JP2007037421A - Gene set for predicting the presence or absence of lymph node metastasis from colorectal cancer - Google Patents

Abstract

The present invention provides a method for predicting the presence or absence of lymph node metastasis of colorectal cancer and a gene set usable in the method.
A method of selecting a gene set comprising the following steps (1) to (4): (1) In a primary colorectal cancer tissue of a patient whose presence or absence of lymph node metastasis is revealed by histopathological determination From the gene group selected by the analysis method used in the steps (2) and (1) of selecting gene groups that can classify the presence or absence of lymph node metastasis with a positive classification rate of 75% or more in each analysis method, A step of selecting a common gene selected in common in any analysis method; (3) classification of the presence or absence of lymph node metastasis from any combination of two or more genes by analyzing the gene expression information; And (4) a logistic cycle in which the presence or absence of lymph node metastasis is set as a response using the common gene and the gene combination as explanatory variables. The step of performing variable selection in the model.
[Selection figure] None

Description

  The present invention relates to a gene group useful for predicting the presence or absence of lymph node metastasis of colorectal cancer, and a method for utilizing the gene expression information.

  Colorectal cancer has a particularly high incidence in developed countries and is increasing year by year in Japan and is one of the leading causes of cancer-related death. According to statistical reports (for example, see Non-Patent Document 1), 37% of colorectal cancer patients are cancers confined to the main lesion without metastasis, and 37% have metastases only in the regional lymph nodes. It is known that the cancer is resident and the rest is accompanied by distant metastasis.

  Currently, the Dukes classification, which is most commonly used as a malignancy classification of colorectal cancer in clinical practice, uses pathological items such as the degree of cancer colon penetration and the degree of metastasis to regional lymph nodes as indicators. There is no doubt about the correlation between this classification and prognosis. However, with regard to the determination of the presence or absence of lymph node metastasis, which greatly affects the above classification results, a classic tissue in which a specimen prepared using a part of a number of excised lymph node tissues is observed under a microscope. The current situation relies on pathological techniques. As shown in a report (for example, see Non-Patent Document 2) that 20 to 40% of patients determined to be negative for lymph node metastasis by such a technique will be found later, refer to conventional lymph node metastasis determination. The method was not always accurate enough.

  Colorectal cancer is one of the most advanced molecular biological studies such as the structure of multistage carcinogenesis, and there have been many reports on individual genes such as APC, K-ras, p53, and DCC. It is done. However, just focusing on one of these genes is not sufficient to express the individuality of colorectal cancer, and in recent years, expression information on a large number of genes can be obtained at once by using a DNA microarray or the like. Attempts have been made to obtain more useful new knowledge.

  Alizadeh et al. Measured the B lymphocytes collected from the peripheral blood of a patient with diffuse large B-cell lymphoma using a DNA microarray as a sample, and performed hierarchical clustering of the obtained gene expression data. The patient's peripheral blood B lymphocytes show a gene expression pattern similar to that of B cells present in the germinal center of the lymphoid tissue and a case of showing a gene expression pattern similar to B cells activated in vitro. It was found that there are types (Non-Patent Document 3). As a result of examining the survival rate of both in the Kaplan-Meier plot, it became clear that patients with B cells showing the latter expression pattern had a worse prognosis than patients with B cells showing the former expression pattern. . In addition, the results obtained by the clustering of gene expression information performed by the authors were more correlated with the prognosis than following the prognosis prediction based on the conventional pathological diagnosis. The results of Alizadeh et al. Are significant in that they can derive clinically useful useful laws from gene expression information. However, it has not been verified whether the law can be applied to completely new clinical cases, and it cannot be denied that it is a result that is valid only within the scope of this paper.

  Khan et al. Reported that four types of cancer belonging to small round blue cell tumors, which are difficult to distinguish histologically, can be accurately distinguished by analysis of gene expression information using artificial neural networks (Non-Patent Literature). 4). In this report, accurate judgment results can be obtained even when test sample data is input to an artificial neural network model derived from a part of the random data extracted from the entire data. It has been verified. Therefore, the artificial neural network model derived here is not limited to the range of data in this paper, but is generally applicable to distinguish four types of cancer belonging to small round blue cell tumor. It is suggested that However, the judgment results obtained with the artificial neural network model are generally not accepted in that the mathematical basis cannot be clearly explained.

  As a recent research example conducted using a DNA microarray for the purpose of identifying a molecular target involved in liver metastasis of colorectal cancer, there is a report by Yanagawa et al. (Non-patent Document 5). The authors performed PCR using human cDNA as a template using oligo DNA designed based on the base sequence of human cDNA registered in public gene databases, and obtained 9,121 kinds of amplified cDNA fragments. It was. Next, using a DNA microarray in which these cDNA fragments were printed as probes, gene expression profiles of colon cancer primary lesions and colon cancer liver metastases isolated from 10 colon cancer patients were examined. As a result, we have identified 40 genes whose expression is increased in liver metastases relative to the primary lesion and 7 genes whose expression is decreased in liver metastasis relative to the primary lesion. We identified a set of candidate genes that may be involved in liver metastasis.

  For the gene set involved in liver metastasis of colorectal cancer, by analyzing the expression information of genes specifically expressed in the colon cancer primary tissue by DNA microarray method, A method for identifying a gene set effective for predicting liver metastasis, a method for predicting liver metastasis of colorectal cancer using the gene set identified by the method and expression information of the gene set in the colon cancer primary tissue is known. (Patent Document 1). The gene set and method provide information useful for predicting metachronous liver metastasis of colorectal cancer, and are preferable as a material for identifying an important gene specifically expressed in colorectal cancer. However, because lymph node metastasis of colorectal cancer is completely pathologically different from liver metastasis, these gene sets and methods for liver metastasis of colorectal cancer can be directly applied to lymph node metastasis of colorectal cancer. Never.

  In addition, an original DNA microarray was prepared using a probe selected from a cDNA library prepared using a colon cancer primary tissue, a colon cancer liver metastasis tissue and a normal colon mucosa tissue as a material. It has also been shown that by performing gene expression analysis, it is possible to identify candidate genes that are considered to be related to the development and progression of colorectal cancer (Non-patent Document 6).

  On the other hand, with regard to lymph node metastasis of colorectal cancer, as described above, the presence or absence of lymph node metastasis is determined by observing a specimen prepared using a part of a number of excised lymph node tissues under a microscope. The current status depends on the classical histopathological technique, and such a lymph node metastasis determination method is not necessarily accurate enough. It is also known that postoperative adjuvant therapy after primary colorectal cancer removal surgery can improve the prognosis of patients with lymph node metastasis, but postoperative adjuvant therapy is anorexia, upper abdominal discomfort, nausea It is necessary to determine whether it is necessary or unnecessary from the viewpoint of quality of life (QOL) and medical costs, taking into account the individual patient's condition and disease state. Therefore, if a more accurate method for determining lymph node metastasis is found, it can be used as a useful index for decision-making when selecting postoperative adjuvant therapy, and ultimately patients receive the appropriate treatment after receiving appropriate treatment. It is thought to lead to profit.

JP 2004-33082 A Troisi R.J. et al., 1999, Cancer, vol. 85, p. 1670-1676 Cohen A.M. et al., 1997, Curr Probl Surg., Vol. 34, p. 601-676 Alizadeh et al., 2000, Nature, vol. 403, p. 503-511 Khan et al., 2001, Nature Medicine, vol. 7, p. 673-679 Yanagawa et al., 2001, Neoplasia, vol. 3, No. 5, p.395-401 Takemasa et al., 2001, Biochem. Biophys. Res. Commun., Vol. 285, p. 1244-1249

  As described above, the conventional method for determining the presence or absence of lymph node metastasis of colorectal cancer involves removing a plurality of lymph nodes around the colorectal cancer and observing them under a microscope. there were.

  Therefore, in order to improve this point, an object of the present invention is to provide a method for predicting the presence or absence of lymph node metastasis of colorectal cancer by examining the gene expression profile of colorectal cancer primary tissue. In order to make it possible to predict the presence or absence of cancer cell metastasis to a lymph node, the present invention also provides a lymph node based on a gene set that can be used to determine lymph node metastasis of colorectal cancer and expression information thereof. An object is to provide a discriminant that can be used to determine the presence or absence of metastasis.

  As a result of intensive studies to achieve the above-mentioned object, the present inventors have selected a probe selected from a cDNA library prepared using colon cancer primary tissue, colon cancer liver metastasis tissue and normal colon mucosa tissue as materials. A gene set that can be used to predict the presence or absence of lymph node metastasis through statistical analysis of gene expression analysis data of the primary lesion of colorectal cancer obtained using the DNA microarray, and Based on the expression level, the present inventors have succeeded in finding a discriminant used for actually predicting the presence or absence of lymph node metastasis, and completed the present invention.

That is, the present invention provides the following gene set selection method for predicting the presence or absence of lymph node metastasis of colorectal cancer.
1. A method for selecting a gene set for predicting the presence or absence of colorectal cancer lymph node metastasis, including the following steps (1) to (4):
(1) Analyzing gene expression information in the primary colorectal cancer tissue of a patient whose presence or absence of lymph node metastasis was revealed by histopathological determination using at least one supervised learning analysis method. Selecting a gene group in each analysis method that can classify the presence or absence of lymph node metastasis with a correct classification rate of 75% or more,
(2) a step of selecting a common gene selected in common in any analysis method from the gene group selected in each analysis method used in (1),
(3) analyzing the gene expression information, instructing the classification of the presence or absence of lymph node metastasis from any combination of two or more genes, and selecting a combination of genes showing an interaction; and (4) A step of selecting a variable in a logistic regression model in which the combination of the common gene and the gene is used as an explanatory variable and the presence or absence of lymph node metastasis is used as a response;
2. (1) Analysis method includes (a) Support Vector Machine, (b) Extension of Principal Component Analysis Artificial Neural Network, (c) Combination of Hierarchical Cluster Analysis and Stepwise Logistic Discrimination, and (d) Classification And Regression Tree and Logistic 1 including at least one selected from the group consisting of a combination of discriminations. The method described in;
3. The analysis method (3) is a combination of (d) Classification And Regression Tree and Logistic Discrimination. Or 2. The method described in;
4). The variable selection method in (4) is a stepwise variable selection method described above in 1. Or 3. The method in any one of.

The present invention also provides the following gene set for predicting the presence or absence of metastasis from colon cancer lymph nodes.
5. Above 1. Or 4. A gene set for predicting the presence or absence of colorectal cancer lymph node metastasis selected by any of the methods of:
6). At least the gene represented by the access number (serial number) of the database of NM_003404 (G1592), NM_002128 (G2645), NM_052868 (G3031), NM_005034 (G3177), NM_001540 (G3753), NM_005722 (G3826), and NM_015315 (G4370) Including the above 5. The gene set described in.

The present invention further provides the following method for predicting the presence or absence of lymph node metastasis of colorectal cancer using the selected gene set.
7). 5. above. Or 6. A method for predicting the presence or absence of lymph node metastasis of colorectal cancer, comprising using the gene set according to any one of the above;
8). 6. The above discriminating formula is used. Method described:
D = 0.23027-2.7132 × NM_003404 (G1592) expression level + 8.9509 × NM_052868 (G3031) expression level + 8.7975 × NM_005722 (G3826) expression level-2.3098 × NM_015315 (G4370) expression level Amount + 3.5126 x NM_002128 (G2645) expression x NM_005034 (G3177) expression-8.8226 x NM_001540 (G3753) expression x NM_005722 (G3826) expression (D> 0, lymph node metastasis) , It is determined that there is no lymph node metastasis when D ≦ 0).

Examples of gene group analysis methods used in the present invention include the following:
1. Hierarchical Cluster Analysis Cluster analysis Logistic Discrimination (including variable selection method) Logistic analysis Classification And Regression Tree Cart 4. 4. Principal Component Analysis Artificial Neural Network (including extended method) PCA / principal component analysis Projection Pursuit for supervised classification Projection Pursuit / Projection Pursuit 6. Support Vector Machine Support Vector Machine / SBM Self Organizing Map SOM8. AdaBoost AdaBoost Gene selection process combining two or more of these methods.

  An example of a method for analyzing gene expression information used in the present invention is logistic analysis.

Examples of variable selection methods used in the present invention include the following:
1. Stepwise (stepwise)
2. Forward selection method (forward)
3. Backward selection method (backward)

  In the present invention, a series of gene sets useful for determining whether or not colon cancer cells are likely to have metastasized to surrounding lymph nodes at the time when the colon cancer patient undergoes primary colorectal cancer resection surgery. And a discriminant for predicting the presence or absence of lymph node metastasis based on the gene expression information. According to the method of the present invention, a good lymph node metastasis determination result can be obtained by analyzing the gene expression information of the gene set in the colon cancer primary tissue with a logistic regression equation. Therefore, it is possible to predict the presence or absence of cancer cell metastasis to lymph nodes at the time of primary colorectal cancer resection.

  The method of the present invention predicts the presence or absence of lymph node metastasis based on the gene set effective in predicting the presence or absence of lymph node metastasis at the time of primary colorectal cancer resection surgery, and the expression level of the gene set. Characterized by a discriminant for

  A gene set useful for predicting the presence or absence of lymph node metastasis is obtained by comprehensively examining gene expression in multiple samples of colon cancer primary tissue and selecting a set of genes that can be used for the determination. It is done. Such comprehensive gene expression analysis methods include microarray, Northern analysis, ATAC-PCR method (Kato et al., Nuc. Acids Res., Vol. 25, p. 4694-4696, 1997) and Taq Various methods such as real-time PCR represented by Man PCR (Applied Biosystems), SAGE (Velculescu et al., Science, vol. 270, p. 484-487, 1995) and the like can be used.

  In a preferred embodiment of the present invention, it is carried out according to the method described in JP-A-2004-33082 using a DNA microarray. More specifically, 63 cases of primary colorectal cancer tissue derived from patients whose metastasis to the lymph node was observed by histopathological observation at the time of removal of the primary lesion, collected through informed consent, and lymph node metastasis Using the DNA microarray, gene expression data were obtained for a total of 150 tissues, including 87 primary lesion tissues derived from patients in which no sigma was observed. As a comparative control, gene expression data obtained from normal large intestine mucosal tissue around the colon cancer primary tissue for 40 cases was used.

  The above gene expression data is obtained by hybridizing fluorescence-labeled cDNA prepared using total RNA extracted from cancer tissue to a DNA microarray, and using the fluorescence signal emitted by the fluorescence-labeled cDNA hybridized to the probe on the DNA microarray. It is obtained by detecting and quantifying with a scanner. A more specific procedure is described below.

  Extraction of total RNA from colon cancer tissue or normal colon mucosa tissue can be performed using a reagent such as TRIzol reagent (GIBCO BRL) or ISOGEN (Nippon Gene) according to the method described in the package insert of each reagent. The total RNA thus prepared can be used as it is for the preparation of the following labeled cDNA. Further, for example, by using a commercially available kit such as mRNA Purification Kit (Amersham BioSciences) and purifying polyadenine-added RNA (hereinafter also referred to as “mRNA”) from the total RNA according to the attached method, It can also be used for the preparation of cDNA.

  Cy3-labeled cDNA derived from colon cancer primary tissue (hereinafter sometimes referred to as “Cy3 cDNA”) is prepared by applying reverse transcriptase to a mixture containing the above total RNA or mRNA, oligo dT primer, dNTP and Cy3-labeled dUTP. After the addition, it is prepared by heating at 37 to 45 ° C. for 1 to 3 hours, preferably at 42 ° C. for 1 hour. Preparation of Cy5-labeled normal colon mucosa-derived cDNA (hereinafter also referred to as “Cy5 cDNA”) used as a comparative control is performed in the same manner using total RNA of normal colon mucosa tissue. The thus obtained Cy3 cDNA and Cy5 cDNA are each heated in a denaturing solution at 65 to 70 ° C. for 10 to 20 minutes, preferably at 70 ° C. for 10 minutes, neutralized, and then mixed in equal amounts (hereinafter referred to as this mixing). The liquid may be referred to as “Cy5 / Cy3 cDNA”). As the denaturing solution, 0.5N NaOH or 1N NaOH containing 50 mM EDTA can be used, but 0.5N NaOH containing 50 mM EDTA is preferably used. Purification of Cy5 / Cy3 cDNA is performed according to the attached method using a commercially available kit such as Microcon-30 (Amicon).

  Hybridization of Cy5 / Cy3 cDNA and the probe printed on the DNA microarray is carried out as follows. First, in order to heat denature the probe, the DNA microarray was heat-treated, and a Cy5 / Cy3 cDNA-containing hybridization solution that had been heat-treated at 100 ° C. for 2 minutes was added dropwise, covered with a cover glass, and the DNA microarray was then placed in a sealed container. And perform hybridization. As hybridization conditions, when the hybridization solution contains formamide, hybridization is performed at 42 ° C. for 12 hours or more, and when it does not contain formamide, hybridization is performed at about 68 ° C. for 12 hours or more. . After completion of hybridization, the fluorescence of Cy3 and Cy5 is scanned with an instrument such as Scan Array 4000 (GSI Lumonics) to obtain a fluorescence pattern as image data. Subsequently, these image data are analyzed using, for example, microarray data dedicated analysis software such as Quantarray software (GSI Lumonics) to obtain the fluorescence intensity of Cy3 and Cy5 for all probes as numerical data in text format. be able to.

  In the preferred embodiment of the present invention, the above-described DNA microarray is used, but the same result can be obtained by using a synthetic DNA having a chain length effective for hybridization instead. For example, based on the gene name or sequence information disclosed in the present invention, a synthetic DNA consisting of a part of the sequence and having a length of about 20 nucleotides or more is used as a probe and is immobilized on a glass substrate or the like. It is also possible.

  In general, data with low fluorescence intensity is greatly affected by the background. For example, data of probes with low fluorescence intensity is rejected by leaving only 3,000 data points from the higher fluorescence intensity. And treated as missing values. Subsequently, an operation for correcting and standardizing the deviation in detection sensitivity adjustment between Cy3 and Cy5 that may occur during scanning is performed. That is, Cy3 / Cy5, which is the ratio of the fluorescence intensity values of Cy3 and Cy5 for each probe, is calculated, converted to a logarithmic value with a base of 2 (hereinafter referred to as “log (Cy3 / Cy5)”), A standardized log (Cy3 / Cy5) value can be obtained by subtracting the median of all log (Cy3 / Cy5) values from the log (Cy3 / Cy5) value for the probe. The standardized log (Cy3 / Cy5) value can be used as the expression level of each gene.

  The standardized numerical data for all cases obtained in this way (hereinafter sometimes referred to as “standardized numerical data”) are once integrated, and probe data containing many missing values are analyzed later. The following selection operation is performed for the purpose of removing from That is, only probe data for which data has been acquired in 128 or more cases, which is 85% or more out of all 150 cases analyzed by the microarray, is selected. As a result, it is possible to select only the probe data containing 15% or less of missing values. In addition, the following selection operation is added for the purpose of excluding personal genetic background factors. That is, for each probe, a variance value in the data of 150 primary colon cancer lesions and a variance value in the data of 12 normal colon mucosa were calculated, and the former exceeded 1.1 times the latter. Only the data of the existing probe is selected.

  Through these series of selection operations, standardized numerical data of 2,121 types of probes is selected as a subsequent analysis target. The presence of missing values included in the standardized numerical data selected in this way causes inconvenience in later statistical analysis, and thus needs to be supplemented by some method.

  Various methods can be used as a complementation method. For example, the average value of all the data for the case containing the missing value is added to the average value of all the cases for the gene containing the missing value. There is a method of complementing with the value obtained by subtracting the average value of the data of all genes for all cases. In addition, in the report of Troyanskaya et al. (Bioinformatics, vol. 17, p. 520-525, 2001), there are three types of complementation methods: K-Nearest Neighbors (KNN) method, Singular Value Decomposition (SVD) based method and row An example of completion by the average method is shown. By applying any of these methods, it is possible to complement all missing values.

  The standardized numerical data (hereinafter also referred to as “standardized gene expression data”) supplemented with the missing values thus prepared is not affected by the background, and an error due to a difference in detection sensitivity between Cy3 and Cy5. And gene expression information that does not include missing values, and the variation range of gene expression in the primary colorectal cancer lesion exceeds the variation range of gene expression due to individual differences compared to normal colon mucosa The reliability of subsequent statistical analysis can be ensured.

  There is no established general method for processing a large amount of gene expression data obtained by DNA microarray measurement using statistical methods and deriving gene sets that meet the objectives. What you need is the reality. In the present invention, analysis is first performed by four different approaches, and for each of these, a gene group that can be used for prediction of the presence or absence of lymph node metastasis is identified.

The four approaches taken in the present invention are:
(A) Support Vector Machine (SVM) (Hastie et al., The Elements of Statistical Learning-Data Mining, Inference, and Prediction, Springer, 2001),
(B) Extension method of Principal Component Analysis / artificial Neural Network (PCA / aNN) (Khan et al., Nature Medicine, vol. 7, p. 673-679, 2001),
(C) Hierarchical Cluster Analysis (HCA) + Stepwise Logistic Discrimination and (d) Classification And Regression Tree (CART) (Breiman et al., Classification and Regression Trees, Wadswarth, 1983) + Logistic Discrimination
It is. When identifying gene groups, analysis is performed by dividing all data into two groups for predicting gene identification and evaluation for the purpose of ensuring statistical reliability. More specifically, the data of 150 cases are divided into 99 cases consisting of 42 cases with lymph node metastasis and 57 cases without lymph node metastasis, 21 cases with lymph node metastasis and 30 cases without lymph node metastasis. The data of 99 cases of the former are used to identify genes and establish discriminants for predicting the presence or absence of lymph node metastasis. The discriminant is evaluated by discriminating the data. In the following description, the former 99 cases of data used for identification of predictive genes and establishment of discriminants are expressed as “training data” and the latter 51 cases used for discriminant evaluation. Data may be expressed as “test data”.

  Regarding three of (a), (c), and (d) among the above four approaches, analysis is performed by randomly dividing the above data into two times 100 times in consideration of sample variation at the time of division. Thus, after identifying 100 types of gene groups for determination, genes that have been identified many times are employed.

  On the other hand, the approach (b) is performed only for the first two divisions in consideration of a huge amount of calculation. However, the data for 99 training examples are randomly divided into 1250 times at a ratio of 2 to 1, and the principal component analysis and the neural network learning are repeated using the data. After learning, rank genes based on their sensitivity to identify the presence or absence of lymph node metastasis, and narrow down the genes. Starting from 2121 genes, the learning is advanced with 1536, 768, 384, 192, 96, 48, and 24 refined numbers.

  By performing the above analysis, the number of genes included in each gene set and the correct classification rate of test data using established discriminants (the discriminant discriminant results and histopathological examination results As an average value of the number of matched cases / number of test data × 100 (%)), (a) has 144 genes and a normal classification rate of 80.2% (standard deviation is 5.6%), ( For b), the number of genes is 192 and the correct classification rate is (90.2%). For (c), the number of genes is 133 and the correct classification rate is 78.6% (standard deviation is 6.2%). As for (d), the number of genes is 138 and the correct classification rate is 86.3% (standard deviation: 4.5%). At this time, 16 types of genes are commonly included in the gene set selected by each approach.

  Next, in order to narrow down the number of genes used for the prediction while keeping the normal classification rate from dropping, the target genes are first set to the above 16 genes, and the contribution of each of these 16 genes (hereinafter referred to as “main effect”). In addition to (to be described), statistical analysis is also performed in consideration of the interaction of two genes. As a result, it is possible to search for a discrimination rule in a wider range including not only the main effect by individual genes but also the interaction between genes, and it is expected that high discrimination performance can be maintained.

  For the interaction search, CART analysis is performed again with the presence or absence of lymph node metastasis as a response for each of the 100 training data used in the above analysis. In this CART analysis, Free software R rpart was used. At that time, default values were used for all operation parameters. By this analysis, 3 to 5 genes can be obtained per analysis as the number of genes appearing as variables for instructing data division.

  For example, when there are three genes that appear, there are three possible gene pairs, and thus all these pairs are considered as interactions. Similarly, when there are 4 genes, 6 pairs, and when 5 genes, 10 pairs are regarded as an interaction. Then, in order to cover as many candidates as possible, 18 gene pairs that appear 12 times or more out of 100 analyzes are selected as interaction candidates.

Next, in order to establish a discriminant, using the data of 150 cases, stepwise in a logistic regression model using the main effects of 16 genes and 18 interactions as explanatory variables and the presence or absence of lymph node metastasis as responses. Perform variable selection. At that time, the p-value of the significance test of the regression coefficient is used as a variable inclusion criterion (less than 0.05) and an exclusion criterion (greater than 0.05). As a result, six variables, namely, G1592, G3031, G3826, G4370, the interaction between G2645 and G3177, and the interaction between G3753 and G3826 are selected. And the discriminant for predicting the presence or absence of lymph node metastasis is
D = 0.23027-2.7132 × “expression amount of G1592”
+ 8.9509 × “expression amount of G3031”
+ 8.7975 × “expression amount of G3826”
-2.398 × “expression amount of G4370”
+ 3.5126 × “expression amount of G2645” × “expression amount of G3177”
−8.8226 × “expression amount of G3753” × “expression amount of G3826”
Thus, a determination rule is derived that lymph node metastasis is present when D> 0 and no lymph node metastasis is present when D ≦ 0. Seven genes appearing in this discriminant, that is, G1592, G2645, G3031, G3177, G3753, G3826, and G4370 are selected as a set of genes that contribute to the identification of the presence or absence of lymph node metastasis. Their gene names are shown in Table 1.

  Finally, the discrimination performance of lymph node metastasis by the selected gene set is evaluated by the LOO method. That is, using the data of the remaining 149 samples excluding one sample, the logistic discriminant including the above six variables is estimated, and the operation of discriminating the sample removed by that is performed for each of the 150 samples. . Thereby, as shown in Table 2, the correct classification rate by the set of selected genes is estimated to be 88.7% (sensitivity: 77.8%, specificity: 96.6%). As described above, in the present invention, it is possible to clarify a gene set necessary for predicting the presence or absence of lymph node metastasis of colorectal cancer with high accuracy.

  EXAMPLES The present invention will be described in detail below with reference to examples, but the present invention is not limited by these examples. The reagents used in the examples were those purchased from Nacalai Tesque, unless otherwise specified.

(1) Preparation of total RNA from colorectal cancer tissue sample As a sample for performing gene expression analysis in colorectal cancer using a DNA microarray, colorectal cancer collected through informed consent and excised at the time of colorectal cancer surgery 150 cases of primary lesion tissue were used. The breakdown consists of 63 patients (hereinafter referred to as “positive lymph node metastasis”) derived from patients whose lymph node metastasis was observed by histopathological observation at the time of primary lesion removal, and lymph node metastasis There are 87 cases (hereinafter referred to as “negative lymph node metastasis cases”) derived from patients who were not recognized. Total RNA was extracted from these colon cancer tissue samples using TRIzol reagent (purchased from GIBCO BRL). The extraction procedure basically followed the manual attached to the reagent. In addition to this, total RNA derived from 40 portions of normal large intestine mucosa was extracted and mixed to obtain standard normal large intestine mucosa total RNA used throughout all experiments. The concentrations of these RNA samples were calculated based on the absorbance at a wavelength of 260 nm measured using a spectrophotometer as usual.

(2) Preparation of fluorescent label target A fluorescent label target to be hybridized with a DNA microarray was prepared by the following procedure. First, 25 μg of total RNA derived from colon cancer sample (hereinafter referred to as “colon cancer RNA”) and 25 μg of standard normal colon mucosa total RNA (hereinafter referred to as “standard colon mucosa RNA”) are put in separate tubes, 2 μg of oligo dT primer consisting of 18 nucleotides was added to each, the volume was made up to 14 μL with sterilized distilled water, heated at 70 ° C. for 10 minutes, then immediately transferred to ice and rapidly cooled. Each tube was then filled with 6 μL of 5 × First Strand Buffer, 3 μL of 0.1 M DTT, 1.5 μL of 20 × dNTPmix (mixture of 10 mM dATP, dCTP, dGTP and 6 mM dTTP) and 0.5 μL RNAguard. Was added.

  Furthermore, 3 μL of dUTP labeled with the fluorescent dye Cy3 (hereinafter referred to as “Cy3-dUTP”; concentration 1 mM) in the tube containing the colon cancer RNA, and Cy5 in the tube containing the standard colon mucosa RNA. 3 μL of labeled dUTP (hereinafter referred to as “Cy5-dUTP”; concentration 1 mM) was added and incubated at 42 ° C. for 2 minutes. Thereafter, 2 μL of Superscript II, which is a reverse transcriptase, was added to each tube and incubated at 42 ° C. for an additional hour to carry out the labeling reaction. By this reaction, when cDNA synthesis occurs using colorectal cancer RNA and standard colorectal mucosal RNA as a template, Cy3-dUTP and Cy5-dUTP are incorporated, respectively, so that a colon cancer label target fluorescently labeled with Cy3 and Cy5, respectively, A standard colon mucosa label target is generated.

  All 5 × First Strand Buffer, 0.1M DTT and Superscript II used in this reaction were purchased from GIBCO BRL. DATP, dCTP, dGTP and dTTP, Cy5-dUTP and Cy3-dUTP, and RNAguard were all purchased from Amersham Biosciences. After the reaction, 5 μL of a denaturing solution (0.5 N NaOH, 50 mM EDTA) is added to each tube, heated at 70 ° C. for 10 minutes, and then 7.5 μL of 1 M Tris-HCl (pH 7.5) is added. Neutralized. At the stage where these treatments were performed, the colon cancer label target and the standard colon mucosa label target were mixed, and 10 μg of human COT-1 DNA (purchased from GIBCO BRL) was added thereto. The buffer solution was adjusted to 500 μL by adding TE buffer, and purified and concentrated using Microcon-30 (purchased from Amicon) to remove unreacted Cy5-dUTP and Cy3-dUTP. The procedure of purification / concentration was according to the manual attached to Microcon-30. Finally, it was concentrated until the total volume became 5 μL, and this was used as a label target to be hybridized to the DNA microarray.

(3) Pretreatment of DNA microarray Masking is performed by immersing the DNA microarray in a masking solution (mixture of 3 g of succinic anhydride, 190 mL of N-methyl-2-pyrrolidone and 21 mL of 0.2 M sodium borate) for 5 minutes. Then, the cDNA printed on the microarray was heat denatured by immersing it in distilled water at 95 ° C. for 3 minutes. Immediately thereafter, it was immersed in 95% or more of ethanol for 1 minute to dehydrate and air dry.

(4) Hybridization of label target and DNA microarray For 5 μL of the label target solution prepared as described above, 2.5 μL of 10 mg / mL polyadenine (purchased from Roche), 0.5 μL of 10% SDS Solution, 3 μL of 20 × PM solution (mixture of 0.4% BSA and 1% SDS), 15 μL of formamide, 3 μL of 20 × SSC (3M sodium chloride, 0.3M sodium citrate, pH 7.0) and sterilization 1 μL of distilled water was added and heated at 100 ° C. for 2 minutes, and then allowed to stand at room temperature for about 30 minutes in the dark. After that, it is dropped on the portion of the DNA microarray where the cDNA is pretreated by the method described in the previous section, covered with a 24 × 40 mm cover glass (purchased from Matsunami Glass Industry), and the microarray is placed in a sealed container. The label target was hybridized to the cDNA on the microarray by placing the entire container in a 42 ° C. incubator for about 16 hours. After hybridization, the microarray was soaked in 2 × SSC containing 0.1% SDS for 10 minutes and then soaked in 0.1 × SSC containing 0.1% SDS for 10 minutes. Further, after being soaked in 0.1 × SSC and washed twice for 5 minutes, the drops were cut and air-dried in the dark.

(5) Microarray scanning and data analysis Microarrays that were air-dried after washing were scanned independently for fluorescence of Cy3 and Cy5 using ScanArray 4000 (GSI Lumonics), a confocal laser scanner dedicated to microarrays. The fluorescence patterns of Cy3 and Cy5 derived from the colon cancer target and standard colon target hybridized with each of the above probes were obtained as 16-bit Tiff format scan image data. Subsequently, the fluorescence intensity of Cy3 and Cy5 for all probes was obtained as numerical data in text format by analyzing the image data using QuantArray software (manufactured by GSI Lumonics), which is analysis software dedicated to microarray data. . For background correction, the fluorescence intensity value of the part where the cDNA was not printed was subtracted from the fluorescence intensity value for each probe. In addition, since the portion where the fluorescence intensity value is low is greatly affected by the experimental error, other data are rejected leaving about 3000 data points from the higher fluorescence intensity value. The ratio of the fluorescence intensity values of Cy3 and Cy5 for each probe, that is, Cy3 / Cy5, was calculated and converted to a logarithmic value with a base of 2 (hereinafter referred to as “log (Cy3 / Cy5)”). In order to correct and standardize the shift in detection sensitivity adjustment between Cy3 and Cy5 that may occur during scanning, the median (median) of all log (Cy3 / Cy5) values is calculated from the log (Cy3 / Cy5) values for each probe. ) Was subtracted to obtain a normalized log (Cy3 / Cy5) value.

  By the above operations, the relative expression intensities of the primary colorectal cancer lesions for 63 cases with lymph node metastasis and 87 cases without lymph node metastasis when using standard colonic mucosa RNA as a standard are logarithmized and standardized. Obtained numerical data. In addition, by the same operation, numerical data for 12 normal colon mucosa samples with reference to standard colon mucosa RNA was also obtained. Among these numerical data, data has been obtained for 128 cases or more, which is 85% of the analyzed 150 primary colorectal cancer lesions, and the variance value in the data of 150 primary colon cancer lesions ( Only the data for a total of 2,121 types of probes whose variance) exceeded 1.1 times the variance in the data for 12 normal colonic mucosa were used for subsequent statistical analysis.

  Among the data of these 2,121 probes, the expression level was low in some cases, and some of them were rejected because they were below the cut-off value, and these data are missing values. There were a total of 8,816 such missing values. Since the total number of data is 150 × 2,121 = 318,150, the content rate of missing values is about 2.8%. The presence of these missing values causes inconveniences in later statistical analyses, so they were supplemented using the k-Nearest Neighbor method. Specifically, in the data matrix of 150 cases × 2,121 probes, the expression level that was a missing value was estimated by the average value of the expression levels of the eight genes closest to that gene in the inter-sample distance. Specifically, in the data matrix of 150 cases × 2,121 probes, the distances between all the paired samples were calculated. And the average value of the gene expression level which should be complemented with the 8 samples which were the closest to the sample with the missing value was obtained and used as the complemented value of the missing value. Here, the number of samples closest to the sample having a missing value was defined as the number with which the root mean square (RMS) is minimized by sequentially increasing the number. All the numerical data obtained by complementing the missing values in this way are hereinafter referred to as standardized gene expression data.

Determination of highly discriminating gene set for predicting lymph node metastasis by statistical analysis of standardized gene expression data In this example, probes printed on a DNA microarray may be referred to as genes.
As an initial attempt, the following four approaches were applied to identify the gene set for judgment. (A) Support Vector Machine (SVM) (Hastie et al., The Elements of Statistical Learning-Data Mining, Inference, and Prediction, Springer, 2001), (b) Principal Component Analysis / artificial Neural Network (PCA / aNN) ( Khan et al., Nature Medicine, vol. 7, p. 673-679, 2001), (c) Hierarchical Cluster Analysis (HCA) + Stepwise Logistic Discrimination and (d) Classification And Regression Tree (CART) (Breiman et al., Classification and Regression Trees, Wadswarth, 1983) + Logistic Discrimination.

  In the identification of the gene group, all data were divided into two groups for determination gene identification and evaluation and analyzed for the purpose of ensuring statistical reliability. More specifically, the data of 150 cases are divided into 99 cases consisting of 42 cases with lymph node metastasis and 57 cases without lymph node metastasis, 21 cases with lymph node metastasis and 30 cases without lymph node metastasis. The data of 99 cases of the former are used to identify genes and establish a discriminant for determining the presence or absence of lymph node metastasis. By discriminating data, gene identification and discriminant evaluation were performed. In the following description, the data used for identification of the gene for determination and establishment of the discriminant like the data for the former 99 cases is expressed as “training data”, and the data for the latter 51 cases In some cases, data used for discriminant evaluation is expressed as “test data”.

  With regard to three of the above four approaches (a), (c) and (d), 100 data groups for determination can be obtained by analyzing the above data by dividing the data into two at random 100 times. After identification, genes that were identified many times were employed. On the other hand, the approach (b) was performed only for the first two divisions. However, the data for 99 cases for training were randomly divided into 1250 times at a ratio of 2: 1, and principal component analysis and neural network learning were repeated using the data. After learning, the genes were ranked based on the sensitivity to identify the presence or absence of lymph node metastasis, and the genes were narrowed down. Starting from 2121 genes, learning proceeded with 1536, 768, 384, 192, 96, 48, and 24 refined numbers.

  As a result of the above examination, a discriminant gene set and a discriminant were established for each approach. The number of genes included in each gene set and the correct classification rate of test data using established discriminants (number of cases in which the discriminant discriminant results match histopathological examination results / test data) The average value of (number x 100 (%)) is 144 genes for (a), the correct classification rate is 80.2% (standard deviation is 5.6%), and the number of genes is 192 for (b). The correct classification rate is (90.2%), the number of genes is 133 for (c), the correct classification rate is 78.6% (standard deviation is 6.2%), and the number of genes is for (d). 138 and the correct classification rate was 86.3% (standard deviation: 4.5%). In addition, there were 16 genes that were commonly included in the gene set selected by each approach.

  As mentioned above, each of the four different approaches was able to achieve a high correct classification rate of over 80%. However, the number of genes used for discrimination exceeds 100 in any approach, and it was considered that there are too many for practical use as an assay method.

  Therefore, we considered establishing a new discrimination rule by narrowing down the number of genes used for judgment while keeping the normal classification rate from dropping. For this purpose, first, the genes of interest were 16 genes that were commonly included in the gene set selected by the above approaches. Next, in addition to the contribution of each of these 16 genes (hereinafter referred to as “main effect”), it was devised to take into account the interaction of 2 genes. We expected that high discrimination performance could be maintained by searching for discrimination rules in a wider range including not only the main effects of individual genes but also interactions between genes.

  In order to search for interaction, CART analysis was performed again with the presence or absence of lymph node metastasis as a response in each of the 100 training data used in the above analysis. As a result, the number of genes that appeared as variables for instructing data division was 3 to 5 per analysis. For example, when there are three genes that appear, there are three possible gene pairs, so all these pairs were considered as interactions. Similarly, when there were 4 genes, 6 pairs and when 5 genes, 10 pairs were considered as interactions. Then, 18 gene pairs that appeared 12 times or more out of 100 analyzes were selected as interaction candidates.

Next, in order to establish a discriminant, using the data of 150 cases, stepwise in a logistic regression model using the main effects of 16 genes and 18 interactions as explanatory variables and the presence or absence of lymph node metastasis as responses. Variable selection was performed. At that time, the p-value of the significance test of the regression coefficient was used as a variable inclusion criterion (less than 0.05) and an exclusion criterion (greater than 0.05). As a result, six variables were selected: G1592, G3031, G3826, G4370, G2645 and G3177 interaction, and G3753 and G3826 interaction. G1592, G3031, G3826, G4370, G2645, G3177 and G3753 are serial numbers assigned to each probe (gene) on the ColonoChip used in the present invention. And the discriminant for judging the presence or absence of lymph node metastasis is
D = 0.2307-2.7132 × G1592 expression level + 8.9509 × G3031 expression level + 8.7975 × G3826 expression level-2.3098 × G4370 expression level + 3.5126 × G2645 expression level × G3177 Expression level−8.8226 × G3753 expression level × G3826 expression level was estimated, and a determination rule was derived that there was lymph node metastasis when D> 0 and no lymph node metastasis when D ≦ 0.

  Seven genes appearing in this discriminant, namely G1592, G2645, G3031, G3177, G3753, G3826, and G4370 were selected as a set of genes that contribute to the identification of the presence or absence of lymph node metastasis. Their gene names are shown in Table 1. In Table 1, the access numbers in the RefSeq database for each of the above genes are also shown. The RefSeq database can be accessed from the National Center for Biotechnology Information (NCBI) website (http://www.ncbi.nlm.nih.gov/RefSeq/index.html).

  Finally, the judgment performance of lymph node metastasis by the selected gene set was evaluated by the LOO method. That is, using the data of the remaining 149 samples excluding one sample, the logistic discriminant including the above six variables was estimated, and the operation of discriminating the sample excluded by that was performed for each of the 150 samples. . The results are shown in Table 2. From Table 2, the correct classification rate by the selected gene set was estimated to be 88.7% (sensitivity: 77.8%, specificity: 96.6%).

  According to the determination of lymph node metastasis enabled by the present invention, it is possible to select a better treatment policy according to the case and to expect a medical economic effect. For example, prognosis improvement can be expected by aggressive treatment for patients with a high possibility of lymph node metastasis, while postoperative patients with a low possibility of lymph node metastasis. Adjuvant therapy can be mild and can reduce the physical and economic burden on the patient.

  Furthermore, since it is assumed that individual genes included in the gene set disclosed in the present invention may function as a cause of lymph node metastasis, drugs targeting these genes and their expression products may be used. It can also be expected to develop and be able to directly suppress lymph node metastasis.

Claims (8)

A method for selecting a gene set for predicting the presence or absence of colorectal cancer lymph node metastasis, including the following steps (1) to (4):
(1) Analyzing gene expression information in the primary colorectal cancer tissue of a patient whose presence or absence of lymph node metastasis was revealed by histopathological determination using at least one supervised learning analysis method. Selecting a gene group in each analysis method that can classify the presence or absence of lymph node metastasis with a correct classification rate of 75% or more,
(2) a step of selecting a common gene selected in common in any analysis method from the gene group selected in each analysis method used in (1),
(3) analyzing the gene expression information, instructing the classification of the presence or absence of lymph node metastasis from any combination of two or more genes, and selecting a combination of genes showing an interaction; and (4) A step of selecting a variable in a logistic regression model in which the combination of the common gene and the gene is used as an explanatory variable and the presence or absence of lymph node metastasis is used as a response.   (1) Analysis method includes (a) Support Vector Machine, (b) Extension of Principal Component Analysis Artificial Neural Network, (c) Combination of Hierarchical Cluster Analysis and Stepwise Logistic Discrimination, and (d) Classification And Regression Tree and Logistic The method of claim 1, comprising at least one selected from the group consisting of a combination of discriminations.   The method according to claim 1 or 2, wherein the analysis method of (3) is a combination of (d) Classification And Regression Tree and Logistic Discrimination.   4. The method according to claim 1, wherein the variable selection method of (4) is a stepwise variable selection method.   A gene set for predicting the presence or absence of colorectal cancer lymph node metastasis, selected by the method according to claim 1.   At least the gene represented by the access number (serial number) of the database of NM_003404 (G1592), NM_002128 (G2645), NM_052868 (G3031), NM_005034 (G3177), NM_001540 (G3753), NM_005722 (G3826), and NM_015315 (G4370) The gene set according to claim 5, comprising:   A method for predicting the presence or absence of colorectal cancer lymph node metastasis, comprising using the gene set according to claim 5. 8. The method of claim 7, wherein the following discriminant is used:
D = 0.23027-2.7132 × NM_003404 (G1592) expression level + 8.9509 × NM_052868 (G3031) expression level + 8.7975 × NM_005722 (G3826) expression level-2.3098 × NM_015315 (G4370) expression level Amount + 3.5126 x NM_002128 (G2645) expression x NM_005034 (G3177) expression-8.8226 x NM_001540 (G3753) expression x NM_005722 (G3826) expression (D> 0, lymph node metastasis) , It is determined that there is no lymph node metastasis when D ≦ 0).