Detection markers, binary classification models, and applications thereof

Abstract

The present application relates to the field of biological detection, and particularly relates to a detection marker, a two-classification model and application thereof.CpG sites, including one, multiple or a combination of cg26492368, cg26608718, cg19516340, cg02554274, cg09341491, cg13788685, cg08383929, cg03526459, cg14304336, cg27039312 or cg26956009.The detection marker provided by the present application provides a new auxiliary means for accurate diagnosis of HCC.

Description

Technical Field

[0001] This invention relates to the field of biological detection, and in particular to detection biomarkers, binary classification models and their applications. Background Technology

[0002] Hepatocellular carcinoma (HCC) is a highly heterogeneous and complex disease. Its development likely stems from the accumulation of genetic and epigenetic variations. Numerous genetic variations have been identified in HCC research, with TP5 exhibiting the highest frequency. Abnormal methylation of many gene promoter regions has also been reported, including the RASSF1A gene, which is hypermethylated in 85% of HCC cases. In tumor diagnosis and monitoring, DNA methylation detection offers advantages over mutation detection. Firstly, it has higher clinical sensitivity and a wider range of target selection. Specifically, each tumor type has many methylation target regions available for diagnosis, and each methylation region contains multiple CpG sites exhibiting abnormal methylation. In the past decade, several genome-level methylation studies have been conducted. The integration of the hepatitis B virus (HBV) genome into the hepatocyte genome leads to abnormal methylation of the hepatocyte genome and a highly unstable genome structure, thus promoting the development and progression of HCC. Based on genome-level DNA methylation data, several methylation diagnostic models have been reported. However, most of these studies are based on methylation microarray data covering hundreds of thousands of CpG sites, while there are nearly 30 million CpG sites on the human genome that can undergo methylation. High-coverage methylation detection of CpG sites on the genome can provide a more comprehensive understanding of the role of DNA methylation in disease development and progression, and can also help us find more effective biomarkers. Current technologies are limited by the small sample size (N=4) to perform analyses of biomarker detection performance and other aspects.

[0003] The application of cell-free DNA (cfDNA) methylation in tumor monitoring and diagnosis has become a research hotspot in recent years. In healthy individuals, plasma cfDNA primarily originates from leukocytes (55%), erythrocyte progenitor cells (30%), vascular endothelial cells (10%), and hepatocytes (1%). Analysis of patient cfDNA has confirmed that DNA from various solid tumors is released into peripheral blood, providing theoretical, data, and methodological basis for monitoring the occurrence and development of solid tumors or other diseases through cfDNA methylation detection. The relatively high absolute concentration of liver-derived DNA in peripheral blood cfDNA increases the possibility of using its methylation for the diagnosis and early monitoring of hepatocellular carcinoma (HCC). Studies on the use of cfDNA for HCC diagnosis, including methylation of promoters of individual genes such as RFSSF1A, cfDNA concentration, fragment length, and terminal sequence characteristics, have been widely reported. Existing technologies utilize 450k methylation data from HCC tissue and peripheral blood leukocyte DNA to screen for 1000 CpG sites. Then, targeted capture sulfite sequencing was used to obtain 401 usable CpG sites. Combined with machine learning, a diagnostic model and a prognostic model based on plasma cfDNA methylation were constructed. A diagnostic model was also constructed using peripheral blood ctDNA from a large cohort (1098 HCC patients and 835 healthy individuals). The feasibility of using cfDNA for early diagnosis (monitoring) of HCC was studied at the genomic level using methylated CpG tandem amplification and sequencing (MCTA-Seq). Furthermore, some studies directly used 450k methylation chips to detect cfDNA from hepatitis and liver cancer patients and constructed a diagnostic model based on the methylation levels of five CpG sites. However, given the current detection technologies, the cost of cfDNA methylation detection at the genomic level for clinical early liver cancer screening or monitoring remains too high. Moreover, due to the complex and variable plasma environment, methylation levels (β values) are easily affected and fluctuate, making the stability of diagnostic models that use methylation levels (β values) at multiple sites for liver cancer monitoring an insurmountable challenge. Summary of the Invention

[0004] In view of this, the present invention provides detection biomarkers, a binary classification model, and their applications. The detection biomarkers provided by the present invention offer a new auxiliary means for the accurate diagnosis of HCC.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] The present invention provides CpG sites, including one, more or a combination of cg26492368, cg26608718, cg19516340, cg02554274, cg09341491, cg13788685, cg08383929, cg03526459, cg14304336, cg27039312 or cg26956009.

[0007] The present invention also provides detection markers, including one, more or a combination of SPAG6, AKAP8L, RRM2, CEND1, SNX20, SDHC, NBPF13P, RNVU1-8, C10orf90 or BAI1.

[0008] In some embodiments of the present invention, the correspondence between the above-mentioned detection markers and the CpG sites is as follows:

[0009] cg26492368 SPAG6;

[0010] cg26608718 AKAP8L;

[0011] cg19516340 RRM2;

[0012] cg02554274 CEND1;

[0013] cg09341491 SNX20;

[0014] cg13788685 SDHC;

[0015] cg03526459 NBPF13P or RNVU1-8;

[0016] cg14304336 C10orf90;

[0017] cg27039312 BAI1.

[0018] The present invention also provides the application of the above-mentioned CpG site or the above-mentioned detection marker in the preparation of products for detecting liver cancer.

[0019] This invention also provides a binary classification model, the calculation formula of which is: mBCS=0.3948785+∑β i E i Where i is the CpG site mentioned above, β is the methylation level of the CpG site mentioned above, and E is the regression coefficient of the CpG site mentioned above.

[0020] In some embodiments of the present invention, in the above binary classification model, the methylation level of the CpG site is obtained by whole-genome methylation sequencing (WGBS).

[0021] In some embodiments of the present invention, in the above binary classification model, the methylation level of the public validation dataset is obtained by a 450K methylation chip.

[0022] In some embodiments of the present invention, in the above binary classification model, the regression coefficient of the CpG site and the correspondence between the CpG site are as follows:

[0023] cg26492368 8.1559394;

[0024] cg26608718 -8.4531282;

[0025] cg19516340 -6.1125863;

[0026] cg02554274 -3.4444817;

[0027] cg09341491 6.8589209;

[0028] cg13788685 -3.7790125;

[0029] cg08383929 7.0882640;

[0030] cg03526459 3.3835278;

[0031] cg14304336 -6.0109634;

[0032] cg27039312 -4.5737240;

[0033] cg26956009 -5.3367444.

[0034] In some embodiments of the present invention, the binary classification model described above is obtained by one or more of the following algorithms: random forest algorithm, 10-fold cross-validation, or elastic network regression algorithm.

[0035] This invention also provides the application of the above-mentioned binary classification model in evaluating the generalization ability of detection markers.

[0036] In some embodiments of the present invention, the above-described applications use biomarkers for detecting liver cancer.

[0037] In some embodiments of the present invention, in the above applications, the evaluation criteria include: a mBCS greater than or equal to 0.5 is considered positive, and a mBCS less than 0.5 is considered negative.

[0038] This invention also provides the application of the above-mentioned binary classification model in the preparation of products for detecting the degree of immune cell infiltration.

[0039] In some embodiments of the present invention, in the above applications, the immune cells include one or more of CD8 positive T cells, dendritic cells, macrophages, natural killer T cells, or regulatory T cells.

[0040] In some embodiments of the present invention, in the above applications, the CD8-positive T cells and the dendritic cells are obtained after activation.

[0041] The present invention also provides a detection method to obtain the methylation level of CpG sites in the sample to be tested, and substitute it into the binary classification model as described above to obtain the detection result.

[0042] In some embodiments of the present invention, the detection criteria for the detection results in the above detection method include: a positive result when the mBCS is greater than or equal to 0.5, and a negative result when the mBCS is less than 0.5.

[0043] The present invention provides CpG sites, including one, more or a combination of cg26492368, cg26608718, cg19516340, cg02554274, cg09341491, cg13788685, cg08383929, cg03526459, cg14304336, cg27039312 or cg26956009.

[0044] The detection markers provided by this invention offer a new auxiliary means for the accurate diagnosis of HCC. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0046] Figure 1 This document outlines the process for constructing a binary classification model for DML site methylation and validating it on an external dataset.

[0047] Figure 2 The correlation between the mBCS binary classification model and the level of immune cell infiltration is shown (Pearson); where "*" indicates that the correlation test P<0.05;

[0048] Figure 3 Principal component analysis and cluster analysis of six independent datasets are shown.

[0049] Figure 4 This demonstrates the characteristics of the optimal model based on random forest 10-fold cross-validation;

[0050] Figure 5 The AUC of the external validation dataset is shown.

[0051] Figure 6 The image shows heatmaps of methylation levels at 11 CpG sites across multiple datasets; where AG represents the corresponding datasets Training, Test, TCGA-LIHC, GSE54503, GSE56588, GSE89852, and GSE113017, respectively. Detailed Implementation

[0052] This invention discloses detection markers, a binary classification model, and their applications.

[0053] It should be understood that the expression “one or more of…” individually includes each of the objects described after the expression, as well as various different combinations of two or more of the described objects, unless otherwise understood from the context and usage. The expression “and / or” combined with three or more described objects should be understood to have the same meaning, unless otherwise understood from the context.

[0054] The terms “including,” “having,” or “containing,” including the use of their grammatical synonyms, should generally be understood as open-ended and non-restrictive, for example, not excluding other unstated elements or steps, unless otherwise specifically stated or understood from the context.

[0055] It should be understood that the order of the steps or the order in which certain actions are performed is not important as long as the invention remains operational. Furthermore, two or more steps or actions can be performed simultaneously.

[0056] The use of any and all instances or exemplary language such as “e.g.” or “including” in this document is merely intended to better illustrate the invention and is not intended to limit the scope of the invention unless the claims are made. No language in this specification should be construed as indicating that any unclaimed element is essential to the practice of the invention.

[0057] Furthermore, the numerical ranges and parameters used to define the present invention are approximate values, and the relevant values ​​in the specific embodiments have been presented as precisely as possible. However, any value inevitably contains standard deviations due to individual test methods. Therefore, unless explicitly stated otherwise, it should be understood that all ranges, quantities, values, and percentages used in this disclosure are modified with the word "approximately." Here, "approximately" generally means an actual value within plus or minus 10%, 5%, 1%, or 0.5% of a particular value or range.

[0058] In Embodiments 1 to 6, Verification Example 1, Verification Example 2, and Effect Examples 1 to 6 of this invention:

[0059] (1) Materials and Methods

[0060] Experimental subjects

[0061] The paired cancer tissue and adjacent normal tissue samples from 33 HCC patients used in this invention were collected from July 2015 to June 2016 at the First Affiliated Hospital of Zhejiang University School of Medicine. During the collection process, doctors, nurses, and researchers who had received standardized training obtained relevant information about the specimens, including age, sex, height, and weight. The tissue samples were examined by senior pathologists, who confirmed that the samples matched the characteristics of cancerous or adjacent normal tissue in terms of histological morphology.

[0062] This study was approved by the Ethics Committee of the First Affiliated Hospital of Zhejiang University School of Medicine, and informed consent was obtained from each participant, who also signed an informed consent form. All experiments were conducted in accordance with the Declaration of Helsinki and its amendments.

[0063] Of the 33 HCC patients, one had alcoholic cirrhosis that progressed to hepatocellular carcinoma, and the remaining 32 had hepatocellular carcinoma caused by hepatitis B. There were more males than females (27 patients), while only 6 were female. Among the eight serum biomarkers, alkaline phosphatase (ALP) had the lowest positive rate, with only 2 patients showing elevated values ​​(6% positivity rate, 2 / 33). Alpha-fetoprotein (AFP) had the highest positive rate (60% positivity rate, 20 / 33). All patients had a Child-Pugh score of A. Clinical data of HCC patients from tissue samples are shown in Table 1.

[0064] Table 1 Clinical information of HCC patients

[0065]

[0066]

[0067] (2) Statistical Analysis: All statistical analyses in this study were performed in R (v.4.0.3). For the WGBS data of 33 pairs of liver cancer and adjacent normal tissue samples, we used paired t-tests to analyze the differences between tumor tissue and adjacent normal tissue. Spearman correlation coefficient was used to calculate the correlation between the two groups of data. An absolute value of the correlation coefficient (|cor|) greater than 3 was considered to be correlated, cor>0 was a positive correlation, and cor<0 was a negative correlation. A p-value less than 0.05 was considered significant.

[0068] All raw materials and reagents used can be purchased from the market.

[0069] The present invention will be further illustrated below with reference to the embodiments:

[0070] Example 1: Public Dataset on Hepatocellular Carcinoma Methylation

[0071] This invention downloaded four HCC-related 450K methylation microarray datasets (GSE54503, GSE56588, GSE89852, and GSE113017) from the Gene Expression Omnibus (GEO) database of the National Center for Biotechnology Information (NCBI). We also downloaded methylation microarray data, corresponding clinical information, and methylation microarray annotation files from the TCGA database from the UCSC Xena website (https: / / xenabrowser.net / datapages / ), as shown in Table 2. We first standardized the five public 450K methylation microarray datasets using the `betaqn` function of the R package `wateRmelon`.

[0072] Table 2 Sample Information from Public Datasets

[0073] Tissue Control(N) HCC(N) Platform GSE54503 66 66 450K GSE56588 10+9 224 450K GSE89852 37 37 450K GSE113017 29 29 450K TCGA-HCC 50 380 450K Total 201 736

[0074] Example 2: Whole-genome methylation sequencing and data processing

[0075] This invention uses the phenol-chloroform method to extract total DNA from liver cancer and adjacent normal tissue samples. DNA purity (OD260 / 280 ratio) is detected using Nanodrop, and DNA concentration is precisely quantified using Qubit 3.0. Then, 200 ng of DNA is sonicated to form DNA fragments of approximately 300 bp. Sequencing libraries are then constructed, and library quality control is performed using a biosample analyzer (Agilent 2100 TapeStation system). The library is converted to bisulfite using the EZ DNA methylation kit (catalog number: D5030), followed by insertion length detection using the Agilent 2100. The effective concentration of the library is accurately quantified using q-PCR to ensure library quality. After passing library quality control, the library is pooled and then subjected to 150 bp paired-end high-throughput sequencing (Illumina Xten platform).

[0076] After sequencing, adapter fragments were removed using Cutadapt software (v.1.18). Low-quality bases or sequences were removed using Trimmomatic software (0.33) to obtain clean reads. Subsequently, the data was aligned to the reference genome (hg38) using Bismark software with default settings. After alignment, information for each CpG was extracted using the bismark_methylation_extractor function in Bismark software. CpGs from the positive and negative strands were then merged to obtain the methylation rate of each CpG site. CpG sites with a sequencing depth less than 5× were set to not available (NA), and CpG sites with a deletion rate exceeding 50% in all samples were removed. Furthermore, CpG sites overlapping with common variants were also removed. Finally, the remaining CpG sites were smoothed using DSS software (v.2.26.0) to obtain the methylation level matrix of the CpG sites.

[0077] Example 3 Identification of Differential Methylation Sites

[0078] Differentially methylated CpG loci (DML) are a fundamental step in methylation research and are crucial for subsequent biomarker identification. DML analysis compares the methylation levels (β values) of tumor tissue and adjacent normal tissue to obtain the difference in β values ​​(Δβ), thereby identifying the differentially methylated sites in the tumor tissue.

[0079] We used the R package BWStest (v.0.2.2) to perform a differential methylation analysis based on the Baumgartner-Weiβ-Schindler test (B-test) to correct for the methylation level matrix of each CpG site obtained above, identifying differentially methylated sites (DMLs). Specifically, for each methylated CG site, a separate BWS test was performed within each age group to obtain the corresponding p-values ​​in two directions (i.e., P-left and P-right). Subsequently, the p-values ​​of each site in the same direction across the three age groups were combined to obtain two new statistical values, T-left and T-right (T-left / T-right = -2 × ∑log10(P-left / P-right)), which were then transformed to obtain a new statistical value T (T = max(T-left, T-right)). This was achieved using a 2.0 × 10⁻⁶ test. 8The empirical distribution of T-values ​​for the joint BWS test was obtained by randomly permuting the sample labels. Finally, the empirical P-value for each methylation site was obtained based on the rank of the T-value in the empirical distribution. The P-value for the joint BWS test was less than 1.0 × 10⁻⁶. 5 The site was identified as DML.

[0080] Example 4: Normalization of Common Methylation Chip Data and its Relationship with DML Sites

[0081] To discover generalizable HCC biomarkers using DML, we intersected the DML genomic coordinates with CpG loci from five 450K public datasets (TCGA-LIHC, GSE54503, GSE56588, GSE89852, and GSE113017) obtained in Example 1. Specifically, we intersected the normalized and preserved CpG loci from the five public datasets with the aforementioned DML. No batch correction was performed between the datasets to ensure their independence.

[0082] Example 5: Principal Component Analysis (PCA) and Cluster Analysis Based on Intersection DML

[0083] For biomarker discovery based on WGBS, it is necessary to confirm that the analyzed loci exhibit universal methylation differences, unaffected by effects such as etiology, ethnicity, or experimental batch, in order to discover biomarkers with generalizable capabilities. Therefore, we performed principal component analysis (PCA) on all samples (N=1003) from six independent datasets based on methylation data from intersection loci. Furthermore, we performed clustering analysis on all samples and visualized the results using the R package "pheatmap".

[0084] Example 6: Construction of a Binary Classification Model for Hepatocellular Carcinoma Based on Machine Learning

[0085] To further verify the generalization ability of the biomarkers discovered from the 33 pairs of WGBS data, we performed machine learning based on DML, constructed a binary classification model, and validated its classification performance using five external datasets. The construction process is as follows: Figure 1 As shown.

[0086] (1) Feature selection of random forest

[0087] Random Forest (RF) works by randomly sampling with replacement from the training set to obtain a new training set. The resulting classification trees form a random forest, and the classification results of the new data are scored by voting among these trees. This scoring system helps the overall model achieve higher accuracy and generalization performance.

[0088] In this embodiment, we use the `createDataPartition` function from the `caret` package in R to group the WGBS samples: 75% of the samples are used as the training set and 25% as the test set. Then, we use the `rfe` function from the `caret` package to filter methylation sites in the training set as features. During the filtering process, 10-fold cross-validation is used. The code is as follows:

[0089] library(caret)

[0090] library(e1071)

[0091] library(kernlab)

[0092] setwd(″ / data / RandomForest″)

[0093] data<-read.table(″MET_train.txt″, sep=\t″, header=T, row.names=1, check.names=F)

[0094] x<-scale(data[,2:ncol(data)])

[0095] x<-x[,-findCorrelation(cor(x),.8)]

[0096] x < -as.matrix(x)

[0097] y < -as.matrix(data[,1])

[0098] subsets < -seq(2, 20, 1)

[0099] set.seed(123)

[0100] RFfile<-rfe(x, y, sizes=subsets, rfeControl=rfeControl(functions=rfFuncs, method="cv", verbose=F, returnResamp="all"))

[0101] RFfile$optVariables

[0102] (2) Construct a classification model using Elastic Net regression.

[0103] Elastic Net (ELNET) is a hybrid of LASSO regression and Ridge regression, a linear regression model that uses both L1 and L2 priors as regularization matrices. Macros' research suggests that for 450K methylation microarray data of brain tumors, the binary classification model built with ELNET outperforms models built with Random Forest (RF) and Support Vector Machines (SVM). This embodiment utilizes methylation data, and similar to the above research, it also employs ELNET to construct the classification model.

[0104] In this embodiment, we use the feature-filtered training set data and the glmnet package in R to build the ELNET model. During model building, 10-fold cross-validation is used for model optimization. The script is as follows:

[0105] set.seed(2020)

[0106] a <- seq(0.1, 0.9, 0.05)

[0107] search<-foreach(i=a,.combine=rbind)%dopar%{

[0108] cv<-cv.glmnet(x, y, family="binomial", nfold=10, type.measure="deviance", parallel=TRUE, alpha=i)

[0109] data.frame(cvm=cv$cvm[cv$lambda==cv$lambda.1se], lambda.1se=cv$lambda.1se, alpha=i)

[0110] }

[0111] cv<-search[search$cvm==min(search$cvm),]

[0112] md<-glmnet(x, y, family="binomial", lambda=cv$lambda.1se, alpha=cv$alpha)

[0113] coef(md3)

[0114] Subsequently, we used the constructed ELNET model to calculate the methylation binary classification score (mBCS) for each sample in the WGBS and public datasets. We validated the model's classification performance using the area under the curve (AUC) of the receiver operator characteristic (ROC) and the confusion matrix.

[0115] Validating the correlation between the binary classification model (mBCS) in Example 1 and clinical characteristics

[0116] We further analyzed the differences in the constructed binary classification model among different genders, age groups, etc., and its correlation with clinical information such as AFP.

[0117] Verification Example 2: Relationship between the binary classification model (mBCS) and the level of immune cell infiltration

[0118] We performed ssGSEA analysis on the mRNA-Seq data of 33 HCC samples to obtain the infiltration levels of 23 immune cell types in each tissue sample, as shown in Tables 3–8. Subsequently, we performed correlation analysis between mBCS and the infiltration level of each immune cell type in each sample. The results showed that mBCS was significantly correlated with the infiltration levels of 17 immune cell types, such as… Figure 2 As shown in the figure, among the significantly correlated immune cell types, only active CD4+ T cells, CD56 bright NK cells, and CD56 dim NK cells showed a positive correlation with mBCS, while the other 14 immune cell types, including activated B cells, activated CD8+ T cells, eosinophils, immature B cells, immature dendritic cells, myeloid-derived suppressor cells, macrophages, mast cells, monocytes, NK cells, neutrophils, regulatory T cells, Th1 and Th17 cells, showed a significant negative correlation with mBCS.

[0119] ssGSEA analysis revealed a significant negative correlation between mBCS and the infiltration levels of various immune cells, partially revealing that the mechanism by which mBCS is involved in tumorigenesis may be related to immune deficiency.

[0120] Table 3. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0121]

[0122]

[0123] Table 4. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0124]

[0125]

[0126] Table 5. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0127]

[0128]

[0129] Table 6. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0130]

[0131]

[0132]

[0133] Table 7. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0134]

[0135]

[0136] Table 8. Enrichment fractions of 23 immune cells in each sample using ssGSEA analysis.

[0137]

[0138]

[0139] Example 1: Identification results of differentially methylated sites (DML)

[0140] The WGBS data obtained in Example 3, after QC processing, achieved an average coverage depth of 12.76×, containing 28,978,826 CpG sites. After smoothing, approximately 34% (9,867,700) of the CpG sites showed significant differences in methylation levels between cancerous and adjacent normal tissues. Of these, 157,320 sites were hypermethylated, and 9,710,380 sites were hypomethylated. Gene coordinate analysis revealed that hypermethylated sites were primarily located in gene promoter regions, while hypomethylated sites were mainly located in the gene body.

[0141] The results above show that the number of DMLs analyzed based on WGBS is at least an order of magnitude higher than the differentially expressed sites found by methylation chips, demonstrating the advantage of WGBS technology in providing a more comprehensive analysis of differential methylation in HCC, and also indicating the quantitative advantage of WGBS-based DML in the discovery of diagnostic biomarkers.

[0142] Example 2: Intersection of differential methylation sites and 450K methylation chip sites

[0143] In Example 4, since only 33 pairs of liver tissues were identified for the aforementioned differentially methylated loci (DMLs), and the etiology of hepatocellular carcinoma is complex and highly heterogeneous, the specificity of these DML datasets is the first question that the biomarkers discovered from these DMLs need to answer. Therefore, we integrated the identified differentially methylated loci (DMLs) with five other independent HCC methylation microarray datasets. The sample size of the five HCC-related 450K methylation microarray public datasets is shown in Table 2.

[0144] Five 450K microarray datasets were standardized and their intersection with the DML dataset from Example 3 was analyzed. The results showed that a total of 11,352 CpG sites were present in all six datasets. The methylation matrix of the 11,352 intersection sites in each dataset was extracted without batch correction to ensure independence among the six datasets. The generalization ability of the biomarkers discovered by DML was then analyzed based on the 11,352 intersection sites from these six independent datasets.

[0145] Example 3: Principal component analysis and cluster analysis of six independent datasets.

[0146] We performed principal component analysis (PCA) on methylation data at 11,352 sites from centralized samples (N=1003) in six independent datasets. The results are as follows... Figure 3As shown in Figure A, tumor samples are relatively dispersed, while normal samples are clustered in the upper left corner of the two-dimensional graph. At the same time, the datasets are mixed together without obvious clustering, indicating that the DML identified by WGBS does not have obvious batch differences and can distinguish between tumor and normal samples.

[0147] Cluster analysis of the samples also showed that normal liver tissue samples clustered on the left side, while tumor tissue clustered on the right side, and there was no significant clustering between datasets, demonstrating that the differences between tissue types in DML are greater than the batch differences between datasets. This result suggests the generalization ability of DML in discovering biomarkers.

[0148] Example 4: Feature Selection in Random Forest

[0149] We split the 33 pairs of WGBS samples into a training set of 25 pairs (75%) and a test set of 8 pairs (25%). Then, we used the random forest (RF) algorithm to perform feature selection on the training set, choosing the subset of CpGs most relevant to HCC identification for the next step of machine learning model training. Ten-fold cross-validation selected the 13 optimal feature CpG sites for the random forest model (e.g., ...). Figure 4 (As shown).

[0150] Example 5: Construction and Validation of a Classification Model Based on Elastic Network Regression

[0151] We used the Elastic Network Regression (ELNET) algorithm and performed 10-fold cross-validation to construct the optimal HCC classification model. The optimal model had α = 0.15 and λ = 0.00057. The methylation levels (β values) of the 11 CpG sites were used as model variables (as shown in Table 9), and the binary classification model scores were calculated. The genes corresponding to the CpG sites are also shown in Table 9.

[0152] Table 9. CpG sites and corresponding genes in the ELNET model and regression coefficients in the model.

[0153] CpG Markers Position Ref Gene Coefficients cg26492368 Chr10: 22345804 SPAG6 8.1559394 cg26608718 Chr19:15419926 AKAP8L -8.4531282 cg19516340 Chr2: 10123569 RRM2 -6.1125863 cg02554274 Chr11:788180 CEND1 -3.4444817 cg09341491 Chr16: 50673884 SNX20 6.8589209 cg13788685 Chr1: 161313856 SDHC -3.7790125 cg08383929 Chr5: 181219754 - 7.0882640 cg03526459 Chr1:147078392 NBPF13P, RNVU1-8 3.3835278 cg14304336 Chr10: 126463525 C10orf90 -6.0109634 cg27039312 Chr8: 142544430 BAI1 -4.5737240 cg26956009 Chr19:2356488 - -5.3367444

[0154] The formula for calculating the score in a binary classification model is as follows:

[0155] mBCS=0.3948785+∑βiEi

[0156] Here, i refers to the 11 CpG sites; β represents the independent variable, namely the methylation level of site i; E represents the regression coefficient, where the methylation level is determined by the intensity ratio between methylated and unmethylated alleles at the CpG site (β value).

[0157] Based on the methylation model scoring formula, the mBCS score of each sample in the training set, test set, and five public datasets was calculated. ROC curve analysis was performed on each dataset to assess the sensitivity and specificity of the binary classification model score. The results show that the AUC of 1 is 1 for both the training and test sets, indicating that the model performs superiorly on the same dataset. The AUCs of this binary classification model on the five public validation datasets are 0.984 for TCGA-LIHC, 0.970 for GSE54503, 0.994 for GSE56588, 0.999 for GSE89852, and 1 for GSE113017 (as shown in Table 10). Figure 5 (As shown). In the five validation datasets, the classification model's AUC was no less than 0.97, demonstrating its superior performance. This result further confirms the consistency of DML across the datasets.

[0158] Table 10 shows the AUC of classification performance in the dataset based on the classification model score (mBCS).

[0159] Dataset Control(N) HCC(N) AUC Trianing 25 25 1.000 Test 8 8 1.000 GSE54503 66 66 0.970 GSE56588 10+9 224 0.994 GSE89852 37 37 0.999 GSE113017 29 29 1.000 TCGA-LIHC 50 380 0.984

[0160] A confusion matrix was constructed, with mBCS values ​​greater than or equal to 0.5 considered positive and values ​​less than 0.5 considered negative. The results showed that the overall accuracy of the binary classification model (mBCS) was 85.6% (as shown in Table 11), with sensitivity and specificity of 81.4% and 99.6%, respectively. Furthermore, we created heatmaps displaying the methylation levels of 11 CpG sites for each sample in different datasets (e.g., ...). Figure 6 (As shown).

[0161] Table 11. Prediction Confusion Matrix of Binary Classification Model Score (mBCS)

[0162]

[0163] Example 6: Correlation between the binary classification model and clinical characteristics

[0164] The above analysis indicates that the sensitivity of the mBCS (modified binary classification model) for HCC sample identification has room for improvement. Since TCGA-LIHC has relatively complete clinical information and a large sample size, we analyzed the clinical information of tumor tissue samples from HCC patients in the TCGA-LIHC cohort and the prediction results of the mBCS model. The results are shown in Table 12. There were no significant differences in the mBCS identification rate for HCC across different pathological stages, age groups, and gender groups.

[0165] Table 12 mBCS Prediction Results and Clinical Characteristics

[0166]

[0167]

[0168]

[0169] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A marker, characterized in that, It is composed of cg26492368, cg26608718, cg19516340, cg02554274, cg09341491, cg13788685, cg08383929, cg03526459, cg14304336, cg27039312 and cg26956009; CpG Markers Position Ref Gene Coefficients cg26492368 Chr10:22345804 SPAG6 8.1559394 cg26608718 Chr19:15419926 AKAP8L -8.4531282 cg19516340 Chr2:10123569 RRM2 -6.1125863 cg02554274 Chr11:788180 CEND1 -3.4444817 cg09341491 Chr16:50673884 SNX20 6.8589209 cg13788685 Chr1:161313856 SDHC -3.7790125 cg08383929 Chr5:181219754 - 7.0882640 cg03526459 Chr1:147078392 NBPF13P , RNVU1-8 3.3835278 cg14304336 Chr10:126463525 C10orf90 -6.0109634 cg27039312 Chr8:142544430 BAI1 -4.5737240 cg26956009 Chr19:2356488 - -5.3367444.

2. The use of the reagent for detecting the biomarker as described in claim 1 in the preparation of products for detecting liver cancer.

3. A binary classification model, characterized in that, The calculation formula is: mBCS=0.3948785+∑β i E i Where i is the biomarker as described in claim 1, β is the methylation level of the biomarker as described in claim 1, and E is the regression coefficient of the biomarker as described in claim 1; The correspondence between the regression coefficient of the marker and the marker is as follows: CpG Markers Position Ref Gene Coefficients cg26492368 Chr10:22345804 SPAG6 8.1559394 cg26608718 Chr19:15419926 AKAP8L -8.4531282 cg19516340 Chr2:10123569 RRM2 -6.1125863 cg02554274 Chr11:788180 CEND1 -3.4444817 cg09341491 Chr16:50673884 SNX20 6.8589209 cg13788685 Chr1:161313856 SDHC -3.7790125 cg08383929 Chr5:181219754 - 7.0882640 cg03526459 Chr1:147078392 NBPF13P , RNVU1-8 3.3835278 cg14304336 Chr10:126463525 C10orf90 -6.0109634 cg27039312 Chr8:142544430 BAI1 -4.5737240 cg26956009 Chr19:2356488 - -5.3367444.

4. The application of the binary classification model as described in claim 3 in the preparation of a detection system or device for liver cancer; The evaluation criteria include: A positive result is defined as an mBCS value greater than or equal to 0.5, and a negative result is defined as an mBCS value less than 0.5.

Images