KR102543757B1 - Method and apparatus for discovering biomarker for predicting cancer prognosis using heterogeneous platform of DNA methylation data - Google Patents

Abstract

It relates to a method and apparatus for discovering biomarkers for predicting cancer prognosis using DNA methylation data of a heterogeneous platform. By improving the accuracy, it is possible to derive methylation markers related to cancer prognosis prediction, and a cancer prognosis prediction kit with higher accuracy can be developed.

Description

Method and apparatus for discovering biomarker for predicting cancer prognosis using heterogeneous platform of DNA methylation data}

The present invention relates to a method and apparatus for discovering biomarkers for predicting cancer prognosis using DNA methylation data of a heterogeneous platform.

Epigenetics is a field that studies the regulation of gene expression in a state where the nucleotide sequence of DNA is not changed. Epigenetics studies the regulation of gene expression through epigenetic alterations such as DNA methylation, miRNA or histone acetylation, methylation, phosphorylation, and ubiquitination. Double DNA methylation is the most studied epigenetic mutation. Epigenetic alterations can lead to mutations in gene function and changes to tumor cells. Therefore, DNA methylation is associated with the expression or suppression of disease-regulating genes in cells, and recently, a number of studies have been reported on methods for diagnosing the risk or prognosis of cancer through DNA methylation measurement (Korean Patent Publication No. 10-2014- 0022293), research using machine learning is in progress.

On the other hand, when creating a model using methylation data, if the number of methylation sites is overwhelmingly greater than the number of data, the training data is over-learned, resulting in excessively high accuracy for the training data and low accuracy in the test data. The problem of overfitting arises. For this reason, only a small number of well-known or experimentally verified gene methylations are used as markers. In order to solve the overfitting problem, it is necessary to secure a larger number of data, but a lot of cost is incurred to secure a large number of data from the platform. Therefore, there is a need for a method for improving the accuracy of a model by solving the overfitting problem without increasing the cost.

It is an object of the present invention to provide a method and apparatus for discovering methylation markers for prognosis of cancer using DNA methylation data of heterogeneous platforms. The technical problem to be achieved by the present embodiment is not limited to the technical problems described above, and other technical problems can be inferred from the following embodiments.

One aspect is a method for a device operated by at least one processor to discover methylation markers for predicting cancer prognosis using machine learning, wherein the at least one processor is important for predicting cancer prognosis among first methylation data. A first step of extracting CpG sites having a significant correlation with variables; a second step of selecting, by the at least one processor, data corresponding to the extracted CpG site from among the second methylation data as input data for final learning; and a third step of generating, by the at least one processor, a cancer prognosis prediction model with the final learning input data; a method for discovering methylation markers for predicting cancer prognosis using machine learning is provided.

In another aspect, before the first step, deriving a variable by reducing the dimension of the first methylation data by the at least one processor; generating, by the at least one processor, a cancer prognosis predictive test model through machine learning using the derived variables as input data for learning and clinical data of a cancer patient as output data for learning; and determining, by the at least one processor, one or more variables important in predicting a prognosis of cancer by evaluating the performance of the test model.

In another aspect, the first methylation data or the second methylation data may be array-based methylation data or bisulfite sequencing data.

In another aspect, the step of deriving variables by reducing the dimensionality includes principal component analysis (PCA), singular value decomposition (SVD), and non-negative matrix factorization (NMF). ) It can be performed using any one method selected from the group consisting of.

In another aspect, the generating of the test model for predicting cancer prognosis may be generated using a bootstrap algorithm.

In another aspect, the step of deriving one or more variables important in predicting cancer prognosis may derive one or more variables important in predicting cancer prognosis from a model exhibiting maximum performance by performing cross-validation.

Another aspect provides a computer-readable recording medium recording a program for executing the method on a computer.

Another aspect is a CpG site extraction unit for extracting CpG sites having a significant correlation with variables important in predicting cancer prognosis from among the first methylation data; an input data selection unit for final learning that selects data corresponding to the extracted CpG site from among the second methylation data as input data for final learning; and a model generation unit configured to generate a cancer prognosis prediction model with the final learning input data.

In another aspect, a variable derivation unit for deriving a variable by reducing the dimension of the first methylation data; a test model generation unit configured to generate a cancer prognosis prediction test model through machine learning using the derived variables as input data for learning and clinical data of a cancer patient as output data for learning; and an important variable determining unit configured to determine one or more variables important in predicting cancer prognosis by evaluating the performance of the test model.

In another aspect, the first methylation data or the second methylation data may be array-based methylation data or bisulfite sequencing data.

In another aspect, the dimensionality reduction is any one selected from the group consisting of principal component analysis (PCA), singular value decomposition (SVD), and non-negative matrix factorization (NMF). This can be done using one method.

In another aspect, the test model generation may be generated using a bootstrap algorithm.

In another aspect, the determination of one or more variables important in predicting cancer prognosis may determine one or more variables important in predicting cancer prognosis from a model exhibiting maximum performance by performing cross-validation.

According to the method and apparatus according to one aspect, by using DNA methylation data of a heterogeneous platform to improve the accuracy of a cancer prognosis prediction model, it is possible to derive methylation markers related to cancer prognosis prediction and cancer prognosis with higher accuracy Prediction kits can be developed.

1 is a block diagram illustrating a hardware configuration of a computing device for discovering methylation markers for predicting prognosis of cancer using machine learning according to an embodiment.
FIG. 2 is a block diagram showing a detailed hardware configuration of the processor of FIG. 1 .
3 is a flowchart of a method for discovering methylation markers for predicting cancer prognosis using machine learning according to an embodiment.
4 is a graph illustrating a process of determining one or more variables important in predicting cancer prognosis through test model performance evaluation according to an embodiment.
5 shows (A) a cancer prognosis prediction model generated by using bisulfite sequencing data as input data for learning, and (B) bisulfite sequencing for CpG sites extracted by processing array-based methylation data according to an embodiment. This graph compares and evaluates the average prediction accuracy for each cancer stage of the cancer prognosis prediction model generated using the data as input data for learning.

The terms used in the present embodiments have been selected from general terms that are currently widely used as much as possible while considering the functions in the present embodiments, but these may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technologies, and the like. can In addition, there are terms selected arbitrarily in certain cases, and in this case, their meanings will be described in detail in the description of the corresponding embodiment. Therefore, the term used in the present embodiments should be defined based on the meaning of the term and the overall content of the present embodiment, not a simple name of the term.

In the descriptions of the embodiments, when a part is said to be connected to another part, this includes not only the case where it is directly connected, but also the case where it is electrically connected with another component interposed therebetween. . In addition, when a part includes a certain component, this means that it may further include other components without excluding other components unless otherwise stated. In addition, the terms "...unit" and "...module" described in the embodiments mean a unit that processes at least one function or operation, which is implemented as hardware or software or a combination of hardware and software. can be implemented

Terms such as "consisting of" or "included" used in the present embodiments should not be construed as necessarily including all of the various components or steps described in the specification, and some of the components or It should be construed that some steps may not be included, or may further include additional components or steps.

The description of the following embodiments should not be construed as limiting the scope of rights, and what can be easily inferred by a person skilled in the art should be construed as belonging to the scope of the embodiments. Hereinafter, embodiments for illustrative purposes only will be described in detail with reference to the accompanying drawings.

1 is a block diagram illustrating a hardware configuration of a computing device for discovering methylation markers for predicting prognosis of cancer using machine learning according to an embodiment.

Referring to FIG. 1, the computing device 50 preprocesses/processes/analyzes the first methylation data 40, the second methylation data 41, and the clinical information data 42 of the cancer patient to determine the cancer As a device for discovering markers for predicting the prognosis of , it may include a data interface 80 , a memory 90 , and a processor 100 . Meanwhile, in the computing device 50 shown in FIG. 1, only components related to the present embodiment are shown in order to prevent obscuring the features of the present embodiment, so the computing device 50 has the configuration shown in FIG. In addition to the elements, other general-purpose components may be further included.

The term "cancer", "tumor" or "malignancy" refers to or describes a physiological condition in mammals that is generally characterized by unregulated cell growth. Examples of cancers may include, but are not limited to, carcinoma, lymphoma, leukemia, blastoma, and sarcoma.

The term "prognosis" means prediction of the course and outcome of a disease, and may include prediction of cancer recurrence and progression.

The term "methylation" refers to changes in gene expression patterns by attaching a methyl group to a base constituting DNA, and may mean that it occurs at cytosine of a specific CpG site of a specific gene.

The first methylation data 40 or the second methylation data 41 may be experimentally obtained. Methylation data can be obtained empirically at any time point by methods or devices known to those skilled in the art. For example, PCR, methylation specific PCR, real time methylation specific PCR, MethyLight PCR, MehtyLight digital PCR, EpiTYPER, PCR using methylated DNA specific binding proteins, quantitative PCR, DNA It may be performed by any one method selected from the group consisting of chip, pyrosequencing, bisulfite sequencing, and microarray, but is not limited thereto.

Furthermore, the first methylation data or the second methylation data may be obtained from an open database (DB). For example, it may be obtained from a database (DB) already known in the art, such as National Center for Biotechnology Information (NCBI), Gene Expression Omnibus (GEO), European Bioinformatics Institute databases, or European Nucleotide Archive. In one embodiment, the first methylation data or the second methylation data may be obtained from published array-based methylation data or bisulfite sequencing data. However, since new DNA methylation data are continuously being discovered and updated due to the development of genetic analysis technology, the DNA methylation data to be described in this embodiment is not limited to what can be obtained from a public database (DB).

Array-based methylation data is one of Illumina's methods for methylation level evaluation, which can be obtained by measuring the intensities of methylated and non-methylated beads of the Illumina Infinium HumanMethylation 450 BeadChip.

Bisulfite sequencing data converts cytosine bases to uracil by treating bisulfite prior to sequencing to confirm methylation patterns. Since it still remains as cytosine, this difference can be used to evaluate the level of methylation.

The data interface 80 acquires first methylation data 40, second methylation data 41, and clinical information data 42 of cancer patients, which are experimentally measured from biological samples or stored in a database (DB). do. That is, the data interface 110 may be implemented as hardware of a wired/wireless network interface through which the computing device 50 communicates with other external devices.

The memory 90 is hardware for storing data to be processed and results of processing in the computing device 50, and includes memory chips such as random access memory (RAM) and read only memory (ROM) or a hard disk (HDD). drive), solid state drive (SSD), and the like. That is, the memory 90 may store the first methylation data 40, the second methylation data 41, and the clinical information data 42 of the cancer patient acquired by the data interface 80, and the processor ( 100) can also store data on methylation markers discovered.

The processor 100 generates a model for predicting cancer prognosis using the first methylation data 40, the second methylation data 41, and the clinical information data 42 of the cancer patient, and derives a methylation marker therefrom It corresponds to the hardware for methylation marker discovery for The processor 100 is a module implemented with one or more processing units, and may be implemented with a combination of a microprocessor having an array of a plurality of logic gates and a memory module storing a program that can be executed by the microprocessor. The processor 100 may be implemented in the form of a module of an application program.

The methylation marker discovered by the processor 100 is transmitted to another external device, eg, a display device, another computing device, etc., through the data interface 80, or through an external network, eg the Internet, an open database ( DB) can be transmitted on.

FIG. 2 is a block diagram showing a detailed hardware configuration of the processor of FIG. 1 .

Referring to FIG. 2, the processor 100 includes a variable derivation unit 101, a test model generation unit 102, an important variable determination unit 103, a CpG site extraction unit 104, and a final training input data selection unit. (105), and a model generator (106). Meanwhile, in the processor 100 shown in FIG. 1, only the components related to the present embodiment are shown in order to prevent the features of the present embodiment from being blurred, so the processor 100 includes the components shown in FIG. In addition, other general-purpose components may be further included. Variable derivation unit 101, test model generation unit 102, important variable determination unit 103, CpG site extraction unit 104, final learning input data selection unit 105, and model generation unit 106 are respectively It is only classified into separate and independent names according to the functions of, variable derivation unit 101, test model generation unit 102, important variable determination unit 103, CpG site extraction unit 104, final learning The input data selection unit 105 and the model generation unit 106 may be implemented by one processor 100 . Alternatively, a variable derivation unit 101, a test model generation unit 102, an important variable determination unit 103, a CpG site extraction unit 104, a final learning input data selection unit 105, and a model generation unit 106 Each may correspond to one or more processing modules in processor 100 . Alternatively, a variable derivation unit 101, a test model generation unit 102, an important variable determination unit 103, a CpG site extraction unit 104, a final learning input data selection unit 105, and a model generation unit 106 may correspond to a separate software algorithm unit classified according to each function. That is, within the processor 100, a variable derivation unit 101, a test model generation unit 102, an important variable determination unit 103, a CpG site extraction unit 104, a final learning input data selection unit 105, and The implementation form of the model generation unit 106 is not limited by any one.

The variable deriving unit 101 receives the first methylation data and derives a variable by reducing the dimension of the received first methylation data. The term “dimensionality reduction” may refer to a data preprocessing process in which data density is increased by compressing data. When machine learning is performed using high-dimensional data, the learning speed slows down and the performance of the model may deteriorate. The device according to an embodiment includes the variable derivation unit 101 to improve the learning speed and model performance. it may be to In one embodiment, the dimensionality reduction method is any one selected from the group consisting of principal component analysis (PCA), singular value decomposition (SVD), and non-negative matrix factorization (NMF). This can be done using one method. The dimensionality reduction method according to an embodiment performs principle component analysis (PCA) to compress array-based methylation data to find an axis that maximizes the variance of data, and finds a principal component (Principle components; By deriving PCs), samples in high-dimensional space can be transformed into low-dimensional space. In this specification, the term “feature” may be used as a meaning including the term “principle components (PCs)”.

The test model generation unit 102 receives the variables derived through dimension reduction from the variable derivation unit 101, uses them as input data for learning, receives clinical data of cancer patients, and uses them as output data for learning through machine learning. Generate a prognostic predictive test model for cancer. Test model generation may be generated using a bootstrap algorithm, but is not limited thereto.

The important variable determining unit 103 evaluates the performance of the test model generated by the test model generating unit 102 and determines one or more variables important in predicting the prognosis of cancer therefrom. In determining one or more variables important in predicting cancer prognosis, cross-validation may be performed to determine one or more variables important in predicting cancer prognosis from a model exhibiting maximum performance.

In the test model generation and determination of important variables according to an embodiment, principle component analysis (PCA) is performed on the array-based methylation data, and the derived principal component is used as input data for learning, and the clinical data of cancer patients is used as input data. An important variable may be determined by generating a cancer prognosis prediction model using output data for training and evaluating the performance of the model. When learning a prognostic prediction model, if the number of features is greater than the number of samples, such as methylation data, overfitting of the model may occur. It can be created using the Elastic net, a feature regularization method that prevents the model from overfitting by adding . In the determination of important variables according to an embodiment, variables that are not important for predicting prognosis by the Elastic net have model coefficients of 0, so variables that appear more than 100 times with non-zero coefficients among 200 cross-validations are important for predicting prognosis. It can be judged and selected by variables. 4 is a graph illustrating a process of determining one or more variables important in predicting cancer prognosis through test model performance evaluation according to an embodiment. As shown in FIG. 4, as a result of evaluating the average C-index obtained by performing 200 cross-validation on the test model generated by varying the number of principal components, the model using PC1 to PC50 as predictors is the most suitable model. , and the principal components selected as the most frequently important variables as a result of 200 tests were selected as important variables for predicting prognosis.

The CpG site extractor 104 receives the first methylation data, receives one or more variables determined as important variables for predicting cancer prognosis from the important variable determiner 103, and determines the prognosis of cancer among the first methylation data. Extract CpG sites with high correlation with important variables for prediction. When performing principal component analysis, each principal component is composed of a weighted sum of CpG sites. CpG sites with higher weights, that is, CpG sites with higher correlations, have more influence in forming the corresponding principal components, which is important for predicting prognosis. CpG sites that greatly affect the main component can be selected. In CpG site extraction according to an embodiment, CpG sites having a correlation of 0.6 or more with each principal component important in predicting cancer prognosis determined through test model creation and evaluation from array-based methylation data may be selected.

The input data selection unit 105 for final learning receives the second methylation data, receives the CpG site extracted from the CpG site extraction unit 104, and extracts data corresponding to the CpG site extracted from the second methylation data. It is selected as input data for final learning.

The model generation unit 106 receives methylation data from the final learning input data selection unit 105 and uses it as learning input data to generate a cancer prognosis prediction model.

Selection of input data for final learning according to an embodiment selects data corresponding to CpG sites extracted from array-based methylation data among bisulfite sequencing data as input data for final learning to generate a prognostic prediction model, 200 It is possible to evaluate whether prognosis can be predicted only with bisulfite sequencing data for extracted CpG sites by measuring the average predictive accuracy (Area under the curve; AUC) of the model by performing cross validation. The prognostic prediction model can be generated by dividing training data and test data using the bootstrap technique and applying the Elastic net.

5 shows the average prediction accuracy for each stage of cancer by performing AUC analysis to determine whether the prediction accuracy of the model generated using methylation data of heterogeneous platforms is improved compared to the model using methylation data of a single platform. It is a comparative evaluation graph. 5A shows methylation data of a single platform (bisulfite sequencing data), and FIG. 5B shows methylation data of a heterogeneous platform (bisulfite sequencing data for CpG sites extracted by processing array-based methylation data). ) is the result of evaluating the average prediction accuracy of the cancer prognosis prediction model. After performing principal component analysis on the methylation data of each single platform or heterogeneous platform in FIGS. 5A and 5B, a model was created using principal components with large variance as predictors, and then accuracy was measured. At this time, modeling was performed while increasing the number of principal components used in units of 10, and among them, the model with the largest AUC was determined to be the optimal model for a specific stage of cancer. As a result, as shown in FIG. 5, when methylation data of heterogeneous platforms were used as input data for final learning, compared to the case of using methylation data of a single platform, the AUC value was approximately 0.16 in stage 1, 0.13 in stage 2, stage The increase was 0.06 in stage 3 and 0.19 in stage 4, confirming that the accuracy of the prognostic prediction model was improved using DNA methylation data from heterogeneous platforms. Therefore, this means that the problem of overfitting in the prognosis prediction model is resolved and it is possible to discover methylation markers related to cancer prognosis.

3 is a flowchart of a method for discovering methylation markers for predicting cancer prognosis using machine learning according to an embodiment. Referring to FIG. 3 , the method for discovering methylation markers for predicting prognosis of cancer using machine learning includes steps that are time-sequentially processed by the computing device 50 described in the previous drawings. Therefore, even if the contents are omitted below, the contents described in the previous drawings can also be applied to the method of discovering methylation markers for predicting cancer prognosis using machine learning of FIG. 3 .

In a first step, at least one processor may derive a variable by reducing the dimension of the first methylation data (S10). For example, the first methylation data may be array-based methylation data. In one embodiment, a method of deriving variables by reducing dimensionality includes principal component analysis (PCA), singular value decomposition (SVD), and non-negative matrix factorization (NMF). It can be performed using any one method selected from the group consisting of.

In a second step, at least one processor may generate a cancer prognosis prediction test model through machine learning using the derived variables as input data for learning and clinical data of a cancer patient as output data for learning (S20). In one embodiment, the test model may be created using a bootstrap algorithm.

In a third step, at least one processor may evaluate the performance of the test model to determine one or more variables important in predicting cancer prognosis (S30). In one embodiment, one or more variables important in predicting cancer prognosis may be derived from a model exhibiting maximum performance by performing cross-validation.

In the fourth step, at least one processor may extract a CpG site having a significant correlation with a variable important in predicting the cancer prognosis from among the first methylation data (S40).

In a fifth step, at least one processor may select data corresponding to the extracted CpG site from among the second methylation data as input data for final learning (S50). For example, the second methylation data may be bisulfite sequencing data.

In a sixth step, at least one processor may generate a cancer prognosis prediction model with the final learning input data (S60).

On the other hand, the above-described embodiments of the present invention can be written as a program that can be executed on a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium. In addition, the structure of data used in the above-described embodiments of the present invention can be recorded on a computer-readable recording medium through various means. The computer-readable recording medium includes storage media such as magnetic storage media (eg, ROM, floppy disk, hard disk, etc.) and optical reading media (eg, CD-ROM, DVD, etc.).

40: first methylation data
41: second methylation data
42: Clinical information data of cancer patients
50: computing device
80: data interface
90: memory
100: processor
101: variable derivation unit
102: test model generation unit
103: important variable determining unit
104: CpG site extraction unit
105: input data selection unit for final learning
106: model generating unit

Claims (13)

A method for a device operated by at least one processor to discover methylation markers for predicting cancer prognosis using machine learning, the method comprising:
deriving, by the at least one processor, a principal component (PC) of the first methylation data by performing principal component analysis (PCA);
generating, by the at least one processor, a cancer prognosis prediction test model through machine learning using the derived principal component as input data for learning and clinical data of a cancer patient as output data for learning;
determining, by the at least one processor, one or more principal components important in predicting the prognosis of cancer by evaluating the performance of the test model;
a first step of extracting, by the at least one processor, a CpG site having a significant correlation with the determined principal component from among the first methylation data;
a second step of selecting, by the at least one processor, data on the extracted CpG site from among the second methylation data as input data for final learning; and
A method for discovering methylation markers for predicting cancer prognosis using machine learning, comprising: a third step of generating, by the at least one processor, a cancer prognosis prediction model with the final learning input data. delete The method of claim 1, wherein the first methylation data or the second methylation data is array-based methylation data or bisulfite sequencing data. delete The method according to claim 1, wherein the generating of the test model for predicting cancer prognosis is generated using a bootstrap algorithm. The method of claim 1 , wherein the determining of one or more principal components important in predicting cancer prognosis comprises determining one or more principal components important in predicting cancer prognosis from a model exhibiting maximum performance by performing cross-validation. A computer-readable recording medium recording a program for executing the method of any one of claims 1, 3, 5, and 6 in a computer. a principal component derivation unit deriving a principal component (PC) of the first methylation data by performing principal component analysis (PCA);
a test model generator configured to generate a cancer prognosis prediction test model through machine learning using the derived principal component as input data for learning and clinical data of a cancer patient as output data for learning;
an important principal component determiner configured to determine one or more principal components important in predicting cancer prognosis by evaluating the performance of the test model;
a CpG site extraction unit extracting CpG sites having a significant correlation with the determined principal component from among the first methylation data;
an input data selection unit for final learning that selects, among the second methylation data, data on the extracted CpG site as input data for final learning; and
An apparatus for discovering methylation markers for predicting cancer prognosis using machine learning, comprising: a model generation unit configured to generate a cancer prognosis prediction model with the final learning input data. delete The apparatus of claim 8, wherein the first methylation data or the second methylation data is array-based methylation data or bisulfite sequencing data. delete The apparatus of claim 8 , wherein the test model is generated using a bootstrap algorithm. The apparatus of claim 8 , wherein the determining of the one or more principal components important in predicting the prognosis of cancer is performed by performing cross-validation to determine one or more principal components important in predicting the prognosis of cancer from a model exhibiting maximum performance.

Images