Esophageal cancer marker screening method based on enzyme transfer methylation sequencing

Abstract

The invention provides an esophageal cancer marker screening method based on enzyme transfer method methylation sequencing, which comprises the following steps: S1, collecting esophageal early cancer and para-carcinoma tissue samples, and carrying out enzyme transfer method whole genome methylation sequencing; s2, integrating methylation data of esophageal cancer and para-carcinoma tissues in databases of TCGA, GEO, MethMarker DB and the like; s3, screening differential methylation sites, preferably sites which are hypermethylated in esophageal cancer but hypomethylated in samples such as leukocytes, so as to reduce plasma background interference; s4, obtaining a candidate marker combination based on an integrated methylation feature screening model of LASSO regression and a random forest algorithm; s5, constructing a classifier model based on the candidate marker combination, carrying out training and reasoning test on the model, and evaluating the prediction performance of the model; and S6, obtaining the esophageal cancer marker combination. According to the method, enzyme transfer sequencing and machine learning are combined, the marker prediction accuracy is remarkably improved, and a new reference is provided for clinical research of esophageal cancer.

Description

Technical Field

[0001] The present invention relates to the field of large models for early screening of esophageal cancer, and in particular to a method for screening esophageal cancer markers based on enzyme transfer methylation sequencing. Background Art

[0002] Among digestive tract cancers, esophageal squamous cell carcinoma (ESCC) is the main pathological type. The overall prognosis of patients is poor, with a 5-year survival rate of only about 20%.

[0003] Since esophageal cancer has an insidious onset and its early symptoms are not obvious, most patients are already in the middle or late stage of esophageal cancer when diagnosed, resulting in a poor prognosis for esophageal cancer patients, with a 5-year survival rate of only 30%. However, the 5-year survival rate of patients with early esophageal cancer can reach over 85%.

[0004] Therefore, early diagnosis of esophageal cancer is an effective way to reduce the burden of esophageal cancer.

[0005] Traditional imaging tests and detection of biomarkers such as cytokeratin fragment 19 (CYFRA21-1), carcinoembryonic antigen (CEA), and squamous cell carcinoma antigen (SCC) can improve the diagnostic accuracy and prognostic assessment of advanced esophageal cancer. However, there are many limitations in the early diagnosis of esophageal cancer, and the sensitivity and specificity are low, and it is still immature.

[0006] Therefore, there is an urgent need to develop new markers that are accurate and have low detection costs for diagnosis, so as to contribute to the early diagnosis and personalized precision treatment of patients with esophageal cancer.

[0007] In recent years, the research on tumor-related epigenetics has been continuously deepened, providing new strategies for early screening and diagnosis, dynamic monitoring, and medication guidance. Compared with other markers, methylation has obvious advantages such as more detectable sites, high signal abundance, and tissue tracing.

[0008] At present, some studies have performed methylation sequencing on esophageal cancer tissue samples to screen for differential methylation sites, but all of them are based on the bisulfite conversion method, which is prone to DNA damage and degradation during the conversion process, resulting in distortion of methylation information. In addition, the sample source is relatively single, and the markers obtained by screening have limited performance for early diagnosis of esophageal cancer. Therefore, there is still a lack of marker screening methods and diagnostic models based on large-scale sample methylation sequencing data from different sources.

[0009] In summary, developing a new esophageal cancer marker screening method and diagnostic model based on enzyme transfer methylation sequencing and expanding the molecular technology means for early diagnosis of esophageal cancer are of great significance to improving the diagnosis rate of early esophageal cancer and prolonging patient survival. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a method for screening esophageal cancer markers based on enzyme transfer methylation sequencing.

[0011] Technical solution of the present invention: The present invention discloses a method for screening esophageal cancer markers based on enzyme transfer methylation sequencing, comprising the following specific steps:

[0012] S1. Collect cancer tissues and paired adjacent tissue samples from patients with early-stage esophageal cancer and perform whole-genome methylation sequencing using a novel enzyme transfer method;

[0013] S2. Integration of samples from different databases, including esophageal cancer and adjacent tissue data from TCGA and GEO databases and methylation site information of esophageal cancer from the MethMarkerDB database;

[0014] S3. The differentially methylated sites were further filtered by combining the methylation levels of whole blood, leukocytes, and peripheral blood mononuclear cell samples. The sites that were easily interfered by plasma background methylation were eliminated based on the standard of high methylation in esophageal cancer tissue relative to adjacent tissues and low methylation in other samples such as leukocytes.

[0015] S4. An integrated methylation feature screening model based on LASSO regression and random forest algorithm extracts key features from the genomic methylation map, obtains candidate methylation marker combinations, and achieves classification prediction of esophageal cancer using a relatively small number of methylation marker site features;

[0016] S5. Construct a classifier model based on the candidate marker combination, train and test the model, and evaluate its predictive performance.

[0017] Furthermore, the step S1 specifically includes the following steps:

[0018] S11. Collect cancer tissue and its paired adjacent tissue samples from early-stage esophageal cancer patients, extract DNA from tissue samples and quantify it;

[0019] S12, using enzyme conversion method to perform methylation conversion treatment on the extracted DNA samples, including oxidation reaction using TET methylcytosine dioxygenase 2 and T4 bacteriophage β-glucosyltransferase and deamination reaction using cytosine deaminase;

[0020] S13. Use single-stranded DNA library construction technology to construct a methylation library for the converted samples, including DNA denaturation, P7 adapter ligation, complementary chain synthesis, P5 adapter ligation, and PCR amplification, to avoid the introduction of non-methylated sites in the end repair process of traditional double-stranded library construction, which will have an adverse effect on data analysis;

[0021] NovaSeq 6000 was used for paired-end 150 bp read length sequencing with a sequencing depth of 30×.

[0022] Furthermore, in steps S1 and S2, the methylation whole genome sequencing data and chip data of samples from different sources need to be quality controlled and differentially methylated CpG sites screened;

[0023] The methylation whole genome sequencing data was quality controlled under the following conditions:

[0024] a. Remove the reads contaminated by the adapter;

[0025] b. Truncate the reads with the first and last quality lower than 3 or containing N, where N means the base information cannot be determined;

[0026] c. Remove the fragments 10bp upstream and downstream of the read segment;

[0027] d. Cut off low-quality reads using a sliding window approach, with 4 bases as a window. If the average base quality value of the window is lower than 15, the read is cut off from there;

[0028] The screening conditions for the differentially methylated CpG sites are: methylation difference ≥ 25% and Q-value ≤ 0.01;

[0029] The methylation chip data were subjected to data quality control and normalization, and the filtering conditions were as follows:

[0030] a. Probes with probe detection P>0.01;

[0031] b. Filter out probes with less than 3 beads in at least 5% of the samples;

[0032] c. Filter out all non-CpG probes;

[0033] d. Eliminate all SNP-related probes;

[0034] e. Eliminate the probe on the sex chromosome;

[0035] Differential methylation probes were screened based on methylation chip data, and the screening conditions were: |Δβ|≥0.25 and Q-value≤0.01.

[0036] Furthermore, in step S3, the differentially methylated sites are further filtered, and the average methylation signal levels of the corresponding CpG sites in whole blood, leukocytes, PBMC and other samples in TCGA and GEO, i.e., the plasma background methylation signal levels, are analyzed, and sites that are highly methylated in esophageal cancer tissues and have a plasma background methylation level of less than 0.1 are retained.

[0037] Furthermore, in step S4, the following steps are included:

[0038] S4.1. Construction of methylation feature fusion matrix: Based on the genomic methylation data, the CpG site methylation level data of all samples are obtained, and data quality control and preprocessing are performed. The methylation information of all CpG sites of each sample in the training set is fused to form a methylation feature fusion matrix;

[0039] S4.2. Perform LASSO regression analysis, including the following steps:

[0040] S4.21, taking the methylation feature fusion matrix as input, and using LASSO regression for feature selection;

[0041] S4.22. Add L1 regularization term to the loss function to control model complexity and select important features; where:

[0042] The loss function of the LASSO regression model is:

[0043] Where ||w||1 represents the L1 norm of w,

[0044] S4.23. Use the gradient descent algorithm to iteratively find the most appropriate value; loss function gradient: Where α is the penalty term, sign(w) is the sign function, and the appropriate α value is set according to the actual number of sites, and the number of features whose coefficient converges to 0 is adjusted to perform feature screening;

[0045] S4.24. The model was iteratively optimized by gradient descent method to find the optimal regularization coefficient α and identify key methylation sites associated with esophageal cancer.

[0046] S4.3, random forest algorithm application, including the following steps:

[0047] S4.31. Build a random forest model. Use the Bootstrap sampling method to build multiple sub-datasets from the training set. For each sub-dataset, build a decision tree and randomly select a subset of feature variables at each node for splitting. After the training is completed, calculate the Gini index of each feature to evaluate the importance of the feature.

[0048] S4.32. Feature importance assessment: score the importance of the features in each tree, using the Gini index difference calculation formula:

[0049] Among them, GI mis the Gini index of the node before splitting, GI l and GI r is the Gini index of the left and right child nodes after splitting;

[0050] The feature importance scores of all trees are normalized to obtain the final feature importance score:

[0051]

[0052] For the methylation feature sites, they were sorted in descending order according to their importance scores, and the feature sites with the highest ranking were selected for further screening.

[0053] S4.4, Intersection feature selection, combining the characteristic sites screened out by LASSO regression and random forest algorithm respectively, taking the intersection of the two as the final candidate marker combination.

[0054] Furthermore, in step S5, based on the candidate marker combination, a classifier model is established using a Logistic regression algorithm, the data set is split into a training set and a validation set in a ratio of 7:3, the methylation level data of the candidate marker sites are used as input, and the sample type is represented by 0 and 1 as the output of the model. The model is trained and tested, and its predictive performance is evaluated. The evaluation indicators include sensitivity, specificity, and ROC curve AUC value.

[0055] Furthermore, the different databases include esophageal cancer tissue and adjacent tissue data in TCGA, GEO, and esophageal cancer related DMR information in MethMarkerDB database.

[0056] Furthermore, the methylation conversion treatment includes using TET methylcytosine dioxygenase 2 and T4 bacteriophage β-glucosyltransferase to perform an oxidation reaction and using cytosine deaminase to perform a deamination reaction.

[0057] Furthermore, NovaSeq 6000 was used for paired-end 150 bp read length sequencing with a sequencing depth of 30×.

[0058] The beneficial effects of the present invention compared with the prior art are as follows:

[0059] 1. High accuracy of marker screening: This method is based on enzyme transfer methylation sequencing, combined with the collection of clinical early esophageal cancer tissue, adjacent tissue sequencing data and sample data from multiple databases (TCGA, GEO, MethMarkerDB, etc.), which can effectively screen out highly specific methylation sites related to esophageal cancer; further filtering through differential methylation sites, retaining sites with high methylation in esophageal cancer and low plasma background methylation levels, effectively reducing background interference and improving screening efficiency. The screened marker combination has significantly improved the diagnostic sensitivity of early esophageal cancer, which is much higher than existing clinical markers.

[0060] 2. Integrated optimization of feature extraction: By combining LASSO regression and random forest algorithm for feature screening, we can extract key linear features and capture complex nonlinear features. This integrated method reduces the number of variables, improves model interpretability, and reduces the risk of overfitting.

[0061] 3. Validity and reliability of the classifier: Build a classification model based on candidate markers and logistic regression, train and test the model, evaluate its predictive performance, and ensure the reliability and stability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a flow chart of the screening method of the present invention;

[0063] Figure 2 The ROC curves of the classifier in the training set of esophageal cancer stage 0-I samples (left figure) and stage II-IV samples (right figure) in the embodiment of the present invention are shown in FIG.

[0064] Figure 3 1 is the ROC curve of the classifier in the test set of esophageal cancer stage 0-I samples (left figure) and stage II-IV samples (right figure) in the embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings. The embodiments are only used to explain the present invention and do not constitute a limitation on the protection scope of the present invention.

[0066] like Figure 1 As shown, a method for screening esophageal cancer markers based on enzyme transfer methylation sequencing comprises the following steps:

[0067] S1. Collect cancer tissues and paired adjacent tissue samples from patients with early-stage esophageal cancer and perform whole-genome methylation sequencing using a novel enzyme transfer method;

[0068] S2. Integration of samples from different databases, including esophageal cancer tissue and adjacent tissue data from TCGA, GEO, and esophageal cancer methylation site information from the MethMarkerDB database;

[0069] S3. Further filtering of differentially methylated sites was performed by combining the methylation levels of whole blood, leukocytes, PBMCs and other samples. The sites that were easily affected by plasma background methylation were eliminated based on the standard of high methylation in esophageal cancer tissues relative to adjacent paracancerous tissues and low methylation in other samples such as leukocytes.

[0070] S4. Based on LASSO regression and random forest algorithm, key features were extracted to obtain candidate methylation marker combinations;

[0071] S5. Build a classifier model based on the candidate marker combination, train and perform inference tests on the model, and evaluate its predictive performance;

[0072] Wherein, S1 comprises the following steps:

[0073] S11. Collect cancer tissue and its paired adjacent tissue samples from early-stage esophageal cancer patients, extract DNA from tissue samples and quantify it;

[0074] S12, using enzyme conversion method to perform methylation conversion treatment on the extracted DNA sample, specifically including using TET methylcytosine dioxygenase 2 (TET2) and T4 bacteriophage β-glucosyltransferase (T4-BGT) for oxidation reaction and using cytosine deaminase (APOBEC) for deamination reaction;

[0075] S13. Single-stranded DNA library construction technology was used to construct the methylation library of the converted samples, including DNA denaturation, P7 adapter ligation, complementary chain synthesis, P5 adapter ligation, PCR amplification and other reactions, which effectively avoided the introduction of non-methylated sites in the end repair process of traditional double-stranded library construction, which had an adverse effect on data analysis; NovaSeq 6000 was used for double-end 150bp read length sequencing with a sequencing depth of 30×.

[0076] The methods for quality control of different types of data and screening of differentially methylated sites in step S2 are as follows:

[0077] Whole genome methylation sequencing data: Use fastp to trim the raw data, and the control conditions are: a. Truncate the reads contaminated by the adapter; b. Truncate the reads with a head and tail quality lower than 3 or containing N (N means that the base information cannot be determined); c. Eliminate the fragments 10bp upstream and downstream of the read; d. Truncate the low-quality reads, using a sliding window method, with 4 bases as a window. If the average base quality value of the window is lower than 15, the read is truncated from there. Use methylKit for differential methylation site analysis, and the screening conditions are: methylation difference ≥ 25% and Q-value ≤ 0.01;

[0078] The chip data was quality controlled and normalized using ChAMP, with the following filtering conditions: a. Probes with P>0.01; b. Probes with less than 3 beads in at least 5% of the samples were filtered out; c. All non-CpG probes were filtered out; d. All SNP-related probes were removed; e. Probes on sex chromosomes were removed. Limma was used to screen differential methylation probes, and the methylation level β between groups was first analyzed. The calculation formula for β was:

[0079]

[0080] Where M is the methylation signal intensity, U is the unmethylation signal intensity; the screening conditions for differential methylation probes are: |Δβ|≥0.25 and Q-value≤0.01

[0081] MethMarkerDB: The MethMarkerDB database contains a large amount of methylation-related data, which integrates 724 DNA methylation biomarker genes, and contains information from 1,425 published articles from PubMed, as well as a large amount of whole-genome methylation sequencing data. Based on a large amount of public methylation data, the database also integrates and analyzes the key differentially methylated regions, providing data support for various methylation-based analyses and the discovery of methylation marker sites.

[0082] Based on the information collected in MethMarkerDB, DMRs associated with esophageal cancer were screened, including a total of 212,895 DMRs, of which 2,389 DMRs had higher methylation levels in cancer tissue samples than in non-cancerous tissue samples, namely Hyper-DMRs, and 210,506 DMRs had lower methylation levels in cancer tissue samples than in non-cancerous tissue samples, namely Hypo-DMRs.

[0083] In step S3, the differentially methylated sites are further filtered, mainly including analyzing the average methylation signal level of the corresponding CpG sites in the sample data of whole blood, leukocytes, PBMC, etc. in the TCGA and GEO databases, that is, the plasma background methylation signal level, and retaining the sites that are highly methylated in esophageal cancer tissues relative to adjacent tissues and have a plasma background methylation level lower than 0.1.

[0084] In step S4, an integrated methylation feature screening model based on LASSO regression and random forest algorithm is used to extract key features and obtain candidate methylation marker combinations. Among them, the LASSO regression algorithm has the advantages of small computational complexity and easy extraction of important linear features in variables, good model interpretability, and the random forest algorithm has the ability to extract complex high-dimensional nonlinear features in variables. Combining the two to construct an integrated methylation feature screening model can achieve the retention of the most important data features while significantly reducing the number of variables, and can also avoid the risk of overfitting.

[0085] The main steps include:

[0086] Based on the genomic methylation data, the methylation level data of the genomic CpG sites are obtained. After preprocessing the data such as quality control, the methylation information of all CpG sites of each sample in the training set is subjected to feature fusion extraction processing to construct a methylation feature fusion matrix.

[0087] LASSO regression algorithm: A regularization term based on the L1 norm is added to the loss function, as follows:

[0088] If there is such a linear regression equation y m×1 =X m×n w n×1 +b n×1 , then its loss function is as follows:

[0089]

[0090] The first term describes the degree of deviation of y, ||w||1 represents the L1 norm of w,

[0091] Use the gradient descent algorithm to iteratively find the most appropriate value; loss function gradient:

[0092]

[0093] Where α is the penalty term, sign(w) is the sign function, and the appropriate α value is set according to the actual number of sites to extract key features.

[0094] The methylation feature fusion matrix was input into the LASSO regression model, and the gradient descent algorithm was used to learn and determine the feature weights under different regularization coefficients α, to balance the model fitting and the number of features, to obtain the optimal regularization coefficient α, and to obtain the corresponding methylation feature marker sites for subsequent further screening.

[0095] Random forest algorithm: The basic principle of random forest is to integrate multiple decision trees built based on Bootstrap sampling sample queues, and randomly select a subset of feature variables at each node. First, use the Bootstrap sampling method to extract samples from the original data to form multiple sub-datasets; for each sub-dataset, build a decision tree, and at each node, randomly select a part of the features for splitting. Repeat the above steps until the specified number of decision trees are generated.

[0096] The sample data set was randomly sampled in a ratio of 7:3 and divided into a training set and a test set. The proportion of different types of samples was balanced in both the training set and the test set. The training set was used to build a random forest model, and the inference test of the trained model was performed on the test set. For each feature of the random forest, the Gini index was used to evaluate the importance of the feature. The larger the Gini index, the higher the importance of the feature, and the top-ranked sites were selected as candidate markers.

[0097] The methylation feature fusion matrix of the training set samples was input to construct a random forest model. According to the importance score of each feature site, all feature sites were sorted in descending order of importance score to obtain the top feature site, which is the most important feature site for the diagnosis of esophageal cancer.

[0098] Feature intersection selection: The key features screened by LASSO regression and random forest algorithm were selected as the final candidate marker combination.

[0099] After obtaining the candidate methylation marker combination based on LASSO regression and random forest in S5, the classifier model was established using the Logistic regression algorithm, and the data set was split into a training set and a validation set in a ratio of 7:3. The methylation level data of these marker sites were used as input, and the sample type was represented by 0 and 1 as the output of the model. The model was trained and tested, and its predictive performance was evaluated. The evaluation indicators included sensitivity, specificity, and AUC value.

[0100] Example 1

[0101] The present invention proposes an esophageal cancer marker screening method and marker combination based on enzyme transfer methylation sequencing, comprising:

[0102] 1. Sequencing and data analysis process of clinical early esophageal cancer-related tissue samples

[0103] (1) 40 pairs of samples of cancer tissue and corresponding adjacent tissue from patients with early esophageal cancer were collected and subjected to enzyme-based methylation sequencing to obtain raw sequencing data. Sequencing was performed using Novasek 6000 with a sequencing depth of 30×.

[0104] (2) Read quality control

[0105] Fastp was used to remove read segment adapters and perform quality control on the raw data obtained from sequencing. The read segments were required to meet the following conditions:

[0106] a) Remove reads containing the uncertain base N

[0107] b) Remove the 10 bp fragments at both ends of the read segment

[0108] c) Use a 4bp sliding window to remove the parts with a head quality lower than 3

[0109] d) After removing the low-quality head fragments, use a 4bp sliding window to remove downstream fragments with a quality lower than 15

[0110] (3) Internal reference comparison and quality control

[0111] Based on the reference sequence data of positive and negative internal references, Bismark was used to align the filtered reads to obtain the CpG methylation ratio of the internal reference, requiring the CpG methylation ratio of the positive internal reference to be higher than 98% and the CpG methylation ratio of the negative internal reference to be lower than 1%.

[0112] (4) Genome alignment and sequence deduplication

[0113] Based on the human hg19 reference genome sequence data, Bismark was used to align the filtered reads to obtain the resulting BAM file; Bismark deduplicate was used to deduplicate the BAM file to avoid the influence of excessive PCR amplification; Bismark methylation extractor was used to extract methylation sequences and calculate the methylation level data of CpG sites in the whole genome for subsequent analysis.

[0114] (5) Differential methylation analysis

[0115] For the methylation whole genome sequencing data, the R package methylKit was used to screen the differential methylation sites for the cancer and adjacent adjacent samples. The screening conditions were: methylation difference ≥ 25% and Q-value ≤ 0.01.

[0116] After differential methylation analysis, 8750 DMCs were found, of which 1358 DMCs had higher methylation levels in cancer tissues than in adjacent tissues (hyper-DMCs), and 7392 DMCs had lower methylation levels in cancer tissues than in adjacent tissues (hypo-DMCs).

[0117] 2. Screening of differentially methylated sites in public databases

[0118] (1) Data Collection

[0119] Methylation data from cancer tissue samples and paracancerous tissue samples were collected. Methylation chip datasets GSE121930, GSE178212, and GSE79366 of cancer and paracancerous tissues were collected from the NCBI GEO (Gene Expression Omnibus) database, and methylation data of esophageal cancer tissues and paracancerous tissues were collected from The Cancer Genome Atlas (TCGA) database.

[0120] At the same time, data from plasma, blood cells, lymphocytes, etc. are collected to filter out background noise in plasma. Data related to plasma, blood cells, etc. are collected from the GEO database. The dataset GSE40279 contains 656 whole blood samples from healthy people of different ages, GSE115278 contains 474 peripheral blood leukocyte samples from healthy people, GSE125105 contains 210 whole blood samples from healthy people, and GSE69270 contains 184 leukocyte samples from healthy people.

[0121] Based on the information collected in MethMarkerDB, DMRs associated with esophageal cancer were screened out, including a total of 212,895 DMRs, of which 2,389 DMRs had higher methylation levels in cancer tissue samples than in non-cancerous tissue samples, namely Hyper-DMRs, and 210,506 DMRs had lower methylation levels in cancer tissue samples than in non-cancerous tissue samples, namely Hypo-DMRs.

[0122] (2) Data quality control and differential methylation analysis

[0123] The chip data from different sample sources were filtered, and the filtering conditions were: a. probes with P>0.01; b. probes with less than 3 beads in at least 5% of the samples were filtered out; c. all non-CpG probes were filtered out; d. all SNP-related probes were eliminated; e. probes on sex chromosomes were eliminated. Using limma to screen differential methylation probes, the methylation level β between groups was first analyzed, and the calculation formula of β was:

[0124]

[0125] Where M is the methylation signal intensity, and U is the unmethylation signal intensity. The screening conditions for differential methylation probes were: |Δβ|≥0.25 and Q-value≤0.01. Finally, 3578 DMCs were found to have higher methylation levels in cancer tissues than in adjacent tissues (hyper-DMCs), and 2565 DMCs were found to have lower methylation levels in cancer tissues than in adjacent tissues (hypo-DMCs).

[0126] 3. Combined analysis of differential methylation and preliminary screening of DMC sites

[0127] The DMC sites obtained by screening data from different sources in 1 and 2 were jointly analyzed, and the intersection of sites with significant differences in different data was selected, and a total of 896 DMCs were obtained for subsequent feature extraction and screening.

[0128] 4. Differential methylation site filtering

[0129] For the DMCs screened out from samples of cancer tissues and adjacent tissues, we further analyzed the average methylation signal levels of the corresponding CpG sites in samples such as plasma, blood cells, and lymphocytes, i.e., the plasma background methylation signal levels. DMCs with plasma background methylation signal levels higher than 0.1 were eliminated, and 266 hyper-DMCs with low plasma background methylation levels were obtained.

[0130] 5. Integrated methylation feature screening model

[0131] An integrated methylation feature screening model based on the LASSO regression algorithm and the random forest algorithm was used to extract key features and obtain candidate methylation marker combinations. Among them, the LASSO regression algorithm has the advantages of small computational complexity and easy extraction of important linear features in variables, and good model interpretability. The random forest algorithm has the ability to extract complex high-dimensional nonlinear features in variables. Combining the two to construct an integrated methylation feature screening model can achieve the retention of the most important data features while significantly reducing the number of variables, while also avoiding the risk of overfitting.

[0132] Based on the genomic methylation data, the methylation level data of the genomic CpG sites are obtained. After preprocessing the data such as quality control, the methylation information of all CpG sites of each sample in the training set is subjected to feature fusion extraction processing to construct a methylation feature fusion matrix.

[0133] 6. Site selection optimization based on LASSO regression machine learning method

[0134] The LASSO regression cross-validation provided by scikit-learn can be used to determine the optimal coefficient α. After cross-validation of samples, the regularization coefficient α = 0.0011895 is selected. At this time, 43 of the 266 candidate sites are selected.

[0135] 7. Use scikit-learn for random forest training and testing

[0136] The sample data set was randomly sampled in a ratio of 7:3 and divided into a training set and a test set. The proportion of different types of samples was balanced in both the training set and the test set. The training set was used to build a random forest model, and the inference test of the trained model was performed on the test set. For each feature of the random forest, the Gini index was used to evaluate the importance of the feature. The larger the Gini index, the higher the importance of the feature. The top 100 loci were selected as candidate markers.

[0137] 8. Using an integrated approach, we combined the methylation feature sites obtained by LASSO regression and random forest screening to obtain a total of 19 candidate marker sites.

[0138] The sites obtained through feature screening can characterize the characteristic sites with the highest importance in the methylation map of the esophageal cancer genome, greatly reducing the demand for site information and reducing the risk of overfitting of the classification model.

[0139] Based on scikit-learn, a classifier model was established using the Logistic regression algorithm. The methylation level data of the marker sites were taken as input, and the sample types were represented by 0 and 1 as the output of the model to train the model.

[0140] 9. Performance Evaluation

[0141] For the classifier model obtained from the two trainings, its prediction performance can be calculated. Figure 2 and Figure 3 As shown in the figure, the Logistic regression classifier model is used to distinguish different stage samples of esophageal cancer (stage 0-I and stage II-IV), and it shows good classification effect on both the training set and the test set. The results are as follows:

[0142]

[0143]

[0144] The AUC for stage 0-I samples is 0.8974, indicating that the model has a higher accuracy in identifying early samples; the AUC for stage II-IV samples is 0.9503, indicating that the model is more effective in identifying late samples.

[0145] Compared with other esophageal cancer markers, the marker combination obtained by the above screening method has a diagnostic sensitivity of over 80% for early esophageal cancer and over 85% for mid- and late-stage esophageal cancer, which is much higher than the existing esophageal cancer markers (the diagnostic sensitivity of early esophageal cancer is usually less than 50%).

[0146] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above specific embodiments. The above specific embodiments and the description in the specification are only for further illustrating the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the claims and their equivalents.

Claims

1. A method for screening esophageal cancer markers based on enzyme transfer methylation sequencing, characterized in that: The specific steps include: S1. Collect cancer tissues and paired adjacent tissue samples from patients with early-stage esophageal cancer and perform whole-genome methylation sequencing using a novel enzyme transfer method; S2. Integration of methylation data of samples from different sources, including collection of clinical esophageal cancer and adjacent tissue data and esophageal cancer methylation site information in public databases; S3. The differentially methylated sites were further filtered by combining the methylation levels of whole blood, leukocytes, and peripheral blood mononuclear cell samples. The sites that were easily interfered by plasma background methylation were eliminated based on the standard of high methylation in esophageal cancer tissue relative to adjacent tissues and low methylation in other samples such as leukocytes. S4. An integrated methylation feature screening model based on LASSO regression and random forest algorithm extracts key features from the genomic methylation map, obtains candidate methylation marker combinations, and achieves classification prediction of esophageal cancer using a relatively small number of methylation marker site features; S5. Construct a classifier model based on the candidate marker combination, train and test the model, and evaluate its predictive performance.

2. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11. Collect cancer tissue and its paired adjacent tissue samples from early-stage esophageal cancer patients, extract DNA from tissue samples and quantify it; S12, using enzyme conversion method to perform methylation conversion treatment on the extracted DNA sample; S13. Single-stranded DNA library construction technology was used to construct the methylation library of the transformed samples, including DNA denaturation, P7 adapter ligation, complementary chain synthesis, P5 adapter ligation, and PCR amplification, to avoid the introduction of non-methylated sites in the end repair process of traditional double-stranded library construction, which would have an adverse effect on data analysis.

3. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 1, characterized in that: In steps S1 and S2, the methylation whole genome sequencing data and chip data of samples from different sources need to be quality controlled and differentially methylated CpG sites screened; The methylation whole genome sequencing data was quality controlled under the following conditions: a. Remove the reads contaminated by the adapter; b. Truncate the reads with the first and last quality lower than 3 or containing N, where N means the base information cannot be determined; c. Remove the fragments 10bp upstream and downstream of the read segment; d. Cut off low-quality reads using a sliding window approach, with 4 bases as a window. If the average base quality value of the window is less than 15, the read is cut off from there; The screening conditions for the differentially methylated CpG sites are: methylation difference ≥ 25% and Q-value ≤ 0.01; The methylation chip data were subjected to data quality control and normalization, and the filtering conditions were as follows: a. Probes with probe detection P>0.01; b. Filter out probes with less than 3 beads in at least 5% of the samples; c. Filter out all non-CpG probes; d. Eliminate all SNP-related probes; e. Eliminate the probe on the sex chromosome; Differential methylation probes were screened based on methylation chip data, and the screening conditions were: |Δβ|≥0.25 and Q-value≤0.

01.

4. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 1, characterized in that: In step S3, the differentially methylated sites are further filtered, and the average methylation signal levels of the corresponding CpG sites in whole blood, leukocytes, PBMC and other samples in TCGA and GEO, i.e., the plasma background methylation signal levels, are analyzed, and the sites with high methylation in esophageal cancer tissues and plasma background methylation levels less than 0.1 are retained.

5. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 1, characterized in that: In step S4, the following steps are included: S4.

1. Construction of methylation feature fusion matrix: Based on methylation data from different sources, the methylation levels of CpG sites of all samples were obtained, data quality control and preprocessing were performed, and the methylation information of all CpG sites of each sample in the training set was fused to form a methylation feature fusion matrix; S4.

2. Perform LASSO regression analysis, including the following steps: S4.21, taking the methylation feature fusion matrix as input, and using LASSO regression for feature selection; S4.

22. Add L1 regularization term to the loss function to control model complexity and select important features; where: The loss function of the LASSO regression model is: Where ||w||1 represents the L1 norm of w, S4.

23. Use the gradient descent algorithm to iteratively find the most appropriate value; loss function gradient: Where α is the penalty term, sign(w) is the sign function, and the appropriate α value is set according to the actual number of sites, and the number of features whose coefficient converges to 0 is adjusted to perform feature screening; S4.

24. The model was iteratively optimized by gradient descent method to find the optimal regularization coefficient α and identify key methylation sites associated with esophageal cancer. S4.3, random forest algorithm application, including the following steps: S4.

31. Build a random forest model. Use the Bootstrap sampling method to build multiple sub-datasets from the training set. For each sub-dataset, build a decision tree and randomly select a subset of feature variables at each node for splitting. After the training is completed, calculate the Gini index of each feature to evaluate the importance of the feature. S4.

32. Feature importance assessment: score the importance of the features in each tree, using the Gini index difference calculation formula: Among them, GI m is the Gini index of the node before splitting, GI l and GI r is the Gini index of the left and right child nodes after splitting; The feature importance scores of all trees are normalized to obtain the final feature importance score: For the methylation feature sites, they were sorted in descending order according to their importance scores, and the feature sites with the highest ranking were selected for further screening. S4.4, Intersection feature selection, combining the characteristic sites screened out by LASSO regression and random forest algorithm respectively, taking the intersection of the two as the final candidate marker combination.

6. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 2, characterized in that: In step S5, based on the candidate marker combination, a classifier model is established using a Logistic regression algorithm, the data set is split into a training set and a validation set in proportion, the methylation level data of the candidate marker sites are used as input, the sample type is represented by 0 and 1 as the output of the model, the model is trained and tested, and its predictive performance is evaluated, and the evaluation indicators include sensitivity, specificity and AUC value.

7. The method for screening esophageal cancer markers based on enzyme-transferred methylation sequencing according to claim 6, characterized in that: The dataset was split into a training set and a validation set in a ratio of 7:

3.

8. The method for screening esophageal cancer markers based on enzyme transfer methylation sequencing according to claim 1, characterized in that: The different databases include esophageal cancer tissue and adjacent tissue data in TCGA and GEO databases and esophageal cancer-related DMR information in the MethMarkerDB database.

9. The method for screening esophageal cancer markers based on enzyme-transferred methylation sequencing according to claim 2, characterized in that: The methylation conversion treatment includes using TET methylcytosine dioxygenase 2 and T4 bacteriophage β-glucosyltransferase to perform an oxidation reaction and using cytosine deaminase to perform a deamination reaction.

10. The method for screening esophageal cancer markers based on enzyme-transferred methylation sequencing according to claim 9, characterized in that: NovaSeq 6000 was used for paired-end 150 bp read length sequencing with a sequencing depth of 30×.

Images