biomarker having a difference in expression in a lung cancer patient and a normal person. have.

The kit distinguishes whether or not the patient is lung cancer and enables medical practitioners, such as doctors, to diagnose and screen lung cancer, as well as monitor the patient's response to the treatment and modify the treatment accordingly. . It can also be used to identify compounds that modulate the expression of one or more biomarkers in vivo or ex vivo in a lung cancer model (eg, an animal model such as mouse, rat, etc.). Thus, the biomarker of the present invention may be further included in the kit as a standard material.

Antibodies that can be used in the kits of the present invention include polyclonal antibodies, monoclonal antibodies, fragments capable of binding epitopes, and the like.

Polyclonal antibodies can be produced by conventional methods of injecting any one of the 13 proteins into an animal and collecting blood from the animal to obtain a serum comprising the antibody. Such polyclonal antibodies can be purified by any method known in the art and can be made from any animal species host, such as goats, rabbits, sheep, monkeys, horses, pigs, cattle, dogs and the like.

Monoclonal antibodies can be prepared using any technique that provides for the production of antibody molecules through the culture of continuous cell lines. Such techniques include, but are not limited to, hybridoma technology, human B-cell hybridoma technology, and EBV-hybridoma technology (Kohler G et al., Nature 256: 495-497, 1975; Kozbor D et al., J Immunol Methods 81: 31-42, 1985; Cote RJ et al., Proc Natl Acad Sci 80: 2026-2030, 1983; and Cole SP et al., Mol Cell Biol 62: 109-120, 1984).

In addition, antibody fragments containing specific binding sites for any of the 13 proteins can be prepared. For example, but not limited to, F (ab ') 2 fragments can be prepared by digesting antibody molecules with pepsin, and Fab fragments can be prepared by reducing the disulfide bridges of F (ab') 2 fragments. Alternatively, a Fab expression library can be constructed to quickly and simply identify monoclonal Fab fragments with the desired specificity (Huse WD et al., Science 254: 1275-1281, 1989).

The antibody can be bound to a solid substrate to facilitate subsequent steps such as washing or separation of the complex. Solid substrates include, for example, synthetic resins, nitrocellulose, glass substrates, metal substrates, glass fibers, paramagnetic beads, microspheres and microbeads. In addition, the synthetic resins include polyester, polyvinyl chloride, polystyrene, polypropylene, PVDF and nylon. In a specific embodiment of the present invention, in order to bind an antibody specifically binding to a protein to a solid substrate, the microspheres are suspended and then transferred to a microtube to remove the supernatant by centrifugation, and then resuspended. N-hydroxy-sulfosuccinimide and 1-ethyl-3- (3-dimethylaminopropyl) -carbodiimide hydrochloride-ethyl-3- (3-dimethylaminopropyl) -carbodiimide hydrochloride ) Was treated sequentially and then the supernatant was removed by centrifugation, washed and stored. In addition, when a sample obtained from a patient is contacted with an antibody capable of specifically binding to any of the 13 proteins of the present invention bound to a solid substrate, the sample may be diluted to a suitable degree prior to contact with the antibody. .

The kit of the present invention may further comprise a detection antibody that specifically binds to the biomarker. The detection antibody may be a conjugate labeled with a detector such as a chromophore, a fluorescent substance, a radioisotope or a colloid, and preferably a primary antibody capable of specifically binding to the biomarker. For example, the chromase may be peroxidase, alkaline phosphatase or acid phosphatase (eg horseradish peroxidase); In the case of fluorescent materials, fluorescein carboxylic acid (FCA), fluorescein isothiocyanate (FITC), fluorescein thiourea (FTH), 7-acetoxycoumarin-3-yl, fluorescein-5-yl , Fluorescein-6-yl, 2 ', 7'-dichlorofluorescein-5-yl, 2', 7'-dichlorofluorescin-6-yl, dihydrotetramethyllosamine-4-yl, Tetramethyllodamine-5-yl, tetramethyllodamine-6-yl, 4,4-difluoro-5,7-dimethyl-4-bora-3a, 4a-diaza-s-indacene-3- Ethyl or 4,4-difluoro-5,7-diphenyl-4-bora-3a, 4a-diaza-s-indacene-3-ethyl, Cy3, Cy5, poly L-lysine-fluorescein iso Thiocyanate (poly L-lysine-fluorescein isothiocyanate (FITC), rhodamine-B-isothiocyanate (RITC), rhodamine (rhodamine), PE (Phycoerythrin), etc. can be used. Do.

In addition, the kit of the present invention may further comprise (1) a detection antibody that specifically binds to the biomarker and (2) a ligand that can specifically bind to the detection antibody. The ligand includes a secondary antibody that specifically binds to protein A or an antibody for detection. In addition, the ligand may be a conjugate labeled with a detector such as a chromophore, a fluorescent substance, a radioisotope or a colloid. The detection antibody is preferably a biotinylated or digoxigenin-treated primary antibody for the ligand, but the method of treating the detection antibody is not limited thereto. In addition, as the ligand, streptavidin, avidin, or the like is preferably used to bind the detection antibody, but is not limited thereto. In a specific embodiment of the present invention, streptavidin (streptavidin) having a fluorescent substance attached thereto was used as a ligand, and a detection antibody biotinylated for the ligand was used.

The diagnostic and screening kit of the present invention can diagnose and screen lung cancer by treating the antibody and biomarker complex with a detection antibody and then searching for the amount of the detection antibody. Alternatively, the antibody and the biomarker complex may be sequentially treated with a detection antibody and a ligand, and then lung cancer may be diagnosed and screened by searching for the amount of the antibody for a detector. In a preferred embodiment of the present invention, the amount of the biomarker can be determined by measuring the antibody for detection by aligning the antibody for detection with the washed antibody-biomarker complex and then washing the antibody. Determination of the amount or detection of the antibody for detection can be made through fluorescence, luminescence, chemiluminescence, absorbance, reflection or transmission.

In addition, as a method of detecting the amount of the antibody or ligand for detection, it is preferable to use a high throughput screening (HTS) system, wherein a fluorescence method or detection performed by detecting a fluorescence by attaching a fluorescent material to the detector Radiation method performed by detecting radiation by attaching a radioisotope into a sieve; It is preferable to use a surface plasmon resonance (SPR) method for measuring the plasmon resonance change of the surface in real time without a label of the detector, or a SPRI (surface plasmon resonance imaging) method for imaging the SPR system.

For example, the fluorescence method uses a fluorescence scanner program to label the detection antibody with a fluorescent material and spot the signal by spotting. This method can be applied to confirm the degree of binding. The fluorescent material is Cy3, Cy5, poly L-lysine-fluorescein isothiocyanate (FITC), rhodamine-B-isothiocyanate (RITC) , Rhodamine, PE (Phycoerythrin) is preferably any one selected from the group consisting of, but is not limited thereto. Unlike the fluorescence method, the SPR system does not require that the sample be labeled with a fluorescent material, but it is possible to analyze the binding degree of the antibody in real time, but it has a disadvantage that it is impossible to simultaneously analyze samples. In the case of SPRI, it is possible to perform simultaneous multiple sample analysis by using the fine alignment method, but the detection strength is low.

In addition, the kit for diagnosis and screening of the present invention may further include a washing solution or an eluent which can remove the substrate and unbound protein and the like to develop a color reaction with the enzyme and retain only the bound biomarker. Samples used for analysis include biological samples capable of identifying disease specific polypeptides that can be distinguished from normal conditions such as serum, urine, and tear saliva. Preferably from a biological liquid sample, for example blood, serum, plasma, more preferably serum. The sample may be prepared to increase the detection sensitivity of the biomarker. For example, a serum sample obtained from a patient may be subjected to anion exchange chromatography, affinity chromatography, size exclusion chromatography, liquid chromatography, Such as, but not limited to, sequential extraction or gel electrophoresis.

In addition, the present invention provides a biochip for lung cancer diagnosis and screening in which a biomolecule capable of specifically binding to any one of the 13 proteins is integrated on a solid substrate.

In a specific embodiment of the present invention, 13 proteins with significantly varying expression levels in the serum of lung cancer patients were selected (see Table 5), and higher accuracy in the classification model consisting of a combination using at least two of the 13 proteins. It was confirmed that lung cancer classification can be performed. Thus, the biochip of the present invention can specifically bind to any one of the 13 proteins for use in measuring one or more of the 13 proteins, such as the difference in expression in lung cancer patients and normal people. Antibodies, or combinations of two or more of the above specific antibodies.

The biomolecule is selected from the group consisting of low molecular weight compounds, ligands, aptamers, peptides, polypeptides, specific binding proteins, high molecular materials and antibodies, and any material that can specifically bind to the protein, It is preferable to use an antibody or aptamer, but is not limited thereto.

The antibody is preferably a polyclonal antibody or a monoclonal antibody, more preferably a monoclonal antibody. Antibodies that specifically bind to the proteins may be prepared by known methods known to those skilled in the art, and commercially known antibodies may be purchased and used. The antibody can be prepared by injecting a protein that is an immunogen into an external host according to conventional methods known to those skilled in the art. External hosts include mammals such as mice, rats, sheep, rabbits. Immunogens are injected by intramuscular, intraperitoneal or subcutaneous injection and can generally be administered with an adjuvant to increase antigenicity. Antibodies can be isolated by collecting blood periodically from an external host and collecting serum showing shaped titers and specificity for the antigen.

In addition, the solid substrate of the biochip of the present invention may be selected from the group consisting of plastics, glass, metals and silicon, and preferably may be chemically treated or a linker molecule is bound to attach the antibody to the surface thereof. It is not limited. The biochip of the present invention can easily and accurately diagnose lung cancer by screening the whole protein from the sample and reacting with the biochip.

The active group coated on the substrate of the biochip serves to bind the material, and may be selected from the group consisting of an amine group, an aldehyde group, a carboxyl group and a thiol group. Any active group known as an activator capable of binding a protein molecule to a substrate may be used by one skilled in the art, but is not limited thereto.

14 relates to an exemplary configuration of a lung cancer diagnosis system.

The lung cancer diagnosis system performs lung cancer diagnosis using information directly or derived from or read from the diagnostic kit. The lung cancer diagnosis system includes an information acquisition module for obtaining expression amount information or expression rate ratio information for each biomarker constituting the biomarker combination measured from blood, plasma, serum or other collected material separated from the subject's body, And a lung cancer diagnostic module configured to process the obtained expression level information or expression level ratio information into a preset lung cancer diagnostic model, and lung cancer diagnostic information generation module configured to generate at least one lung cancer diagnostic information from the lung cancer diagnostic module. The lung cancer diagnostic module may further include at least one or more conversion modules preset to the expression level information or expression level ratio information, and the conversion module may further include expression level conversion information or expression level ratio information for the expression level information. First, the expression rate ratio conversion information is generated.

Meanwhile, the lung cancer diagnostic model receives the generated expression level conversion information or the expression level ratio conversion information as input values, and the conversion module expresses the expression using a partial dependence plot or partial dependency function relationship of an ensemble technique using a tree. Quantity conversion information or expression rate ratio conversion information is generated. This is described above. The lung cancer diagnostic model is a logistic model, and the logistic model estimates a probability value classified as lung cancer by receiving the expression level conversion information or the expression level ratio conversion information.

The CP information generation unit 1310 of the lung cancer diagnosis information generation module 1300 may further generate information on the disease diagnosis contribution rate for each biomarker, and the disease diagnosis contribution rate for each biomarker may be determined by the biomarker combination. The degree of effect on lung cancer is generated in the form of a coefficient plot using a predetermined discriminant function obtained from a logistic model for the biomarkers included in.

The information obtaining module obtains expression amount information or expression amount ratio information for each biomarker by the lung cancer diagnosis system directly obtained from the diagnosis kit, the lung cancer diagnosis system and the diagnostic kit connected through a wired or wireless network. The method of obtaining the expression information of each biomarker is transmitted from a third system capable of reading the method and the biomarker is transmitted from the computer of the person receiving the expression information of the biomarker connected to the wire and wireless network with the lung cancer diagnosis system A method of obtaining in a manner may be used. When the lung cancer diagnosis system can directly read expression level information of a biomarker of a diagnostic kit, expression level information can be directly obtained from the diagnostic kit. However, when it cannot read directly, it can also obtain it by the method of receiving from the 3rd system, such as a machine, apparatus, and apparatus which reads the expression quantity information. On the other hand, when the third system and the lung cancer diagnosis system are not connected to the wired / wireless network or cannot directly transmit or receive information, the expression level information read directly or indirectly from the computer of the person who read the expression level information is wired or wireless. It is possible to transmit to the lung cancer diagnosis system through a network.

The lung cancer diagnosis system obtains expression amount information or expression amount ratio information for each biomarker constituting a biomarker combination measured from blood, plasma, serum or other collected material separated from the subject's body (S41). The expression level information or expression level ratio information is processed into a lung cancer diagnostic module including a preset lung cancer diagnostic model (S42) to generate at least one lung cancer diagnostic information from the lung cancer diagnostic module (S43).

Meanwhile, the lung cancer diagnosis system may store a plurality of lung cancer diagnosis models in a lung cancer diagnosis model unit and perform lung cancer diagnosis service for those who use a plurality of different lung cancer diagnosis biomarker combinations. For example, hospital A performs lung cancer diagnosis using a lung cancer diagnostic kit associated with a + b + c + d complex biomarker, and hospital B uses a lung cancer diagnostic kit associated with a + c + e + f complex biomarker. When lung cancer diagnosis is performed using different biomarker combinations for each diagnostic kit, different lung cancer diagnostic models should be used. In this case, the information obtained by the lung cancer diagnosis system should include the sample ID and expression information for each biomarker. Therefore, the lung cancer diagnosis model selection unit of the lung cancer diagnosis system extracts the biomarker combination used in the diagnostic kit through the plurality of biomarkers corresponding to the expression amount information from the biomarker expression amount information obtained, The marker combination information is used to determine which lung cancer diagnostic model to select. In other words, lung cancer diagnosis is performed using a lung cancer diagnostic model associated with a + b + c + d complex biomarker for hospital A, and lung cancer diagnostic model associated with a + c + e + f complex biomarker for hospital B. To perform lung cancer diagnosis.

The present invention can be utilized in the medical industry, the medical information processing industry, and industries related to cancer diagnosis and prevention.

1000: Lung Cancer Diagnosis System
1100: information acquisition module
1200: Lung Cancer Diagnostic Module
1210: Conversion Module
1211: Partial Dependency Plot / function relationship generator
1220: lung cancer diagnostic model generator
1221: lung cancer diagnostic model
1300: lung cancer diagnostic information generation module
1310: CP information generation unit
1320: lung cancer diagnostic model selection unit
2000: Biomarker expression level provider
2100: Diagnostic Kits
2200: Third system capable of reading expression level information for each biomarker of a diagnostic kit
2300: Computer of a person who obtains expression level information for each biomarker
3000: wired and wireless network

Claims (5)

In the method for using the combined biomarker information for lung cancer diagnosis of the lung cancer diagnosis system, the lung cancer diagnosis system,
(A) at least one first biomarker selected from the group of first biomarkers consisting of individual biomarkers IGF-1 and RANTES measured from blood, plasma, serum or other material taken from the body of a subject diagnosed with lung cancer Determination of the amount of expression of each biomarker of the group and the amount of expression of the second biomarker of the individual biomarkers A1AT, CYFRA21-1, proApoA1, AFP, EGFR, PAI-1, TTR, CEA, CA19-9 and ApoA1 Obtaining information;
(B) processing the expression amount information for each biomarker of the first biomarker group and the expression amount information for each biomarker of the second biomarker group and input the biomarker into a predetermined lung cancer determination model; And
(C) generating lung cancer determination information from the lung cancer determination model; lung cancer diagnostic composite biomarker information using method comprising a. The method of claim 1,
Processing the expression level information for each biomarker in the step (B), when there is expression information of ApoA1 and expression level of proApoA1 in the second biomarker group, generates a ratio value of ApoA1 expression amount and proApoA1 expression amount The method for using the complex biomarker information for lung cancer diagnosis, characterized in that the lung cancer determination model, the expression amount of ApoA1, the expression amount of proApoA1, and the ratio of the ratio of ApoA1 expression amount and proApoA1 expression amount are injected. The method of claim 1,
Processing the expression level information for each biomarker is to generate the expression level information for each biomarker converted using partial dependency plot or partial dependency function relationship of the ensemble method using the decision tree. Method for using complex biomarker information for lung cancer diagnosis, characterized in that. The method of claim 1,
The lung cancer determination model is a method for using complex biomarker information for lung cancer diagnosis, characterized in that the logistic regression model. 5. The method of claim 4,
The logistic regression model is a method for using complex biomarker information for diagnosing lung cancer, characterized by using a ridge penalty function.

Images