Cancer risk prediction method and system

Abstract

The invention discloses a cancer risk prediction method and system, and relates to the technical field of biological information. A construction method of a cancer risk prediction model involved in the cancer risk prediction method comprises the following steps: S1, acquiring a CpG cluster conforming to expectation on a whole genome based on a methylation probe in a methylation chip; s2, grouping according to the tumor purity of the tumor tissue sample to be analyzed; s3, identifying cancer specific CpG clusters (DPCC) of each group according to a difference mark screening method; s4, generating a reference fragment of the DPCC; s5, simulating the methylation state of the reference fragment, and generating a pseudo fragment database; and S6, training a fragment dichotomy model based on DNA methylation and sequence features of the pseudo fragment data. The method is suitable for analysis of various cancer types, and can process different DNA methylation sequencing data.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a cancer risk prediction method and system. Background Art

[0002] At present, malignant tumors have become one of the most serious diseases that endanger human health. Most cancers have no obvious symptoms in the early stages, and when they are discovered, they are already in the late stages, missing the best treatment period. If the tumor is discovered and intervened in the early stages, it can effectively reduce the mortality rate of cancer and prolong survival. Early screening, early diagnosis, and early treatment of tumors are the key to tumor prevention and treatment, and can effectively change the severe situation of tumor diagnosis and treatment. At present, the few available biomarkers have limited effects on early diagnosis, analysis, and postoperative follow-up of tumors. We still need to further find biological methods that are minimally invasive, easy to detect, and highly sensitive and specific.

[0003] Terminology explanation:

[0004] High-throughput sequencing technology: High-throughput sequencing technology is also called "next generation" sequencing technology (Next generationsequencing, NGS) or deep sequencing. It can sequence hundreds of thousands to millions of DNA molecules in parallel at one time, and the read length is generally short. According to the development history, influence, sequencing principles and technologies, it mainly includes Massively Parallel Signature Sequencing (MPSS), Polony Sequencing, 454 pyrosequencing, Illumina (Solexa) sequencing, ABI SOLiD sequencing, Ion semiconductor sequencing, DNA nanoball sequencing, etc.

[0005] DNA methylation: DNA methylation is an epigenetic modification that is usually catalyzed by DNA methyltransferase enzymes (DNMTs), using S-adenosylmethionine as a methyl donor to transfer a methyl group to the 5th carbon atom (5C) of cytosine in genomic DNA to form 5' methylcytosine. In mammals, DNA methylation mainly occurs on cytosine-phosphate-guanine (CpG) dinucleotides. The DNA methyltransferase family consists of three highly conserved proteins: DNMT1, DNMT3A, and DNMT3B. DNMT3A and DNMT3B are responsible for de novo methylation of unmethylated DNA duplexes, while DNMT1 maintains methylation during semiconservative DNA replication. DNMT1 performs methylation modification at the corresponding position on the newly synthesized chain based on the specific methylation site on the parent chain, which is the basis for the inheritance and clinical application of DNA methylation.

[0006] 450K Methylation Chip (Illumina Infinium HumanMethylation450 BeadChip): 450K Methylation Chip is a high-throughput DNA methylation analysis platform. It can simultaneously detect the methylation status of approximately 450,000 CpG sites across the entire genome. By covering gene promoter regions, CpG islands, intergenic regions, and other functional regions in the genome, the 450K chip provides researchers with a comprehensive DNA methylation map.

[0007] Methylated CpG short tandem amplification sequencing (MCTA sequencing, MCTA-Seq): MCTA-Seq is a highly sensitive technology used to detect the DNA methylation status of specific genomic regions. The principle is to first enrich the CpG regions of interest through targeted capture, then amplify these regions using PCR, and analyze the methylation status of each CpG site through high-throughput sequencing. MCTA-Seq is widely used in cancer research and epigenetic research to help identify disease-related methylation markers and gain a deeper understanding of gene expression regulation mechanisms.

[0008] Whole Genome Bisulfite Sequencing (WGBS): WGBS is a high-throughput DNA sequencing technology used to comprehensively study the pattern and distribution of DNA methylation. WGBS is mainly used to analyze the methylation status of cytosine bases in DNA, which is crucial for understanding processes such as gene expression regulation, cell differentiation, and disease occurrence.

[0009] cfDNA (cell-free DNA): cfDNA is a DNA fragment that exists in the human blood circulation and is free from cells. These DNA fragments mainly originate from the process of cell apoptosis or necrosis and exist in the blood circulation in the form of small fragments. cfDNA has important clinical applications in oncology, because cfDNA released by tumor cells can reflect the genomic characteristics of the tumor, which can be used for non-invasive cancer detection and monitoring. Analyzing cfDNA through high-sensitivity sequencing technology can achieve early detection of the disease, evaluation of treatment effects and recurrence monitoring.

[0010] Stratified sampling: It is a sampling method that keeps the distribution of each stratum (or category) in the sample subset similar to that of each stratum in the overall sample set.

[0011] Segment Binary Classification Model: The segment binary classification model is a machine learning model whose purpose is to classify an instance into one of two mutually exclusive categories given a set of input features. The model assigns each input sample to one of two possible categories (e.g., tumor tissue and non-tumor tissue) by learning patterns and features in the training data.

[0012] Deep learning model: A deep learning model is a machine learning method based on artificial neural networks, usually composed of multiple layers (i.e., "depth") of neurons. These layers can automatically learn and extract features from large amounts of data without relying on hand-designed rules. By abstracting and combining simple features layer by layer, deep learning models are able to recognize complex patterns and structures and are widely used in image recognition, natural language processing, speech recognition and other fields. In biomedicine, deep learning models are used to analyze complex biological data, such as genomic data, to identify key features of diseases.

[0013] “Not less than” means higher than or equal to, and should be understood as including the number itself. Summary of the invention

[0014] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method for constructing a cancer risk prediction model.

[0015] The present invention also provides a method for predicting cancer risk.

[0016] The present invention also provides a cancer risk prediction system.

[0017] The invention also provides a computer device.

[0018] The present invention also provides a computer-readable storage medium.

[0019] A cancer risk prediction method according to a first aspect of an embodiment of the present invention comprises the following steps:

[0020] S1. Based on the methylation probes in the methylation chip, the expected CpG clusters in the whole genome are obtained and recorded as PCC;

[0021] S2. grouping according to the tumor purity of the tumor tissue samples to be analyzed, wherein the tumor tissue samples to be analyzed have methylation probe detection data;

[0022] a1) when a tumor tissue sample has a paired adjacent paracancerous tissue sample, calculating the tumor purity of the tumor tissue sample, and selecting a desired number of the tumor tissue samples as a PH group;

[0023] a2) when the tumor tissue sample does not have a paired adjacent cancer tissue sample, calculating the tumor purity of the tumor tissue sample, sorting the samples according to the tumor purity, and selecting a desired number of tumor tissue samples with the highest tumor purity as the TH group;

[0024] a3) When the tumor tissue sample does not have a paired adjacent cancer tissue sample, the tumor purity of the tumor tissue sample is calculated, the samples are sorted according to the tumor purity, and the whole sample is stratified and sampled to obtain the desired number of samples as the MH group;

[0025] S3, identifying cancer-specific CpG clusters in the PCC, denoted as DPCC;

[0026] For the PH group, the first methylation state difference of the PCC between each paired tumor tissue sample and adjacent paracancerous tissue sample in the PH group and the second methylation state difference of the PCC between the tumor tissue sample and healthy peripheral blood are calculated, and the PCC whose first methylation state difference and the second methylation state difference are greater than the first expected methylation state difference threshold are screened as DPCC;

[0027] For the TH group and / or MH group: calculating the third methylation state difference between the PCC in the tumor tissue sample and the healthy peripheral blood, and screening the PCC whose third methylation state difference is greater than the second expected methylation state difference threshold as DPCC;

[0028] S4, generating a reference segment of the DPCC:

[0029] The methylation status of the DPCC in the tumor tissue sample is obtained. Given the sequencing depth and the generated fragment length, the base composition of the DPCC is obtained. After a start site is generated by a randomizer, a reference fragment covering the DPCC is generated according to the fragment length. The number of generated DNA fragments S is calculated using the following formula:

[0030]

[0031] In the formula, D is the sequencing depth, R is the DPCC length, and L is the length of the generated DNA fragment;

[0032] S5. Simulate the methylation status of the reference fragment and generate a pseudo-fragment database:

[0033] Fill the methylation status of the reference fragment obtained in step S4 into a matrix with a length of PCC and a width of the number of fragments to simulate the methylation status; and distinguish and mark the methylated CpG sites from the remaining sites according to the methylation status of the DPCC in the tumor tissue sample to obtain pseudo-fragment data and generate a pseudo-fragment database of the DPCC;

[0034] S6. Training fragment binary classification model based on DNA methylation and sequence features of pseudo-fragment data:

[0035] The fragment information located at the DPCC in the healthy sample is obtained to obtain a control fragment database; the pseudo-fragment database and the control fragment database are input into the training deep learning model to construct a fragment binary classification model, thereby obtaining the cancer risk prediction model.

[0036] The construction method according to the embodiment of the present invention has at least the following beneficial effects:

[0037] The construction method of the embodiment converts whole genome methylation chip data into pseudo-fragment data similar to WGBS sequencing data, which greatly reduces the cost of data acquisition and increases the amount and diversity of training data, so that relevant DNA methylation features can be effectively identified in the diagnosis of cancers such as liver cancer, gastric cancer and pancreatic cancer, and improves the accuracy, robustness and generalization ability of the model.

[0038] The construction method of the embodiment is applicable to the analysis of various cancer types (such as liver cancer, gastric cancer, pancreatic cancer, etc.), and also has the ability to process various DNA methylation sequencing data (such as WGBS and MCTA, etc.), showing adaptability to data generated from different sources, different technical platforms, and overcoming the problems of cancer heterogeneity and specificity. It does not require a lot of specific tuning or retraining, ensuring wide applicability and reliability. The obtained model is more valuable in practical applications, can adapt to various sequencing scenarios and data sets, and has great application potential in the medical field.

[0039] The model constructed by this construction method can not only distinguish cancer patients from healthy controls, but also make accurate classifications according to the stage of cancer progression, and is more patient-friendly. The model showed high AUC values ​​in the classification of different cancer types (such as liver cancer, gastric cancer, and pancreatic cancer, etc.), especially the low methylation features in peripheral blood, which showed the potential of this method in non-invasive cancer diagnosis. It can be applied to a wide range of DNA methylation sequencing data (such as MCTA sequencing data), and has good tumor-derived DNA methylation fragment recognition and cancer diagnosis effects.

[0040] According to some embodiments of the present invention, in step S1, by a sliding window method, if the first flanking region of the first methylation probe overlaps with the second flanking region of the second methylation probe, the DNA sequences corresponding to the first flanking region and the second flanking region on the reference genome are classified into one CpG cluster; and the CpG cluster that meets the expectation is selected, namely, PCC;

[0041] The first flanking region includes the first methylation probe and DNA sequences of set window sizes upstream and downstream thereof;

[0042] The second flanking region includes the second methylation probe and DNA sequences with set window sizes upstream and downstream thereof.

[0043] According to some embodiments of the present invention, the desired CpG cluster satisfies the following conditions:

[0044] b1) There are at least 3 CpG sites covered by methylation probes;

[0045] b2) capable of providing methylation signals for calculating methylation status;

[0046] b3) contains the portion spanning the gene functional region.

[0047] According to some embodiments of the present invention, the flanking region of the methylation probe is confirmed based on a window size set by a sliding window method, wherein the window size set by the sliding window method is 80 bp to 120 bp, for example, 100 bp.

[0048] According to some embodiments of the present invention, the methylation chip includes at least one of Infinium Human Methylation450 (450K), Methylation EPIC v1.0 BeadChips (850k), and Infinium MethylationEPIC v2.0BeadChip (935k).

[0049] According to some embodiments of the present invention, in step S2, those skilled in the art may select a desired number of tumor tissue samples for further analysis according to computing power. The desired number is not less than 10. For example, it may be 30 or 50.

[0050] According to some embodiments of the present invention, in step S3, for the PH group, the desired PCC satisfies the following conditions:

[0051] c1) the expected number of the first methylation state difference and the second methylation state difference is greater than the first expected methylation state difference threshold;

[0052] c2) The difference between the first median of the methylation state of the methylation region in the PCC that meets the condition shown in c1 and the second median of the methylation state in the healthy peripheral blood is greater than the first expected methylation state difference threshold.

[0053] According to some embodiments of the present invention, in c1), those skilled in the art can set the expected number as needed, such as: the expected number is at least 50% of the paired samples in the PH group.

[0054] According to some embodiments of the present invention, for the source data of the methylation status of PCC in the tumor tissue sample, paracancerous tissue sample, and healthy peripheral blood:

[0055] If it is methylation chip data, the methylation status of the probe in the PCC is obtained by fluorescence intensity and its mean is calculated, which is the methylation status of the PCC;

[0056] If it is sequencing data, the methylation status of a single CpG site covered in the PCC is calculated first, and then the average of the methylation status of all CpG sites covered by the PCC is calculated, which is the methylation status of the PCC; the methylation status of a single CpG site refers to the average of the methylation status of the CpG sites of multiple DNA fragments.

[0057] According to some embodiments of the present invention, in step S4, the reference fragment includes its position information, base composition and methylation status in the reference genome.

[0058] According to some embodiments of the present invention, those skilled in the art may use different indicator symbols as needed to achieve the following purposes:

[0059] In step S4, the C base of the CpG site is distinguished from the other bases;

[0060] And / or, in step S5, the methylated CpG sites are distinguished from the remaining sites.

[0061] According to some embodiments of the present invention, the indicator includes but is not limited to numbers, letters or graphics. For example: in step S4, the C base of the CpG site is marked as 2, and the remaining bases are marked as 0; in step S5, the C base of the methylated CpG site is marked as 1, and the remaining bases are marked as 0.

[0062] According to some embodiments of the present invention, in step S5, the pseudo-fragment database includes position information, base composition and methylation signal data of the reference fragment.

[0063] According to some embodiments of the present invention, in step S6, the fragment information located in the DPCC interval in the healthy sample includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome.

[0064] According to some embodiments of the present invention, in step S6, training the segment binary classification model includes at least one round (for example, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 or 13 rounds may be repeated), and each round includes the following steps:

[0065] d1) randomly dividing the control fragment database and the pseudo fragment database into two parts, respectively used for training and verification of the fragment binary classification model (e.g., 75% of the samples are used for training and 25% of the samples are used for verification);

[0066] d2) Encoding the DNA fragments into a one-hot matrix of the fragment length * 5 in step S4 as the input matrix of the model to train the deep learning model, and using the sigmoid function as the last layer activation function of the model to obtain the cancer origin probability of each fragment; establishing the fragment binary classification model.

[0067] According to some embodiments of the present invention, step S6 further includes: obtaining information on the fragments located in the DPCC in the peripheral blood samples of tumor patients and / or healthy peripheral blood samples, inputting the fragments into the binary classification model to obtain the cancer source probability of all fragments and calculating the maximum posterior probability In order to evaluate the prediction effect of cancer risk, a fragment binary classification model that meets the expectations is obtained, which is the cancer risk prediction model.

[0068] A cancer risk prediction method according to a second aspect of an embodiment of the present invention comprises the following steps:

[0069] E1) constructing the cancer risk prediction model according to the construction method described in the embodiment of the first aspect of the present invention;

[0070] E2) Obtaining the fragment information located in the DPCC in the peripheral blood sample to be predicted, inputting it into the cancer risk prediction model, and calculating the maximum posterior probability That is, the cancer risk index of the peripheral blood sample to be predicted.

[0071] According to some embodiments of the present invention, in step E2), the fragment information located in the DPCC interval in the peripheral blood sample to be predicted includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome.

[0072] According to some embodiments of the present invention, in step E2), the data input into the cancer risk prediction model also includes a fragment information package of a healthy peripheral blood sample located in the DPCC interval. The fragment information located in the DPCC interval includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome.

[0073] According to some embodiments of the present invention, the fragment information located in the DPCC interval in the peripheral blood sample to be predicted includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome.

[0074] According to some embodiments of the present invention, sequencing data of the peripheral blood sample to be predicted is obtained by sequencing, and the fragment information located in the DPCC interval in the peripheral blood sample to be predicted is obtained by the samtools tool. The sequencing method is selected from at least one of WGBS and MCTA.

[0075] According to some embodiments of the present invention, the greater the cancer risk index, the greater the risk of acquiring cancer.

[0076] According to some embodiments of the present invention, the prediction method is for non-disease diagnosis and treatment purposes.

[0077] According to some embodiments of the present invention, the cancer includes at least one of liver cancer, gastric cancer, and pancreatic cancer.

[0078] According to some embodiments of the present invention, the maximum a posteriori probability The calculation formula is as follows:

[0079]

[0080] Where, d i It refers to the probability that each DNA fragment comes from the tumor, n refers to the number of DNA fragments, and r refers to the proportion of DNA fragments derived from the tumor in the peripheral blood sample to be predicted.

[0081] A cancer risk prediction system according to a third aspect of an embodiment of the present invention includes:

[0082] A PCC acquisition module, used to acquire CpG clusters on the whole genome based on the methylation probes in the methylation chip, namely the PCC;

[0083] A grouping module, grouping according to the tumor purity of the tumor tissue samples to be analyzed, wherein the tumor tissue samples to be analyzed have methylation probe detection data; a1) when the tumor tissue samples have paired adjacent cancer tissue samples, calculating the tumor purity of the tumor tissue samples, and selecting a desired number of the tumor tissue samples as the PH group; a2) when the tumor tissue samples do not have paired adjacent cancer tissue samples, calculating the tumor purity of the tumor tissue samples, sorting them according to the tumor purity, and selecting a desired number of the tumor tissue samples with the highest tumor purity as the TH group; a3) when the tumor tissue samples do not have paired adjacent cancer tissue samples, calculating the tumor purity of the tumor tissue samples, sorting them according to the tumor purity, and performing stratified sampling on the overall samples to obtain a desired number of samples as the MH group;

[0084] A DPCC identification module is used to identify cancer-specific CpG clusters in the PCC to obtain the DPCC; for the PH group, a first methylation state difference of the PCC between each paired tumor tissue sample and adjacent tissue sample in the PH group and a second methylation state difference of the PCC between the tumor tissue sample and healthy peripheral blood are calculated, and PCCs whose first methylation state difference and second methylation state difference are greater than a first expected methylation state difference threshold are screened as DPCCs; for the TH group and / or MH group: a third methylation state difference of the PCC in the tumor tissue sample and healthy peripheral blood is calculated, and PCCs whose third methylation state difference is greater than a second expected methylation state difference threshold are screened as DPCCs;

[0085] The reference fragment generation module is used to obtain the methylation status of the DPCC in the tumor tissue sample, obtain the base composition of the DPCC given the sequencing depth and the generated fragment length, generate a starting site through a randomizer, generate a reference fragment covering the DPCC according to the fragment length, and calculate the number of generated fragments S, the calculation formula is as follows:

[0086]

[0087] In the formula, D is the sequencing depth, R is the DPCC length, and L is the length of the generated DNA fragment;

[0088] A pseudo-fragment database generation module is used to fill the methylation status of the reference fragment into a matrix with a length of PCC and a width of the number of fragments to simulate the methylation status; and according to the methylation status of the DPCC in the tumor tissue sample, distinguish and mark the methylated CpG sites from the remaining sites to obtain pseudo-fragment data, and generate a pseudo-fragment database of the DPCC;

[0089] A model generation module is used to obtain the fragment information located at the DPCC in the healthy sample to obtain a control fragment database; the pseudo-fragment database and the control fragment database are input into the training deep learning model to construct a fragment binary classification model;

[0090] The cancer risk index prediction module is used to obtain the fragment information located in the DPCC in the peripheral blood sample to be predicted, input it into the cancer risk prediction model, obtain the cancer source probability of all fragments and calculate the cancer risk index.

[0091] According to some embodiments of the present invention, the cancer risk prediction system is used to implement the cancer risk prediction method described in the embodiment of the second aspect of the present invention.

[0092] A computer device according to an embodiment of the fourth aspect of the present invention comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the construction method described in the embodiment of the first aspect of the present invention or the cancer risk prediction method described in the embodiment of the second aspect of the present invention.

[0093] According to a computer-readable storage medium of the fifth aspect of the present invention, a computer program is stored. When the computer program is executed by a processor, the processor executes the steps of the construction method described in the embodiment of the first aspect of the present invention or the cancer risk prediction method described in the embodiment of the second aspect of the present invention.

[0094] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0096] Figure 1 Schematic diagram of the process of predicting cancer risk based on DNA methylation of the present invention;

[0097] Figure 2 The schematic diagram of the cancer risk prediction method based on DNA methylation of the present invention;

[0098] Figure 3 It is a flow chart of a training fragment binary classification model of the cancer risk prediction method based on DNA methylation of the present invention;

[0099] Figure 4 This is a schematic flow chart of model training and testing in Application Example 1;

[0100] Figure 5 The whole genome methylation landscape of cancer tissue, peripheral blood and healthy human peripheral blood of the liver cancer group in Application Example 1; Figure A shows the distribution of PCC on the whole genome, and Figure B shows the distribution of the variation of methylation status calculated on each PCC for different groups (measured by the region-specific standard deviation (SD));

[0101] Figure 6 This is the analysis of the whole genome LIHC specific markers in Application Example 1;

[0102] Figure 7 This is the methylation status of the different PCCs in different types of samples in Application Example 1;

[0103] Figure 8 The results of the analysis of all tumor and healthy cfDNA samples on co-DPCC in Application Example 1; Figure A is the result of principal component analysis (PCA), Figure B is the result of UMAP analysis, and Figure C is the result of t-distributed stochastic neighbor embedding (t-SNE) analysis;

[0104] Fig. 9 This is the GO enrichment analysis result of co-DPCC in Application Example 1;

[0105] Fig.10 The KEGG enrichment analysis results of co-DPCC in Application Example 1;

[0106] Fig.11 The performance of the two-classification model of different groups of fragments in Application Example 1 on the training set and the validation set;

[0107] Fig.12 The performance of the low methylation features in application example 1 in different classification tasks and cohorts;

[0108] Fig.13 The sensitivity of the fragment binary classification model for different groups in Application Example 1 for early cancer diagnosis;

[0109] Fig.14 The cancer risk scores of different types of samples estimated by the PH_02 model in Application Example 1;

[0110] Fig.15 The results of the correlation analysis between tumor size and cancer risk score in Application Example 1;

[0111] Fig.16 Cancer prediction scores for two liver cancer patients sampled at different time points in Application Example 1;

[0112] Fig.17 The results of the analysis of various classification evaluation indicators such as accuracy (ACC), F1 score, KAPPA, receiver operating characteristic (ROC) curve, positive predictive value (PPV), negative predictive value (NPV), sensitivity (SEN), specificity (SPE) and total proportion agreement (TPA) of different groups of models in application example 1 in WGBS data;

[0113] Fig.18 The performance of different groups of models in Application Example 1 in the MCTA dataset; Figures A and B are high methylation features, and Figures C and D are low methylation features;

[0114] Fig.19 Comparison of the diagnostic performance of the method of the present invention in Application Example 1 and other cancer blood markers;

[0115] Fig. 20 This is the discovery of specific markers for gastric cancer in the whole genome in Application Example 2; Figures A and B are the common high-methylation and low-methylation features identified by the two schemes, and Figure C is the genomic annotation of the identified common features;

[0116] Fig.21 The classification effects of different groups of models in application example 2 in gastric cancer;

[0117] Fig. 22 The cancer risk scores predicted by the TH model for gastric cancer samples and healthy control samples in Application Example 2;

[0118] Fig.23 The correlation between clinical information of gastric cancer and cancer risk score in application example 2;

[0119] Fig.24 For the discovery of genome-wide pancreatic cancer-specific markers in Application Example 3;

[0120] Fig.25 The classification effect of different groups of models in application example 3 in pancreatic cancer;

[0121] Fig.26 This is the cancer risk score predicted by the TH model for pancreatic cancer samples and healthy control samples in Application Example 3. DETAILED DESCRIPTION

[0122] The following will be combined with the embodiments to clearly and completely describe the concept of the present invention and the technical effects produced, so as to fully understand the purpose, characteristics and effects of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0123] If the specific conditions are not specified in the examples, the experiments were carried out under conventional conditions or conditions recommended by the manufacturer. If the manufacturers of the reagents or instruments are not specified, they are all conventional products that can be purchased commercially.

[0124] In the description of the present invention, if there is a description of first, second, third, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the sequence of the indicated technical features.

[0125] When a numerical range is disclosed herein, the above range is considered to be continuous and includes the minimum and maximum values ​​of the range, as well as every value between such minimum and maximum values. Further, when a range refers to an integer, every integer between the minimum and maximum values ​​of the range is included. In addition, when multiple ranges are provided to describe features or characteristics, the ranges can be combined. In other words, unless otherwise indicated, all ranges disclosed herein should be understood to include any and all subranges included therein.

[0126] "And / or" is used to indicate that one or both of the stated situations may occur, for example, A and / or B includes (A and B) and (A or B).

[0127] Unless otherwise specified, "DeepLPseudo" refers to a model constructed using the cancer risk prediction method based on DNA methylation of the present invention.

[0128] Example 1 (A method for predicting cancer risk based on DNA methylation)

[0129] See also Figure 1 In this embodiment, a cancer risk prediction method based on DNA methylation is provided, wherein steps S1 to S6 are used to construct and evaluate a cancer risk prediction model, and step S7 uses the constructed cancer risk prediction model to predict and evaluate cancer risk. Specifically, the steps include:

[0130] S1. Define potential CpG clusters (PCC) on the whole genome:

[0131] Potential CpG clusters were defined genome-wide based on the CpG sites of the methylation probes in the methylation array using a 100 bp sliding window.

[0132] Specifically, for a methylation probe, the methylation probe and the 100 bp DNA sequences upstream and downstream thereof are defined as flanking regions, and it is assumed that all CpG sites in the flanking regions have the same average methylation status.

[0133] If the flanking regions of the two methylation probes overlap, the two flanking regions are grouped into one CpG cluster. If the undetermined CpG cluster meets the following conditions, it can be used as a PCC for subsequent analysis:

[0134] ① There are at least 3 CpG sites covered by methylation probes;

[0135] ② Able to provide sufficient methylation signals to facilitate the calculation of methylation status;

[0136] ③Contains parts that span the functional range of the gene (such as methylation islands).

[0137] S2. Grouping according to the tumor purity of the tumor tissue samples to be analyzed with methylation probe detection data:

[0138] Due to tumor heterogeneity, the number of tumor tissue samples in the PH, TH, and MH groups should be no less than 10 to ensure sample diversity.

[0139] a1) When tumor tissue samples have paired adjacent paracancerous tissue samples, the InfiniumPurify R package is used to calculate the tumor purity of the tumor tissue samples, and a samples are randomly selected as the PH group.

[0140] Those skilled in the art may adjust the value of a according to actual needs and computing capabilities, such as adjusting a to 50.

[0141] a2) When the tumor tissue sample does not have a paired adjacent adjacent tissue sample, the InfiniumPurify R package is used to calculate the tumor purity of the tumor tissue sample, and all samples are sorted according to the tumor purity, and the b samples with the highest tumor purity are selected as the TH group.

[0142] Those skilled in the art may adjust the value of b according to actual needs and computing capabilities, such as adjusting b to 30.

[0143] a3) When the tumor tissue sample does not have a paired adjacent adjacent tissue sample, the InfiniumPurify R package is used to calculate the tumor purity of the tumor tissue sample, and stratified sampling is performed to obtain c samples as the MH group. The distribution of the tumor purity of the c samples is similar to that of all samples.

[0144] Those skilled in the art may adjust the value of c according to actual needs and computing capabilities, such as adjusting c to 30.

[0145] S3. Identification of cancer-specific CpG clusters (DPCC):

[0146] For the methylation chip, the methylation status of the probe in the PCC is obtained by the fluorescence intensity and the mean is calculated, which is the methylation status of the PCC;

[0147] For sequencing data, the methylation status of a single CpG site covered in PCC (referring to the average of the methylation status of CpG sites in multiple DNA fragments) is first calculated, and the average of the methylation status of all covered CpG sites is calculated to represent the methylation status of PCC.

[0148] (1) For PH group:

[0149] Define the minimum cutoff value (cutoff, i.e., the expected methylation status difference threshold) of the methylation status difference between tumor tissue and adjacent tissue. Calculate the methylation status difference of each PCC between each paired tumor tissue sample and adjacent tissue sample in the PH group and between tumor tissue sample and healthy peripheral blood sample, and screen out PCCs that meet the following two conditions, i.e., DPCCs.

[0150] ① The PCC with a methylation status difference greater than the cutoff in more than 50% of paired samples is FDMR;

[0151] Paired samples refer to "tumor tissue samples and adjacent tissue samples".

[0152] ② The median difference in methylation status between PCC (FDMR) with differential methylation status between tumor tissue and adjacent tissue is greater than the cutoff.

[0153] (2) For TH group / MH group:

[0154] The methylation status difference of each PCC in each tumor tissue sample and healthy peripheral blood was calculated, and the PCC with a methylation status difference greater than the cutoff was defined as DPCC.

[0155] S4. Generate reference fragment of DPCC:

[0156] The methylation status of DPCC in tumor tissue samples is obtained from the chip data. Given the sequencing depth and the length of the generated DNA fragment, the base composition (i.e., sequence information) of DPCC is obtained from the reference genome. After a randomizer generates the starting site, a reference fragment covering the DPCC is generated according to the length of the generated DNA fragment. The reference fragment contains its position information, base composition, and initialized methylation status in the reference genome (the C base of the CpG site is recorded as 2, and the rest of the sites are recorded as 0). The number of generated DNA fragments (S) is calculated as follows:

[0157]

[0158] Where D is the sequencing depth, R is the DPCC length, and L is the length of the generated DNA fragment.

[0159] The sequencing depth can be selected by a person skilled in the art according to the data volume requirement and data processing capability. The deeper the sequencing depth, the greater the data volume.

[0160] The length of the generated DNA fragment can be determined by a person skilled in the art as needed. For example, the sequencing fragments of peripheral blood cfDNA have two modes: 75 bp and 150 bp. When the lowest common divisor 75 is selected, the length of the generated DNA fragment can be set to 66 after removing the 9 bp at the head, taking into account the length of the connector.

[0161] S5. Simulate the methylation status of the reference fragment and generate a pseudo-fragment database:

[0162] The methylation status of the reference fragment obtained in step S4 is filled into a matrix with a length of PCC and a width of the number of fragments to simulate the methylation status. Based on the assumption that similar CpG sites have similar methylation states, the methylation states of CpG sites in the same PCC are consistent. According to the methylation status of DPCC in the tumor tissue sample, the labels of the corresponding number of sites are changed, wherein the unmethylated CpG sites are marked as 2 and the methylated CpG sites are marked as 1; then the label 2 of the CpG site is replaced with 0, and the methylation signal of the final fragment only retains the methylated CpG site 1 and the remaining sites 0 as subsequent inputs to obtain the pseudo fragment data (pseudo fragment) of the DPCC. Finally, a pseudo fragment database of DPCC is generated (including the position information, base composition and methylation signal (0 / 1) of each reference fragment). An exemplary process is as follows Figure 1 shown.

[0163] S6. Training fragment binary classification model based on DNA methylation and sequence features of pseudo-fragment data:

[0164] The sequencing data of healthy volunteer samples (such as WGBS data of healthy human cfDNA) were analyzed, and the position information, base composition and methylation signal (0 / 1) of the sequenced fragments in the reference genome were obtained by the samtools tool, and the fragment information located in the DPCC interval was screened to obtain a control fragment database. The control fragment database was randomly divided into two groups, one containing 75% of the fragments, which formed a training set with the tumor pseudo-fragment database obtained in step S5; and one containing 25% of the fragments, which were used to form a validation set with other data in step S7. 75% of the data in the training set were used for training, and 25% of the data were used for validation. Each fragment in the control fragment database and the pseudo-fragment database was encoded into a one-hot matrix of fragment length (L, the same as the DNA fragment length generated in step S4)*5 as the input matrix of the model to train the deep learning model. The sigmoid function was used as the last layer activation function of the model to obtain a fragment binary classification model.

[0165] S7. Predicting the cancer risk index (Cancer Scores) of the peripheral blood sample to be predicted:

[0166] The sequencing data of peripheral blood samples of tumor patients were obtained by sequencing (WGBS / MCTA), and the true fragment data (including the location information of the fragment, base composition and methylation signal (0 / 1)) in the DPCC interval were obtained by the samtools tool, and the 25% fragments obtained by grouping in step S6 constituted the validation set. The data was input into the fragment binary classification model constructed in step S6, and the posterior probability d-scores (i.e. P) of each fragment was output. The maximum posterior probability was used to calculate the proportion of DNA fragments derived from tumor tissue in the peripheral blood sample as the cancer risk index (Cancer Scores) of the sample.

[0167] When the proportion of DNA fragments derived from tumors in the peripheral blood sample to be predicted is given as r, and assuming that the dscore of each read is independent, the calculation formula of P is as follows:

[0168]

[0169] Where, d i It refers to the probability that each DNA fragment comes from the tumor, n refers to the number of DNA fragments, and r refers to the proportion of DNA fragments derived from the tumor in the peripheral blood sample to be predicted.

[0170] When P is maximized, the estimated ratio of tumor-derived reads (denoted as That is, the cancer risk index (Cancer Scores) is used to evaluate the risk of cancer in the subject whose peripheral blood sample is to be predicted).

[0171]

[0172] To avoid information leakage, the normal control cfDNA sample cohort was randomly divided into 75% and 25%, respectively, for the training of the fragment binary classification model and the evaluation of the effect of cancer risk index prediction. This process was repeated 10 times, that is, steps S6 and S7 were repeated 10 times. The best performing model was selected as the final fragment binary classification model.

[0173] The performance of the model classification samples was evaluated by the reportROC package. According to the comparison between the predicted results and the actual results, indicators such as ACC, F1 score, KAPPA, ROC curve, PPV, NPV, SEN, SPE and TPA were used. The ROC curve can comprehensively evaluate the sensitivity and specificity of the method. The area under the ROC curve (AUC) is often used to measure the performance of the classifier. The value range of AUC is 0.5 to 1. 0.5 means that the performance of the classifier is the same as random guessing, and 1 means a perfect classifier. According to different classification samples, the performance of different models in distinguishing between cancer and healthy people (Tumor vs healthy, TvsH), cancer and non-cancer patients (Tumor vs Non-Tumor, TvsNT) and cancer and non-cancer disease patients (Tumor vs Non-Tumor Disease, TvsNTD) can be evaluated.

[0174] Application Example 1

[0175] This application example uses liver cancer as an example to demonstrate the effectiveness of the cancer risk prediction method based on DNA methylation in Example 1.

[0176] The 450K methylation chip data of liver cancer (TCGA-LIHC) were downloaded from the TCGA database, with a total of 377 samples, and the sample type was tissue samples, of which 50 samples were paired adjacent tissues of tumor tissues of liver cancer patients. Three WGBS datasets (EGAS00001000566, CRA1001537 and CRA001142) were downloaded from the European Genome-Phenome Archive Database (EGA) and the National Genome Data Center (NGDC), and one MCTA dataset (GSE63775) was downloaded from the GEO database.

[0177] Nine healthy volunteers were recruited from Hospital A, blood samples (12 mL / person) were collected, and clinical information was collected. All recruited subjects signed informed consent forms, and the project has passed the ethical application. cfDNA was extracted from blood samples of healthy volunteers. Using the Acegen Bisulfite-Seq Library Prep Kit, 1 μg of cfDNA was taken and spiked with 1 ng of unmethylated Lambda DNA for ultrasonic shearing to a main band range of 200 bp to 500 bp. 100 ng to 500 ng was taken for end repair, adding "A" and connecting methylated illumina adapters. The ligation product was converted to sulfite and amplified with double-end 8 bp index primers by no more than 10 PCRs to obtain the amplified library product. The amplified library product was purified by 0.85×AMPure XP, and then the fragment range and effective concentration of the library were detected by Agilent 2100 and qPCR. Qualified libraries were sequenced on Illumina HiSeq X10 or Nova6000 for DNA library sequencing, and the sequencing strategy selected double-end 150 bp sequencing. The obtained data set is recorded as Inhouse1.

[0178] The relevant information of each data is shown in Table 1.

[0179] Table 1

[0180]

[0181]

[0182] The risk of liver cancer is predicted according to the "a method for predicting cancer risk based on DNA methylation" in the above Example 1. The analysis parameters are set as follows:

[0183] In step S1, the methylation chip is a 450K chip probe (Infinium Human Methylation 450 (450K)), and the whole genome is the reference genome GRCh38.

[0184] In step S2, the tumor tissue samples to be analyzed are cancer tissue samples in TCGA-LIHC; the samples in the PH group are 50 cancer tissue samples with adjacent adjacent cancer tissues; the samples in the TH group are the top 30 samples in terms of tumor purity among the 325 cancer tissue samples; since the tumor purity distribution of the 50 liver cancer tissue samples in the PH group is similar to that of the stratified sampling samples, the 50 tumor tissues in the PH group are directly used as the tumor tissue samples in the MH group, that is, the MH group and the PH group only differ in the method of screening the difference interval, and the samples are the same.

[0185] In step S3, different numbers of DPCCs were screened under different minimum cutoff values, and four minimum cutoff values ​​of PCC numbers between 2000 and 4500 were selected for subsequent analysis. The minimum cutoff values ​​of each group were different; among them, the minimum cutoff values ​​of the PH group were 0.2, 0.25, 0.3, and 0.35 (the cutoffs for the comparison of tumor tissue samples with paired adjacent tissue samples and the comparison of tumor tissue samples with healthy peripheral blood were consistent); the minimum cutoff values ​​of the MH group were 0.3, 0.35, 0.4, and 0.45, and the minimum cutoff values ​​of the TH group were 0.25, 0.3, 0.35, and 0.4.

[0186] In step S4, the sequencing depth is set to 3X and the length of the generated DNA fragment is set to 66.

[0187] In step S6, the sample sources of the reference fragment database are CRA001142 and Inhouse1; the schematic process of model training and testing is as follows Figure 4 shown.

[0188] In step S7, the cfDNA data in EGAS00001000566, CRA1001537, and GSE63775 are input into the fragment binary classification model constructed in step S6 to calculate the maximum posterior probability as the cancer risk index of the sample.

[0189] The length of the CpG clusters analyzed was between 204 bp and 3724 bp, and each PCC contained at least 3 probes and at most 79 probes. The length of the PCC was between 0 and 1000 bp, and the number of CpG sites was between 0 and 50, mainly located in the promoter and exon regions ( Figure 5 (Figure A).

[0190] In statistics, the degree of variation describes the distribution width or dispersion of a set of data. The degree of variation SD of each PCC of samples in different types and groups is observed. The calculation formula of the degree of variation SD is:

[0191] In the formula, x i is the methylation status of the PCC of the sample in the group, Mean is the methylation status of the PCC of all samples in the group, N is the number of samples, is the correction factor for the sample standard deviation, used for unbiased estimation.

[0192] The analysis found that at the whole genome level, whether in tissue or peripheral blood, the degree of variation in tumor tissue is higher than that in adjacent tissue, and the degree of variation in WGBS data of tumor tissue is higher than that in methylation chip data. Even the degree of variation in peripheral blood of healthy people is greater than that in healthy tissue, and these variations may interfere with the identification of tumor cfDNA ( Figure 5(Fig. B).

[0193] The analysis obtained the liver cancer-specific PCC clusters (common Different PCC, co-DPCC) commonly identified by the MH group, TH group, and PH group. Hypermethylated co-DPCC refers to PCCs in tumor tissues with significantly higher methylation levels than those in adjacent cancer tissues and / or healthy peripheral blood; hypomethylated co-DPCCs refer to regions in tumor tissues with significantly lower methylation levels than those in adjacent cancer tissues and / or healthy peripheral blood. Among them, there were 141 hypermethylated DPCCs and 1094 hypomethylated DPCCs. Further annotation showed that most of the hypermethylated PCCs were located in the promoter region, while the hypomethylated PCCs were located in the promoter and intron regions ( Figure 6 ).

[0194] By displaying the methylation status of different PCCs in different types of samples using heat maps, it was found that liver cancer samples with different tumor purities showed different methylation intervals. The higher the purity, the more obvious the hypomethylation and hypermethylation of the samples ( Figure 7 ). The results of dimensionality reduction analysis (PCC, UMAP and tSNE) showed that tissue samples could be well clustered, but most cfDNA was separated from the tissue clusters to form relatively loose and mixed clusters ( Figure 8 ), which proves that the samples are reasonably clustered by disease and sample type.

[0195] Functional analysis of co-DPCC was performed (including GO and KEGG enrichment analysis). The results of GO enrichment analysis showed that genes related to high methylation were mainly enriched in embryonic organ development, cell fate commitment, leading edge membrane, and DNA binding transcription activator, while genes related to low methylation were mainly enriched in miRNA-mediated gene silencing, transcription-related by inhibiting translation, RISC complex, and plasma channel-related items such as single-atom cation channels ( Fig. 9 ). The results of KEGG enrichment analysis showed that genes associated with high methylation were mainly enriched in signaling pathways regulating stem cells and PI3K-Akt signaling pathway pluripotency, and genes associated with low methylation were mainly enriched in pathways such as olfactory transduction, PI3K-Akt signaling pathway, MAPK signaling pathway, and Th17 cell differentiation ( Fig.10 ). These results are consistent with the enrichment results of differential intervals found in the past.

[0196] Referring to the published literature (Li J, Wei L, Xianglin ZhangWei ZhangHaochen WangBixiZhongZhen XieHairong LvXiaowo Wang.DISMIR: Deep learning-based noninvasive cancer detection by integrating DNA sequence and methylation information ofindividual cell-free DNA reads[J].Briefings in bioinformatics,2021,22(6):bbab250-1-bbab250-11.), a deep learning network framework (CNN+LSTM+CNN) with good performance on liver cancer was constructed on pseudo-fragment data as the control group (org) and tested in the same sequencing data.

[0197] The accuracy of the models in the three training schemes of the MH group, TH group, and PH group in the two-category classification of fragments (cancer fragments vs healthy peripheral blood fragments) was above 0.7 ( Fig.11 ).

[0198] First, we observed the performance of the model with low methylation features in the WGBS data. The classification tasks were divided into five types: distinguishing between HCC and non-HCC patients (T vs NT) in the same cohort (CUHK and CRA cohorts), distinguishing between HCC patients and healthy people (T vs H), HCC patients and non-HCC patients (T vs NT), and HCC patients and non-HCC liver disease patients (T vs NTD) in different cohorts. The difficulty of classification gradually increased. In the same cohort, the AUC values ​​of the three models were higher than those of the control group model, and the average AUC value in the CUHK cohort was above 0.9. When samples from different cohorts were added, the mean AUC of the control group was significantly reduced. Although the AUC value of the model in the experimental group decreased with the difficulty of classification, it still had a high accuracy. And the smaller the threshold and the more features, the more stable the classification effect ( Fig.12 ). For early screening of the disease, it tends to improve sensitivity and find more suspicious patients. Among the three models, the MH model with low methylation characteristics (0.95) and the TH model with high methylation characteristics (0.97) have the highest sensitivity in diagnosing early liver cancer, demonstrating the advantages of the method of the present invention in early cancer screening. When it is necessary to minimize the misdiagnosis rate, the PH model with low methylation characteristics (0.998) and the MH model with high methylation characteristics (0.895) have the best specificity ( Fig.13 ).

[0199] From the cancer risk scores of different samples of the best-performing PH_02 model, it can be seen that the scores of patients with advanced liver cancer are significantly higher than those of other groups, and early liver cancer can also be well distinguished from other groups ( Fig.14 ). Combined with clinical information, it was found that tumor size and cancer risk score have a strong correlation ( Fig.15 , R=0.73, p<0.001). For patients numbered TBR34 and TBR36, the cancer scores of their peripheral blood samples at different time points were observed and analyzed in combination with their actual diagnosis. The cancer risk score of patient TBR34 decreased significantly after surgery, but increased again two months after surgery. Subsequently, at 3 months after surgery, multifocal tumor deposits were found in the remaining unresected liver and multiple lung metastases were found 4 months after surgery. Similarly, the cancer risk score of patient TBR36 decreased significantly after surgery and maintained a low cancer risk until 12 months after surgery. In fact, the patient was still in clinical remission (a period of time when cancer cells are reduced to a low level that no longer causes symptoms or requires treatment) 20 months after tumor resection ( Fig.16 ). These results suggest that for individual patients, the dynamic changes of the score can sensitively reflect postoperative recurrence or sustained remission, demonstrating its potential in monitoring treatment effects.

[0200] Observe other types of indicators ( Fig.17 ), in general, the low methylation features in liver cancer performed better, and the model constructed using the features screened by paired adjacent cancer tissues (PH) performed more robustly, with higher specificity and positive predictive value (PPV), that is, it could more reliably identify true positive cases. Similarly, the model with high methylation features also had good classification ability. In the high methylation model, the TH group using the top 30 samples of tumor purity had higher sensitivity and negative predictive value (NPV), and was more able to exclude true negative cases.

[0201] MCTA-Seq technology is a targeted sulfite sequencing technology that can enrich highly methylated CpG islands in highly fragmented cfDNA, obtain rich information, high sensitivity and cost-effectiveness. This method performs well in WGBS datasets. In order to test its generalization ability in other types of data, the best performing hypermethylation-related model in the MH group, PH group, and TH group was directly used in the MCTA data. The results showed that the AUC of the PH model could reach more than 0.9, and there was a significant difference in the cancer risk scores of patients with liver cancer and those without liver cancer ( Fig.18 Figure A and Figure B in the middle). Although the performance of low methylation features is not as good as that of high methylation features, the PH and TH models still capture weak cancer signals ( Fig.18The predicted results of the method of the present invention were compared with the detection results of common markers of liver cancer in peripheral blood, such as α-fetoprotein (AFP), glutamyl transferase (GGT) and alanine aminotransferase (ALT) ( Fig.19 The accuracy, sensitivity and specificity of the method of the present invention in predicting liver cancer samples are superior to those of three common biomarkers ( Fig.19 (Figure B), 42% of liver cancer patients showed false negative results in AFP, and 55% of early liver cancer showed false negatives in AFP, while the false negative rates of liver cancer and early liver cancer of the method of the present invention were only 8% and 9%, respectively, significantly reducing the missed diagnosis rate. In addition, the cancer scores of multiple tumors were significantly higher than the cancer prediction scores of patients with a single lesion. This result shows that the model trained by the method of the present invention on whole gene data can be well applied to targeted data, further demonstrating its potential in clinical applications. The method of the present invention can also serve as an effective supplement to traditional biomarkers, and play an important role in improving the accuracy and reliability of liver cancer diagnosis.

[0202] In summary, the model constructed by the cancer risk prediction method based on DNA methylation of the present invention can identify effective liver cancer methylation features for cancer diagnosis, and the pseudo-fragment data generated by the converter can provide effective information for the deep learning model, thereby identifying tumor fragments for the classification of cancer patients. In addition, the information identified within the whole genome can be applied to other types of methylation data, thereby expanding the application potential. More importantly, this method demonstrates the ability to convert data, overcomes the high cost and technical challenges of actually acquiring WGBS data, and provides a low-cost solution with wide adaptability for the early detection and accurate classification of various cancers. This result not only provides a new approach for liver cancer diagnosis, but also lays a solid foundation for the development of diagnostic technologies for other cancers in the future, showing great prospects for clinical application.

[0203] Application Example 2

[0204] This application example uses gastric cancer as an example to demonstrate the effectiveness of the cancer risk prediction method based on DNA methylation in Example 1.

[0205] 19 gastric cancer patients (inhouse2 cohort) and 9 healthy volunteers (inhouse1 cohort) were enrolled from Hospital B, and their clinical information is shown in Table 2. In the cfDNA-Seq data (WGBS sequencing data) of gastric cancer patients and healthy volunteers, an average of about 320,339,683 reads (Total reads) were obtained per sample, and the number of bases that met the quality of Q20 accounted for about 96.04%, of which GC bases accounted for about 22.30%. After quality control, the alignment rate of reads aligned to the human genome was about 80.86%, the ratio of PCR duplicates was about 24.98%, and the final number of reads obtained and used for downstream analysis was about 194,564,760.

[0206] Table 2

[0207]

[0208] The methylation chip data of gastric cancer was downloaded from TCGA, and a total of 395 samples were obtained, of which only 2 had paired adjacent cancer tissues. Since there are relatively few paired samples of gastric cancer samples in TCGA, the risk of gastric cancer was predicted according to the "a method for predicting cancer risk based on DNA methylation" in Example 1 above, and only the TH and MH group models were used for verification in gastric cancer. The analysis parameters were set as follows:

[0209] In step S1, the methylation chip is a 450K chip probe (Infinium Human Methylation 450 (450K)), and the whole genome is the reference genome GRCh38.

[0210] In step S2, the tumor tissue samples to be analyzed are cancer tissue samples in TCGA; the samples in the TH group are the top 30 samples in terms of tumor purity among 395 cancer tissue samples; and the samples in the MH group are 30 samples of stratified sampling among the 395 cancer tissue samples.

[0211] In step S3, different numbers of DPCCs were screened under different minimum cutoff values, and a minimum cutoff value of 3000-4000 PCCs was selected for subsequent analysis. The minimum cutoff values ​​of each group were different; among them, the minimum cutoff value of the MH group was 0.3, and the minimum cutoff value of the TH group was 0.45.

[0212] In step S4, the sequencing depth is set to 3X and the length of the generated DNA fragment is set to 66.

[0213] In step S6, the sample source of the reference fragment database is the CRA001142 and Inhouse1 cohorts in Application Example 1, and 37 samples are randomly selected from them to construct a fragment binary classification model.

[0214] In step S7, WGBS sequencing data of cfDNA of 21 gastric cancer patients enrolled in Hospital B, CRA001142 from the sample source of the control fragment database, the remaining 12 healthy samples in the Inhouse1 cohort, and healthy samples in the EGAS00001000566 and CRA1001537 cohorts, totaling 48 samples, are input into the fragment binary classification model constructed in step S6 to calculate the maximum posterior probability as the cancer risk index of the sample.

[0215] The gastric cancer-specific PCC clusters (commonDifferent PCC, co-DPCC) identified by the two training schemes of TH and MH groups were obtained, of which there were 1977 hypermethylated DPCCs and 147 hypomethylated DPCCs ( Fig. 20 After annotation, it was found that most DPCCs were located in the promoter region ( Fig. 20 (Figure C).

[0216] The model framework is the same as that in Application Example 1. The model trained with pseudo-fragment data is used to predict the WGBS sequencing data of cfDNA of gastric cancer patients and healthy controls, and the cancer risk index of each sample is obtained to classify the samples. The classification effect is shown in the following figure. Fig.21 As shown, the TH model with low methylation features performed best (AUC = 0.873), with a classification sensitivity of 0.895 and a specificity of 0.787. Compared with the MH model, the TH model has better specificity, a higher TPV value, and the model has a higher overall classification accuracy for samples. It is worth noting that although more hypermethylation features were found, the hypermethylation features were not significantly effective in the classification of cancer in peripheral blood. This may be because specific hypermethylation signals in peripheral blood are more easily masked by signals from normal cells. From the cancer risk scores of gastric cancer samples and healthy control group samples, it can be observed that the scores predicted by the low methylation features were significantly different between the two groups ( Fig. 22 ). Combined with the clinical information of gastric cancer, it can be observed that the later the cancer stage and the larger the tumor, the higher the predicted cancer risk score will be ( Fig.23 ).

[0217] In summary, the model constructed by the cancer risk prediction method based on DNA methylation of the present invention can be used for the diagnosis of gastric cancer, and low methylation is more suitable for identifying signals of gastric cancer occurrence from peripheral blood than high methylation features. In addition, the potential for wide application of this prediction method in different types of cancer, especially its effectiveness in early diagnosis of cancer, was further verified. The performance of low methylation features in peripheral blood, especially its significant differences in detecting and distinguishing gastric cancer patients from healthy controls, provides a new direction for non-invasive cancer screening. This discovery suggests how to better utilize methylation feature combinations in actual clinical applications to optimize the sensitivity and specificity of cancer detection, thereby improving the accuracy and reliability of early diagnosis.

[0218] Application Example 3

[0219] This application example uses pancreatic cancer as an example to demonstrate the effectiveness of the cancer risk prediction method based on DNA methylation in Example 1.

[0220] Seven pancreatic cancer patients were enrolled from Hospital A (inhouse3 cohort). In the pancreatic cancer cfDNA-Seq data, an average of about 320,339,683 reads (Total reads) were obtained for each sample, and the number of bases that met Q20 quality accounted for about 95.485%, of which GC bases accounted for about 22.75%. After quality control, the alignment rate of reads aligned to the human genome was about 75.63%, the ratio of PCRduplicates was about 48.84%, and the final number of reads obtained and used for downstream analysis was about 124,897,013.

[0221] 184 tumor chip data were downloaded from TCGA, of which only 10 samples had paired adjacent cancer tissues. Since there were relatively few paired samples of pancreatic cancer samples in TCGA, the risk of gastric cancer was predicted according to the "a method for predicting cancer risk based on DNA methylation" in Example 1 above, and only the TH and MH group models were used for verification in pancreatic cancer.

[0222] The analysis parameters are set as follows:

[0223] In step S1, the methylation chip is a 450K chip probe (Infinium Human Methylation 450 (450K)), and the whole genome is the reference genome GRCh38.

[0224] In step S2, the tumor tissue samples to be analyzed are cancer tissue samples in TCGA; the samples in the TH group are the top 30 samples in terms of tumor purity among 184 cancer tissue samples; and the samples in the MH group are 30 samples of stratified sampling among the 184 cancer tissue samples.

[0225] In step S3, different numbers of DPCCs were screened under different minimum cutoff values, and a minimum cutoff value of 3000-4000 PCCs was selected for subsequent analysis. The minimum cutoff values ​​of each group were different; among them, the minimum cutoff value of the MH group was 0.3, and the minimum cutoff value of the TH group was 0.35.

[0226] In step S4, the sequencing depth is set to 3X and the length of the generated DNA fragment is set to 66.

[0227] In step S6, the sample sources of the control fragment database are 37 healthy samples randomly selected from the CRA001142 and Inhouse1 cohorts.

[0228] In step S7, the WGBS sequencing data of cfDNA of 7 pancreatic cancer patients enrolled in Hospital A (inhouse3 cohort), the remaining 12 healthy samples in CRA001142 and Inhouse1 cohorts, and the healthy samples in EGAS00001000566 and CRA1001537 cohorts, totaling 48 samples, are input into the fragment binary classification model constructed in step S6 to calculate the maximum posterior probability as the cancer risk index of the sample.

[0229] The gastric cancer-specific PCC clusters (commonDifferent PCC, co-DPCC) identified by the two training schemes of TH and MH groups were obtained, of which 1615 were hypermethylated DPCCs and 200 were hypomethylated DPCCs ( Fig.24 After annotation, it was found that most DPCCs were located in the promoter region ( Fig.24 (Figure C).

[0230] The model framework is the same as that in Application Example 1. The model trained with pseudo-fragment data is used to predict the WGBS sequencing data of cfDNA of pancreatic cancer patients and healthy controls, and the cancer risk index of each sample is obtained to classify the samples. The classification effect is shown in Figure 2. Fig.25 As shown in the figure, the mean AUC values ​​of the two training models (TH and MH) for hypermethylation and hypomethylation features are both around 0.9, and the model for hypomethylation features has higher sensitivity and specificity, despite identifying more hypermethylation features. The TH model has significant differences in the cancer risk scores of pancreatic cancer and healthy controls ( Fig.26 ).

[0231] In summary, the cancer risk prediction method based on DNA methylation of the present invention also has good application potential in the diagnosis of pancreatic cancer, especially in the use of low methylation features to identify pancreatic cancer signals from peripheral blood. Despite the limited number of pancreatic cancer samples and the lack of detailed clinical information, the prediction method can still effectively identify pancreatic cancer patients from a large number of healthy control samples, highlighting its high sensitivity in diagnosis. In addition, these healthy samples come from different cohorts, and the model still maintains good classification performance, further demonstrating the generalization ability and robustness of the method of the present invention. This result not only verifies the effectiveness of this method in pancreatic cancer detection, but also shows its adaptability in processing complex and highly heterogeneous data sets. In the future, with the accumulation of more clinical data and further model optimization, it is expected to become a powerful tool for early screening of various cancers such as pancreatic cancer, and promote the realization of personalized medicine.

[0232] The embodiments of the present invention are described in detail above in conjunction with the embodiments, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for constructing a cancer risk prediction model, characterized in that: The following steps are involved: S1. Based on the methylation probes in the methylation chip, the expected CpG clusters in the whole genome are obtained and recorded as PCC; S2. grouping according to the tumor purity of the tumor tissue samples to be analyzed, wherein the tumor tissue samples to be analyzed have methylation probe detection data; a1) when a tumor tissue sample has a paired adjacent paracancerous tissue sample, calculating the tumor purity of the tumor tissue sample, and selecting a desired number of the tumor tissue samples as a PH group; a2) when the tumor tissue sample does not have a paired adjacent cancer tissue sample, calculating the tumor purity of the tumor tissue sample, sorting the samples according to the tumor purity, and selecting a desired number of tumor tissue samples with the highest tumor purity as the TH group; a3) When the tumor tissue sample does not have a paired adjacent cancer tissue sample, the tumor purity of the tumor tissue sample is calculated, the samples are sorted according to the tumor purity, and the whole sample is stratified and sampled to obtain the desired number of samples as the MH group; S3, identifying cancer-specific CpG clusters in the PCC, denoted as DPCC; For the PH group, the first methylation state difference of the PCC between each paired tumor tissue sample and adjacent paracancerous tissue sample in the PH group and the second methylation state difference of the PCC between the tumor tissue sample and healthy peripheral blood are calculated, and the PCC whose first methylation state difference and the second methylation state difference are greater than the first expected methylation state difference threshold are screened as DPCC; For the TH group and / or MH group: calculating the third methylation state difference between the PCC in the tumor tissue sample and the healthy peripheral blood, and screening the PCC whose third methylation state difference is greater than the second expected methylation state difference threshold as DPCC; S4, generating a reference segment of the DPCC: The methylation status of the DPCC in the tumor tissue sample is obtained. Given the sequencing depth and the length of the generated fragment, the base composition of the DPCC is obtained. After a start site is generated by a randomizer, a reference fragment covering the DPCC is generated according to the fragment length, and the number of generated fragments S is calculated. The calculation formula is as follows: In the formula, D is the sequencing depth, R is the DPCC length, and L is the length of the generated DNA fragment; S5. Simulate the methylation status of the reference fragment and generate a pseudo-fragment database: Fill the methylation status of the reference fragment obtained in step S4 into a matrix with a length of PCC and a width of the number of fragments to simulate the methylation status; and distinguish and mark the methylated CpG sites from the remaining sites according to the methylation status of the DPCC in the tumor tissue sample to obtain pseudo-fragment data and generate a pseudo-fragment database of the DPCC; S6. Training fragment binary classification model based on DNA methylation and sequence features of pseudo-fragment data: The fragment information located at the DPCC in the healthy sample is obtained to obtain a control fragment database; the pseudo-fragment database and the control fragment database are input into the training deep learning model to construct a fragment binary classification model, thereby obtaining the cancer risk prediction model.

2. The construction method according to claim 1, characterized in that: In step S1, by using a sliding window method, if the first flanking region of the first methylation probe overlaps with the second flanking region of the second methylation probe, the DNA sequences corresponding to the first flanking region and the second flanking region on the reference genome are classified into one CpG cluster; and the CpG cluster that meets the expectation is selected, namely, PCC; The first flanking region includes the first methylation probe and DNA sequences of set window sizes upstream and downstream thereof; The second flanking region includes the second methylation probe and DNA sequences of set window sizes upstream and downstream thereof; Preferably, the desired CpG cluster meets the following conditions: b1) There are at least 3 CpG sites covered by methylation probes; b2) capable of providing methylation signals for calculating methylation status; b3) contains the portion spanning the gene functional region.

3. The construction method according to claim 1, characterized in that: The desired PCC meets the following conditions: c1) the expected number of the first methylation state difference and the second methylation state difference is greater than the first expected methylation state difference threshold; c2) The difference between the first median of the methylation state of the methylation region in the PCC that meets the condition shown in c1 and the second median of the methylation state in the healthy peripheral blood is greater than the first expected methylation state difference threshold.

4. The construction method according to claim 1, characterized in that: Source data of the methylation status of PCC in the tumor tissue sample, adjacent tissue sample, and healthy peripheral blood: If it is methylation chip data, the methylation status of the probe in the PCC is obtained by fluorescence intensity and its mean is calculated, which is the methylation status of the PCC; If it is sequencing data, the methylation status of a single CpG site covered in the PCC is calculated first, and then the average of the methylation status of all CpG sites covered by the PCC is calculated, which is the methylation status of the PCC; the methylation status of a single CpG site refers to the average of the methylation status of the CpG sites of multiple DNA fragments.

5. The construction method according to claim 1, characterized in that: In step S4, the reference fragment includes its position information, base composition and methylation status in the reference genome; and / or, in step S5, the pseudo-fragment database includes the position information, base composition and methylation signal data of the reference fragment; And / or, in step S6, the fragment information located in the DPCC interval in the healthy sample includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome.

6. The construction method according to claim 1, characterized in that: In step S6, training the segment binary classification model includes at least one round, and each round includes the following steps: d1) randomly dividing the control fragment database and the pseudo fragment database into two parts, which are used for training and verification of the fragment binary classification model respectively; d2) Encoding the DNA fragment into a one-hot matrix of the fragment length * 5 in step S4 as the input matrix of the model to train the deep learning model, and using the sigmoid function as the last layer activation function of the model to output the probability of cancer origin of the fragment; establishing the fragment binary classification model.

7. A method for predicting cancer risk, characterized in that: The following steps are involved: E1) constructing the fragment binary classification model according to the construction method according to any one of claims 1 to 6; E2) Obtaining information of the fragments located in the DPCC in the peripheral blood sample to be predicted, inputting the fragments into the binary classification model, obtaining the cancer source probability of all fragments, and calculating the maximum posterior probability That is, the cancer risk index of the peripheral blood sample to be predicted; Preferably, the fragment information located in the DPCC interval in the peripheral blood sample to be predicted includes the position information, base composition and methylation signal of the DNA fragment located in the DPCC interval in the reference genome; Preferably, the maximum a posteriori probability The calculation formula is as follows: Where, d i It refers to the probability that each DNA fragment comes from the tumor, n refers to the number of DNA fragments, and r refers to the proportion of DNA fragments derived from the tumor in the peripheral blood sample to be predicted.

8. A cancer risk prediction system, characterized in that: include: A PCC acquisition module, used to acquire CpG clusters on the whole genome based on the methylation probes in the methylation chip, namely the PCC; A grouping module, grouping according to the tumor purity of the tumor tissue samples to be analyzed, wherein the tumor tissue samples to be analyzed have methylation probe detection data; a1) when the tumor tissue samples have paired adjacent cancer tissue samples, calculating the tumor purity of the tumor tissue samples, and selecting a desired number of the tumor tissue samples as the PH group; a2) when the tumor tissue samples do not have paired adjacent cancer tissue samples, calculating the tumor purity of the tumor tissue samples, sorting them according to the tumor purity, and selecting a desired number of the tumor tissue samples with the highest tumor purity as the TH group; a3) When the tumor tissue sample does not have a paired adjacent cancer tissue sample, the tumor purity of the tumor tissue sample is calculated, the samples are sorted according to the tumor purity, and the whole sample is stratified and sampled to obtain the desired number of samples as the MH group; A DPCC identification module is used to identify cancer-specific CpG clusters in the PCC to obtain the DPCC; for the PH group, a first methylation state difference of the PCC between each paired tumor tissue sample and adjacent tissue sample in the PH group and a second methylation state difference of the PCC between the tumor tissue sample and healthy peripheral blood are calculated, and PCCs whose first methylation state difference and second methylation state difference are greater than a first expected methylation state difference threshold are screened as DPCCs; for the TH group and / or MH group: a third methylation state difference of the PCC in the tumor tissue sample and healthy peripheral blood is calculated, and PCCs whose third methylation state difference is greater than a second expected methylation state difference threshold are screened as DPCCs; The reference fragment generation module is used to obtain the methylation status of the DPCC in the tumor tissue sample, obtain the base composition of the DPCC given the sequencing depth and the generated fragment length, generate a starting site through a randomizer, generate a reference fragment covering the DPCC according to the fragment length, and calculate the number of generated fragments S, the calculation formula is as follows: In the formula, D is the sequencing depth, R is the DPCC length, and L is the length of the generated DNA fragment; A pseudo-fragment database generation module is used to fill the methylation status of the reference fragment into a matrix with a length of PCC and a width of the number of fragments to simulate the methylation status; and according to the methylation status of the DPCC in the tumor tissue sample, distinguish and mark the methylated CpG sites from the remaining sites to obtain pseudo-fragment data, and generate a pseudo-fragment database of the DPCC; A model generation module is used to obtain the fragment information located at the DPCC in the healthy sample to obtain a control fragment database; the pseudo-fragment database and the control fragment database are input into the training deep learning model to construct a fragment binary classification model; The cancer risk index prediction module is used to obtain the fragment information located in the DPCC in the peripheral blood sample to be predicted, input it into the cancer risk prediction model, obtain the cancer source probability of all fragments and calculate the cancer risk index.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the construction method according to any one of claims 1 to 6 or the cancer risk prediction method according to claim 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the construction method according to any one of claims 1 to 6 or the steps according to claim 7.

Images