WGS methylation level prediction method and tissue origin inference method

Through the analysis of Gaussian mixed model and hidden Markov model of WGS data, combined with UXM algorithm, the low sensitivity and complex operation of cfDNA methylation detection in the existing technology are solved, and the methylation level and tissue origin of cfDNA is achieved. Accurate inference improves the correlation between data utilization and detection.

CN118983003APending Publication Date: 2024-11-19GENESEEQ TECH INC +1

Patent Information

Application Number
CN202410835078.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-25
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Existing DNA methylation detection technologies such as bisulfite conversion sequencing have problems such as incomplete transformation, DNA breakage and base damage, resulting in low diagnostic sensitivity and complex operation, and the inability to effectively infer tissue origin through cfDNA, limiting its non-invasiveness. Applications in tumor detection and other fields.

Method used

The cfDNA methylation level prediction was performed using WGS data. By extracting fragment characteristics, building Gaussian mixed model and non-uniform hidden Markov model, combining Viterbi algorithm and Baum-Welch algorithm, the methylation mode state of each fragment was calculated. The UXM algorithm was used to estimate the relative abundance of cell types and tissues.

Benefits of technology

It is possible to accurately infer cfDNA methylation level and tissue origin from WGS data, overcome technical obstacles in bisulfite conversion, improve data utilization, and is highly correlated with the true methylation results of WGBS detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118983003A_ABST
    Figure CN118983003A_ABST
Patent Text Reader

Abstract

The invention relates to a WGS methylation level prediction method and a tissue origin inference method, and belongs to the technical field of gene detection. And performing whole genome sequencing (WGS) on the cfDNA by using a next-generation sequencing technology. The method comprises the following steps: extracting fragment characteristics of each cfDNA fragment through a bioinformatics means, marking CpG points covered by each fragment of WGS data as an unmethylated state and a methylated state by using a Gaussian mixture model, then constructing a non-uniform hidden Markov model, decoding the optimal methylated state of each CpG in each fragment, and finally obtaining the optimal methylated state of each CpG in each fragment. And summarizing the state of each CpG point on the reference genome, and calculating the continuous methylation level. In contrast to the disclosed human specific reference methylation profiles, methylation signals are deconvolved at the fragment level, estimating the relative abundance of cell types and tissues contributing to cfDNA. According to the method, the methylation level and the origin tissue state of the cfDNA can be directly deduced from WGS data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for inferring WGS cfDNA methylation level and tissue of origin, and belongs to the technical field of gene detection. Background Art

[0002] DNA methylation is an important component of epigenetics and plays an important role in gene regulation during disease progression and embryonic development. The genome-wide DNA methylation level of plasma free DNA (cfDNA) has been widely studied for disease diagnosis and prognosis. Currently, bisulfite conversion sequencing is the gold standard for DNA methylation detection. However, bisulfite treatment has certain limitations: 1. Incomplete conversion. One of the technical difficulties is to ensure that unmethylated cytosine can be completely converted to uracil. If the conversion is incomplete, the analysis results may overestimate the methylation level. 2. DNA breakage and base damage may occur during bisulfite treatment, which may limit the sensitivity of diagnostic tests and analysis. 3. The operation steps are complicated, the processing time is long, and a large amount of sample input is required. The development of enzymatic conversion and long-read sequencing methods has partially alleviated these problems, but has not yet been widely used.

[0003] Each cell type has its own unique DNA methylation profile, which is formed during cell differentiation and can reflect the cell's developmental history and tissue origin to a certain extent. By analyzing the cfDNA methylation pattern, the tissue type of these fragments can be inferred, which has broad application prospects in non-invasive tumor detection, pregnancy monitoring, organ transplantation and other fields.

[0004] Unlike genomic DNA (gDNA), cfDNA is not randomly fragmented, and its fragmentation pattern is highly correlated with the local epigenetic background. Some studies have found significant differences in DNA fragmentation patterns between methylated and unmethylated cfDNA molecules, and these findings suggest the possibility of computationally inferring DNA methylation levels from the fragmentation patterns of cfDNA. Another study provides a conceptual solution to predict the binary state of DNA methylation in ultra-high-depth WGBS using a deep learning model.

[0005] WGS (Whole Genome Sequencing) and WGBS (Whole Genome Bisulfite Sequencing) are two different types of whole genome sequencing technologies. Although both are widely used in biomedical research and clinical diagnosis, WGS is more focused on methylation than WGBS. WGS can not only detect various types of genetic variations including single nucleotide polymorphisms (SNPs), insertions and deletions (Indels), copy number variations (CNVs) and structural variations (SVs), which provides extremely important information for explaining the genetic basis of diseases, genetic differences between individuals, and evolutionary studies. Moreover, no bisulfite conversion treatment is required, which can avoid sample loss or degradation caused by chemical treatment. In short, WGS has a wider application prospect in the field of genomic research due to its wide coverage, ability to detect various genetic variations, cost-effectiveness and wide range of applications. Therefore, academic institutions and commercial entities around the world have tested hundreds of thousands of cfDNA WGS samples for early cancer detection and other purposes.

[0006] Therefore, how to predict methylation status through cfDNA WGS, obtain genetic variation and epigenetic information simultaneously through a single test, and improve the efficiency of data utilization are problems that need to be solved in existing technologies. Summary of the invention

[0007] In order to obtain DNA methylation information, the present invention has developed a new cfDNA gene detection technology to infer cfDNA methylation levels from WGS data and estimate the relative abundance of cell types and tissues contributing to cfDNA.

[0008] A method for predicting WGS cfDNA methylation level, wherein the method is applied for non-therapeutic and diagnostic purposes, comprising the following steps:

[0009] Step 1: cfDNA extraction, library construction, and sequencing;

[0010] Step 2: Remove the adapters and low-quality sequences from the sequencing data to obtain high-quality data;

[0011] Step 3: Align the high-quality data to the human reference genome, remove duplicate sequences, and obtain a sequence alignment file (bam);

[0012] Step 4: Extract the fragment characteristics of each cfDNA fragment from bam: fragment length, normalized sequencing depth, and the distance from each CpG site to the center of the fragment;

[0013] Step 5: Use a Gaussian mixture model to mark the CpG sites covered by each fragment of the WGS data as unmethylated (u) and methylated (m) states, and initialize the non-uniform hidden Markov model (HMM) parameters;

[0014] Step 6: Apply the fragment features obtained in step 4 and the unmethylated (u) and methylated (m) states of the CpG site markers as HMM parameters to the HMM model, construct a Markov chain for each fragment, perform model parameter fitting and solve, and obtain the methylation pattern state of each cfDNA fragment.

[0015] The method also includes step 7, calculating the average methylation level of each window of each sample, and the calculation method is: the sum of the number of reads of all CpG sites marked as m in the window divided by the sum of the total number of reads of all CpG sites m and u in this interval.

[0016] In step 1, the sequencing method is WGS.

[0017] In the above step 5, it is always assumed that the average distance between the CpG site in the m state and the fragment center is larger.

[0018] In step 6, an HMM model is constructed using all fragments covering at least 7 CpGs; in the parameter fitting solution, the Viterbi algorithm is used to decode the optimal methylation state (u or m) of each CpG in each fragment, the Baum-Welch algorithm is used to estimate the maximum likelihood parameters of the WGS data set, and the Kullback-Leibler distance is used to estimate whether the model converges during the iteration process; and the weighted healthy leukocyte DNA methylation prior is used to calculate the CpG posterior probability of the hidden state.

[0019] A method for inferring the origin of tumor tissue based on methylation levels comprises the following steps:

[0020] Step a, obtaining the methylation pattern status of each cfDNA fragment in the sample by the above method;

[0021] Step b, labeling each fragment as U (unmethylated), M (methylated), or X (mixed) according to the number of methylated and unmethylated CpGs;

[0022] Step c, using a public human-specific reference methylation map, calculate the ratio of U / X / M fragments with at least k CpGs in each marker region and all cell types, assuming that the reference methylation map has N marker regions, Ai,j represents the U ratio of the i-th marker in the j-th cell type;

[0023] Step d: given a methylation pattern state obtained from the WGS sequencing result of a cfDNA as an input sample, within each marker interval, screen out cfDNA fragments that cover at least 4 CpGs, and then mark each cfDNA segment as U / X / M according to the set threshold, and then count the proportion of U fragments in all fragments within this marker interval to generate an N×1 vector b; by minimizing |A*Xb| 2 , and satisfy non-negative X, and normalized to Σ j X j =1, thereby calculating the coefficient vector x; x is the relative abundance of cell types and tissues that contribute to cfDNA, and x is between 0 and 1. The larger the x, the greater the possibility that the cfDNA comes from a certain tissue.

[0024] In the step c, k is set to 4, and the ratio of U / X / M segments is determined by: for U segments, methylated CpG ≤ 25%; for M segments, methylated CpG ≥ 75%; for X segments, methylated CpG 25% <X<75%。

[0025] The coefficient vector x is solved using the least non-negative squares method (NNLS).

[0026] A WGS cfDNA methylation level prediction device, comprising:

[0027] An extraction module, used for extracting cfDNA samples;

[0028] A sequencing module, used to sequence the extracted cfDNA samples;

[0029] The feature value acquisition module is used to extract the fragment features of each cfDNA fragment from the sequence alignment file: fragment length, normalized sequencing depth, and the distance from each CpG site to the center of the fragment;

[0030] The label module is used to label the CpG sites covered by each fragment of the WGS data as unmethylated (u) and methylated (m) states using a Gaussian mixture model, and initialize the parameters of the non-uniform hidden Markov model (HMM);

[0031] The methylation pattern state prediction module is used to apply the obtained fragment features and the unmethylated (u) and methylated (m) states of the CpG site markers as HMM parameters to the HMM model, construct a Markov chain for each fragment, perform model parameter fitting and solution, and obtain the methylation pattern state of each cfDNA fragment.

[0032] It also includes: an average methylation level calculation module, which is used to calculate the average methylation level of each window of each sample, and the calculation method is: the sum of the number of reads of all CpG sites marked as m in the window divided by the sum of the total number of reads of all CpG sites m and u in this interval.

[0033] The sequencing module uses WGS sequencing.

[0034] The average distance from the CpG site to the fragment center is larger when the m state is set in the tag module.

[0035] All fragments covering at least 7 CpGs were used to construct the HMM model.

[0036] In the parameter fitting solution in the methylation pattern state prediction module, the Viterbi algorithm is used to decode the optimal methylation state (u or m) of each CpG in each fragment, the Baum-Welch algorithm is used to estimate the maximum likelihood parameters of the WGS data set, and the Kullback-Leibler distance is used to estimate whether the model converges during the iteration process; and the weighted healthy leukocyte DNA methylation prior is used to calculate the CpG posterior probability of the hidden state.

[0037] A device for inferring the origin of tumor tissue based on methylation level, comprising the above-mentioned WGS cfDNA methylation level prediction device, further comprising:

[0038] A fragment methylation status labeling module is used to label each fragment as U (unmethylated), M (methylated), or X (mixed) according to the number of methylated and unmethylated CpGs;

[0039] A methylation ratio calculation module is used to calculate the ratio of U / X / M fragments with at least k CpGs in each marker region and all cell types using a public human-specific reference methylation map. Assuming that the reference methylation map has N marker regions, Ai,j represents the U ratio of the i-th marker in the j-th cell type;

[0040] The tissue tracing inference module is used to take the methylation pattern state obtained from the WGS sequencing result of a cfDNA as an input sample, screen out cfDNA fragments covering at least 4 CpGs in each marker interval, and then mark each cfDNA segment as U / X / M according to the set threshold, and then count the proportion of U fragments in all fragments in this marker interval to generate an N×1 vector b; by minimizing |A*Xb| 2 , and satisfy non-negative X, and normalized to Σ j X j=1, thereby calculating the coefficient vector x; x is the relative abundance of cell types and tissues that contribute to cfDNA, and x is between 0 and 1. The larger the x, the greater the possibility that the cfDNA comes from a certain tissue.

[0041] In the methylation ratio calculation module, k is set to 4, and the ratio of U / X / M fragments is determined as follows: for U fragments, methylated CpG ≤ 25%; for M fragments, methylated CpG ≥ 75%; for X fragments, methylated CpG 25% <X<75%。

[0042] A computer-readable medium having recorded thereon a program capable of executing the above-mentioned WGS cfDNA methylation level prediction method or the tumor tissue origin inference method based on methylation level.

[0043] Beneficial Effects

[0044] This invention provides for the first time a technology that infers cfDNA methylation levels from WGS data, deconvolutes methylation signals at the fragment level, and estimates the relative abundance of cell types and tissues that contribute to cfDNA using the UXM algorithm. This technology can not only accurately predict methylation status and tissue traceability, but also overcome the technical barriers of bisulfite conversion, allowing epigenetic analysis using WGS data, greatly improving data utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The flow chart of this patent is shown.

[0046] Figure 2 The correlation between the WGS-inferred methylation rate and the true methylation rate detected by WGBS is shown.

[0047] Figure 3 The distribution of WGS-inferred methylation rates and the true methylation rates detected by WGBS are shown.

[0048] Figure 4 The input file format for fragment-level deconvolution of methylation signals is shown.

[0049] Figure 5a The abundance distribution of the top three tissues contributing to cfDNA inferred from healthy and HCC WGBS samples is shown.

[0050] Figure 5b The top two tissue types inferred to contribute to cfDNA in HCC WGBS samples after filtering the hematopoietic system are shown.

[0051] Figure 6a The abundance distribution of the top three tissues contributing to cfDNA inferred from healthy and liver cancer WGS samples is shown.

[0052] Figure 6b The top two tissue types contributing to cfDNA inferred from liver cancer WGS samples after filtering out the hematopoietic system are shown. DETAILED DESCRIPTION

[0053] The present invention provides a method for inferring cfDNA methylation levels and tissue tracing from WGS data. The cfDNA methylation levels estimated by this method are highly correlated with the true methylation detected by WGBS. This prediction method is reliable, and the more dense the CpG sites, the higher the correlation. The inferred methylation signal is deconvoluted at the fragment level, and the relative abundance of cell types and tissues contributing to cfDNA is estimated using the UXM method. The traceability spectrum is highly consistent with the true methylation traceability spectrum detected by WGBS, and can correctly decompose the cell type sources of healthy and liver cancer cfDNA.

[0054] The present invention will be further described below in conjunction with specific embodiments.

[0055] Example 1

[0056] In the example, 30 plasma samples were used, 10 from healthy adults and 20 from liver cancer patients. The data were derived from clinical samples actually obtained by the applicant, and the plasma was separated on the same day of blood collection. One plasma separated from the same blood sample was tested for WGBS and the other for WGS, which was used to evaluate the consistency between the inferred methylation and the actual methylation.

[0057] WGBS Testing Process

[0058] The process involves plasma cfDNA extraction, bisulfite conversion, library construction, whole genome sequencing, and bioinformatics analysis. The steps are as follows:

[0059] 1) Centrifuge the whole blood sample to obtain plasma containing cfDNA, centrifuge the plasma at 4°C for 10 min, and obtain the supernatant;

[0060] 2) Add lysis binding solution, proteinase K and magnetic beads to the supernatant, vortex for 30 seconds, and incubate at room temperature for 10 minutes;

[0061] 3) Place on a magnetic stand for 3 min, remove the supernatant, add wash buffer 1, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0062] 4) Place on a magnetic stand for 3 min, remove the supernatant, add wash buffer 2, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0063] 5) Place on a magnetic stand for 3 min, remove the supernatant, add 100% ethanol, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0064] 6) Place on a magnetic rack and let stand for 3 minutes, remove the supernatant, dry at room temperature for 45-55 minutes, elute with enzyme-free water, and incubate at room temperature for 5 minutes;

[0065] 7) Place on a magnetic stand and let stand for 2 minutes, then transfer the supernatant to a new 1.5 ml centrifuge tube;

[0066] 8) Add an internal control to the extracted cfDNA sample, and then use the EZ DNA Methylation-Gold Kit (Zymoresearch) to perform bisulfite conversion on the DNA sample containing the internal control. For specific extraction steps, refer to the kit operating instructions;

[0067] 9) The converted cfDNA was used as a template for library construction. The library was constructed using the Accel-NGS Methyl-SeqDNA Library Kit (Swift Biosciences). For specific extraction steps, please refer to the kit manual;

[0068] 10) For the cfDNA library obtained above, whole genome methylation sequencing was performed using Illumina's HiSeq Xten sequencer. After sequencing was completed and the machine was off, bcl2fastq was used to generate a fastq file;

[0069] 11) FastQC software was used to perform quality control on the data, and Trimmomatic software was used to remove adapters and low-quality sequences. The obtained high-quality data was aligned to the human reference genome hg19 using Bismark to generate a bam file;

[0070] 12) Execute the bismark_methylation_extractor operation to extract the number of methylated (m) and unmethylated (u) reads for each CpG site in the whole genome.

[0071] WGS Testing Process

[0072] The process involves plasma cfDNA extraction, library construction, whole genome sequencing, and bioinformatics analysis. The steps are as follows:

[0073] 1) Centrifuge the whole blood sample to obtain plasma containing cfDNA, centrifuge the plasma at 4°C for 10 min, and obtain the supernatant;

[0074] 2) Add lysis binding solution, proteinase K and magnetic beads to the supernatant, vortex for 30 seconds, and incubate at room temperature for 10 minutes;

[0075] 3) Place on a magnetic rack and let stand for 3 min, remove the supernatant, add wash buffer 1, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0076] 4) Place on a magnetic rack and let stand for 3 min, remove the supernatant, add wash buffer 2, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0077] 5) Place on a magnetic rack and let stand for 3 min, remove the supernatant, add 100% ethanol, vortex mix for 10 s, and incubate at room temperature for 2 min;

[0078] 6) Place on a magnetic rack and let stand for 3 minutes, remove the supernatant, dry at room temperature for 45-55 minutes, elute with enzyme-free water, and incubate at room temperature for 5 minutes;

[0079] 7) Place on a magnetic stand and let stand for 2 minutes, then transfer the supernatant to a new 1.5 ml centrifuge tube;

[0080] 8) Add Qiagen end repair reagent to the cfDNA sample, mix well (operate on ice), incubate until stable at 4°C, then add IDT 8nt UDI Adapter, mix well, and incubate at 20°C for 15 minutes;

[0081] 9) Perform 0.85X magnetic bead purification on the above reaction solution, mix well, and incubate at room temperature for 5 minutes to allow the DNA to bind to the magnetic beads. Place the 96-well PCR plate on a magnetic stand and incubate at room temperature for 5 minutes until the magnetic beads adhere to the wall and the supernatant is clear, and discard the supernatant;

[0082] 10) Add 80% ethanol, incubate at room temperature for 30 seconds, remove the ethanol, dry at room temperature for 2-5 minutes, and repeat twice;

[0083] 11) Add enzyme-free water to elute the DNA bound to the magnetic beads, incubate at room temperature for 5 minutes until the supernatant is clear, and transfer the supernatant for loading onto the machine;

[0084] 12) The cfDNA library obtained above was sequenced using the Illumina Hiseq Xten sequencer for whole genome sequencing. After the sequencing was completed, bcl2fastq was used to generate a fastq file;

[0085] 13) Fastp software was used to perform quality control on the data, remove adapters and low-quality sequences, and the obtained high-quality data was aligned to the human reference genome hg19 using bwa to generate a bam file.

[0086] Example 2

[0087] This example is used to illustrate the process of using WGS to infer cfDNA methylation levels.

[0088] Since DNA methylation and nucleosome occupancy are closely related, and naked DNA away from the center of the nucleosome is more susceptible to methylation than DNA wrapped in the nucleosome, it is hypothesized that the fragmentation pattern observed for each cfDNA molecule should be able to reveal its associated methylation pattern and thus its tissue of origin.

[0089] Assume that each CpG in each cfDNA fragment has only two states: methylated ("m") or unmethylated ("u"), which cannot be observed in WGS (i.e., "hidden"). Assume that these states are affected by three fragment characteristics: fragment length (l), distance from the CpG point to the center of the fragment (d), and normalized sequencing depth at this position (c), and assume that the state of each CpG point in each fragment is a multivariate Gaussian distribution of the above three characteristics.

[0090] Extract the l, d, c fragment features of each CpG of each fragment from the WGS bam file of Example 1, and use Gaussian mixture model (GMM) to mark all CpG sites as m or u, and initialize parameters. According to the nucleosome occupancy principle, on the same cfDNA fragment, compared with unmethylated CG, the methylated CG point is farther from the center of the fragment. Because cfDNA is essentially DNA wrapped in nucleosomes (naked DNA is digested by nucleases during apoptosis because there is no protection), the closer to the nucleosome center, due to the influence of histone space occupancy, the methyl group is not easy to bind to DNA, and the DNA at the edge of the nucleosome is not histone-entangled, and the spatial state is more favorable for methyl groups, so the present invention always assumes that the average distance of the CpG point in the m state to the center of the fragment is farther. The parameters after initialization are applied to the hidden Markov model (HMM), and a Markov chain is constructed for each cfDNA fragment, wherein the parameters used are the l, d, c fragment features obtained by actual detection, and the m / u labels generated by GMM. Since the state of each CpG point is affected by the adjacent CpG in the same fragment in the Markov process, the Baum-Welch algorithm is used to estimate the maximum likelihood parameters in the WGS data set. After parameter estimation, a weighted healthy leukocyte DNA methylation prior is used to calculate the CpG posterior probability of the hidden state, and the Viterbi algorithm is used to estimate the optimal state (m or u) of each CpG in each fragment. Finally, all fragments are summarized and the methylation state of each CpG is obtained according to the genomic coordinates. Unlike the traditional HMM model that assumes that the transition probabilities between CpGs are equal, the present invention uses a non-uniform model to estimate different transition probability matrices given the distances between different CpGs. The Kullback-Leibler distance is used to estimate whether the model converges during the iteration process, and the number of CpGs covered by the fragments during the decoding process is unlimited.

[0091] The above calculation process can be implemented by the Java open source program FinaleMe, for example, using the function org.cchmc.epifluidlab.finaleme.utils.CpgMultiMetricsStats to extract fragment features and GMM initialization parameters, and the function org.cchmc.epifluidlab.finaleme.hmm.FinaleMe to perform HMM modeling and decoding, and the model parameters are adjusted according to actual conditions. For the software installation process and operation commands, see ReadMe (https: / / github.com / epifluidlab / FinaleMe).

[0092] The whole genome is divided into windows of equal length according to the gradient, and the average methylation level of each window of each sample is calculated by dividing the sum of the number of reads of all CpG sites marked as m in the window by the sum of the total number of reads of all CpG sites m and u in this interval. The methylation spectrum inferred by each WGS is highly consistent with the actual methylation spectrum detected by WGBS, and the denser the CpG sites, the higher the correlation ( Figure 2 ), the number of CpGs ≥ 40, and the Pearson correlation coefficient of the 2kb window was higher than 0.95 ( Figure 3 ).

[0093] Example 3

[0094] Deconvoluting DNA methylation signals at the fragment level

[0095] The published human-specific methylation reference atlas A (Loyfer N, Magenheim J, Peretz A, et al. A DNA methylation atlas of normal human cell types [J]. Nature, 2023, 613 (7943): 355-364) was used, which includes 39 specific cell types and 8785 marker intervals. Each cfDNA fragment was labeled as U (unmethylated), M (methylated) or X (mixed) according to the number of methylated and unmethylated CpGs, and the proportion of U / X / M fragments covering at least 4 CpGs at each marker was calculated (i.e., within each marker interval, cfDNA fragments covering at least 4 CpGs were screened, and then each cfDNA segment was labeled as U / X / M according to the set threshold, and then the proportion of U fragments in all fragments within this marker interval was calculated). Threshold setting: For U fragments, the number of methylated CpGs is ≤25%; for M fragments, the number of methylated CpGs is ≥75%. For the reference atlas A, generate a 8785×39 matrix that records the U ratio, A i,jRepresents the U ratio of the i-th marker in the j-th cell type. For an input sample, a 8785×1 vector b is generated to record the U ratio.

[0096] For vector b, non-negative least squares (NNLS) is applied by minimizing |A×xb| 2 , and satisfies non-negative x, and is normalized to Σ j x j =1, thus inferring the coefficient vector x. x is the estimated relative abundance of cell types and tissues that contribute to cfDNA. x is between 0 and 1. The larger the value, the greater the possibility that the cfDNA originates from the corresponding cell or tissue.

[0097] For WGBS data, the software wgbstools (Loyfer N, et al. Nature, 2023) is used to generate pat files directly from bam. The pat file contains 4 columns: chromosome, CpG index on the genome (1-28,217,448), the specific methylation pattern of CpG covered by each fragment (C-methylated, T-unmethylated, T-unknown status), the number of occurrences of this pattern fragment ( Figure 4 ). This file line represents a cfDNA fragment. Deconvolution of the pat file using UXM software: The top three sources of cfDNA in healthy people are white blood cells, megakaryocytes, and erythroid progenitor cells. The ranking of contributions to cfDNA in liver cancer is white blood cells, endothelial cells, hepatocytes, megakaryocytes, etc. ( Figure 5a ), this map is consistent with the published results. In view of the interference of the hematopoietic system on the tracing of cancer tissues, after removing the background of the hematopoietic system, the first classification accuracy of liver cancer tissue tracing was 90% (18 / 20), and the second classification accuracy was 100% (20 / 20) ( Figure 5b ).

[0098] Since WGS is not bisulfite-converted, the original alignment file does not have methylation information, and wgbstools cannot be used to generate pat files directly from bam. Therefore, it is necessary to derive pat files from the methylation patterns inferred by HMM. The derivation method is as follows: First, the CpG sites in the whole genome are indexed according to the coordinates. Secondly, the CpG coordinates covered by each cfDNA fragment are extracted from the GMM output file and its methylation pattern is preliminarily determined. The methylation pattern is corrected using the HMM output file, and then the coordinates are replaced by the index, and the pat format is output, and finally UXM deconvolution is performed. It is worth noting that for CpG points that are empty a priori in leukocytes and not covered by sequencing, placeholder processing is required when generating pat files.

[0099] The WGS traceability spectrum is highly consistent with the WGBS: the hematopoietic system cells such as white blood cells, megakaryocytes, and erythrocyte progenitor cells contribute most to healthy cfDNA, and the solid tissues are endothelial tissue and liver. The results are consistent with the reported conclusions. The top cells or solid tissues that are the sources of liver cancer cfDNA are white blood cells, endothelial cells, megakaryocytes, hepatocytes, etc. ( Figure 6a After removing the hematopoietic system background, the first classification accuracy of liver cancer tissue tracing was 85% (17 / 20), and the second classification accuracy was 95% (19 / 20) ( Figure 6b ).

Claims

1. A method for predicting WGScfDNA methylation level, wherein the method is applied for non-therapeutic and diagnostic purposes, characterized in that: The steps include: Step 1: cfDNA extraction, library construction, and sequencing; Step 2: Remove the adapters and low-quality sequences from the sequencing data to obtain high-quality data; Step 3: Align the high-quality data to the human reference genome, remove duplicate sequences, and obtain a sequence alignment file (bam); Step 4: Extract the fragment characteristics of each cfDNA fragment from bam: fragment length, normalized sequencing depth, and the distance from each CpG site to the center of the fragment; Step 5: Use a Gaussian mixture model to mark the CpG sites covered by each fragment of the WGS data as unmethylated (u) and methylated (m) states, and initialize the non-uniform hidden Markov model (HMM) parameters; Step 6: Apply the fragment features obtained in step 4 and the unmethylated (u) and methylated (m) states of the CpG site markers as HMM parameters to the HMM model, construct a Markov chain for each fragment, perform model parameter fitting and solve, and obtain the methylation pattern state of each cfDNA fragment.

2. The WGScfDNA methylation level prediction method according to claim 1, characterized in that: The method also includes step 7, calculating the average methylation level of each window of each sample, and the calculation method is: the sum of the number of reads of all CpG sites marked as m in the window divided by the sum of the total number of reads of all CpG sites m and u in this interval.

3. The WGScfDNA methylation level prediction method according to claim 1, characterized in that: In the step 1, the sequencing method is WGS; in the step 5, it is always assumed that the average distance from the CpG point in the m state to the center of the fragment is larger.

4. The WGScfDNA methylation level prediction method according to claim 1, characterized in that: In step 6, an HMM model is constructed using all fragments covering at least 7 CpGs; in the parameter fitting solution, the Viterbi algorithm is used to decode the optimal methylation state (u or m) of each CpG in each fragment, the Baum-Welch algorithm is used to estimate the maximum likelihood parameters of the WGS data set, and the Kullback-Leibler distance is used to estimate whether the model converges during the iteration process; the weighted healthy leukocyte DNA methylation prior is used to calculate the CpG posterior probability of the hidden state.

5. A method for inferring the origin of tumor tissue based on methylation levels, wherein the method is applied for non-therapeutic and diagnostic purposes, characterized in that: The steps include: Step a, obtaining the methylation pattern status of each cfDNA fragment in the sample by the above method; Step b, labeling each fragment as U (unmethylated), M (methylated), or X (mixed) according to the number of methylated and unmethylated CpGs; Step c, using a public human-specific reference methylation map, calculate the ratio of U / X / M fragments with at least k CpGs in each marker region and all cell types, assuming that the reference methylation map has N marker regions, Ai,j represents the U ratio of the i-th marker in the j-th cell type; Step d: given a methylation pattern state obtained from the WGS sequencing result of a cfDNA as an input sample, in each marker interval, screen out cfDNA fragments that cover at least 4 CpGs, and then mark each cfDNA segment as U / X / M according to the set threshold, and then count the proportion of U fragments in all fragments within this marker interval to generate an N×1 vector b; by minimizing |A*Xb|2, and satisfying non-negative X, and normalizing to Σ j X j =1, thereby calculating the coefficient vector x; x is the relative abundance of cell types and tissues that contribute to cfDNA, and x is between 0 and 1. The larger the x, the greater the possibility that the cfDNA originates from a certain tissue; in the step c, k=4 is set, and the method for determining the ratio of U / X / M fragments is: for U fragments, methylated CpG ≤ 25%; for M fragments, methylated CpG ≥ 75%; for X fragments, methylated CpG 25% <X<75%。 6. A WGS cfDNA methylation level prediction device, characterized in that: include: An extraction module, used for extracting cfDNA samples; A sequencing module, used to sequence the extracted cfDNA samples; The feature value acquisition module is used to extract the fragment features of each cfDNA fragment from the sequence alignment file: fragment length, normalized sequencing depth, and the distance from each CpG site to the center of the fragment; The label module is used to label the CpG sites covered by each fragment of the WGS data as unmethylated (u) and methylated (m) states using a Gaussian mixture model, and initialize the parameters of the non-uniform hidden Markov model (HMM); The methylation pattern state prediction module is used to apply the obtained fragment features and the unmethylated (u) and methylated (m) states of the CpG site markers as HMM parameters to the HMM model, construct a Markov chain for each fragment, perform model parameter fitting and solution, and obtain the methylation pattern state of each cfDNA fragment.

7. The WGS cfDNA methylation level prediction device according to claim 6, characterized in that: The sequencing module uses WGS sequencing; the average distance from the CpG point in the m state to the fragment center is set to be larger in the label module; the HMM model is constructed using all fragments covering at least 7 CpGs; in the parameter fitting solution in the methylation pattern state prediction module, the Viterbi algorithm is used to decode the optimal methylation state (u or m) of each CpG in each fragment, the Baum-Welch algorithm is used to estimate the maximum likelihood parameters of the WGS data set, and the Kullback-Leibler distance estimation model converges during the iteration process; and the weighted healthy leukocyte DNA methylation prior is used to calculate the CpG posterior probability of the hidden state.

8. A device for inferring the origin of tumor tissue based on methylation level, characterized in that: The WGScfDNA methylation level prediction device according to claim 6 further comprises: A fragment methylation status labeling module is used to label each fragment as U (unmethylated), M (methylated), or X (mixed) according to the number of methylated and unmethylated CpGs; A methylation ratio calculation module is used to calculate the ratio of U / X / M fragments with at least k CpGs in each marker region and all cell types using a public human-specific reference methylation map. Assuming that the reference methylation map has N marker regions, Ai,j represents the U ratio of the i-th marker in the j-th cell type; The tissue tracing inference module is used to take the methylation pattern state obtained from the WGS sequencing result of a cfDNA as the input sample, screen out cfDNA fragments covering at least 4 CpGs in each marker interval, and then mark each cfDNA segment as U / X / M according to the set threshold, and then count the proportion of U fragments in all fragments in this marker interval to generate an N×1 vector b; by minimizing |A*Xb|2, and satisfying non-negative X, and normalizing to Σ j X j =1, thereby calculating the coefficient vector x; x is the relative abundance of cell types and tissues that contribute to cfDNA, and x is between 0 and 1. The larger the x, the greater the possibility that the cfDNA comes from a certain tissue.

9. The device for inferring tumor tissue origin based on methylation level according to claim 8, characterized in that: In the methylation ratio calculation module, k is set to 4, and the ratio of U / X / M fragments is determined as follows: for U fragments, methylated CpG ≤ 25%; for M fragments, methylated CpG ≥ 75%; for X fragments, methylated CpG 25% <X<75%。 10. A computer-readable medium recording a program capable of executing the WGS cfDNA methylation level prediction method of claim 1 or the tumor tissue origin inference method based on methylation level of claim 5.

Citation Information

Patent Citations

  • Methods and systems for high-depth sequencing of methylated nucleic acid

    CN112823213A

  • Methods for genome characterization

    US11788135B2

  • Methods for genome characterization

    WO2018027176A1

Cited By

  • Method, device, apparatus and program product for tumor detection

    CN120148625A