KR20200057664A - Gene expression marker screening method using neural network based on gene selection algorithm - Google Patents

Abstract

The present invention relates to a gene expression marker screening method using a neural network based gene selection algorithm. According to the present invention, the method comprises: a step of collecting each biopsy tissue from a plurality of patients and collecting a plurality of gene expression information experimentally measured from each collected biopsy tissue; a step of calculating a discriminative index (DI) for each gene by applying the collected plural gene expression information to a pre-trained neural network based gene selection algorithm; a step of listing the genes according to the calculated discriminative index (DI) and selecting a plurality of specific genes having a large discriminative index (DI) from the listed genes; and a step of predicting whether or not cancer occurs by using expression values of the selected specific genes. According to the present invention, genes are ranked on the basis of the discriminative index (DI), which indicates the classification ability to distinguish the normal group from the cancer patient group, and the optimal gene set among the highest ranking genes through the list of genes ranked by the discriminative index (DI) can be selected.

Description

Gene expression marker screening method using neural network based on gene selection algorithm}

The present invention relates to a method for selecting a gene expression marker using a neural network based gene selection algorithm, and more specifically, a gene expression for selecting a gene expression marker associated with a plurality of cancers from gene expression information using a neural network based gene selection algorithm It relates to a marker selection method.

Next-generation sequencing (NGS) or massively parallel sequencing is a method of massively parallelizing sequencing to increase sequencing data production.

NGS has been applied in many research fields because it can convert molecular information into numerical values. But. The approach using NGS had to select the appropriate gene (or locus) to direct the next step in a given study. For example, in the case of the human genome, selecting a rational gene (function) from a list of expression levels above about 50,000 genes (or up to 190,000 transcripts) has become a major bottleneck.

Many researchers selected genes from the differentially expressed gene list (DEG) using a differentially expressed gene (DEG) identification algorithm with a p value adjusted to 0.05 (or less) in various tests. However, as the number of samples increased, the number of DEGs increased to thousands. Accordingly, there has been a need for a method of automatically recommending an ideal gene set for a biomarker candidate.

In an embodiment of the present invention, an optimal biomarker is selected using a neural network-based gene selection algorithm.

The background technology of the present invention is disclosed in Republic of Korea Patent Publication No. 10-1489536 (2015.02.04 announcement).

An object of the present invention is to provide a gene expression marker selection method for selecting gene expression markers associated with 12 cancers from gene expression information using a neural network-based gene selection algorithm.

According to an embodiment of the present invention for achieving such a technical problem, in the method for selecting a gene expression marker using a neural network-based gene selection algorithm, each biopsy tissue is collected from a plurality of patients and experimentally obtained from each collected biopsy tissue. Collecting a plurality of gene expression information measured by, applying the collected plurality of gene expression information to a pre-trained neural network-based gene selection algorithm to calculate a differential index (DI) for each gene, Genes are listed according to the calculated differential index (DI), selecting a plurality of specific genes having a large differential index (DI) from the listed genes, and using the expression values of the selected plurality of specific genes to develop cancer And predicting whether or not to include.

Further comprising the step of constructing and learning the neural network based gene selection algorithm, the step of constructing and learning the neural network based gene selection algorithm is a plurality of cancer types from the Cancer Genome Atlas (TCGA) program. Receiving gene expression information for, grouping the received gene expression information into a cancer patient group and a normal group, and randomly extracting gene information obtained from each group to form a data set, the formed data set And constructing a gene selection algorithm for extracting a plurality of specific genes that are classified into a normal group and a cancer patient group.

The plurality of cancer types include bladder urinary tract carcinoma (BLCA), breast invasive carcinoma (BRCA), adenocarcinoma (COAD), squamous cell carcinoma of the head and neck (HNSC), renal chromophore (KICH), and kidney transparent cell carcinoma (KIRC). , Renal papillary cell carcinoma (KIRP), liver cancer (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), prostate cancer (PRAD) and thyroid carcinoma (THCA).

The gene selection algorithm calculates each discrimination index (DI) value for all the expression genes included in a plurality of cancer types that are always received, and uses the calculated ranking of the discrimination index (DI) values to select a plurality of upper multiples. Specific genes can be extracted.

The discrimination index (DI) value may be calculated through the following equation.

는 j번째 유전자에 대응하는 암조직의 유전자 발현값들의 총합을 나타내고,

는 j번째 유전자에 대응하는 정상 조직의 유전자 발현값들의 총합을 나타내며, W는 가중치를 나타낸다.here,

Denotes the sum of gene expression values of cancer tissues corresponding to the j-th gene,

Denotes the sum of gene expression values of normal tissues corresponding to the jth gene, and W denotes a weight.

In the generating of the data set, the expression gene information may be randomly extracted regardless of the ratio of different cancer samples and normal samples for each of the plurality of cancer types to generate a data set.

In the generating of the data set, a learning data set, a verification data set and an evaluation data set are generated at a predetermined rate using the entire cancer gene expression data, and the generated learning data set, the verification data set, and the evaluation data, respectively. The three can form the same ratio of the cancer sample and the normal sample.

The plurality of specific genes may include FN1, ALB, EEF1A1, SFTPC, GAPDH, P4HB, DCN, A2M, MGP, UMOD, GPX3, FTL, ACPP and CTSD.

As described above, according to the present invention, genes are ranked based on a discrimination index (DI) indicating a classification ability for distinguishing between a normal group and a cancer patient group, and the highest ranking is obtained through a list of genes ranked by the discrimination index (DI) It is possible to select the optimal gene set among the genes.

1 is a view schematically showing a gene expression marker selection device according to an embodiment of the present invention.
2 is a flowchart schematically showing a method of selecting a specific gene using a neural network-based gene selection algorithm according to an embodiment of the present invention.
3 is a view for explaining the step S210 shown in FIG.
4 is a graph showing classification accuracy according to the number of genes given in step S213 shown in FIG. 3.
5 is a view showing that the differential index is calculated for each gene in step S230.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this process, the thickness of the lines or the size of components shown in the drawings may be exaggerated for clarity and convenience.

In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to a user's or operator's intention or practice. Therefore, the definition of these terms should be made based on the contents throughout the specification.

Hereinafter, a recurrence prediction apparatus according to an embodiment of the present invention will be described in more detail with reference to FIG. 1.

1 is a view schematically showing a gene expression marker selection device according to an embodiment of the present invention.

1, the gene expression marker selection device 100 according to an embodiment of the present invention includes a collection unit 110, an algorithm generation unit 120, a differential index calculation unit 130, and a specific gene selection unit ( 140) and the prediction unit 150.

First, the collection unit 110 collects genetic information extracted from the tissues of a plurality of subjects. In other words, the collection unit 110 extracts RNA of tissue collected from a plurality of subjects. In addition, the collection unit 110 obtains gene expression data by applying the extracted RNA to the nCounter®Analysis System. Here, the obtained gene expression data includes approximately 20,000 gene information.

Then, the collection unit 110 transmits the obtained gene expression data to the differential index calculation unit 130.

The algorithm generator 120 generates a gene selection algorithm using the received gene expression data. Here, the gene selection algorithm is a model for selecting a specific gene that affects the classification of cancer patients and normal persons from expression gene information.

In other words, the algorithm generating unit 120 acquires gene expression data for 12 different cancer types disclosed in the cancer genome atlas (TCGA). The obtained gene expression data consists of a total of 6,226 samples (5,609 cancer samples and 617 normal samples).

In addition, the algorithm generating unit 120 randomly selects n of the 12 gene expression data of 12 different cancer types and generates a plurality of combined data sets. The algorithm generating unit 120 divides the generated plurality of data sets in a ratio of 7: 2: 1, and the data set corresponding to 7 is used for learning, and the data set corresponding to 2 is used for evaluation. In addition, the data set corresponding to 1 is used for final evaluation. That is, the algorithm generator 120 generates a gene selection algorithm that calculates a differential index (DI) for a gene by performing learning and evaluation using a data set.

The discrimination index calculation unit 130 applies the expression gene information obtained from the subject to the pre-trained gene selection algorithm. In addition, the discrimination index calculation unit 130 acquires a discrimination index (DI) score for each of the input expression genes. Here, the discrimination index (DI) represents a value calculated in order to evaluate the classification ability of how well a specific gene distinguishes a given group.

The specific gene selection unit 140 lists all genes according to the calculated differential index (DI). In addition, the specific gene selection unit 140 selects genes corresponding to the top 14RO among all the genes listed. In addition, the specific gene selection unit 140 selects genes corresponding to the top 14 selected genes as specific genes.

Finally, the prediction unit 150 determines whether cancer has occurred using specific genetic information corresponding to the top 14 selected.

Hereinafter, a method of selecting a specific gene using the gene expression marker selection device will be described in more detail with reference to FIGS. 2 to 5.

2 is a flowchart schematically showing a method of selecting a specific gene using a neural network-based gene selection algorithm according to an embodiment of the present invention.

As illustrated in FIG. 2, first, the algorithm generating unit 120 receives gene expression information for a plurality of cancers from the Cancer Genome Atlas (TCGA) program. In addition, the algorithm generator 120 constructs a gene selection algorithm using the gene expression information for the 12 types of cancer received (S210).

Hereinafter, step S210 will be described in more detail with reference to FIGS. 3 and 4.

FIG. 3 is a diagram for explaining step S210 shown in FIG. 2, and FIG. 4 is a graph showing classification accuracy according to the number of genes given in step S213 shown in FIG. 3.

As shown in FIG. 3, the algorithm generator 120 receives gene expression information for 12 types of cancers disclosed in the Cancer Genome Atlas (TCGA) program (S211).

Table 1 above shows cancers in 12 types received through the Cancer Genome Atlas (TCGA) program, and shows cancer tissue samples and normal tissue samples obtained for each cancer.

Here, the names of the 12 cancers are bladder urinary tract carcinoma (BLCA), breast invasive carcinoma (BRCA), adenocarcinoma (COAD), squamous cell carcinoma of the head and neck (HNSC), renal chromophore (KICH), and kidney transparent cell carcinoma (KIRC). , Kidney papillary cell carcinoma (KIRP), liver cancer (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), prostate cancer (PRAD) and thyroid carcinoma (THCA).

Next, the algorithm generating unit 120 groups cancer tissues and normal breakfasts from 12 types of cancers obtained, and then randomly extracts gene expression information obtained from each group to form a data set (S212).

More specifically, as shown in Table 1 above, the 12 types of cancer received include a total of 6210 samples. Here, 5,609 are samples for cancer tissue, and 617 are samples for normal tissue.

Therefore, the algorithm generator 120 distributes a total of 6210 at a ratio of 7: 2: 1 to generate a training data set, an evaluation data set, and a final evaluation data set.

At this time, each cancer is composed of a ratio of different cancer samples and normal samples. Therefore, the algorithm generator 120 randomly extracts the expression gene information regardless of the ratio to generate a data set. However, the training data set, the evaluation data set, and the final evaluation data set are generated while maintaining the ratio between the cancer sample and the normal sample.

Next, the algorithm generator 120 learns the gene selection algorithm using the generated training data set (S213).

On the other hand, the genetic selection algorithm is based on a neural network method for training the weights of the network using a training data set. At this time, since the trained weight is highly dependent on an arbitrary value assigned to the initial value, the result may be slightly different. Therefore, in order to reduce irregularities in the results, genes are ranked by the average value of the differential index (DI) for each gene calculated by repeating the gene selection algorithm 10,000 times. Then, the gene set with the highest differential index (DI) was selected as a specific gene.

On the other hand, in order to calculate how many genes should be designated as specific genes in the ranking of the score of the discrimination index (DI) so that there is no degradation in classification performance, the optimal number of genes is first calculated from the gene list sorted by the DI score. To this end, the average accuracy of cancer and normal sample classification of the training data set is calculated by setting each set to one specific gene set from 1,000 genes to 1,000 genes while increasing the number.

As a result, as shown in FIG. 4, when the top 100 genes were composed of one specific gene set, the highest average accuracy was shown, and even if more genes were added, the average accuracy was not increased.

When step S213 is completed, the algorithm generating unit 120 performs an intermediate evaluation by inputting the intermediate evaluation data set to the gene selection algorithm (S214).

At this time, the algorithm generating unit 120 calculates the differential index (DI) for each gene with different weights.

Then, the algorithm generator 120 performs a final evaluation by inputting the final evaluation data set into the gene selection algorithm (S215).

As a result of performing steps S213 to S215, the algorithm generator 120 selected 14 specific genes.

Here, 14 specific genes include FN1, ALB, EEF1A1, SFTPC, GAPDH, P4HB, DCN, A2M, MGP, UMOD, GPX3, FTL, ACPP and CTSD.

In a state in which the gene selection algorithm is constructed through step S210, the collection unit 110 collects tissues from a plurality of subjects and obtains genetic information from the collected tissues (S220).

Here, the plurality of subjects includes a cancer patient group and a normal person group, and is composed of 50 people each. Then, the collection unit 110 acquires tissues separated from 100 subjects, and extracts RNA from the tissues.

Then, the collection unit 100 analyzes the extracted RNA through the nCounter®Analysis System. In the nCounter®Analysis System, a digital analyzer acquires genetic information by capturing and counting the color of each molecule contained in RNA. Meanwhile, the collection unit 100 acquires approximately 20,000 gene expression data for a plurality of subjects.

Then, the collection unit 100 transmits the obtained 20,000 gene expression data to the differential index calculation unit 130.

Next, the discrimination index calculation unit 130 calculates the discrimination index for each gene by inputting the received 20,000 gene expression data into a pre-built gene selection algorithm (S230).

5 is a view showing that the differential index is calculated for each gene in step S230.

)과 정상조직의 유전자 발현값의 총합(

)을 산출한다. As shown in Figure 5, the differential index calculation unit 130 is the sum of gene expression values of cancer tissues for every 20,000 gene expression data received (

) And the sum of gene expression values in normal tissue (

).

Next, the differential index calculation unit 130 calculates the differential index (DI) using the following equation.

는 j번째 유전자의 차별지수이고,

는 j번째 유전자에 대응하는 암조직의 유전자 발현값의 총합을 나타내고,

는 j번째 유전자에 대응하는 정상 조직의 유전자 발현값의 총합을 나타내며, W는 가중치를 나타낸다.here,

Is the differential index of the jth gene,

Denotes the sum of gene expression values of cancer tissues corresponding to the j-th gene,

Denotes the sum of gene expression values of normal tissues corresponding to the j-th gene, and W denotes a weight.

값을 더하여 산출된다. That is, the influence of a specific gene, that is, the discrimination index (DI), is a different pair of input data.

It is calculated by adding the values.

Here, the different pairs represent the sum of gene expression samples for the tumor, and the other represent the sum of normal gene expression samples.

When the step S230 is completed, the specific gene selection unit 140 lists 20,000 genes in the order in which the calculated differential index (DI) is large. Then, the specific gene selection unit 140 selects genes corresponding to the top 14 from the listed genes (S240).

Finally, the prediction unit 130 compares a specific gene obtained through a pre-built gene selection algorithm with a gene selected in step S230 to predict whether cancer has occurred (S250).

Here, specific genes represent FN1, ALB, EEF1A1, SFTPC, GAPDH, P4HB, DCN, A2M, MGP, UMOD, GPX3, FTL, ACPP and CTSD.

Hereinafter, the classification accuracy of a specific gene extracted through the gene expression marker selection device according to an embodiment of the present invention will be described in more detail.

In the embodiment of the present invention, the top 14 specific genes with a high differential index were selected as gene expression markers. As shown in Table 2, the gene expression markers in previous studies (Peng et al. Or Martinez-Ledesma et al.) Consist of 7 or 14 genes. However, it can be seen that the gene expression marker obtained through the gene selection algorithm and the gene expression marker in the previous study do not overlap. Therefore, as a result of evaluating how accurate the gene expression marker selected through the gene selection algorithm has to classify cancer, as shown in Table 3 below, high classification accuracy for 5 out of 7 cancer types It was shown.

As described above, the gene expression marker selection method according to the present invention ranks genes on the basis of the discrimination index (DI) indicating the classification ability to distinguish the normal group from the cancer patient group, and the genes ranked by the discrimination index (DI) The list allows you to choose the best set of genes from the top ranking genes.

The present invention has been described with reference to the embodiment shown in the drawings, but this is only exemplary, and those skilled in the art to which the art belongs understand that various modifications and other equivalent embodiments are possible therefrom. will be. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the claims below.

100: gene expression marker selection device
110: collection unit
120: algorithm generator
130: discrimination index calculation department
140: specific gene selection unit
150: prediction unit

Claims (8)

In a method for selecting a gene expression marker based on a neural network using a gene expression marker selection device,
Collecting each biopsy tissue from a plurality of patients and collecting a plurality of gene expression information experimentally measured from each collected biopsy tissue,
Calculating the differential index (DI) for each gene by applying the collected plurality of gene expression information to a pre-trained neural network-based gene selection algorithm,
Gene sorting according to the calculated differential index (DI), and selecting a plurality of specific genes having a large differential index (DI) among the listed genes, and
Gene expression marker selection method comprising the step of predicting whether or not the occurrence of cancer by using the expression values of the selected plurality of specific genes. According to claim 1,
Further comprising constructing and learning the neural network based gene selection algorithm,
The step of constructing and learning a genetic selection algorithm based on the neural network,
Receiving gene expression information for a plurality of cancer types from the Cancer Genome Atlas (TCGA) program,
Grouping the received gene expression information into a cancer patient group and a normal group, and randomly extracting the gene information obtained from each group to form a data set, and
Gene expression marker selection method comprising the step of constructing a gene selection algorithm for extracting a plurality of specific genes to be classified into a normal patient group and a cancer patient group using the formed data set. According to claim 2,
The plurality of cancer types,
Bladder urinary tract carcinoma (BLCA), breast invasive carcinoma (BRCA), adenocarcinoma (COAD), squamous cell carcinoma of the head and neck (HNSC), renal chromophore (KICH), renal transparent cell carcinoma (KIRC), renal papillary cell carcinoma (KIRP) ), Liver cancer (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), prostate cancer (PRAD) and thyroid carcinoma (THCA). According to claim 2,
The gene selection algorithm,
Each differential index (DI) value for all the expression genes included in a plurality of cancer types that are always received is calculated,
Gene expression marker selection method for extracting a plurality of specific genes by using the ranking of the calculated discrimination index (DI) value. The method of claim 4,
The discrimination index (DI) value,
Gene expression marker selection method calculated through the following equation:

here,

Denotes the sum of gene expression values of cancer tissues corresponding to the j-th gene,

Denotes the sum of gene expression values of normal tissues corresponding to the j-th gene, and W denotes a weight. According to claim 2,
Generating the data set,
Gene expression marker selection method for generating a data set by extracting expression gene information at random regardless of the ratio of cancer samples and normal samples that are different for each of the plurality of cancer types. The method of claim 6,
Generating the data set,
Using the entire cancer gene expression data, a learning data set, a verification data set, and an evaluation data set are generated at a predetermined rate.
The generated training data set, verification data set, and evaluation data set, respectively, are gene expression marker selection methods that form the same ratio of cancer samples and normal samples. The method according to any one of claims 1 to 7,
The plurality of specific genes,
Methods for selecting gene expression markers including FN1, ALB, EEF1A1, SFTPC, GAPDH, P4HB, DCN, A2M, MGP, UMOD, GPX3, FTL, ACPP and CTSD.

Images