HK40112829A - Cancer detection and classification using methylome analysis - Google Patents

Description

Cancer detection and classification using methylome analysis

This application is a divisional application of Chinese patent application No. 201880059089.5, entitled "Cancer Detection and Classification Using Methylome Analysis" (the corresponding PCT application was filed on July 11, 2018, with application number PCT/CA2018/000141).

Related applications

This application claims priority to U.S. Provisional Patent Application No. 62/531527, filed July 12, 2017, which is incorporated herein by reference.

Technical Field

This invention relates to cancer detection and classification, and more specifically, to the use of methylome analysis for cancer detection and classification.

Background Technology

The use of circulating cell-free DNA (cfDNA) as a source of biomarkers has been rapidly developed in oncology [1]. The use of DNA methylation mapping of cfDNA as a biomarker could have a significant impact on the field of liquid biopsy, as it can identify the tissue of origin in a minimally invasive manner [2] and can be used to classify and stratify cancer patients by cancer type and subtype [3]. Furthermore, the use of whole-genome DNA methylation mapping of cfDNA can overcome the critical sensitivity problem of detecting circulating tumor DNA (ctDNA) in patients with early-stage cancer without radiographic evidence of disease. Existing ctDNA detection methods are based on sequencing mutations and have limited sensitivity, partly due to the limited number of recurrent mutations that can be used to distinguish tumor and normal circulating cfDNA [4, 5]. On the other hand, whole-genome DNA methylation mapping takes advantage of a large number of epigenetic alterations that can be used to distinguish circulating tumor DNA (ctDNA) from normal circulating cell-free DNA (cfDNA). For example, certain types of tumors (e.g., ependymoma) may have extensive DNA methylation aberrations without any obvious recurrent somatic mutations [6].

Some methods for capturing cell-free methylated DNA are described in WO 2017/190215, which is incorporated herein by reference.

Summary of the Invention

In one aspect, a method is provided for detecting the presence of DNA from cancer cells in a subject, comprising: providing a sample of cell-free DNA from the subject; performing library preparation on the sample to allow subsequent sequencing of the cell-free methylated DNA; adding a first amount of filler DNA to the sample, wherein at least a portion of the filler DNA is methylated, and then optionally denaturing the sample; capturing the cell-free methylated DNA using a binding agent selective for methylated polynucleotides; sequencing the captured cell-free methylated DNA; comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; and identifying the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals.

In one aspect, a method is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: receiving sequencing data of cell-free methylated DNA from a subject sample; comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identifying the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals; and, if DNA from cancer cells is identified, further identifying the cancer cell tissue of origin and cancer subtype based on the comparison.

In one aspect, a computer-implemented method is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: receiving sequencing data of cell-free methylated DNA from a subject sample by at least one processor; at at least one processor, comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identifying the presence of DNA from cancer cells at at least one processor if the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from cancer individuals; and if DNA from cancer cells is identified, further identifying the cancer cell tissue of origin and cancer subtype based on the comparison.

In one aspect, a computer program product is provided for use in conjunction with a general-purpose computer having a processor and memory connected to the processor, the computer program product including a computer-readable storage medium thereon having computer mechanisms encoded thereon, wherein the computer program mechanisms can be loaded into the computer's memory and cause the computer to perform the methods described herein.

In one aspect, a computer-readable medium is provided having a data structure thereon for storing the computer program product described herein.

In one aspect, an apparatus is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the apparatus comprising: at least one processor; and an electronic memory in communication with the at least one processor, the electronic memory storing processor-executable code that, when executed at the at least one processor, causes the at least one processor to: receive sequencing data of cell-free methylated DNA from a subject sample; compare the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identify the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals; and, if DNA from cancer cells is identified, further identify the cancer cell tissue of origin and cancer subtype based on the comparison.

In one aspect, a method is provided for detecting the presence of DNA from cancer cells and determining the location of cancer cells from two or more possible organs, the method comprising: providing a sample of cell-free DNA from a subject; capturing cell-free methylated DNA from the sample using a binding agent selective for methylated polynucleotides; sequencing the captured cell-free methylated DNA; comparing the sequence pattern of the captured cell-free methylated DNA with the DNA sequence patterns of two or more groups of a control individual, each of the two or more groups having localized cancer in a different organ; and determining which organ the cancer cells originated from based on a statistically significant similarity between the methylation pattern of the cell-free DNA and one of the two or more groups.

Attached Figure Description

These and other features of the preferred embodiments of the invention will become more apparent in the following detailed description, in which reference is made to the accompanying drawings, wherein:

Figures 1A-1G show that cfDNA methylome analysis is a highly sensitive method for enriching and detecting ctDNA in small amounts of input DNA. A) Computer simulation of the probability of detecting at least one epigenetic mutation as a function of the number of DMRs studied (rows) and sequencing depth (x-axis) as ctDNA concentration (columns). B) Genome-wide Pearson correlations between DNA methylation signals from input DNA of the HCT116 cell line fragmented to simulate plasma cfDNA, from 1 to 100 ng. Each concentration has two biological replicates. C) DNA methylation maps (green tracks) obtained from cfMeDIP-seq of input DNA from HCT116 at different concentrations, plus RRBS (Reduced Representative Bisulfite Sequencing) HCT116 data obtained from ENCODE (ENCSR000DFS) and WGBS (Whole Genome Bisulfite Sequencing) HCT116 data obtained from GEO (GSM1465024). For the heatmap (RRBS tracks), yellow indicates methylation, blue indicates unmethylation, and gray indicates no coverage. DE) Serial dilutions of the CRC cell line HCT116 were performed in the multiple myeloma (MM) cell line MM1.S. cfMeDIP-seq was performed on pure HCT116 DNA (100% CRC), pure MM1.S DNA (100% MM), and 10%, 1%, 0.1%, 0.01%, and 0.001% CRC DNA diluted in MM DNA. All DNA fragments were simulated to mimic plasma cfDNA. We observed near-perfect linear correlations between the observed and expected values of the number of DMRs (D) and (E) DNA methylation signals (expressed as RPKM) in these DMRs ( ^r² = 0.99, p < 0.0001). F) In the same dilution series, known somatic mutations were detected at a fraction of 1/100 alleles only by ultra-deep (>10,000X) targeted sequencing beyond the background sequencer and polymerase error rates. The read fractions for each mutation site or insertion/deletion are shown in the CRC cell line. G) Percentage of ctDNA (human) in the plasma of mice carrying patient-derived xenografts (PDX) from two colorectal cancer patients to total cfDNA (human + mouse).

Figures 2A-2D show that methylome analysis of plasma cfDNA can be used for tumor classification. A) Schematic illustration of the machine learning classifier construction method used for cancer classification. B) Heatmap of DMR contained in a multi-class elastic net machine learning classifier. Classifiers were trained on plasma DNA samples from healthy donors (n=24), lung cancer (n=25), breast cancer (n=25), colorectal cancer (n=23), acute myeloid leukemia (AML) (n=28), and glioblastoma multiforme (GBM) (n=71). Hierarchical clustering method: Ward. C) 2D visualization of cancer type-related DMRs identified in 10% or 25% of the models using tSNE (t-distributed random neighborhood embedding). D) Performance metrics of the multi-cancer classifier based on plasma cfDNA methylation. Area under the recipient operating curve (auROC) is shown on the Y-axis after generating a 50x elastic net machine learning classifier.

Figures 3A-3B show the validation of the multi-cancer classifier against individual cohorts. A) shows the ROC curves for the individual validation of the multi-cancer classifier against lung cancer (LUC) (n=55 LUCs vs. others n=97), AML (n=35 AMLs vs. others n=117), and a healthy donor cohort (n=62 healthy donors vs. others n=90). B) shows the ROC curves for the individual validation of the multi-cancer classifier against early-stage LUC (n=32 stage I-II LUCs vs. others n=97) and late-stage LUC (n=23 stage III-IV LUCs vs. others n=97).

Figures 4A-4E demonstrate that methylome analysis of plasma cfDNA can be used for tumor subtype classification. A) 2D visualization of cancer subtype-related DMRs using tSNE (t-distributed random neighborhood embedding). Breast cancer subtypes showed the ability to distinguish patients with tumors exhibiting different gene expression patterns and transcription factor activities (ER status) and different tumor copy number aberrations (HER2 status). AML subtypes showed the ability to distinguish patients with tumors exhibiting different rearrangements (FLT3 status). Glioblastoma multiforme (GBM) subtypes showed the ability to distinguish patients with tumors exhibiting different point mutations (IDH gene mutation status). Lung cancer subtypes showed the ability to distinguish patients with tumors exhibiting different histological features (adenocarcinoma vs. squamous cell carcinoma vs. small cell carcinoma), which has prognostic and therapeutic significance. B) A heatmap shows that the top DMR accurately distinguishes the three breast cancer subtypes in breast cancer plasma samples. C) A heatmap shows that the top DMR accurately distinguishes the FLT3-ITD status in AML patient plasma samples. D) The heatmap shows that the top DMR can accurately distinguish the IDH gene mutation status in plasma samples from patients with glioblastoma multiforme (GBM). E) The heatmap shows that the top DMR can accurately distinguish three lung cancer histologies in lung cancer plasma samples.

Figure 5 illustrates a suitably configured computer device and associated communication networks, devices, software, and firmware to provide a platform for implementing one or more embodiments described herein.

Figures 6A-6C show the sequencing saturation analysis and quality control. A) This figure shows the saturation analysis results of cfMeDIP-seq data from each replicate of HCT116 DNA fragmented to mimic plasma cfDNA, analyzed using the Bioconductor software package MEDIPS. B) Two replicate tests were performed for four starting DNA concentrations (100, 10, 5, and 1 ng) of the HCT116 cell line. The specificity of the reaction was calculated using methylated and unmethylated Arabidopsis DNA. The enrichment fold ratio was calculated using genomic regions of fragmented HCT116 DNA (primers for methylated testis-specific H2B, TSH2B0, and unmethylated human DNA regions (GAPDH promoter)). The horizontal dashed line indicates an enrichment fold ratio threshold of 25. Error bars represent ±1 s.e.m. C) The CpG enrichment scores of the sequencing samples show that CpG was significantly enriched in genomic regions from the immunoprecipitated samples compared to the input control. CpG enrichment scores are obtained by dividing the relative frequency of CpG in a region by the relative frequency of CpG in the human genome. Error bars represent ±1 s.e.m.

Detailed Implementation

In the following description, numerous specific details are set forth to provide a thorough understanding of the invention. However, it should be understood that the invention may be practiced without these specific details.

DNA methylation profiling is cell type specific and is disrupted in cancer. Using a robust and sensitive method designed for methylome analysis of trace amounts of circulating cell-free DNA (cfDNA), we identified thousands of differentially methylated regions (DMRs) that distinguish multiple tumor types from each other and from healthy individuals. Methylome analysis of cfDNA is highly sensitive and suitable for detecting circulating tumor DNA (ctDNA) in early-stage patients. A machine learning-derived classifier using cfDNA methylome analysis was able to correctly classify 196 plasma samples from patients with five cancer types and healthy donors based on cross-validation. In individual validation, using the same DMRs identified in plasma cfDNA, the classifier was able to correctly classify AML, lung cancer, and healthy donors, as well as early and late-stage lung cancer. Therefore, methylome analysis of cfDNA can be used for non-invasive early detection of ctDNA and reliably classify cancer types.

In one aspect, a method is provided for detecting the presence of DNA from cancer cells in a subject, comprising: providing a sample of cell-free DNA from the subject; performing library preparation on the sample to allow subsequent sequencing of the cell-free methylated DNA; adding a first amount of filler DNA to the sample, wherein at least a portion of the filler DNA is methylated, and then optionally denaturing the sample; capturing the cell-free methylated DNA using a binding agent selective for methylated polynucleotides; sequencing the captured cell-free methylated DNA; comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; and identifying the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals.

The applicant’s jointly owned applications, U.S. Provisional Patent Application No. 62/331,070, filed May 3, 2016, and International Patent Application No. PCT/CA2017/000108, filed May 3, 2017, describe methods for capturing cell-free methylated DNA and are incorporated herein by reference.

Traditionally, cancers have been classified according to the tissue of origin—for example, colorectal cancer, breast cancer, lung cancer, etc. In modern clinical oncology practice, the ability to differentiate cancer subtypes based on various molecular, developmental, and functional bases has become increasingly important. Treatment decisions often depend on the precise subtype of the cancer, and clinicians may need to identify the subtype before initiating treatment. Examples of cancer subtypes that may influence treatment decisions include (but are not limited to) stage (e.g., early-stage lung cancer treated surgically versus advanced lung cancer treated with chemotherapy), histology (e.g., small cell carcinoma versus adenocarcinoma versus squamous cell carcinoma in lung cancer), gene expression patterns or transcription factor activity (e.g., ER status in breast cancer), copy number aberrations (e.g., HER2 status in breast cancer), specific rearrangements (e.g., FLT3 in AML), specific gene point mutation status (e.g., IDH gene point mutation), and DNA methylation patterns (e.g., MGMT gene promoter methylation in brain cancer).

The methods described in this article are applicable to a variety of cancers, including but not limited to adrenal cancer, anal cancer, bile duct cancer, bladder cancer, bone cancer, brain/CNS tumors, breast cancer, Castleman's disease, cervical cancer, colon/rectal cancer, endometrial cancer, esophageal cancer, Ewing family tumors, eye cancer, gallbladder cancer, gastrointestinal carcinoid tumors, gastrointestinal stromal tumors (GIST), gestational trophoblastic disease, Hodgkin's disease, Kaposi's sarcoma, kidney cancer, laryngeal and hypopharyngeal cancer, leukemia (acute lymphoblastic, acute myeloid, chronic lymphoblastic, chronic myeloid, chronic myelomonocytic), liver cancer, and lung cancer (non-small cell, small cell, lung cancer). Carcinoid tumors, lymphomas, cutaneous lymphomas, malignant mesotheliomas, multiple myeloma, myelodysplastic syndromes, nasal and paranasal sinus carcinomas, nasopharyngeal carcinomas, neuroblastomas, non-Hodgkin lymphomas, oral and oropharyngeal carcinomas, osteosarcomas, ovarian cancers, penile cancers, pituitary adenomas, prostate cancers, retinoblastomas, rhabdomyosarcomas, salivary gland carcinomas, sarcomas—adult soft tissue cancers, skin cancers (basal and squamous cell, melanoma, Merkel cell), small bowel cancers, gastric cancers, testicular cancers, thymic carcinomas, thyroid cancers, uterine sarcomas, vaginal cancers, vulvar cancers, Waldenstrom's macroglobulinemia, Wilms' tumors.

Various sequencing technologies are known to those skilled in the art, such as polymerase chain reaction (PCR) followed by Sanger sequencing. Next-generation sequencing (NGS) technologies, also known as high-throughput sequencing, can also be used, encompassing a variety of sequencing technologies including Illumina (Solexa) sequencing, Roche 454 sequencing, Ion torrent: proton/PGM sequencing, and SOLiD sequencing. NGS can sequence DNA and RNA faster and cheaper than the previously used Sanger sequencing. In some embodiments, the sequencing is optimized for short-read sequencing.

As used herein, the term “subject” means any member of the animal kingdom, preferably a human, and most preferably a human who has, has had, or is suspected of having prostate cancer.

Cell-free methylated DNA is DNA that circulates freely in the blood and is methylated in various known regions of the DNA. Samples (e.g., plasma samples) can be collected to analyze cell-free methylated DNA. Therefore, in some embodiments, the sample is the blood or plasma of a subject.

As used herein, “library preparation” includes list end repair, A-tailing, adaptor ligation, or any other preparation of cell-free DNA to allow for subsequent sequencing of the DNA.

As used in this article, “filler DNA” can be non-coding DNA or can consist of amplicon.

For example, sufficient heat can be used to denature the DNA sample.

In some implementations, the comparison step is based on fitting using a statistical classifier. A statistical classifier using DNA methylation data can be used to assign samples to specific disease states, such as cancer types or subtypes. For the purpose of cancer type or subtype classification, the classifier will consist of one or more DNA methylation variables (i.e., features) within a statistical model, and the output of the statistical model will have one or more thresholds to distinguish different disease states. The specific features and thresholds used in the statistical classifier can be derived from prior knowledge of cancer types or subtypes, prior knowledge of the most potentially useful features, machine learning, or a combination of two or more of these methods.

In some implementations, the classifier is obtained through machine learning. Preferably, the classifier is an elastic net classifier, lasso, support vector machine, random forest, or neural network.

The analyzed genomic space can be whole-genome or preferably limited to regulatory regions (i.e., FANTOM5 enhancers, CpG islands, CpG shores, and CpG frameworks).

Preferably, the percentage of recovered spike-in methylated DNA is included as a covariate to control for variations in pull-down efficiency.

For a classifier capable of distinguishing multiple cancer types (or subtypes) from one another, the classifier will preferably consist of differentially methylated regions from pairwise comparisons of each type (or subtype) of interest.

In some implementations, control cell-free methylated DNA sequences from healthy and cancer individuals are included in a database of differentially methylated regions (DMRs) between healthy and cancer individuals.

In some embodiments, the control cell-free methylated DNA sequences from healthy and cancer individuals are limited to those control cell-free methylated DNA sequences that are differentially methylated between healthy and cancer individuals in cell-free DNA derived from bodily fluids, such as those from serum, cerebrospinal fluid, urine, sputum, pleural fluid, ascites, tears, sweat, Pap smears, endoscopic brush fluid, etc., preferably from plasma.

In some implementations, the sample contains less than 100 ng, 75 ng, or 50 ng of cell-free DNA.

In some embodiments, the first amount of filler DNA comprises about 5%, 10%, 15%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 100% methylated filler DNA, with the remainder being unmethylated filler DNA, preferably 5% to 50%, 10% to 40%, or 15% to 30% methylated filler DNA.

In some embodiments, the first amount of filling DNA is 20 ng to 100 ng, preferably 30 ng to 100 ng, and more preferably 50 ng to 100 ng.

In some embodiments, the cell-free DNA from the sample and the first amount of filler DNA together contain at least 50 ng of total DNA, preferably at least 100 ng of total DNA.

In some embodiments, the length of the filled DNA is 50 bp to 800 bp, preferably 100 bp to 600 bp, and more preferably 200 bp to 600 bp.

In some embodiments, the filler DNA is double-stranded. The filler DNA is double-stranded. For example, the filler DNA can be junk DNA. The filler DNA can also be endogenous or exogenous DNA. For example, the filler DNA is non-human DNA, and in a preferred embodiment, it is λDNA. As used herein, "λDNA" refers to Enterobacterial bacteriophage λDNA. In some embodiments, the filler DNA is not aligned with human DNA.

In some embodiments, the binder is a protein containing a methyl-CpG binding domain. One exemplary protein of this kind is the MBD2 protein. As used herein, a “methyl-CpG binding domain (MBD)” refers to a domain of a protein or enzyme that is approximately 70 residues long and binds to DNA containing one or more symmetrically methylated CpGs. The MBDs of MeCP2, MBD1, MBD2, MBD4, and BAZ2 mediate binding to DNA, with the latter preferentially binding to methylated CpGs in the case of MeCP2, MBD1, and MBD2. Human proteins MECP2, MBD1, MBD2, MBD3, and MBD4 comprise the nucleoprotein families associated with each methyl-CpG binding domain (MBD). Each of these proteins, except MBD3, is capable of specifically binding to methylated DNA.

In other embodiments, the binder is an antibody, and capturing cell-free methylated DNA involves immunoprecipitating the cell-free methylated DNA with an antibody. As used herein, “immunoprecipitation” refers to a technique of precipitating an antigen (e.g., peptides and nucleotides) from solution using an antibody that specifically binds to that particular antigen. This process can be used to isolate and concentrate specific proteins or DNA from a sample and requires, at some point in the process, the antibody to be conjugated to a solid substrate. Solid substrates include, for example, beads, such as magnetic beads. Other types of beads and solid substrates are known in the art.

An exemplary antibody is a 5-MeC antibody. For the immunoprecipitation process, in some embodiments, at least 0.05 μg of antibody is added to the sample; in more preferred embodiments, at least 0.16 μg of antibody is added to the sample. To confirm the immunoprecipitation reaction, in some embodiments, the method described herein further includes the step of adding a second amount of control DNA to the sample.

In some embodiments, the method further includes the step of adding a second amount of control DNA to the sample to confirm the immunoprecipitation reaction.

As used in this article, “control” may include positive and negative controls, or at least positive controls.

In some embodiments, the method further includes the step of adding a second amount of control DNA to the sample to confirm the capture of cell-free methylated DNA.

In some implementations, identifying the presence of DNA from cancer cells also includes identifying the cancer cell tissue of origin.

In some cases, tumor tissue sampling can be challenging or carry significant risks, in which case it may be desirable to diagnose and/or subtype cancer without requiring tumor tissue sampling. For example, lung tumor tissue sampling may require invasive procedures such as mediastinoscopy, open thoracotomy, or percutaneous biopsy; these procedures may require hospitalization, chest tubes, mechanical ventilation, antibiotics, or other medical interventions. Some individuals may not undergo the invasive procedures required for tumor tissue sampling due to medical comorbidities or personal preference. In some cases, the actual procedure for obtaining tumor tissue may depend on the suspected cancer subtype. In other cases, the cancer subtype may progress over time within the same body. Continuous evaluation via invasive tumor tissue sampling procedures is often impractical and poorly tolerated by patients. Therefore, non-invasive cancer subtyping via blood tests may have many advantageous applications in the implementation of clinical oncology.

Therefore, in some embodiments, identifying the origin of cancerous tissue also includes identifying cancer subtypes. Preferably, cancer subtypes are distinguished based on stage (e.g., early-stage lung cancer treated surgically versus late-stage lung cancer treated with chemotherapy), histology (e.g., small cell carcinoma versus adenocarcinoma versus squamous cell carcinoma in lung cancer), gene expression patterns or transcription factor activity (e.g., ER status in breast cancer), copy number aberrations (e.g., HER2 status in breast cancer), specific rearrangements (e.g., FLT3 in AML), specific gene point mutation status (e.g., IDH gene point mutation), and DNA methylation patterns (e.g., MGMT gene promoter methylation in brain cancer).

In some implementations, the comparison in step (f) is performed on a genome-wide scale.

In other implementations, the comparison in step (f) ranges from genome-wide restriction to specific regulatory regions, such as, but not limited to, the FANTOM5 enhancer, CpG islands, CpG shores, CpG scaffolds, or any combination thereof.

In some implementations, certain steps are performed by a computer processor.

In one aspect, a method is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: receiving sequencing data of cell-free methylated DNA from a subject sample; comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identifying the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals; and, if DNA from cancer cells is identified, further identifying the cancer cell tissue of origin and cancer subtype based on the comparison step.

In one aspect, a method is provided for detecting the presence of DNA from cancer cells and determining the location of cancer cells from two or more possible organs, the method comprising: providing a sample of cell-free DNA from a subject; capturing cell-free methylated DNA from the sample using a binding agent selective for methylated polynucleotides; sequencing the captured cell-free methylated DNA; comparing the sequence pattern of the captured cell-free methylated DNA with the DNA sequence patterns of two or more groups of a control individual, each of the two or more groups having localized cancer in a different organ; and determining which organ the cancer cells originated from based on a statistically significant similarity between the methylation pattern of the cell-free DNA and one of the two or more groups.

The systems and methods of the present invention can be implemented in various embodiments. Appropriately configured computer devices, along with associated communication networks, devices, software, and firmware, can provide a platform for implementing one or more embodiments as described above. By way of example, FIG5 illustrates a general-purpose computer device 100, which may include a central processing unit (“CPU”) 102 connected to a storage unit 104 and random access memory 106. The CPU 102 may process an operating system 101, application programs 103, and data 123. The operating system 101, application programs 103, and data 123 may be stored in the storage unit 104 and loaded into the memory 106 as needed. The computer device 100 may further include a graphics processing unit (GPU) 122 operatively connected to the CPU 102 and the memory 106 to offload intensive image processing computations from the CPU 102 and to run these computations in parallel with the CPU 102. Using a video display 108 connected via video interface 105 and various input/output devices connected via I/O interface 109, such as a keyboard 115, mouse 112, and disk drive or solid-state drive 114, an operator 107 can interact with computer device 100. The mouse 112 can be configured in known ways to control cursor movement within the video display 108 and to use mouse buttons to operate various graphical user interface (GUI) controls displayed on the video display 108. The disk drive or solid-state drive 114 can be configured to accept computer-readable media 116. Computer device 100 can form part of a network via network interface 111, allowing computer device 100 to communicate with other appropriately configured data processing systems (non-display). One or more different types of sensors 135 can be used to receive input from various sources.

The systems and methods of the present invention can be implemented on virtually any type of computer device, including desktop computers, laptop computers, tablet computers, or wireless handheld devices. The systems and methods of the present invention can also be implemented as computer-readable/usable media comprising computer program code, enabling one or more computer devices to implement each of the various processing steps in the methods according to the invention. In cases where more than one computer device performs the entire operation, the computer devices are networked to distribute the various steps of the operation. It should be understood that the terms computer-readable medium or computer-usable medium include one or more of any type of physical embodiment of program code. In particular, computer-readable/usable media may include program code embodied on one or more portable storage articles (e.g., optical discs, magnetic disks, magnetic tapes, etc.) or on a data storage portion (e.g., memory associated with a computer and/or storage system) of one or more computing devices.

In one aspect, a computer-implemented method is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: receiving sequencing data of cell-free methylated DNA from a subject sample by at least one processor; at at least one processor, comparing the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identifying the presence of DNA from cancer cells at at least one processor if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals; and further identifying the originating cancer cell tissue and cancer subtype based on the comparison step if DNA from cancer cells is identified.

In one aspect, a computer program product is provided for use in conjunction with a general-purpose computer having a processor and memory connected to the processor, the computer program product including a computer-readable storage medium thereon having computer mechanisms encoded thereon, wherein the computer program mechanisms can be loaded into the computer's memory and cause the computer to perform the methods described herein.

In one aspect, a computer-readable medium is provided having a data structure thereon for storing the computer program product described herein.

In one aspect, an apparatus is provided for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the apparatus comprising: at least one processor; and an electronic memory in communication with the at least one processor, the electronic memory storing processor-executable code that, when executed at the at least one processor, causes the at least one processor to: receive sequencing data of cell-free methylated DNA from a subject sample; compare the sequence of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; identify the presence of DNA from cancer cells if there is statistically significant similarity between the captured cell-free methylated DNA and one or more sequences of cell-free methylated DNA from cancer individuals; and, if DNA from cancer cells is identified, further identify the cancer cell tissue of origin and cancer subtype based on the comparison step.

As used herein, “processor” can be any type of processor, such as, for example, any type of general-purpose microprocessor or microcontroller (e.g., Intel ^™ x86, PowerPC ^™ , ARM ^™ processors, etc.), digital signal processing (DSP) processor, integrated circuit, field-programmable gate array (FPGA), or any combination thereof.

As used herein, “memory” can include any suitable combination of computer memories, whether internal or external, such as, for example, random access memory (RAM), read-only memory (ROM), optical disc read-only memory (CDROM), electro-optical memory, magneto-optical memory, erasable programmable read-only memory (EPROM), and electrically erasable programmable read-only memory (EEPROM). A portion of memory 102 can be organized using a conventional file system and controlled and managed by an operating system that controls the overall operation of the device.

As used herein, a "computer-readable storage medium" (also referred to as a machine-readable medium embodying computer-readable program code, a processor-readable medium, or a computer-usable medium) is a medium capable of storing data in a computer- or machine-readable format. A machine-readable medium can be any suitable tangible, non-transitory medium, including magnetic, optical, or electrical storage media, including magnetic disks, optical disc read-only memories (CD-ROMs), storage devices (volatile or non-volatile), or similar storage mechanisms. The computer-readable storage medium can contain various sets of instructions, code sequences, configuration information, or other data that, when executed, cause a processor to perform the steps in the methods according to embodiments of this disclosure. Those skilled in the art will recognize that other instructions and operations necessary to implement the described implementations can also be stored on the computer-readable storage medium. Instructions stored on the computer-readable storage medium can be executed by a processor or other suitable processing device and can interact with circuitry to perform the described tasks.

As used in this article, a "data structure" is a specific way of organizing data in a computer so that it can be used efficiently. A data structure can implement one or more specific abstract data types (ADTs), which specify the operations that can be performed on the data structure and the computational complexity of those operations. In contrast, a data structure is a concrete implementation of the specification provided by an ADT.

The advantages of this invention are further illustrated by the following examples. The examples and specific details described herein are for illustrative purposes only and should not be construed as limiting the scope of the claims of this invention.

This invention provides, but is not limited to, the following embodiments:

1. A method for detecting the presence of DNA from cancer cells in a subject, comprising:

(a) Provide a sample of cell-free DNA from the subject;

(b) Perform library preparation on the sample to allow subsequent sequencing of the cell-free methylated DNA;

(c) Add a first amount of filler DNA to the sample, wherein at least a portion of the filler DNA is methylated, and then optionally denature the sample;

(d) Capturing cell-free methylated polynucleotides using a binding agent selective for methylated polynucleotides.

Matrixed DNA;

(e) Sequencing of the captured cell-free methylated DNA;

(f) Compare the sequences of the captured cell-free methylated DNA with those from healthy and cancerous individuals.

Individual control cell-free methylated DNA sequences;

(g) The presence of DNA from cancer cells is identified if the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual.

2. The method according to embodiment 1, wherein the sample is derived from the subject's blood or plasma.

3. The method according to embodiment 1 or 2, wherein the comparison step (f) is based on fitting using a statistical classifier.

4. The method according to embodiment 3, wherein the classifier is derived by machine learning.

5. The method according to embodiment 4, wherein the classifier is an elastic net classifier, a lasso, a support vector machine, a random forest, or a neural network.

6. The method according to any one of embodiments 1 to 5, wherein the control cell-free methylated DNA sequences from healthy and cancer individuals are included in a database of differentially methylated regions (DMRs) between healthy and cancer individuals.

7. The method according to any one of embodiments 1 to 6, wherein the control cell-free methylated DNA sequences from healthy and cancer individuals are limited to those control cell-free methylated DNA sequences that are differentially methylated between healthy and cancer individuals in DNA derived from cell-free DNA.

8. The method according to embodiment 7, wherein the control cell-free methylated DNA sequence is differentially methylated in plasma-derived DNA between healthy and cancer individuals.

9. The method according to any one of embodiments 1 to 8, wherein the sample has a content of less than 100

ng, 75ng or 50ng of cell-free DNA.

10. The method according to any one of embodiments 1 to 9, wherein the first amount of filling DNA

It contains about 5%, 10%, 15%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 100% methylated filler DNA, with the remainder being unmethylated filler DNA, preferably 5% to 50%, 10% to 40%, or 15% to 30% methylated filler DNA.

11. The method according to any one of embodiments 1 to 9, wherein the first amount of filling DNA

The amount is 20 ng to 100 ng, preferably 30 ng to 100 ng, and more preferably 50 ng to 100 ng.

12. The method according to any one of embodiments 1 to 11, wherein the cell-free DNA from the sample and the first amount of filler DNA together contain at least 50 ng of total DNA, preferably at least 100 ng of total DNA.

13. The method according to any one of embodiments 1 to 12, wherein the length of the filled DNA is 50 bp to 800 bp, preferably 100 bp to 600 bp, and more preferably 200 bp.

Up to 600bp.

14. The method according to any one of embodiments 1 to 13, wherein the filling DNA is double-stranded.

15. The method according to any one of embodiments 1 to 2, wherein the filling DNA is junk DNA.

16. The method according to any one of embodiments 1 to 3, wherein the filling DNA is endogenous or exogenous DNA.

17. The method according to embodiment 16, wherein the filling DNA is non-human DNA, preferably λ DNA.

18. The method according to any one of embodiments 1 to 17, wherein the filled DNA is not compared with human DNA.

19. The method according to any one of embodiments 1 to 18, wherein the binding agent is a protein containing a methyl-CpG binding domain.

20. The method according to any one of embodiments 1 to 19, wherein the protein is MBD2 protein.

21. The method according to any one of embodiments 1 to 20, wherein step (d) includes immunoprecipitation of the cell-free methylated DNA using an antibody.

22. The method according to embodiment 21, comprising adding at least 0.05 μg of the antibody to the sample for immunoprecipitation, and preferably at least 0.16 μg.

23. The method according to embodiment 21, wherein the antibody is a 5-MeC antibody.

24. The method according to embodiment 21 further includes the step of adding a second amount of control DNA to the sample after step (c) to confirm the immunoprecipitation reaction.

25. The method according to any one of embodiments 1 to 23 further includes the step of adding a second amount of control DNA to the sample after step (c) to confirm the capture of the cell-free methylated DNA.

26. The method according to any one of embodiments 1 to 25, wherein identifying the presence of DNA from cancer cells further includes identifying the cancer cell tissue of origin.

27. The method according to embodiment 26, wherein identifying the cancer cell tissue of the origin further includes identifying a cancer subtype.

28. The method according to embodiment 27, wherein the cancer subtype is distinguished based on stage, histology, gene expression pattern, copy number aberration, rearrangement or point mutation status.

29. The method according to any one of embodiments 1 to 28, wherein the comparison in step (f) is performed on a genome-wide scale.

30. The method according to any one of embodiments 1 to 28, wherein the comparison in step (f) is restricted from the whole genome to a specific regulatory region.

31. The method according to embodiment 30, wherein the adjustment region is a FANTOM5 enhancer,

CpG islands, CpG coasts, CpG shelves, or any combination thereof.

32. The method according to any one of embodiments 1 to 31, wherein steps (f) and (g) are executed by a computer processor.

33. A method for detecting the presence of DNA from cancer cells and identifying cancer subtypes, said method

include:

a. Receiving sequencing data of cell-free methylated DNA from subject samples;

b. Compare the sequences of the captured cell-free methylated DNA with those from healthy and cancer patients.

The control group of cell-free methylated DNA sequences;

c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, then

Identifying the presence of DNA from cancer cells; and

d. If DNA from cancer cells is identified in step c, the origin of the cancer cell tissue and cancer subtype is further identified based on the comparison in step b.

34. A computer-implemented method for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising:

a. At least one processor receives sequencing of cell-free methylated DNA from the subject sample.

data;

b. At the at least one processor, the sequence of the captured cell-free methylated DNA is compared with the sequence of control cell-free methylated DNA from healthy and cancer individuals;

c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, the presence of DNA from cancer cells is identified at the at least one processor; and if DNA from cancer cells is identified in step c, the origin of the cancer cell tissue and cancer subtype is further identified based on the comparison in step b.

35. A computer program product for use in conjunction with a general-purpose computer having a processor and a memory connected to the processor, the computer program product comprising a computer-readable storage medium thereon having a computer mechanism encoded thereon, wherein the computer program mechanism can be loaded into the memory of the computer and cause the computer to perform the method of embodiment 34.

36. A computer-readable medium having a data structure thereon for storing the computer program product of embodiment 35.

37. An apparatus for detecting the presence of DNA from cancer cells and identifying cancer subtypes, wherein...

The equipment includes:

At least one processor; and

An electronic memory communicating with the at least one processor, the electronic memory storing processor-executable code, which, when executed at the at least one processor, causes the at least one processor to...

One less processor:

a. Receiving sequencing data of cell-free methylated DNA from subject samples;

b. Compare the sequences of the captured cell-free methylated DNA with those from healthy and cancer patients.

The control group of cell-free methylated DNA sequences;

c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, the presence of DNA from cancer cells is identified, and if DNA from cancer cells is identified in step c, the cancer cell tissue of origin and cancer subtype are further identified based on the comparison in step b.

38. A method for detecting the presence of DNA from cancer cells and determining the location of cancer cells from two or more possible organs, the method comprising:

(a) Provide a sample of cell-free DNA from the subject;

(b) Using a binding agent selective for methylated polynucleotides from said sample

Capture cell-free methylated DNA;

(c) Sequencing of the captured cell-free methylated DNA;

(d) Compare the sequence patterns of the captured cell-free methylated DNA with the DNA sequence patterns of two or more populations of control individuals, each of the two or more populations in which the sequence patterns are...

Localized cancers can occur in different organs;

(e) Based on the statistically significant similarity between the methylation pattern of the cell-free DNA and one of the two or more populations, determine which organ the cancer cells originated from.

Example

Methods and Materials

Donor recruitment and sample collection

CRC, breast cancer, and GBM samples were obtained from the University Health Network Biobank; AML samples were obtained from the University Health Network Leukemia Biobank; and finally, healthy controls were recruited through the Family Medicine Centre at Mount Sinai Hospital (MSH) in Toronto, Canada. All samples collected with patient consent received institutional approval from the research ethics committee, the University Health Network, and Mount Sinai Hospital in Toronto, Canada.

Sample processing – cfDNA

EDTA and ACD plasma samples were obtained from the biobank and the Family Medicine Centre at Mount Sinai Hospital (MSH) in Toronto, Canada. All samples were stored at -80°C or in liquid nitrogen until use. Cell-free DNA was extracted from 0.5–3.5 ml of plasma using the QIAamp Cyclic Nucleic Acid Kit (Qiagen). The extracted DNA was quantified using Qubit before use.

Sample preparation – PDX cfDNA

With the approval of the University Health Network Research Ethics Committee and with patient consent, human colorectal tumor tissue obtained from the University Health Network Biobank was digested into single cells using collagenase A. The single cells were subcutaneously injected into 4–6 week old NOD/SCID male mice. Mice were euthanized by carbon dioxide inhalation prior to cardiac blood collection and stored in EDTA tubes. Plasma was separated from the collected blood samples and stored at -80°C. Cell-free DNA was extracted from 0.3–0.7 ml of plasma using the QIAamp Cyclic Nucleic Acid Kit (Qiagen). All animal work was conducted in accordance with the ethical guidelines approved by the University Health Network Animal Protection Committee.

cfMeDIP-seq

A schematic diagram of the cfMeDIP-seq protocol is shown in WO2017/190215. Prior to cfMeDIP, DNA samples were prepared into libraries using the KapaHyper Prep kit (Kapa Biosystems). Some modifications were made to the manufacturer's protocol. In short, the DNA of interest was added to a 0.2 mL PCR tube, and end repair and A-tailing were performed. Intercalators were ligated using NEBNext adaptors (NEBNext Multiplex Oligos for Illumina kit from New England Biolabs) to a final concentration of 0.181 μM, incubated at 20°C for 20 min, and purified using AMPure XP beads. The eluted library was digested with USER enzyme (New England Biolabs, Canada) and then purified using the Qiagen MinElute PCR Purification Kit before MeDIP.

The prepared library was combined with the combined methylated/unmethylated λPCR products to a final DNA volume of 100 ng, and MeDIP was performed using the protocol of Taiwo et al. 2012 [7] after some modifications. In short, for MeDIP, the Diagenode MagMeDIP kit (number C02010021) was used according to the manufacturer’s protocol after some modifications. After adding 0.3 ng of control methylated and 0.3 ng of control unmethylated Arabidopsis DNA, filler DNA (to make the total amount of DNA [cfDNA + filler + control] 100 ng) and buffer to the PCR tube containing the adapter-ligated DNA, the sample was heated to 95 °C for 10 min and then immediately placed in an ice water bath for 10 min. Each sample was divided into two 0.2 mL PCR tubes: one for the 10% input control and the other for the sample to be immunoprecipitated. The 5-mC monoclonal antibody 33D3 (C15200081) included in the MagMeDIP kit was diluted 1:15 to generate a diluted antibody mixture, which was then added to the sample. Washed magnetic beads (according to the manufacturer's instructions) were also added, followed by incubation at 4°C for 17 hours. The sample was purified using the DiagenodeiPure kit and eluted with 50 μl of buffer C. The success of the reaction was verified by qPCR (QC1) to detect the presence of added Arabidopsis DNA, ensuring a recovery percentage of unmethylated added DNA <1% and a reaction specificity >99% (calculated as 1 - [recovery of unmethylated control DNA exceeding recovery of methylated control DNA]), before proceeding to the next step. The optimal number of cycles for amplifying each library was determined using qPCR, and the sample was then amplified using the KAPAHiFi Hotstart Mastermix with NEBNext multiplex oligonucleotides added to a final concentration of 0.3 μM. The PCR setup for library amplification was as follows: activation at 95°C for 3 minutes, followed by predetermined cycles of 98°C for 20 seconds, 65°C for 15 seconds, and 72°C for 30 seconds, with a final extension at 72°C for 1 minute. The amplified library was purified using a MinElute PCR purification column, followed by gel size selection on a 3% Nusieve GTG agarose gel to remove any adaptor dimers. Prior to sequencing, the enrichment folds of methylated human DNA regions (testis-specific H2B, TSH2B) and unmethylated human DNA regions (GAPDH promoter) in MeDIP-seq and cfMeDIP-seq libraries generated from HCT116 cell line DNA (obtained from ATCC, mycoplasma-free) that had been cleaved to mimic cell-free DNA were determined. The final library was submitted for analysis using BioAnalyzer and then sequenced using Illumina HiSeq 2000 at the UHN Princess Margaret Genomic Centre.

Ultra-deep targeted sequencing for point mutation detection

We used the QIAgen Cyclic Nucleic Acid Kit to isolate cell-free DNA from approximately 20 mL of plasma (4-5 x 10 mL EDTA blood tubes) from patients with matched tumor tissue molecular profiling data generated at Princess Margaret Cancer Centre prior to enrollment in early clinical trials. DNA was extracted from cell lines (CRC and MM cell line dilutions) using the PureGene Gentra Kit, fragmented to approximately 180 bp using Covaris sonication, and larger fragments were excluded using Ampure beads to mimic the fragment size of cell-free DNA. DNA sequencing libraries were constructed from 83 ng of fragmented DNA using the NEXTflex-96 DNA barcoding adaptor (Bio Scientific, Austin, TX) using the KAPA Hyper Prep Kit (Kapa Biosystems, Wilmington, MA). To isolate DNA fragments containing known mutations, we designed biotinylated DNA capture probes (xGen Lockdown Custom Probes Mini Pool, Integrated DNA Technologies, Coralville, IA) targeting mutation hotspots in 48 genes tested by clinical laboratories using Illumina TruSeq amplicon cancer assays. Barcode libraries were merged, and then the custom-mixed capture library was applied following the manufacturer's instructions (IDT xGEN Lockdown protocol version 2.1). These fragments were sequenced to >10,000X read coverage using an Illumina HiSeq 2000 instrument. The resulting reads were aligned using bwa-mem, and mutations were detected using samtools and muTect version 1.1.4.

Modeling the relationship between the number of tumor-specific features and the probability of detection via sequencing depth.

We created 145,000 simulated genomes, with cancer-specific methylation DMRs set to proportions of 0.001%, 0.01%, 0.1%, 1%, and 10%, and composed of 1, 10, 100, 1000, and 10000 independent DMRs, respectively. From these raw mixtures, we sampled 14,500 diploid genomes (representing 100 ng of DNA) and further sampled 10, 100, 1000, and 10000 reads at each locus to represent sequencing coverage at these depths. This process was repeated 100 times for each combination of coverage, abundance, and feature number. We estimated the frequency of successful detection of at least one DMR for each parameter combination and plotted probability curves (Figure 1A) to visually assess the impact of feature number on the probability of successful detection depending on sequencing depth.

Derivation of organizational distinguishing features, development of a multi-organization classifier, and validation on 450k data sets.

The cfDNAMeDIP profiles were quantified using the MEDIPS R package [8], converted to RPKM, and then to log2 per million counts. Subsequently, a linear model was fitted to a feature matrix mapped to the FANTOM5 enhancer, CpG islands, CpG shores, and CpG scaffolds using limma-trend [9], with the percentage of recovered spike-in methylated DNA included as a covariate to control for pull-down efficiency variations. Pairwise comparisons were evaluated for each tissue type, and the top 150 and bottom 150 DMRs were selected for training an elastic net classifier and validation of cancer type specificity. Performance metrics were derived by majority vote of the out-of-bounds call of the model with the highest Kappa value in cross-validation, a heuristic previously employed by Chakravarthy et al. [10].

Machine learning analytics are used to evaluate classification accuracy.

Model training and evaluation for discovery queues

To evaluate the performance of cfMeDIP data in tumor classification without high computational cost, we reduced the initial set of possible candidate features to a window containing CpG islands, CpG shores, CpG frames, and FANTOM5 enhancers (labeled here as "modulation features"), resulting in a matrix of 196 samples and 505,027 features. We then used the caret R package to split the discovery cohort data into 50 independent training and test sets in an 80%–20% manner (Figure 2A). This splitting was performed while maintaining the class proportions within the discovery cohort. We then used limma-trend for each class relative to the others, selecting the top 300 DMRs by a moderate t-statistic (150 high methylations, 150 low methylations) on the training data partitions. A binomial GLMnet was then trained using these DMRs (up to 300 DMRs x 7 other classes = 2100 features), and optimized with 3 iterations of 10x cross-validation (CV) to optimize the values of the mixture parameters (α, values = 0, 0.2, 0.5, 0.8, and 1) and the penalty (λ, values = 0–0.05 in 0.01 increments), using Cohen's Kappa as the performance metric. For each training set, a set of 6 binomial classifiers for one class against the other classes was generated.

We then used AUROC (the region below the acceptor working characteristic curve) to estimate the classification performance on the retained test set. These estimates represent an unbiased measure of classification, since the retained test set samples were not used for DMR pre-selection or GLMnet training and tuning. The 50 independent training and test sets also allow for minimizing optimistic estimates due to training set bias.

Model evaluation of the verification queue

For each validation cohort cfMeDIP sample, we estimated the classification probabilities of AML, LUC, and normal one-to-many binomial classifiers trained on 50 different training sets within the discovery cohort. The probabilities from the 50 models were averaged to produce a single score, which was then used for AUROC estimation. We also left early (stages I and II) or late LUC samples (stages III and IV) to the one-to-many classifier to assess whether disease stage affected performance by estimating AUROC.

Results and discussion

We used bioinformatics methods to simulate mixtures of different proportions of ctDNA, ranging from 0.001% to 10% (Fig. 1A, column plane). We also simulated cases where ctDNA had 1, 10, 100, 1000, or 10000 DMRs (differential methylation regions) compared to normal cfDNA (Fig. 1A, row plane). Readings were then sampled at different sequencing depths (10X, 100X, 1000X, and 10000X) at each locus (Fig. 1A, x-axis). We found that the probability of detecting at least one cancer-specific event increased with increasing DMR count (Fig. 1A), even with low abundance of cancer ctDNA and shallow coverage.

Furthermore, whole-cancer data from the Cancer Genome Atlas (TCGA) show that there is a significant amount of DMR between tumor and normal tissue in almost all tumor types[11]. Therefore, these findings highlight that a assay that successfully recovers cancer-specific DNA methylation alterations from ctDNA could serve as a highly sensitive tool for detecting, classifying, and monitoring malignant diseases at low sequencing-related costs.

However, whole-genome mapping of DNA methylation in plasma cfDNA is challenging due to the limited quantity and fragmentation of circulating DNA [12]. Therefore, previous efforts to map methylation in cfDNA have largely been limited to locus-specific PCR assays [2, 3], such as the FDA-approved SEPT9 methylation assay for colorectal cancer screening [13]. Although whole-genome bisulfite sequencing of fragmented cfDNA has recently been performed [14-16], the low abundance of CpGs across the genome may reduce the amount of useful methylation-related information available from sequencing. Thus, the main problems with WGBS for plasma DNA are high cost, low efficiency, and DNA loss associated with bisulfite conversion. On the other hand, methods that selectively enrich readily methylated CpG-rich features could maximize the amount of useful information available per read, while reducing cost and DNA loss.

Genome-wide approach for cfDNA methylation mapping

We have developed a novel approach called cfMeDIP-seq (cell-free methylated DNA immunoprecipitation and high-throughput sequencing) for whole-genome DNA methylation mapping using cell-free DNA. The cfMeDIP-seq approach described herein was developed by modifying an existing low-input MeDIP-seq protocol [7], which, in our experience, remains very robust with input DNA as low as 100 ng. However, most plasma samples produce far less than 100 ng of DNA. To overcome this challenge, we added exogenous λDNA (filler DNA) to the adaptor-linked cfDNA library to artificially increase the amount of starting DNA to 100 ng. This minimizes the amount of nonspecific binding of antibodies and also minimizes the amount of DNA loss due to binding to plastics. The filler DNA consists of amplicons of similar size to the adaptor-linked cfDNA library and consists of unmethylated and in vitro methylated DNA with varying CpG densities. This addition of filler DNA also has practical uses, as different patients produce different amounts of cfDNA, thus normalizing the amount of input DNA to 100 ng. This ensures that the downstream protocols for all samples remain exactly the same, regardless of the amount of cfDNA available.

We first validated the cfMeDIP-seq protocol using DNA from the human colorectal cancer cell line HCT116, which was cut to fragment sizes similar to those observed in cfDNA. HCT116 was chosen because public DNA methylation data were available. We simultaneously performed the gold standard MeDIP-seq protocol using 100 ng of cut cell line DNA [7] and the cfMeDIP-seq protocol using 10 ng, 5 ng, and 1 ng of the same cut cell line DNA. This was performed in two biological replicates. Under all conditions, we obtained more than 99% response specificity (1 - [the recovery rate of the added unmethylated control DNA exceeded the recovery rate of the added methylated control DNA]) and very large enrichment of known methylated regions relative to unmethylated regions (TSH2B0 and GAPDH, respectively) (Fig. 6B).

The libraries were sequenced to saturation with approximately 30 to 70 million reads per library (Supplementary Table 1) (Figure 6A). The raw reads were aligned to the human genome and the λ genome, and almost no alignment was found with the λ genome (Supplementary Table 1). Therefore, adding exogenous λ DNA as filler DNA did not interfere with the generation of sequencing data. Finally, we calculated the CpG enrichment score as a quality control metric for the immunoprecipitation step [8]. All libraries showed similar CpG enrichment, while the input control was not enriched as expected (Figure 6C), validating our immunoprecipitation even at very low input (1 ng).

Comparison of genome-wide correlation estimates at different input DNA levels showed that both MeDIP-seq (100 ng) and cfMeDIP-seq (10, 5, and 1 ng) methods were highly robust, with Pearson correlations of at least 0.94 between any two biological replicates (Figure 1B). The analysis also showed that cfMeDIP-seq with 5 and 10 ng input DNA robustly generalized the methylation profiles obtained by conventional MeDIP-seq with 100 ng (Pairwise Pearson correlations of at least 0.9) (Figure 1B). cfMeDIP-seq with 1 ng input DNA showed slightly lower performance compared to 100 ng MeDIP-seq, but still exhibited strong Pearson correlations >0.7 (Figure 1B). We also observed that the cfMeDIP-seq protocol generalized the DNA methylation profile of HCT116 using the gold standard RRBS (Reduced Representative Bisulfite Sequencing) and WGBS (Whole Genome Bisulfite Sequencing) (Figure 1C). In summary, our data demonstrate that cfMeDIP-seq is a reliable protocol for whole-genome methylation mapping of fragmented and low-input DNA material (e.g., circulating cfDNA).

cfMeDIP-seq showed high sensitivity for detecting tumor-derived ctDNA.

To evaluate the sensitivity of the cfMeDIP-seq protocol, we serially diluted colorectal cancer (CRC) HCT116 cell line DNA into multiple myeloma (MM) MM1.S cell line DNA, both of which were cleaved to a simulated cfDNA size. We diluted CRC DNA from 100%, 10%, 1%, 0.1%, 0.01%, 0.001% to 0% and performed cfMeDIP-seq on each dilution. We also performed ultra-deep (10,000X median coverage) targeted sequencing to detect three-point mutations in the same sample. The number of DMRs identified at each CRC dilution relative to pure MM DNA, observed using a 5% false discovery rate (FDR) threshold, was almost perfectly linear with the expected number of DMRs based on the dilution factor reduced to 0.001% ( ^r² = 0.99, p < 0.0001) (Figure 1D). Furthermore, the DNA methylation signals in these DMRs showed a near-perfect linear relationship between the observed and expected signals ( ^r² = 0.99, p < 0.0001) (Figure 1E; Supplementary Table 2B). In contrast, beyond 1% dilution, ultra-deep targeted sequencing could not reliably distinguish between CRC-specific variants and pseudovariants due to PCR or sequencing errors (Figure 1F; Supplementary Table 2A). Therefore, cfMeDIP-seq demonstrated superior sensitivity for detecting cancer-derived DNA, surpassing the performance of ultra-deep targeted sequencing using standard protocols for variant detection.

Cancer DNA is frequently hypermethylated in CpG-rich regions [17]. Since cfMeDIP-seq specifically targets methylated CpG-rich sequences, we hypothesized that ctDNA would be preferentially enriched during immunoprecipitation. To test this, we generated patient-derived xenografts (PDXs) from two colorectal cancer patients and collected mouse plasma. Tumor-derived human cfDNA was present in the total cfDNA pool at a frequency of less than 1% in the input samples and its abundance increased by 2-fold after immunoprecipitation (Fig. 1G; Supplementary Table 3). These results suggest that the cfMeDIP procedure can further improve the sensitivity of ctDNA detection by biased sequencing of ctDNA.

Circulating plasma cfDNA methylation profiles can differentiate between various cancer types and healthy donors.

DNA methylation patterns are tissue-specific and have been used to classify cancer patients into clinically relevant disease subgroups for many cancer types, including glioblastoma[18], ependymoma[6], colorectal cancer[19], and breast cancer[20, 21]. We asked whether cfDNA-related mapping could be used to identify the tissue of origin for multiple tumor types. To this end, we analyzed 196 samples from 5 different tumor types and normal controls from early and late-stage tumors. We used linear modeling to identify the top 300 DMRs mapped to CpG shores, CpG scaffolds, CpG islands, and FANTOM5 enhancers in each pairwise comparison, resulting in a total of 2,100 unique DMRs (Fig. 2A). Density clustering based on t-distribution random neighborhood embedding (tSNE) of the 196 plasma samples[22] based on the methylation status of these features showed different clusters of samples based on tissue of origin and tumor type (Fig. 2B, Fig. 2C). Using an elastic mesh multi-cancer classifier suited to these characteristics (Fig. 2A), we observed highly accurate differentiation between different tumor types (Fig. 2D).

Differentiation of disease subtypes

We evaluated the ability of cfDNA MeDIP maps to distinguish disease subtypes in five different cases—gene expression patterns (ER status in breast cancer), copy number aberrations (HER2 status in breast cancer), rearrangements (FLT3 ITD status in AML), point mutations (IDH mutations in GBM), and finally, histology in lung cancer. In each case, as previously described, a linear model was used to select features and rank them. In each case, hierarchical clustering was used to evaluate the grouping of samples. Density clustering based on t-distributed random neighborhood embedding (tSNE) based on the methylation status of the selected features [22] showed that the samples were clustered differently based on each of the different instances of these five cancer subtype classifications.

Using machine learning to detect and classify cancer types

To rigorously evaluate the ability of the cfMeDIP map to detect cancer and further classify cancer types, we subsequently performed a set of machine learning analyses on the discovery cohort. To accelerate computational analysis, we initially reduced our cfMeDIP discovery cohort to features mapped to CpG islands, CpG shores, CpG frames, and FANTOM5 enhancers (n = 505,027 windows). We then implemented a strategy on our discovery cohort samples to derive an unbiased estimate of performance while taking into account training set bias.

In this paper, we find that the queue is divided into balanced training and test sets (80% training, 20% test). Using only samples from the training set, we select the top 300 DMRs for each class (sample type) relative to other classes based on limma-trend test statistics, and use these features to train a series of one-class versus other-class GLMnets on the training set data. The training procedure consists of 3 rounds of 10x cross-validation (CV) on grids of α and λ values, optimized using Cohen's Kappa. We hope to leverage more randomization for more general model tuning, thus encouraging the use of multiple rounds of 10x CV.

Performance was then evaluated using the AUROC (the region under the receiver operating characteristic curve) derived from the test set samples (preserved during DMR selection and subsequent GLMnet training/tuning steps). This process was repeated 50 times from the discovery queue to different training and test sets to mitigate training set bias. A total of 50 models were ultimately collected for comparisons of each class against the others (out of a total of 480 models). Therefore, we refer to this set of models as the E50.

We then evaluated the performance of each batch by generating a validation cohort of 152 additional plasma samples: AML (n=35), lung cancer (n=55), and healthy controls (n=62). For each category, we averaged the class probabilities of the model output in E50 and estimated the AUROC of one category relative to all other categories (Figure 3A). The classifier showed high AUROC values for the classification of AML relative to others (0.993), LUC relative to others (0.943), and normal relative to others (1.000). This further confirms the ability of combining cfMeDIP-seq with machine learning methods to accurately detect and classify tumor types. Finally, we observed that the classifier was as accurate in early samples (0.950) as in late samples (0.934) (Figure 3B), indicating that the method is suitable for early cancer detection as well as the detection of both early and late-stage cancers.

Other advantages of using cfMeDIP-seq for cfDNA methylome mapping analysis

The ability of cfDNA methylation patterns to accurately represent the tissue of origin also overcomes the limitations of mutation-based assays, which may have low specificity for the tissue of origin due to the recurrent nature of potential driver gene mutations in many cancers across different tissues [23]. Mutation-based assays may also become insensitive due to the clonal structure of tumors, in which case subclonal drivers may be more difficult to detect due to their low abundance in ctDNA [24]. Mutation-based ctDNA methods are also susceptible to potential confounding effects from driver mutations in benign tissues, which has been observed [25] and there is evidence of positive selection in the literature [26].

In summary, we found that based on the largest cancer cfDNA methylome obtained to date, cfMeDIP-seq can be established as an effective and cost-efficient tool with the potential to impact cancer management and early detection. The accuracy and versatility of cfMeDIP-seq can be used to provide treatment decisions in contexts associated with drug resistance and epigenetic alterations, such as sensitivity to androgen receptor inhibition in prostate cancer [27]. The potential opportunities for early diagnosis and screening are particularly evident in lung cancer, where such screening has shown clinical utility, but there are significant limitations to existing screening tests (i.e., low-dose CT scans), such as ionizing radiation exposure and high false positive rates.

In summary, our findings highlight the practicality of cfDNA methylation mapping as a basis for non-invasive, cost-effective, sensitive, and highly accurate early tumor detection, multi-cancer classification, and cancer subtype classification.

Although preferred embodiments of the invention have been described herein, those skilled in the art will understand that variations may be made without departing from the spirit of the invention or the scope of the appended claims. All documents disclosed herein, including those in the following list of references, are incorporated herein by reference.

Table 1. Read counts and mapping efficiencies of sequencing MeDIP-seq (100 ng Rep 1 and Rep 2) and cfMeDIP-seq (10 ng, 5 ng, and 1 ng, Rep 1 and Rep 2) libraries prepared using HCT116 cell line DNA spliced to mimic cfDNA, compared with human (Hg19) genome and λ genome. Two biological replicates were used for the starting input DNA. For starting inputs smaller than 100 ng, exogenous λDNA was added to the sample before MeDIP to artificially increase the starting amount to 100 ng.

Table 2A: Average coverage of ultra-deep targeted variant sequencing using serial dilutions of CRC cell line HCT116 DNA in MM cell line MM1.sDNA

Table 2B: DMR and DNA methylation signals observed in serial dilutions of CRC cell line HCT116 DNA in MM cell line MM1.sDNA

Table 3: Read count and mapping efficiency of cfMeDIP-seq libraries of PDX and input control samples after alignment with the human (Hg19) genome.

References

1.Diaz, L.A., Jr.and A.Bardelli, Liquid biopsies:genotyping circulatingtumor DNA.J ClinOncol, 2014.32(6):p.579-86.

2.Lehmann-Werman,R.,et al.,Identification of tissue-specific celldeath using methylationpatterns of circulating DNA.Proc Natl Acad Sci U S A, 2016.113(13):p.E1826-34.3.Visvanathan,K.,et al. ,Monitoring of Serum DNAMethylation as an Early IndependentMarker of Response and Survival inMetastatic Breast Cancer:TBCRC 005ProspectiveBiomarker Study.J Clin Oncol,2016:p.JCO2015662080.

4.Newman,A.M.,et al.,An ultrasensitive method for quantitatingcirculating tumor DNAwith broad patient coverage.Nat Med,2014.20(5):p.548-54.

5. Aravanis, A.M., M.Lee, and R.D.Klausner, Next-Generation Sequencing ofCirculating Tumor DNA for Early Cancer Detection.Cell, 2017.168(4):p.571-574.

6.Mack, S.C., et al., Epigenomic alterations define lethal CIMP-positive ependymomas ofinfancy. Nature, 2014.506(7489): p.445-50.

7.Taiwo,O.,et al.,Methylome analysis using MeDIP-seq with low DNAconcentrations.NatProtoc, 2012.7(4):p.617-36.

8.Lienhard,M.,et al.,MEDIPS:genome-wide differential coverageanalysis of sequencingdata derived from DNA enrichmentexperiments.Bioinformatics,2014.30(2):p.284-6.9.Law,C.W.,et al.,voom:Precision weights unlock linear model analysis tools for RNA-

seq read counts.Genome Biol,2014.15(2):p.R29.

10.Chakravarthy,A.,et al.,Human Papillomavirus Drives TumorDevelopment Throughoutthe Head and Neck:Improved Prognosis Is Associated Withan Immune ResponseLargely Restricted to the Oropharynx.J Clin Oncol,2016.34(34):p.4132-4141.11.Hoadley,K.A.,et al.,Multiplatform analysis of 12cancertypes reveal molecularclassification within and across tissues oforigin.Cell,2014.158(4):p.929-44.

12.Fleischhacker,M.and B.Schmidt,Circulating nucleic acids(CNAs)andcancer--a survey.

Biochim Biophys Acta,2007.1775(1):p.181-232.

13.Potter, N.T., et al., Validation of a real-time PCR-based qualitative assay for the detection of methylated SEPT9 DNA in human plas ma.Clin Chem,2014.60(9):p.1183-91.14.Legendre,C.,et al.,Whole-genome bisulfite sequencing of cell-free DNA identifysignature ass associated with metastatic breastcancer.Clin Epigenetics,2015.7:p.100.15.Sun,K.,et al.,Plasma DNA tissuemapping by genome-wide methy lation sequencing for noninvasive prenatal, cancer, and transplantation assessments. Proc Natl Acad Sci U SA, 2015.112(40):p.E5503-12.

16.Chan,K.C.,et al.,Noninvasive detection of cancer-associatedgenome-widehypomethylation and copy number aberrations by plasma DNAbisulfite sequencing.ProcNatl Acad Sci U S A, 2013.110(47):p.18761-8.

17. Sharma, S., T. K. Kelly, and P. A. Jones, Epigenetics incancer. Carcinogenesis, 2010.

31(1):p.27-36.

18.Sturm, D., et al., Hotspot mutations in H3F3A and IDH1 defined epigenetic and biological subgroups of glioblastoma. Cancer Cell, 2012.22(4):p.425-37.

19.Hinoue, T., et al., Genome-scale analysis of aberrant DNA methylationin colorectalcancer. Genome Res, 2012.22(2):p.271-82.

20.Stirzaker,C.,et al.,Methylome sequencing in triple-negative breastcancer revealsdistinct methylation clusters with prognostic value.Nat Comm un,2015.6:p.5899.21.Fang,F.,et al.,Breast cancer methylomes establish anepigenomic foundation formetastasis.Sci Transl Med,2011.3(75):p.75ra25.

22. Laurens van der Maaten, G.H., Visualizing Data using t-SNE. Journal of Machine Learning Research, 2008.9: p.2579-2605.

23.Kandoth,C.,et al.,Mutational landscape and significance across 12major cancer types.

Nature,2013.502(7471):p.333-9.

24.McGranahan,N.,et al.,Clonal status of actionable driver events and the timing ofmutational processes in cancer evolution.Sci Transl Med,2015.7(283):p.283ra54.25.Zauber,P.,S.Ma rotta, and M.Sabbath-Solitare, KRAS genemutations are more common in colorectal villous adenomas and in situcarcinomas than in carcinomas.Int JMol Epidemiol Genet, 2013.4(1):p.1-10.

26.Martincorena,I.,et al.,Tumor evolution.High burden and pervasivepositive selection ofsomatic mutations in normal human skin.Science,2015.348(6237):p.880-6.

27. Beltran, H., et al., Divergent clonal evolution of castration-resistant neuroendocrineprostate cancer. 2016.22(3): p.298-305.

Claims (10)

1. A method for detecting the presence of DNA from cancer cells in a subject, comprising: (a) Provide a sample of cell-free DNA from the subject; (b) Perform library preparation on the sample to allow subsequent sequencing of the cell-free methylated DNA; (c) Add a first amount of filler DNA to the sample, wherein at least a portion of the filler DNA is methylated, and then optionally denature the sample; (d) Capturing cell-free methylated DNA using a binding agent that is selective for methylated polynucleotides; (e) Sequencing of the captured cell-free methylated DNA; (f) Compare the sequences of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; (g) The presence of DNA from cancer cells is identified if the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual. 2. The method of claim 1, wherein the sample is derived from the subject's blood or plasma. 3. The method according to claim 1 or 2, wherein the comparison step (f) is based on a fit using a statistical classifier. 4. The method of claim 3, wherein the classifier is derived from machine learning. 5. A method for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: a. Receiving sequencing data of cell-free methylated DNA from subject samples; b. Compare the sequences of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, the presence of DNA from cancer cells is identified; and d. If DNA from cancer cells is identified in step c, the origin of the cancer cell tissue and cancer subtype is further identified based on the comparison in step b. 6. A computer-implemented method for detecting the presence of DNA from cancer cells and identifying cancer subtypes, the method comprising: a. At least one processor receives sequencing data of cell-free methylated DNA from a subject sample; b. At the at least one processor, the sequence of the captured cell-free methylated DNA is compared with the sequence of control cell-free methylated DNA from healthy and cancer individuals; c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, the presence of DNA from cancer cells is identified at the at least one processor; and if DNA from cancer cells is identified in step c, the origin of the cancer cell tissue and cancer subtype is further identified based on the comparison in step b. 7. A computer program product for use in conjunction with a general-purpose computer having a processor and a memory connected to the processor, the computer program product comprising a computer-readable storage medium thereon having a computer mechanism encoded thereon, wherein the computer program mechanism can be loaded into the memory of the computer and cause the computer to perform the method of claim 6. 8. A computer-readable medium having a data structure thereon for storing the computer program product of claim 7. 9. An apparatus for detecting the presence of DNA from cancer cells and identifying cancer subtypes, said apparatus comprising: At least one processor; and An electronic memory communicating with the at least one processor, the electronic memory storing processor-executable code, which, when executed at the at least one processor, causes the at least one processor to: a. Receiving sequencing data of cell-free methylated DNA from subject samples; b. Compare the sequences of the captured cell-free methylated DNA with control cell-free methylated DNA sequences from healthy and cancer individuals; c. If the captured cell-free methylated DNA is statistically significantly similar to one or more sequences of cell-free methylated DNA from a cancer individual, the presence of DNA from cancer cells is identified, and if DNA from cancer cells is identified in step c, the cancer cell tissue of origin and cancer subtype are further identified based on the comparison in step b. 10. A method for detecting the presence of DNA from cancer cells and determining the location of cancer cells from two or more possible organs, the method comprising: (a) Provide a sample of cell-free DNA from the subject; (b) Capture cell-free methylated DNA from the sample using a binding agent that is selective for methylated polynucleotides; (c) Sequencing of the captured cell-free methylated DNA; (d) Compare the sequence patterns of the captured cell-free methylated DNA with the DNA sequence patterns of two or more groups of control individuals, each of which has localized cancer in different organs; (e) Based on the statistically significant similarity between the methylation pattern of the cell-free DNA and one of the two or more populations, determine which organ the cancer cells originated from.