Ns126:Manuscript
Abstract
[Background]Circulating cell-free DNA methylation in plasma have been demonstrated to be powerful potential in non-invasive cancer early diagnosis. However, the genome-wide profile of DNA methylation for the free circulating DNA methylation has not been depicted in a signal base resolution. [Method] In the present study, we carried out a genome-wide survey of single-base resolution methylome across 45 plasma from cancer patients, 30 normal plasma and solid cancer tissues with RRBS, SeqCap and BSPP assays. [Result] In the discovery stage, both RRBS and SeqCap dataset identified large number classification potential biomarkers (N=1592 and 516, respectively) and high level prediction ability with Random Forest model (accuracy=100% and 98.18%, respectively). The above biomarkers were validated in the BSPP dataset and we found the methylation status of 30 derived regions in plasma could explain 93.75% cancers incidence and the specificity is 100%. What’s more, the pattern of the DNA methylation fragment transferring were depicted with paired tissue-plasma methylation dataset. We identified 1590 fragments significantly prefer-selected in the releasing process of the DNA methylation fragment from solid tissues to plasma (P<1.98*10-6, binomial test, Bonferroni correction). These fragments were located in the regions of 1190 genes. Function enrichment analysis showed these genes were significantly associated with cancer relevant biological functions, including embryonic morphogenesis, regulation of transcription, neuron differentiation, regionalization, tissue morphogenesis, transcription factor activity, sequence-specific DNA binding, transcription regulator activity. [Conclusion] Methylation haplotype loading based DNA methylation biomarker would be potential cancer diagnosis biomarker and DNA methylation fragment releasing process were regulated rather than a random event.
Background
DNA methylation is one of most important inherited epigenetic modification in the human genome. It is involved in most human biological and cellular, physiological, pathological changes. Current evidence shows DNA methylation could be powerful tag for cell differentiation, aging estimation, forensic identification and disease status, especially in cancer. DNA methylation can be used as the sensitive biomarker for diagnosis, prognosis surveillance and chemo-response tracking. More and more genome-wide DNA methylation profile for cancers has been completed with different strategies (1). Most of the current method is preferred to microarray or capture-based sequencing technology. However, the coverage is limited and the signal is a fuzzy evaluation of DNA methylation for the population of a specific tissue. Although capture-based DNA methylation sequencing technology, such as MBD-seq and MeDIP-seq have low cost, they cannot obtain single-based methylation status, to decrease the cost of the design. In addition, majority of these profiles are based on solid tissues rather than circulating free-cell DNA, which is considered as the most powerful media for non-invasive diagnosis.
The survival time is highly dependent on the stage of the cancer in which the patient were diagnosed, especially for NSCLC, pancreatic cancer and colon cancer(2). For example, while the overall 5-year survival rates for late stage III and IV of NSCLC patients were just 5%-14% and 1% respectively, the rate could come up to 50% for the early stage of the NSCLC patients who are typically treated with surgery (3). However, early diagnosis of cancer also require another 3 prerequisites, including early biomarkers, non-invasive detection and high specificity. The methylation detection in circulating cell-free DNA do provide such platform for cancer non-invasive early diagnosis.
Human peripheral blood contains low levels of DNA molecules from other tissues or cell types, such as circulating cancer stem cells or cell-free DNA (cf-DNA) from apoptotic cancer cells in cancer patients. Analysis of DNA epigenetic mutations in the circulating cell-free DNA is becoming to a prospective trend for creation of noninvasive methods for the diagnosis and treatment efficiency monitoring in cancer. In the past few year, the basic characteristics of cfDNA has been depicted. Plasma rather than serum was considered to be the perfect media to collect the cfDNA within 8 hours of the storage (4). Two main fragment components of cfDNA, 180bp and 350bp, could be found in non-white-cell contaminated plasma. The yield of cf-DNA in the plasma of cancer patients is a very low concentration which ranged from 1.0-100 ng/ml while it ranges from 1.0-10 ng/ml in healthy individuals. To detect and quantify such low abundance DNA molecules, some significant regions, hyper-methylated in the circulating DNAs derived from cancer cell while non-methylated in white blood cell (WBC), should be identified.
In the present study, we carried out a comprehensive genome-wide DNA methylation analysis across 45 plasma from cancer patients, 30 normal plasma and ** solid cancer tissues with RRBS, SeqCap and umi-BSPP assays. 68 cancer and 25 normal samples were enrolled in RRBS assay. 40 cancer and 25 normal samples were enrolled in SeqCap assay. 16 cancer plasma and 16 normal plasma samples were enrolled in umi-BSPP assay. Methylation haplotype was constructed as our previous method and Methylation haplotype loading (MHL) was proposed to assess the level or the proportion of the DNA methylation. Diagnostic biomarker based on MHL were identified and validated by RRBS, SeqCap and umi-BSPP dataset. In addition, 15 solid cancer tissues and corresponding plasma samples were collected and then the methylome were detected by RRBS assay to evaluate the difference of the genome-wide DNA methylation between solid tissues and circulating cell free DNA.
Method and Materials
Clinical sample and DNA collection
NSCLC samples and corresponding normal lung tissues for validation study in Chinese population were obtained from 150 patients who underwent pulmonary resection for primary NSCLC at Changhai Hospital, Shanghai, China. The study was approved by Fudan University and Changhai Hospital and Informed consents were obtained from the patients. Exclusion criteria included subjects with a family history of lung cancer, previous radiotherapy, and chemotherapy or adjuvant therapy before surgery. All tissues were immediately frozen at -80℃ after surgical resection. Histological examination and tumor-node-metastasis classification were conducted according to World Health Organization classification criteria (5) and AJCC Cancer Staging Manual, 7th Edition (6), respectively. Age, gender, smoking status, histology type, TNM stage and differentiation status were collected as the covariates when conducting association between DNA methylation and disease status. Smoking status was assigned to binary status: never and ever smoking. TNM stage was assigned to early stage (I and II) or late stage (III and IV) when it is necessary so that the sample size can be big enough to get the efficient statistic power.
Umi-BSPP and Methylation Haplotype Loading (MHL)
To detect and quantify such low abundance DNA molecules, we focus on regions in the genome in which there are major differences in DNA methylation between whole blood and cancer cells which can be used for assay and detection. For example, a region containing 6 CpG sites might be completely unmethylated in whole blood, and fully methylated in cancer cells. If the whole blood sample contain 3% of cancer DNA, then we would detect a 3% methylation with an ideal assay. However, all methylation assays have technical errors, such as incomplete bisulfite conversion, incomplete enzyme digestion, sequencing error. Typically all the technical errors combined can contribute to ~1-2%. With the presence of these errors, a 3% methylation cannot be confidently detected. Such technical errors greatly compromised the sensitivity and confidence in detecting and quantifying fetal DNA molecules. Methylation haplotyping analysis can dramatically improve the discriminating power, as technical errors typically occur independently on all DNA molecules at random locations. In contrast, the 3% fetal DNA molecules are fully (or almost fully) methylated at all CpG sites, whereas all maternal DNA molecules are not methylated. In other words, the methylation status at multiple CpG sites on the same molecules are all “linked”. Using methylation haplotypes of four or more CpG sites, one can confidently identify rare DNA molecules at 0.01% or even lower, at least two orders of magnitude below the technical errors (1-2% per site). We were the first group that developed an analytical framework for methylation haplotype and linkage disequilibrium analysis (Shoemaker et al, 2009). In this invention, we developed an assay for digital quantification of methylation (see below), and greatly extended the previous framework to haplotype-based methylation marker analysis. In addition, combining information from multiple markers will further improve the sensitivity and robustness in the presence of biological variability. Finally, cell-free DNA typically come from apoptotic cells, and are in small fragments (Chan et al. 2004). In contrast, whole blood DNA typically have larger sizes (at least kilobases) even after DNA extraction. Using targeted methylation sequencing to analyze haplotypes from DNA molecules of different sizes adds another level of stringency in separating rare cancer or fetal DNA from whole blood DNA. In summary, this invention can achieve an ultra-high sensitivity for detecting rare DNA species from mixed DNA (such as whole blood DNA) using some combinations of the three concepts: (i) multi-locus methylation haplotype analysis (Figure 1); (ii)integrative analysis of multiple marker regions; and (iii) differential haplotype analysis of DNA fragments with different sizes (Figure 2). The invention can be implemented as genetic screening or diagnostic tests for non-invasive prenatal diagnosis, non-invasive monitoring of tumor loads in cancer patients after treatments, or early-stage cancer detection. We have implemented this invention using a target methylation sequencing technology (Bisulfite Padlock Probes, or BSPP, Deng et al, 2008; Diep et al. 2012) developed by the Zhang lab. Note that alternative technologies, such as micro-droplet PCR (Komori et al. 2011) or Selector probes (Johansson et al. 2011), can also potentially be used with some differences in the requirement of input materials and/or cost. MeDiP is another alternative experimental method for data collection (Papageorgiou et al. 2010), with disadvantages including low efficiency, high cost, and low reliability (Tong et al. 2012). Therefore, this invention should cover the aforementioned three key concepts, not a specific implementation based on the BSPP technology.
Regardless of the specific sample preparation methods or sequencing platforms used, our method takes the bisulfite sequencing reads (single-ends or paired-ends) as the input. We derived methylation haplotypes and their abundance from the raw sequencing reads. Each haplotype represent the combination of binary methylation status (methylated or unmethylated) at multiple CpG sites of one sequencing read. For sample preparation methods (such as umi-RRBS, or hybridization-based target capture) that allow identifying multiple clonal sequencing reads originated from the same template DNA molecules, we also derive the consensus haplotypes from the clonal reads to improve the accuracy and avoid over-dispersion of haplotype counts. The probability that a methylation haplotype is present in a sample (or a pool of samples) is determined by the frequency if the exact haplotype is observed, or estimated from the methylation levels of individual CpG sites and the technical error rates (sequencing errors, bisulfite conversion errors) assuming no linkage between adjacent sites (or P(M1M2)=P(M1)*P(M2) where M1M2 is the probability of two-locus methylated haplotype, P(M1) and P(M2) are the methylation level at the two loci). For each methylation haplotype from the patient plasma of interest, we determined the likelihoods of it originating from the pooled tumor primary biopsies data and from the pooled normal plasma data, and calculated the negative log likelihood ratio. A methylation haplotype is classified as the tumor haplotype when the negative log likelihood ratio is above 3. To improve the signal-to-noise ratios, we focus on methylation haplotypes that contain four or more CpG sites.
We have demonstrated the proof-of-concept for detecting low-abundant tumor DNA in blood DNA based on methylation haplotypes, both in synthetic DNA mixtures and clinical plasma samples of three cancer types. For the first demonstration using synthetic DNA mixtures, we searched marker regions by extensive analyses of published and unpublished DNA methylation data on whole blood, cancers, and placenta. We have identified 12,446 candidate regions that exhibit large methylation difference between whole blood and three types of cancer cell lines (pancreatic cancer, glioblastoma multiforme, lung cancer), as well as 1,230 regions different between whole blood and placenta. We designed umi-BSPP (Figure 3), which is an improved version of Bisulfite Padlock Probes (BSPP, Deng et al, 2009; Diep et al. 2012) for targeted methylation sequencing of these candidate regions of whole blood and a panel of five cancer (pancreatic cancer and glioblastoma) cell lines. These probes also contain built-in Unique Molecule Identifier (Kivioja et al. 2011), so that we can perform deep sequencing and true single-molecule counting to avoid quantification artefacts due to DNA amplification. We have designed and synthesize 19,109 long oligonucleotides for the candidate targets.
Computational analysis of methylation haplotypes starts with bisulfite sequencing reads mapped to the reference genome (using common bisulfite read mapping algorithms, such as bisReadMapper, bisMark) in the format of bam files. We derived consensus sequencing reads based on UMI (if available), then determined the methylation haplotypes on multiple CpG sites in single consensus sequencing reads (Figure 4). The haplotypes and their counts in all genomic regions assayed are reported. For selection of the marker set that can classify plasma samples, we define a methylated haplotype load (MHL) for each candidate region, which is the normalized fraction of methylated haplotypes at different length:
The characteristics of the free circulating DNA derived from apoptotic or necrotic were quite different.
In RRBS, DNA fragments of 40-220 base pair are representative of the majority of promoter sequences and CpG islands
The methylation status of target genes in circulating DNA could be evaluated by two kinds of methods including bisulfite assistant methylation assay (RRBS, BSPP) and bisulfite-free methylation assay (MeDIP, MBD-seq).
The technique of DNA methylation has been widely applied to identify clinical associated biomarker or monitor biomarkers for disease prognosis. Genome-wide DNA methylation profiles has been in the solid tissues (1).
The main contribution of our study includes as the following:
The methylation haplotype loading is significantly associated with cancer progress, such as TNM stages.
In present study, single-base genome-wide DNA methylation analysis were conducted in 75 solid cancer, adjacent tissues, cell-free circulating DNA from cancer and normal samples. Methylation haplotype based DNA methylation diagnosis to cancer was evaluated and most powerful biomarker were identified
Bioinformatics and Statistics
We need to look into each cancer individually on the list, maybe rank the cancer samples from most preferable to least preferable. Once we get optimistic results, we can further ask for samples from other centers. We cannot realistically ask specifically for each stage of cancer, although having samples from multiple stages might be better. prevalence treatment options for early detection (stage 1) current diagnostic methods We need to ask for buffy coat (1 tube) and matched serum samples (2 tubes) per patient We need to ask for primary tumor samples, they might be Formalin-fixed paraffin-embedded (FFPE) samples check Blueprint data to see how successful we were at capturing DNA purified from FFPE samples look into the kit for fixing degraded DNA Noi said we successfully captured with 50 ng converted DNA before, and it is possible to obtain this much from 2 tubes of serum samples, however, we can also look into the amplified DNA from Illumina (can we perform capture on these samples?)
RRBS option Dinh need to check the bayes classification algorithm Can it be used on RRBS data? We can downsample high coverage data to lower the effect of clonal reads on the analysis. Noi to check the input requirement for performing RRBS, can we perform RRBS in parallel on the test samples.
In the biomarker discover stage, the biomarker whose methylation were 100% un-methylated in normal plasmas were enrolled while any regions which were detected to be methylated in normal plasmas were filter out from the candidates. In the procedure of differential methylation test, regions whose variance in total samples were stage at lowest 30% quantile were removed to decrease the burden of multi-test correction. For the random forest prediction, optimal parameters were tuned before the tanning of the model the best number of tries and tress were determined by the grid search method with lowest out of beg prediction error. Intersection analysis of the genome position were conducted by BEDTOOLS and the regions whose distance less than 25bp were taken as the same biomarker region.
R packages of IRanges, Biostrings, stringr, randomForest, impute, rpart, e1071, biclust were used in the the statistic and bioinformatics analysis.
Potential biomarker based on published papers from NCBI were extracted for lung cancer (74 papers), colon cancer (8 papers) and pancreatic cancer (13 papers) with the strategy of title including “methylation” and corresponding cancer symbols as well as abstract including “Diagnosis”.